Spatial Modeling Principles in Earth Sciences

Zekai Şen

Spatial Modeling Principles

in Earth Sciences

123
Prof. Zekai Şen
İstanbul Technical University
Civil Engineering Faculty
Campus Maslak
34469 Istanbul
Turkey
[email protected]

ISBN 978-1-4020-9671-6 e-ISBN 978-1-4020-9672-3

DOI 10.1007/978-1-4020-9672-3
Springer Dordrecht Heidelberg London New York

Library of Congress Control Number: 2009920941

c Springer Science+Business Media B.V. 2009

No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by
any means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written
permission from the Publisher, with the exception of any material supplied specifically for the purpose
of being entered and executed on a computer system, for exclusive use by the purchaser of the work.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

This work is dedicated to my students all over
the world with the hope that they will
produce more advanced scientific works in
different aspects of earth sciences
Preface

Earth sciences phenomena have evolved in time and space jointly, but in practical
applications their records are available as temporal records, spatial measurements,
or both. If the records are available only spatially, then they are one realization from
the regionalized variable (ReV), which has geometrical locations as longitudes (or
easting) or latitudes (or northing), with a record of the earth sciences phenomena
at the same location, Hence, in practical applications a set of triplicate values (lon-
gitude, latitude, record) provides the realization out of many realizations from the
ReV concerned. The worth of data in earth sciences and geology is very high since
most of the interpretations and decisions are based on their qualitative and quantita-
tive information content. This information is hidden in representative field samples,
which are analyzed for the extraction of numerical or descriptive characteristics.
These characteristics are referred to as data. Data collection in earth sciences is
difficult, expensive, and requires special care for accurately representing the ge-
ological phenomenon. After all various parameters necessary for the description
and modeling of the geological event, such as bearing capacity, effective strength,
porosity, hydraulic conductivity, chemical contents, are hidden within each sample,
they individually represent a specific point in space and time. Change of locations
leads to another realization, which is different than the others, but they are statis-
tically indistinguishable from each other. This property provides a common basis
for the development of convenient prediction models for the ReVs. In general, the
collection of methodologies for modeling such a set of triplicates falls within the
geostatistical domain, which has been in practical use for the past four decades.
Kriging is the methodology that is used invariably in earth sciences for the regional
(spatial) prediction of spatial variability. Prior to its application, a descriptive func-
tion of spatial variability as the change of square-differenced spatial variation at
two sides with the distance is constructed under the name of semivariogram (SV),
which describes the spatial variation in a quantitative manner. This is rather similar
to covariance in the classical time series analysis of stochastic processes. Since its
origin in the 1960s, few other alternatives as the cumulative SV (CSV), point CSV
(PCSV), and spatial dependence function based on these conceptions are developed
and applied in different aspects of earth sciences. Each one of these techniques is ex-
plained in this book, and their various uses in modeling several earth sciences events,
such as earthquake, meteorology, hydrology, are presented. Various alternatives of

vii
viii Preface

the Kriging methodology are presented and the necessary steps in their applications
are exposed in a rather simple manner. Simple spatial variation prediction method-
ologies are also revised with up-to-date literature, and their connections to the most
advanced spatial modeling methodologies are explained with basic conceptions.
Spatial simulation methodologies in earth sciences are necessary to explore in-
herent variabilities such as those in fracture frequencies, spacing, rock quality des-
ignation, grain size distribution, and many similar random behaviors of the rock
and porous medium. Innovative methodologies are presented for simulation purpose
with convenient applications.
The purpose of the book is to provide a comprehensive presentation of up-to-date
models for utilization in earth sciences and their applications. Some portions of the
textbook will deal with the material already covered in different books and in recent
literature like the top journals related to various topics of earth sciences. However,
a significant part of the book will consist of the original techniques developed and
presented into open literature by the author. Additionally, many unique physical
approaches, field cases, and sample interpretations will be presented prior to the
application of different models.
I could not have completed this work without the love, patience, support, and as-
sistance of my wife, Fatma Şen. I also extend my appreciation to my stay at the Fac-
ulty of Earth Sciences, Hydrogeology Department, and field experience for many
years at this university and later on to my stay with the Saudi Geological Survey
(SGS), where every facility was provided at my disposal for scientific achievements
within different aspects of earth sciences, so that many parts of this book have been
materialized. In the mean time, I am also grateful to the Istanbul Technical Univer-
sity for giving me every opportunity to work in different aspects of earth sciences,
including meteorology, atmospheric sciences, hydrology, and alike topics.

İstanbul, Turkey Zekai Şen

Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Earth Sciences Phenomena . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Variability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Determinism Versus Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Earth, Environment, and Atmospheric Researches . . . . . . . . . . . . . . . 16
1.6 Random Field (RF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.7 Regionalized Variable (ReV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 Data Types and Logical Processing Methods . . . . . . . . . . . . . . . . . . . . . . 21

2.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Numerical Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.5 Number of Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.1 Small Sample Length of Independent Models . . . . . . . . . . . 33
2.5.2 Small Sample Length of Dependent Models . . . . . . . . . . . . . 35
2.6 Regional Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.6.1 Variability Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.6.2 Inverse Distance Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.7 Sub-areal Partition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.7.1 Triangularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.8 Polygonizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.8.1 Delaney, Varoni, and Thiessen Polygons . . . . . . . . . . . . . . . . 52
2.8.2 Percentage-Weighted Polygon (PWP) Method . . . . . . . . . . . 55
2.9 Areal Coverage Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.9.1 Theoretical Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.9.2 Extreme Value Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.10 Spatio-Temporal Drought Theory and Analysis . . . . . . . . . . . . . . . . . 73
2.10.1 Drought Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

ix
x Contents

3 Classical Spatial Variation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.2 Spatio-Temporal Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.3 Spatial Pattern Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.4 Spatial Data Analysis Needs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.5 Simple Uniformity Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.6 Random Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.7 Cluster Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3.8 Nearest Neighbor Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.9 Search Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.9.1 Geometric Weighting Functions . . . . . . . . . . . . . . . . . . . . . . . 103
3.10 Trend Surface Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.10.1 Trend Model Parameter Estimations . . . . . . . . . . . . . . . . . . . 108
3.11 Multisite Kalman Filter Methodology . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.11.1 One-Dimensional Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . 112
3.11.2 Kalman Filter Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

4 Spatial Dependence Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

4.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.2 Isotropy, Anisotropy, and Homogeneity . . . . . . . . . . . . . . . . . . . . . . . 129
4.3 Spatial Dependence Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.4 Spatial Correlation Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.4.1 Correlation Coefficient Drawback . . . . . . . . . . . . . . . . . . . . . 136
4.5 Semivariogram Regional Dependence Measure . . . . . . . . . . . . . . . . . 140
4.5.1 SV Philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.5.2 SV Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.5.3 SV Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.6 Sample SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.7 Theoretical SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
4.7.1 Simple Nugget SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
4.7.2 Linear SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
4.7.3 Exponential SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
4.7.4 Gaussian SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
4.7.5 Quadratic SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
4.7.6 Rational Quadratic SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
4.7.7 Power SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
4.7.8 Wave (Hole Effect) SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
4.7.9 Spherical SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
4.7.10 Logarithmic SV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
4.8 Cumulative Semivariogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
4.8.1 Sample CSV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4.8.2 Theoretical CSV Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
4.9 Point Cumulative Semivariogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
4.10 Spatial Dependence Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Contents xi

5 Spatial Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

5.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
5.2 Spatial Estimation of ReV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
5.3 Optimum Interpolation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
5.3.1 Data and Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
5.4 Geostatistical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
5.4.1 Kriging Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
5.5 Geostatistical Estimator (Kriging) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
5.5.1 Kriging Methodologies and Advantages . . . . . . . . . . . . . . . . 230
5.6 Simple Kriging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
5.7 Ordinary Kriging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
5.8 Universal Kriging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
5.9 Block Kriging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
5.10 Triple Diagram Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
5.11 Regional Rainfall Pattern Description . . . . . . . . . . . . . . . . . . . . . . . . . 256
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266

6 Spatial Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271

6.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
6.2 3D Autoregressive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
6.2.1 Parameters Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
6.2.2 2D Uniform Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . 276
6.2.3 Extension to 3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
6.3 Rock Quality Designation Simulation . . . . . . . . . . . . . . . . . . . . . . . . . 281
6.3.1 Independent Intact Lengths . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
6.3.2 Dependent Intact Lengths . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
6.4 RQD and Correlated Intact Length Simulation . . . . . . . . . . . . . . . . . . 300
6.4.1 Proposed Models of Persistance . . . . . . . . . . . . . . . . . . . . . . . 303
6.4.2 Simulation of Intact Lengths . . . . . . . . . . . . . . . . . . . . . . . . . . 305
6.5 Autorun Simulation of Porous Material . . . . . . . . . . . . . . . . . . . . . . . . 310
6.5.1 Line Characteristic Function of Porous Medium . . . . . . . . . 312
6.5.2 Autorun Analysis of Sandstone . . . . . . . . . . . . . . . . . . . . . . . 312
6.5.3 Autorun Modeling of Porous Media . . . . . . . . . . . . . . . . . . . . 316
6.6 CSV Technique for Identification of Intact Length Correlation
Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
6.6.1 Intact Length CSV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
6.6.2 Theoretical CSV Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
6.7 Multidirectional RQD Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
6.7.1 Fracture Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
6.7.2 RQD Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
6.7.3 RQD Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
Chapter 1
Introduction

Abstract Earth sciences events have spatial, temporal, and spatio-temporal vari-
abilities depending on the scale and purpose of the assessments. For instance, in
geology, geophysics, rock mechanicals, and alike aspects the spatial variability is
dominant, whereas in hydrology, meteorology, etc., aspects the temporal variability
has preference; but in general, irrespective of time scale, all of the earth sciences
events have spatio-temporal variability. Plate tectonics has temporal variations dur-
ing millions of years whereas spatial variability exists at any time instant. Hence,
variability is one of the major factors in the modeling of earth sciences events for
their future behavior predictions or estimations at any given location where there is
no sampling information. Variability implies similarity at some instances and there-
fore similar representative patterns play significant role in earth sciences modeling.

Keywords Atmosphere · Geometry · Dynamism · Kinematic · Model ·

Randomness · Random field · Regionalization · Regionalized variable · Similarity ·
Variability

1.1 General

Mathematical, statistical, probabilistic, stochastic, and alike procedures are

applicable only in the case of spatial or temporal variability in natural or artificial
phenomena. It is not possible to consider any approach of earth sciences phenomena
without the variability property, which is encountered everywhere almost explicitly
but also implicitly at times and places. Explicit variability is the main source of
reasoning, but implicit variability leads to thousands of imagination with different
geometrical shapes on which one is then able to ponder and generate ideas and
conclusions. It is possible to sum up that variability is one of the fundamental ingre-
dients of philosophical thinking, which can be separated into different rational com-
ponents by mental restrictions based on the logic. Almost all social, physical, eco-
nomical, and natural events in small scales and phenomena in large scales include
variability at different levels (micro, meso, or macro) and types (geometry, kinemat-
ics, and dynamics). Hence, the very word “variability” deserves detailed qualitative
understanding for the construction of our future quantitative models.

Z. Şen, Spatial Modeling Principles in Earth Sciences, 1

DOI 10.1007/978-1-4020-9672-3_1,  C Springer Science+Business Media B.V. 2009
2 1 Introduction

Proper understanding of earth sciences phenomenon is itself incomplete, rather

vague, and cannot provide a unique or precise answer. However, in the case of data
availability the statistical methods support the understanding of the phenomenon
and deducing meaningful and rational results. These methods suggest the way to
weight the available data so as to compute the best estimates and predictions with
acceptable error limits. Unfortunately, statistics and use of its methods are taken as
cookbook procedures without fundamental rational background in problem solving.
There are many software programs available to use, but the results cannot be inter-
preted in a meaningful, plausible, and rational manner for the service of practical
applications or further researches.
It is, therefore, the purpose of this book to provide fundamentals, logical basis,
and insights into spatial statistical techniques that are frequently used in engineering
applications and scientific research works. In this manner, prior to cookbook pro-
cedure applications and software use, the professional can give his/her expert view
about the phenomenon concerned and the treatment alternatives of the available
data. The major problems in the spatial analysis are the point estimation from a set
of data sampling points where the measurements are not found, areal average calcu-
lations, and contour mapping of the regionalized variables. The main purpose of this
chapter is to layout the basic spatial variability ingredients, their simple conceptual
grasp and models.

1.2 Earth Sciences Phenomena

Geology, hydrology, and meteorology are the sciences of lithosphere, hydrosphere,
and atmosphere, which consist of different rock, water, and air layers, their occur-
rences, distribution, physical and chemical genesis, mechanical properties, and
structural constitutions and interrelationships. It is also one of the significant sub-
jects of these phenomena to deal with the historical evolution of rock, water, and
atmosphere types and masses concerning their positions, movement, and internal
contents. These features indicate the wide content of earth sciences studies. Unfor-
tunately, these phenomena cannot be simulated under laboratory conditions and,
therefore, field observations and measurements are the basic information sources of
information.
In large scales geological, hydrological, and meteorological compositions are
very anisotropic and heterogeneous; in small scales their homogeneity and isotropy
increase, but the practical representativity decreases. It is therefore necessary to
study them in rather medium scales that can be assumable as homogeneous and
isotropic. In any phenomenon study, the general procedure is to have its behavioral
features and properties at small locations and by correlation to generalize to larger
scales. Unfortunately, most often the description of earth sciences phenomena pro-
vides linguistic and verbal qualitative interpretations that are most often subjective
and depend on the common consensus. The more the contribution to such consen-
sus of experts the better the conclusions and interpretations, but even then it is not
1.2 Earth Sciences Phenomena 3

possible to estimate or derive general conclusions on objective basis. It is, therefore,

necessary to try and increase the effect of quantitative approaches and methodolo-
gies in earth sciences, especially by the use of personal computers, which help to
minimize the calculation procedures and time requirements by software. However,
software programs are attractive in appearance and usage, but without fundamental
handling of procedures and methodologies, the results from these programs cannot
be done properly with useful interpretations and satisfactory calculations that are
acceptable by common expertise. This is one of the very sensitive issues that are
mostly missing in software program training or ready software uses.
It is advised in this book that without knowing the fundamentals of earth sciences
procedure, methodology, or data processing, one must refrain the use of software
programs. Otherwise, mechanical learning of software from colleagues, friends, or
during tutorials with missing fundamentals does not lead the user to write proper
reports or even to discuss the results of software with somebody who is an expert in
the area but may not know software use.
Phenomena in earth sciences are of multitude types and each needs at times spe-
cial interpretation, but whatever the differences there is one common basis, which
is data processing. Hence, any earth science equipped with proper data-processing
technique with fundamentals can help others and makes significant contributions
to open literature, which is of great need for such interpretations. Besides, what-
ever the techniques and technological levels reached, these phenomena show espe-
cially spatial (length, area, volume) differences at many scales, areas, locations, and
depths and, consequently, development of any technique cannot cover the whole of
these variations simultaneously, and there is always an empty space for further inter-
pretations and researches. It is possible to consider the earth sciences all over the
world in two categories, namely conventionalists that are more descriptive working
groups with linguistic and verbal interpretations and conclusions with very basic
and simple calculations. The second group includes those who are well equipped
with advanced quantitative techniques starting with simple probabilistic, statistics,
stochastic, and deterministic mathematical models and calculation principles. The
latter group might be very addicted to the quantitative approaches with little linguis-
tic, verbal, i.e., qualitative interpretations. The view taken in this book remains in
between where the earth scientist should not ignore the verbal and linguistic descrip-
tive information or conclusions, but he/she also looks for quantitative interpretation
techniques to support the views and arrive at general conclusions. Unfortunately,
most often earth scientists are within these two extremes and sometimes cannot
understand each other although the same phenomena are considered. Both views
are necessary but not exclusive. The best conclusions are possible with a good com-
bination of two views. However, the priority is always with the linguistic and ver-
bal descriptions, because in scientific training these are the first stones of learning
and their logical combinations rationalistically constitute quantitative descriptions.
Even during the quantitative studies, the interpretations are based on descriptive,
qualitative, linguistic, and verbal statements. It is well known in scientific arena that
the basic information is in the forms of logical sentences, postulates, definitions,
or theorems, and in earth sciences even speculations might be proposed prior to
4 1 Introduction

quantitative studies and the objectivity increases with shifting toward quantitative
interpretations.
Any geological phenomenon can be viewed initially without detailed information
to have the following general characteristics:
1. It does not change with time or at least within the lifetime of human, and
therefore geological variations have spatial characters and these variations can be
presented in the best possible way by convenient maps. For example, geological
description leads to lithological variation of different rock types. The simplest of
such descriptions is the spatial variation of three basic rock types in any study area.
For instance, in Fig. 1.1 the sedimentary, igneous, and metamorphic rock lithology
is shown. In this simplest classification of rocks, the researcher is not able to look
for different possibilities or ore reserves, water sources, oil field possibilities, etc.
Oil reserves cannot be in igneous or metamorphic rock units and, therefore, he/she
has to restrict the attention on the areas of sedimentary rocks.
On the other hand, a more detailed geological map can be obtained for different
rock types as shown in Fig. 1.2 for the Arabian Peninsula where the subunits of three
basic rock types are shown. Now, it is possible to look for regions of groundwater,
which is in connection with atmosphere, i.e., the rainfall leads to direct infiltra-
tion. This information implies shallow groundwater possibilities and consequently
quaternary sediments (wadi alluviums of present epoch geological activity) can be
delimited from such a map. Each of the above mentioned maps shows spatial varia-
tion of geological phenomena.
2. Geological units are neither uniform nor isotropic nor heterogeneous in hori-
zontal and vertical extends. It is obvious from the maps in Figs. 1.1 and 1.2 that in
any direction (NS, EW, etc.) the geological variation is not constant. Of course, the
vertical direction will include succession of different rock types and subunits, which
are referred to as stratigraphic variation. Such a stratigraphic section is shown in
Fig. 1.3, where neither the thickness of each unit nor their slopes are constant.
It is possible to conclude from these two points that the spatial geometry of geo-
logical phenomena is not definite and, furthermore, its description is not possible
with Euclidean geometry, which is based on lines, planes, and volumes of regular
shapes. However, in practical calculations, the geometric dimensions are simplified

Igneous
Metamorphic

Sedimentary
Fig. 1.1 Three rock type
map
1.2 Earth Sciences Phenomena 5

Fig. 1.2 Rock subunit map

so as to use Euclidean representations and make simple calculations. Otherwise,

just the geometry can be intangible, hindering calculations. This indicates that in
calculations besides other earth sciences quantities that will be mentioned later in
this book, initially geometry causes approximations in the calculations, and in any
calculation geometry is involved. One can conclude from this statement that even
if the other spatial variations are constant, approximation in geometry gives rise
to approximate results. It is very convenient to mention here that the only definite
Euclidean geometry exists at rock crystalline elements. However, their accumulation
leads to irregular geometry that is not expressible by classical geometry.
3. Rock materials are not isotropic and heterogeneous even at small scales of
few centimeters, except at crystalline or atomic levels. Hence, the compositional
variation within the geometrical boundaries appears from a point to another, which
makes the quantitative appreciation almost impossible. In order to alleviate the situ-
ation mentally, researchers enter an idealistic world by simplifications for grasping
the features as isotropic and homogeneous. If one views the thin section of any
rock as shown in Fig. 1.4, he/she will appreciate detailed minute variations from
point to point and conclude that the rock cannot be considered as isotropic and
homogeneous.
6 1 Introduction

Fig. 1.3 Stratigraphic Qal

Qal ALLUVIUM – Along wadi; sand, gravel.
sequence

Qes EOLIAN SAND

.

Qsb SABKHAH DEPOSITS – Silt, clay and gypsum

Qu SAND, GRAVEL, SILT

Tb BASALT FLOW

Tsg TERRACE SAND, GRAVEL, GYPSUM

UNCONFIRMITY

QXs
Xs
Q DIFFERENTIATED SANDSTONE

Fig. 1.4 Thin-section spatial

variability

4. Isotropy implies uniformity along any direction, i.e., directional property con-
stancy. The homogeneity means constancy of any property at each point. These
properties can be satisfied in the artificial material produced by man, but natural
material such as rocks and any natural phenomenon in hydrology and meteorology
1.2 Earth Sciences Phenomena 7

cannot have these properties in the absolute sense. However, provided that the direc-
tional or point-wise variations are not very appreciably different from each other,
then the geological medium can be considered as homogeneous and isotropic on the
average. In this last sentence, the word “average” is the most widely used parame-
ter in quantitative descriptions, but there are many other averages that are used in
the earth sciences evaluations. If there is not any specification with this world, then
it will imply arithmetic average. Arithmetic average does not attach any weight or
priority to any point or direction, i.e., it is an equal weight average.
5. It can be concluded from the previous points that spatial variations cannot
be deterministic in the sense of isotropy and homogeneity, and therefore they can
be considered as non-deterministic, which implies uncertainty and in turn it means
that the spatial assessments and calculations cannot be adopted as the final crisp
value. At this stage, rather than well-founded deterministic mathematical rules of
calculation and evaluation, it is possible to deal with spatial assessments and eval-
uations by uncertainty calculations, which are probability, statistics, and stochastic
processes.
6. Apart from the geometry and material features, the earth sciences event media
also includes in a non-deterministic way the tectonic effects such as fissures, frac-
tures, faults, folds, or chemical solution cavities, which appear rather randomly.
Figure 1.5 indicates the appearance of these features in some of the geological
media.
From these explanations, it is a scientific truth that the earth sciences phenom-
ena cannot be studied with deterministic methodologies for meaningful and useful
interpretations or applications. The non-deterministic, i.e., uncertainty techniques
such as probability, statistics, and stochastic, methodologies are more suitable for
the reflection of any spatial behavior.

Fig. 1.5 Structural

non-determinacy
8 1 Introduction

1.3 Variability
Spatial and temporal variability are the common occurrences in nature. The patterns
caused by the spatial variability of a phenomenon occur at many scales, ranging
from simple chemical reactions between activator and inhibitor reagents to the large-
scale structure of the universe. For example, spiral wave patterns are well known
in chemical reactions, hurricanes, galaxies, and simulated predator–prey systems.
Because common mechanisms may be responsible for generating patterns among
diverse physical, chemical, and biological systems, new pattern formation phenom-
ena are potentially of great scientific interest. The mechanisms that cause patterns
to evolve, however, are not well understood. An understanding of the causes and
consequences of pattern formation in nature will increase our understanding of pro-
cesses in nature such as succession, spread, and persistence of species and manage-
ment. With the ability to detect and describe different patterns comes the power to
discover the determinants of patterns and the mechanisms that generate, maintain,
modify, and destroy those patterns.
Spatial variability from one site to another leads to the concept of regional vari-
ability within the area. This variability determines the regional behavior as well as
the predictability of the precipitation amounts on the basis of which the interpreta-
tions are derived, provided that suitable techniques and models are identified. For
spatial variability the classical time series techniques yield useful interpretations, but
for equal-distance sampling only. However, a great deal of progress has been made
in the adaptation of statistical techniques to unevenly sampled data (North et al.,
1982). These techniques do indeed yield useful information, which is significantly
different from the information obtained by the use of semivariogram technique in
geostatistics (Chapter 4). Regular scatter of sites might not provide enough regional
information as irregular sites since earth sciences agents and surface features are
almost always heterogeneous and anisotropic. Consequently, the following signifi-
cant questions remain unanswered so far in the literature on spatial (regional) earth
sciences assessments:

1) How to quantify from irregular site measurements whether the regional distribu-
tion is homogeneous, continuous, dependent, etc.?
2) How to model the heterogeneity so as to represent continuous variability within
the area concerned?
3) How to construct maps concerning the regional variability such that the estimates
are unbiased?

Earth sciences variability amounts in any area show considerable spatial and
temporal variations. The variation is brought about by differences in the type and
scale of development in event-producing processes and also influenced strongly by
local or regional factors such as topographic elevations and atmospheric conditions
(Wilson and Atwater, 1972). In practice, however, the variation of event is con-
sidered to be significantly site- or, at the very least, area-dependent. In addition,
1.3 Variability 9

for most areas in the world each individual measurement site is assumed to be
representative of a very considerable area around it. Logically, measurement at any
individual site will have an area of influence around it, but there is no physical or
data-based objective criterion for the definition of such an area (Chapters 2 and 4).
The assumption of very considerable spatial variation is a very dangerous one, espe-
cially at short distances and for severe conditions. For instance, there is no guarantee
that point rainfall will in any way produce a reliable guide to the rainfall of imme-
diate surrounding areas (Summer, 1988).
The spatial distribution of any event has always been a factor important in many
earth sciences analyses. Reliable estimations of average values are essential for
adequate planning. The same is true for water balance calculations, groundwater
recharge boundaries, urban drainage purposes, and for climatic conditions. Hence,
in practice the basic problem is that of point representativeness and, subsequently,
further problem is how to derive pictures of spatial relationships, which are reason-
ably close to reality. In overcoming this problem, it is important to determine the
spatial pattern based on a sparse and uneven network of measuring stations.
Variability is a general term used for a fundamental property of all events, which
is boiled down in a more quantitative term “difference.” The equivalence of these
two words at least in the implications provides a bridge between the philosophical
thinking and the rational quantitative modeling of social, economical, and natural
events. The word difference as an expression for variability gives rise to various
algorithms that are currently in use in many disciplines. For instance, the global
numerical values cannot explain the internal or external features of the concerned
event except its scale, but the comparison of any two leads to additional and detailed
information such as the difference (with dimension of the numerical values) and unit
difference (slope). There are many categories of variability such as geometric, kine-
matics, and dynamic types. In addition, the variability may be in regular (certain) or
irregular (uncertain) domains.
The geometrical similarity can be split into two types, namely size corresponding
to scale and shape.
In Fig. 1.6a, three shapes are exactly the same but their scales are different. How-
ever, they cannot be considered as having different variability from each other. Espe-
cially, if the last shape in Fig. 1.6a is further downscaled, it approaches a point.
Downscaling or upscaling each one of these shapes helps to convert each to others
easily, and therefore there is no real variability between them. The only variability
is within each shape, because each point on the figure has different coordinates or
slopes (tangents). The variability in shape is different than scale type, because the
overall appearance of different objects may look like almost the same with local
differences as in Fig. 1.6b. Since these shapes are not exactly the same, there are
relative deviations between the two. For the relative difference either of the shapes
may be taken as the basis and the other is considered as deviating from the first
one. Hence, it is possible that the differences are relative not absolute. This implies
the significance of relative variability. The geometric difference is referred to as the
spatial variability in this book.
10 1 Introduction

Fig. 1.6 Geometric similarities: (a) size (b) shape

Kinematics variability is concerned with the similarity in the time variations,

especially concerning the physical definition of the velocity. In a more general
term, any similarity not concerning the force is referred to as the kinematics vari-
ability. Such a variability and consequent similarity or geometry is very signifi-
cant in physical modeling of natural or man-made events. For instance, the labora-
tory modeling of runoff variability in the drainage basin should be similar to each
other.
The most significant value of the variability in natural events is the velocity varia-
tions after the geometrical variability. Figure 1.7 indicates the kinematics variability
and its correspondence between the natural and the laboratory units. In this figure, in
addition to the geometrical similarity there is also kinematics similarity, because at
each subsection the ratio of velocities is constant, that is v1 /V1 = v2 /V2 = v3 /V3 =
v4 /V4 = v5 /V5 = c where c is a constant, less than 1. This constant is the guarantee
for the temporal variability at spatially distributed points.

V2 V1 V2
V1

V4 V3 V4
V3
V5
V5
Fig. 1.7 Kinematics
variability and similarity
1.3 Variability 11

Finally, the dynamic similarity is similar to the kinematics similarity where

instead of velocities forces are taken into consideration. After all, in any spatial
variability, depending on the type of program the variability must be considered
from geometric, kinematics, and dynamics points of view.
During the discussion of the variability, in addition to the basic word “differ-
ence” another one “similarity” emerged rather unintentionally. Similarity is a type
of difference and consequent variability measure between two events. Depending
on the quantitative calculations as will be mentioned in the following chapters, it
is possible to conclude that two events are different or similar. Difference or simi-
larity implies the difference or similarity in the variability. Variability is a general
term that is used for the comparison of multitude of points, but the difference can
be quantified between two points only.
The variability may have not only very regular forms like waves, functions, static
maps but also irregular behaviors that cannot be described by the mathematical or
classical physical rules where uncertainty techniques such as statistics, probability,
and stochastic become helpful (see Fig. 1.8). Uncertain variations are referred to as
the random variables, RV, or regionalized variables, ReV, which assume any value
randomly at any time or space. Whatever the degree of uncertainty, each RV or ReV
may include regular variations, and in general the regional variability may be con-
sidered as the summation of deterministic and stochastic (statistical or probability)
components. Although mathematical concepts, operations, and functions are suffi-
cient for the depiction of deterministic part, they fail to emulate the stochastic part,
which is the subject of this book.
The following points are significant in the quantitative consideration of the vari-
ability in any modeling study:

1) Quantitative variability is possible by pair-wise consideration of two events,

objects, points, numbers, etc.
2) The subtraction operation is the only one among the four arithmetical operations
that represent the variability.
3) It is preferable to use dimensionless values between two different events, say,
groundwater levels and transmissivity map in order to compare their variability.
It is not possible to compare soil moisture content with runoff directly because
they are different in dimensions. In order to bring various events into equal
dimensional footing, it is necessary to use division operator, which cancels the
dimension.

Variability Uncertain (stochastic) part

Certain (deterministic) part

Time

Fig. 1.8 Deterministic and stochastic variations

12 1 Introduction

After the foregoing logical explanations the two most significant explanations for
the variability measurement appear as difference and division. Furthermore, these
operations can be applied between two points for a given ReV or between two ReVs
at a single point. Hence, if there are n measurements or sampling points, there are
then n(n–1)/2 different ways of distance variability samplings. The variability of
ReVs is represented most often by the arithmetic or preferably weighted-average
procedure. This indicates for a multitude of point variability, the summation opera-
tion provides a general representative value for the regional variability of ReVs.

1.4 Determinism Versus Uncertainty

In general, any field, office, or theoretical study relating to natural phenomena that
occur within the lithosphere, hydrosphere, or atmosphere is referred to as the earth
sciences event. For instance, the lithological compositions in terms of stratigraphy,
historical evolutions, structural and tectonic dynamism are under continuous spa-
tial and temporal changes, and therefore their analysis, control, and predictions are
quite difficult. The difficulty arises not only from the insufficiency of the techniques
only, but inherently from the random behavior of the earth sciences elements in an
area. In earth sciences concerning lithology, there are no significant changes tem-
porally within the average human life, but the lithological characteristics change
even after, say, 1 cm apart. For instance, occurrence of fracture systems includ-
ing fracture number, direction, dip, length, aperture, frequency, etc., are all very
complicated and sophisticated for measurement; and, consequently, right from the
field studies, there arise uncertainties either due to the human ignorance or due to a
significant extent as a result of inherent randomness that exists in nature. Another
example is the rate, time, and direction of individual pebble movement along a river,
and this phenomenon cannot be predicted with any reasonable degree of determin-
ism. It is possible to state different premises and then conclusions deduced logically
from a set of premises, but all these are in terms of linguistic variables and state-
ments. It is the role of quantification methods to translate these verbal statements
into mathematical equations or inequalities so as to come closer to final conclusions
or decisions. Idealization, assumption, and simplification in earth sciences are more
difficult than the atmospheric sciences where the sky does not have, say, any fracture
or discontinuity as in the subsurface. Hence, in earth sciences more than any other
physical or atmospheric sciences, the element of randomness is embedded in the
field, office, and theoretical studies. Unfortunately, geologists have worked so many
years in a descriptive manner and did not care about the uncertainty techniques such
as probability and statistics in their quantitative studies until 1960. However, then
onward many statistical and specifically developed statistics in geology as geostatis-
tics started to be used in many ore reserve, groundwater, petroleum, etc., studies.
The movement of the pebbles, sand, silt, and clay particles may be described as ran-
dom within the limits of environmental parameters. This randomness of movement
is interpreted as a result of chance events inherent in the transporting medium and
the transported material.
1.4 Determinism Versus Uncertainty 13

Deterministic phenomena are those in which outcomes of the individual events

are predictable with complete certainty under any given set of circumstances, if the
required initial conditions are known. In all the physical and astronomical sciences,
traditionally deterministic nature of the phenomena is assumed. It is, therefore, nec-
essary in the use of such approaches the validity of the assumption sets and initial
conditions. In a way, with idealization concepts, assumptions, and simplifications,
deterministic scientific researches yield conclusions in the forms of algorithms,
procedures, or mathematical formulations, which should be used with caution for
restrictive circumstances. The very essence of determinism is the idealization and
assumptions so that uncertain phenomenon becomes graspable and conceivable to
work with the available physical concepts and mathematical procedures. In a way,
idealization and assumption sets render uncertain phenomenon into conceptually
certain situation by trashing out the uncertainty components. A significant question
that may be asked at this point is whether there is any benefit from the determin-
istic approaches in the earth and atmospheric studies if the general behavior of the
events is uncertain? The answer to this question is affirmative because in spite of
the simplifying assumptions and idealizations the skeleton (trend, seasonality) of
the uncertain phenomenon is captured by the deterministic methods. For instance,
the diurnal temperature change at any location has more or less a sinusoidal fluc-
tuation as shown in Fig. 1.9, and the idealization of temporal temperature changes
within a day as a sinusoidal regular curve provides a skeleton on which the uncertain
parts appear as deviations in a random manner.
At the very core of scientific theories lies the notion of “cause” and “effect” rela-
tionship as an absolute certainty in scientific studies. One of the modern philoso-
pher of science, Popper (2002) stated that “to give a causal explanation of a certain
specific event means deducing a statement describing this event from two kinds of
premises: from some universal laws, and from some singular or specific statements
which we may call the specific initial conditions.” According to him, there must be
a very special kind of connection between the premises and the conclusions in a
causal explanation and it must be deductive. In this manner, the conclusion follows
necessarily from the premises. Prior to any mathematical formulation, the premises
and theconclusion consist of verbal (linguistic) statements. It is necessary to justify
at every step of deductive argument by citing a logical rule that is concerned with
the relationship among statements. On the other hand, the concept of “law” lies at
the heart of deductive explanation, and therefore at the heart of the certainty of our
knowledge about specific events.

Rainfall

Day
Fig. 1.9 Diurnal temperature variation
14 1 Introduction

ReV phenomena evolve with time and space, and their appreciation as to the
occurrence, magnitude, and location is possible by observations and still better by
measurements along the time and/or space reference systems. The basic information
about the phenomenon is obtained from the measurements. Any measurement can
be considered as the trace of the phenomenon at a given time and location. Hence,
it should be specified by time and location, but its magnitude is not at the hand of
researcher. Initial observance of the phenomenon leads to non-numerical form of
descriptive data that cannot be evaluated with certainty.
The use of uncertainty techniques such as the probability, statistics, and stochas-
tic methods in earth, environmental, and atmospheric sciences has increased rapidly
since 1960s and most of the researchers as students and teachers seek more train-
ing in these disciplines for dealing with uncertainty in a better quantitative way.
Many professional journals, books, and technical reports in the earth, environment,
and atmospheric science studies include significant parts on the uncertainty tech-
niques in dealing with uncertain natural phenomena. Yet, relatively few scientists
and engineers in these disciplines have a strong background in school mathematics
and the question is then how can they obtain sufficient knowledge of uncertainty
methods including probability, statistics, and stochastic processes in describing nat-
ural phenomena appreciating the arguments, which they must read and then digest
for successful applications in making predictions and interpretations.
The inability to predict specific events might stem from the fact that nature is
intrinsically random and randomness is used in a statistical sense to describe any
phenomenon that is unpredictable with any degree of uncertainty for a specific
event; and deterministic phenomena, on the contrary, are those in which outcomes
of the individual events are predictable with complete certainty under any given set
of circumstances, if the required initial conditions are known. In general, the nature
is considered as random. Consequently, randomness has been suggested as the ulti-
mate and most profound physical concept of nature.
Moreover, it is almost true to claim that classical approaches (analytical,
empirical, etc.) are not deterministic theory if the claim simply means that actual
measurements confirm the predictions of the theory only approximately or only
within certain statistically expressed limits. Any theory formulated in terms of mag-
nitudes capable of mathematically continuous variation must, in the nature of the
case, be statistical and not quite deterministic in this sense, for the numerical val-
ues of physical magnitudes (such a permeability) obtained by laboratory or field
measurement never form a mathematically continuous series; any set of values
so obtained will show some dispersion around the values calculated from theory.
Nevertheless, a theory is labeled properly as a deterministic one if analysis of its
internal structure shows that the theoretical state of a system at one instant logically
determines a unique state of that system for any other instant. In this sense, and
with respect to the theoretically defined mechanical states of systems, mechanics
is unquestionably a deterministic theory. Consequently, when a predictor is intro-
duced into a philosophical discussion of determinism, it is not a human being but
a “superhuman intelligence.” Human predictors cannot settle the issue of determin-
ism because they are unable to predict physical events no matter what the world is
really like.
1.4 Determinism Versus Uncertainty 15

The uncertainty in the earth and atmospheric knowledge arises out of the
conviction that earth and atmospheric generalizations are immensely complicated
instantiations of abstract, and often universal, physical laws. Earth and atmospheric
generalizations always contain the assumptions of boundary and initial condi-
tions. In a way, the uncertainty in the predictions arises from the ignorance of the
researcher to know the initial and boundary conditions in exactness. They cannot
control these conditions with certainty. On the assumptions of physical theory, earth
and atmospherically significant configurations are regarded as highly complex. This
is true whether or not the “world” is deterministic. Physical laws, which are not
formulated as universal statements, may impose uncertainty directly upon earth and
atmospheric events as in the case of inferences based on the principles of radioactive
disintegration.
Earth and atmospheric sciences deal with spatial and temporal behaviors of nat-
ural phenomena at every scale for the purpose of predicting the future replicas
of the similar phenomena, which help to make significant decisions in planning,
management, operation, and maintenance of natural occurrences related to social,
environmental, and engineering activities. Since any of these phenomena cannot be
accounted by measurements, which involve uncertain behaviors, their analysis, con-
trol, and prediction need to use uncertainty techniques for significant achievements
for future characteristic estimations. Many natural phenomena cannot be monitored
at desired instances of time and locations in space and such restrictive time and loca-
tion limitations bring additional irregularity in the measurements. Consequently, the
analysis, in addition to uncertain temporal and spatial occurrences, has the prob-
lem of sampling the natural phenomenon at irregular sites and times. For instance,
floods, earthquakes, car accidents, fracture occurrences are all among the irregu-
larly distributed temporal and spatial events. Uncertainty and irregularity are the
common properties of natural phenomena measurements in earth and atmospheric
researches, but the analytical solutions through numerical approximations all require
regularly available initial and boundary conditions that cannot be obtained by lying
regular measurement sites or time instances. In an uncertain environment any cause
will be associated with different effects, each with different level of possibilities.
Herein possibility means some preference index for the occurrence of each effect.
The greater the possibility index the more frequent the event occurrence. Figure 1.10
indicates the deterministic and uncertain cause–environment–effect system in a sim-
plified manner. In the deterministic case any single cause will lead to a certain sin-
gle effect, whereas in the uncertain domain single cause (effect) might be associated

Input GEOLOGICAL
GEOLOGICAL Output
ENVIRONMENT

Outputs
Fig. 1.10 Cause, Input GEOLOGICAL
environment, and effect ENVIRONMENT
triplicate
16 1 Introduction

Outputs
???????? Earth science
environment Filtration

Outputs
Input ??????????
???????? Identification

????????
Input Earth science
environment Prediction

Fig. 1.11 Estimation, identification, and prediction

with a multitude of effect (cause). In practice, however, the case of single-input

multiple-output situations is most valid and the prediction is, in fact, deciding on
which one of the effects will occur under a certain cause.
In this figure any one of the triplicate components might be uncertain. Further-
more, any two or collectively all three components may be uncertain. In all these
cases, the uncertainty plays significant role in the evolution and prediction of the
phenomenon concerned. In the practical studies, given any two of these compo-
nents the third one might be estimated, identified, or predicted. These three cases are
shown in Fig. 1.11. In the case of unknown causal effects, the problem is referred to
as the estimation or filtration; given the cause and effect knowledge, the search for
the functioning environmental system is identification; and, finally, given the cause
as input and the functional duty of the environment, the search for knowledge on
the effect is known as prediction.
Uncertainty is introduced into any problem through the variation inherent in
nature, through man’s lack of understanding of all the causes and effects in physical
systems, and through insufficient data in practice. Even with long history of data,
one cannot predict the natural phenomenon except within an error band. As a result
of uncertainties, the future can never be predicted entirely by the researchers. Con-
sequently, the researcher must consider the uncertainty methods for the assessment
of the occurrences and amounts of particular events and then try to determine their
likelihood of occurrences.

1.5 Earth, Environment, and Atmospheric Researches

Recently, the scientific evolution of methodologies has shown that the more the
researchers try to clarify the boundaries of their domain of interest the more they
become blurred with other domains of research. For instance, a hydrogeologist tries
to model the groundwater pollution as one of the modern nuisances of humanity so
1.6 Random Field (RF) 17

far as the water resources are concerned; he/she needs information about the geo-
logical environment of the aquifers, meteorological and atmospheric conditions for
the groundwater recharge, in addition to social and human settlement environmental
issues for the pollution sources. Hence, many common philosophies, logical basic
deductions, methodologies, and approaches become common to different disciplines
and the data processing is among the most important common topics, which include
the same methodologies applicable to diversity of disciplines. The way that earth,
environmental, and atmospheric scientists frame their questions varies enormously,
but the solution algorithms may include the same or at least similar procedures.
Any natural phenomenon or its similitude occurs extensively over a region, and
therefore its recordings or observations at different locations pose some questions
as, for instance, are there relationships between phenomena in various locations? In
such a question, the time appears as if it is frozen and the phenomenon concerned
is investigated over the area with its behavioral occurrences between the locations.
Answer to this question may be provided descriptively in linguistic, subjective, and
vague terms, which may be understood even by non-specialists in the discipline.
However, their quantification necessitates objective, regionally applicable method-
ologies to ReV, which is one of the purposes of this book.
Another question that may be stated right at the beginning of the research in
the earth, environment, and atmospheric sciences are places different in terms of
the phenomena present there? Such questions are the source of many researches’
interest in the subject. Scientific treatment and interpretation of even error-laden
data lead to significant practical knowledge concerning the oceans and atmosphere.
It is the prime duty of the earth scientists to filter out the meaningful portions of the
data and to model randomly the error part.

1.6 Random Field (RF)

Precipitation evolves temporally as well as spatially. Under given specific circum-
stances, precipitation has only a single value at each instant of time. Such a space–
time distribution is referred to as a field quantity. Since meteorological quantities
cannot be predicted with certainty, they are assumed to be random and hence it
is necessary to study a new class of fields, i.e., random fields. Quantitatively, a
random field is characterized, ξ(s,t), called a random function at a point, s, and
time, t. It is obvious that random field concept is the generalization of a stochastic
process for which time is the scale variable. The expression “the random function
of the coordinate (s,t)” must be understood in the sense that at each point (s,t) of
four-dimensional space–time, the value ξ(s,t) is a random variable and, conse-
quently, cannot be predicted exactly. In addition, the values ξ(s,t) are subject to a
certain law of probability. A complete description of a random field can be achieved
by constructing all the finite-dimensional probability distribution functions (pdf) of
the field at different points in space. A random field ξ(s) at a fixed time instant with
the given finite dimensional pdf is called homogeneous, if this function is invariant
18 1 Introduction

with respect to a shift of the system of points (Chapter 4). The same random field
is called statistically homogeneous and isotropic (if it is homogeneous in the sense
indicated above), while the pdfs are invariant with respect to an arbitrary rotation of
the system of points such as of a solid body and to a mirror reflection of this sys-
tem with respect to the arbitrary plane passing through the origin of the coordinate
system. In other words, statistical moments depend upon the configuration of the
system of points for which they are formed, but not upon the position of the system
in space.

1.7 Regionalized Variable (ReV)

When the variable is distributed in space, it is regionalized. An ReV has the char-
acteristics of a certain phenomena. For example, ore grade, rainfall, transmissivity,
and water level data are regionalized variables. This theory was first applied in min-
ing, rapidly widening its range in the earth sciences. The procedure of the theory of
regionalized variables follows two main steps. The first step establishes the theoreti-
cal ground for expressing the structural properties of the phenomenon (experimental
semivariogram, SV, covariance, or cumulative SV, CSV, see Chapter 5). The second
step provides a model that uses a combination of functions, which guarantee a solu-
tion to the estimation problem (Kriging), and deals with the ReV by using the prob-
abilistic theory of random functions. The ReV is denoted by Z(x) and interpreted as
a realization of RF denoted by z(x); the random function being an infinite collection
of random variables z(x)s. A well-defined unique numerical value is rendered into a
realization of random process. From the field of observations or measurements one
can have only a single realization of that random function. The problem consists of
finding the characteristics of random function, z(x), to make the estimation of the
unknown point possible, and statistical inference will require further hypothesis of
stationarity (David, 1977; Delhomme, 1979; Isaaks and Srivastava, 1989).
As a result of spatial variability in the regionalized variable, it is not possible to
obtain exact and error-free estimates of unknowns. However, spatial analysis proce-
dures weight the available evidence so as to make the best estimates in an unbiased
manner. Sample of any ReV at distinct sites is considered to be a single realiza-
tion of a random function. Let Pi (X) and Hi (X) denote the average areal amount
and elevation, respectively, at station i. Many methods have been proposed for inter-
pretation and mapping. Most of them are quite empirical and many are intuitively
reasonable, but they lack intellectual rigor. However, in order to provide interpreta-
tions on a sound theoretical basis, the spatial attributes are treated as mathematical
variables that are the functions of their positions on or above the earth’s surface.
This is tantamount to saying that their values are treated as functions of the spatial
coordinates. Similar problems of estimating and mapping properties arose in meteo-
rology (Cressman, 1959; Gandin, 1965) and in mining (Matheron, 1963), where the
ore concentrations vary in space. Although Gandin (1965) described the applica-
tion of optimum interpolation developed by Kolmogorov (1941) for estimating the
1.7 Regionalized Variable (ReV) 19

values of atmospheric pressure and rainfall at sites between the recording stations,
he did not propose a definite procedure for the distance, i.e., radius of influence.
However, in mining geology rapid and effective solutions are needed because of the
enormous costs incurred, and therefore it was in mining geology that the advances
in spatial analysis were made.

References
Cressman, G. P., 1959. An operational objective analysis system. Mon. Wea. Rev. 87(10), 367–374.
David, M., 1977. Geostatistical Ore Reserve Estimation. Elsevier Scientific Publishing Company,
Amsterdam, 364 pp.
Delhomme J. P., 1979. Kriging in the hydrosciences. Adv. Water Resour. 1, 251.
Gandin, L. S., 1965. Objective Analysis, Lectures on Numerical Short-range Weather Prediction,
WMO, Regional Training Seminar, pp. 633–677.
Isaaks, E. H., and Srivastava, R. M., 1989. An Introduction to Applied Geostatistics. Oxford Uni-
versity Press, Oxford, 561 pp.
Kolmogorov, A. N., 1941. Interpolation and extrapolation von stationaren zufalligen folgen. Bull.
Acad. Sci. USSR, Ser. Math. 5, 3–14.
Matheron, G., 1963. Principles of geostatistics. Econ. Geology 58, 1246–1266.
North, G. R., Bell, T.L., and Cahalan, F., 1982. Sampling errors in the estimation of Empirical
orthogonal functions. Monthly Weather Rev. 110, 699–706.
Popper, K., 2002. The Logic of Scientific Discovery. Routledge Publishing Company, London and
New York, 479 pp.
Summer, G., 1988. Precipitation Process and Analysis. John Wiley and Sons, New York, 455 pp.
Wilson, J. W., and Atwater, M. A., 1972. Storm rainfall variability over Connecticut. J. Geophys.
Res. 77(21), 3950–3956.
Chapter 2
Data Types and Logical Processing Methods

Abstract Not only numerical but also linguistic data are necessary in the modeling
of earth sciences events. Measurements are sources of numerical data whereas
observations lead to linguistic data. Numerical data include randomness and errors
but linguistic data are rather fuzzy, which means that there are uncertainties in both
data types. Accordingly, the final model results as predictions or estimations include
errors that must be confined within ±5% limits in practical applications. Spatial
estimations can be obtained either on point basis or on sub-areal basis depend-
ing on the refinement of the problem at hand and purpose. In general, longitude
(easting), latitude (northing), and regionalized variable (ReV) value at this location
are necessary for a complete description and establishment of a point-wise spatial
model, where these three values are referred to as triplicate; but in the case of pixel
location its size is also necessary, which leads to four variables (quadruple) for the
description of ReV. Simple classical triangularization, polygonalization techniques
are used in addition to innovative percentage polygon methodology. Droughts are a
kind of spatial earth sciences with coverage area that can be modeled by probabilis-
tic approaches.

Keywords Droughts · Observation · Measurement · Linguistic data · Numerical

data · Sampling · Small sample · Triangularization · Polygonalization · Percentage
polygon

2.1 General

Scientific and engineering solutions can be given about any earth sciences phe-
nomena through relevant spatial modeling techniques, provided that representa-
tive data are available. However, in many cases it is difficult and expensive to
collect the field data, and therefore it is necessary to make the best use of avail-
able linguistic information, knowledge, and numerical data to estimate the spatial
(regional) behavior of the event with relevant parameter estimations and suitable
models. Available data provide numerical information at a set of finite points, but

Z. Şen, Spatial Modeling Principles in Earth Sciences, 21

DOI 10.1007/978-1-4020-9672-3_2,  C Springer Science+Business Media B.V. 2009
22 2 Data Types and Logical Processing Methods

the professional must fill in the gaps using information, knowledge, and understand-
ing about the phenomena with expert views. Data are the treasure of knowledge
and information leading to meaningful interpretations. Observations are also poten-
tial source of information, which provides linguistic rational and logical expres-
sions in the form of premises. Data imply numerical measurements using different
instruments either in the field or in the laboratory. Observations are not numer-
ical but rather verbal data that assist to describe and identify the phenomenon
concerned.
The development of data estimation methods can be traced back to Gauss (1809),
who suggested the technique of deterministic least-squares approach and employed
it in a relatively simple orbit measurement problem. The next significant contri-
bution to the extensive subject of estimation theory occurred more than 100 years
later when Fisher (1912), working with pdf, introduced the approach of maximum
likelihood estimation. However, Wiener (1942, 1949) set forth a procedure for the
frequency domain design of statistically optimal filters. The technique addressed
the continuous-time problem in terms of correlation functions and the continuous
filter impulse response. Moreover, the Wiener solution does not lend itself very
well to the corresponding discrete-data problem, nor it is easily extended to more
complicated time-variable, multiple-input/output problems. It was limited to sta-
tistically stationary processes and provided optimal estimates only in the steady-
state regime. In the same time period, Kolmogorov (1941) treated the discrete-time
problem.
In this chapter, observation and data types are explained and their preliminary
simple logical treatments for useful spatial information deductions are presented
and applied through examples.

2.2 Observations
They provide information on the phenomenon through sense organs, which can-
not provide numerical measurements but their expressions are linguistic (verbal)
descriptions. In any study, the collection of such information is unavoidable and they
are very precious in the construction of conceptions and models for the control of the
phenomenon concerned. Observations may be expressed rather subjectively by dif-
ferent persons, but experts may deduce the best set of verbal information. Depending
on the personal experience and background, an observation may instigate differ-
ent conceptualization and impression on each person. In a way observations pro-
vide subjective information about the behavior of the phenomenon concerned. In
some branches of scientific applications, observational descriptions are the only
source of data that help for future predictions. Even though observations may be
achieved through some instruments as long as their description remains in verbal
terms, they are not numerical data. Observations were very significant in the early
developments of the scientific and technological developments especially before the
2.2 Observations 23

17th century, but they became more important in modern times including linguistic
implications and logical deductions explaining the fundamentals of any natural or
man-made event (Zadeh, 1965). For instance, in general, geological description of
rocks can be made by field observations and concise linguistic categorizations are
then planned for others to understand again linguistically. It is important to stress
at this point that linguistic expressions of observations help to categorize the event.
Although such a categorization set forward crisp and mutually exclusive classes
according to classical Aristotelian logic, recently, fuzzy logic classification includ-
ing mutually inclusive classes is suggested by Zadeh (1973) and it is most frequently
used in every discipline in an increasing rate. In many disciplines observations are
extremely important such as in geological sciences, medicine, social studies, phys-
iology, military movements, economics, social sciences, meteorology, engineering,
etc. At times they are more valuable than numerical data, but unfortunately, their
role is almost forgotten due to recent mechanical and software programs that work
with numerical data.

Example 2.1
What type of observational information can be obtained when one takes hand speci-
men from a rock? Even a non-specialist in geology try to deduce the following basic
linguistic information based on his/her observation and inspection of the specimen
through a combined use of his/her sense organs.

1) Shape: Regular, irregular, round, spiky, elongated, flat, etc. This information
can be supported for detailed knowledge, with the addition of adjectives such as
“rather,” “quite,” “extremely,” “moderately.” Note that these words imply fuzzy
information.
2) Color: Any color can be attached to whole specimen or different colors for dif-
ferent parts. Detailed information can be provided again by fuzzy adjectives
such as “open,” “dark,” “gray.”
3) Texture: The words for the expression of this feature are “porous,” “fissured,”
“fractured,” “sandy,” “gravelly,” “silty,” etc.
4) Taste: The previous descriptions are through the eye but the tongue can also
provide information as “saline,” “sour,” “sweet,” “brackish,” and so on.
5) Weight: It is possible to judge approximate weight of the specimen and have
description feelings as “light,” “heavy,” “medium,” “very heavy,” “floatable,”
and likewise other descriptions can also be specified.
6) Hardness: The relative hardness of two minerals is defined by scratching each
with the other and seeing which one is gouged. It is defined by an arbitrary
scale of ten standard minerals, arranged in Mohr’s scale hardness and subjec-
tively numbered in scale based on degrees of increasing hardness from 1 to 10.
The hardness scale provides guidance for the classification of hand specimen
according to Table 2.1, where the verbal information is converted to a scale
through numbers.
24 2 Data Types and Logical Processing Methods

Table 2.1 Mohr’s scale of hardness

Hardness scale Mineral Example

1 Talc Hydrated magnesium silicate

2 Gypsum Hydrated calcium sulfate
3 Calcite Calcium carbonate
4 Fluorspar Calcium fluoride
5 Apatite Calcium phosphate
6 Feldspar Alkali silicate
7 Quartz Silicate
8 Topaz Aluminum silicate
9 Corundum Alumina
10 Diamond Carbon

Example 2.2
Earthquake effect on structures can be described according to Table 2.2 guidance,
which is referred to Mercalli scale. The following is an abbreviated description of
the 12 scales of Modified Mercalli intensity.

Table 2.2 Mercalli earthquake intensity scales

Scales Description

I Not felt except by a very few under especially favorable conditions.

II Felt only by a few persons at rest, especially on upper floors of buildings. Delicately
suspended objects may swing.
III Felt quite noticeably by persons indoors, especially on upper floors of buildings.
Many people do not recognize it as an earthquake. Standing motor cars may rock
slightly. Vibration similar to the passing of a truck. Duration estimated.
IV I. Felt indoors by many, outdoors by few during the day. At night, some awakened.
Dishes, windows, doors disturbed; walls make cracking sound. Sensation like
heavy truck striking building. Standing motor cars rocked noticeably.
V Felt by nearly everyone; many awakened. Some dishes, windows broken. Unstable
objects overturned. Pendulum clocks may stop.
VI Felt by all, many frightened. Some heavy furniture moved; a few instances of fallen
plaster. Damage slight.
VII Damage negligible in buildings of good design and construction; slight to moderate
in well-built ordinary structures; considerable damage in poorly built or badly
designed structures; some chimneys broken.
VIII Damage slight in specially designed structures; considerable damage in ordinary
substantial buildings with partial collapse. Damage great in poorly built
structures. Fall of chimneys, factory stacks, columns, monuments, walls. Heavy
furniture overturned.
IX Damage considerable in specially designed structures; well-designed frame
structures thrown out of plumb. Damage great in substantial buildings, with
partial collapse. Buildings shifted off foundations.
X Some well-built wooden structures destroyed; most masonry and frame structures
destroyed with foundations. Rail bent.
XI Few, if any (masonry), structures remain standing. Bridges destroyed. Rails bent
greatly.
XII Damage total. Lines of sight and level are distorted. Objects thrown into the air.
2.3 Numerical Data Types 25

It is important to notice that the linguistic descriptions and scales are neither
time- nor space-dependent, but they have event basis. The reference to any system
is not required apart from logical rules.

2.3 Numerical Data Types

Evolution of any event takes place both in time and in space, but depending on
the practical purposes they can be viewed either temporally or spatially or spatio-
temporally. Accordingly, instruments yield numerical data based on either time or
space reference system. In this book, only spatial data interpretations and treatment
processes are presented. It is assumed herein that the spatial phenomena continu-
ously cover the whole study area. Since the most elemental part of space is a point
in earth sciences, the basic sampling locations are points that may be scattered in
the study area either regularly or irregularly. Theoretically, there is infinite number
of points, but the sampling of all points is not conceivable practically. There are two
ways of sampling for conceivable studies. These are as follows:

1) Point sampling: The size of points is considered as dimensionless or infinitesi-

mally small, with no practically considerable influence area. This is referred to
as sink or source in classical analytical solutions (see Fig. 2.1).
2) Pixel sampling: The size of each point is appreciable and, therefore, the whole
study area is covered with a mesh of very small squares (see Fig. 2.2).

In general, any point or pixel has a system of longitudes and latitudes. Hence,
the whole earth surface is covered by quadrangles, which may be regarded as large-
scale pixels. In practice, for Cartesian distance, area, and volume calculation pur-
poses, longitudes and latitudes are converted to “northing” and “southing” values
with respect to an initial reference point (see Figs. 2.1 and 2.2). This means that the
elements of spatial point data include triple variables, namely, easting, e, northing,
n, and the spatial variable measured at this location, say, z. In short, we can show

Northing

• Point

Study area boundary

Easting
0

Fig. 2.1 Point sampling in an area

26 2 Data Types and Logical Processing Methods

Northing

Pixel

Study area boundary

Easting
0

Fig. 2.2 Pixel sampling in an area

the point data as a triple, {e, n, z}. Likewise, in addition to these triple values any
pixel includes its resolution size, r, which can be represented by a quadruple, {e, n,
z, r}. Even though the pixel size is small, one can practically calculate the number
of pixels that is necessary to cover a given area.

Example 2.3
If a study region has an area of 45 km2 and pixel dimension is 100×100 m2 , what
is the number of pixels for the representation of the whole region? The number can
be calculated as 45×106 /104 = 4,500 pixels. This simple calculation indicates that
each pixel has an area of influence defined by its square area. However, it is not
possible to make the same calculation for point data, since a point does not have an
area of influence by itself. However, it is possible to define the area of influence for
each point based on a set of neighboring points, as will be explained later in this
chapter.
Irregularity is not necessarily related to randomness and has its own connota-
tion in earth and atmospheric sciences. A very brief distinction between these two
terms is that although randomness is inherent in the behavior of natural events out
of human control, irregularity implies human implications in the measurement and
description or definition of natural events. For instance, the locations of meteorolog-
ical stations or groundwater wells are irregularly scattered in the area or space, and
consequently they do not comply by any regular or random pattern. Once the irreg-
ularity is established, then it does not change with time and space easily until there
are other additional inferences by human. Another good distinction in the earth and
atmospheric sciences between regularity and irregularity can be considered in the
solution of differential equations by numerical techniques. Finite element method
requires a definite and regular mesh to be laid over the solution domain of interest
(study area) with regular mesh. Boundary and initial conditions must be defined at
regular set of nodes. Various types of regularity are shown in Fig. 2.3. In practi-
cal studies, measurements as initial conditions are available at a set of irregularly
scattered station locations and hence there is not a desirable match between these
locations and consequent regular nodes.
2.4 Sampling 27

Fig. 2.3 Different

regularities (a) rectangular;
(b) triangular; (c) pentagonal;
(d) radial

(a) (b)

(c) (d)

Many irregularities appear in practical applications and there is beauty even in

the irregularity of the measurement points. Random sampling may be regarded as
another type of irregular sampling procedure. All the meteorology station loca-
tions, water or oil well sites, soil-sampling points, etc., have irregular patterns in
character. The transfer of knowledge from irregular points to regular nodes will
be dealt with in detail later in this chapter and other chapters. Without such a
knowledge and information transfer, none of the numerical solution models such
as Kriging in earth sciences and especially General Circulation Models (GCM) in
atmospheric sciences and others alike can be successfully applicable for practical
solutions.
In theoretical and practical studies, randomness is accounted by the statistical
and especially probabilistic methods whereas irregularity does not need uncertainty
methodology for their treatment, but for the treatment of random events measured
at these irregular points.

2.4 Sampling

The very significant prospect of sampling is the distribution of measurement sites

within the study area. Depending on the prevailing conditions, sometimes the scatter
of sampling points is already set up due to previous human activities such as water
well locations, oil drillings, settlement areas, roads, etc. However, in detailed studies
28 2 Data Types and Logical Processing Methods

Fig. 2.4 Areal sampling

• • • • • •
patterns • • • • •
• • • • • •
• • • • •
• • • • • •
• • • • •
• • • • • •
• • • • •
• • • • • •
• • • • •
• • • • • •
a b
• •
• • • •
• • • •
• •
• • • • • •
• • •
• • • •
• • • • • • •
• • • • •
• • • • •
• • • • •
• • • • • •
• • •
• • • • • • • •
c d

• • • • •
• • • •
• • • • • •
•
• •
• • •
•
• • • • •
• • •
• • • •• • •• •
•
• • • • • • •
• • ••
• • • • •
e f

at smaller scales the researches have to lay down the set of points so as to sample the
concerned phenomenon in a representative manner. There are different techniques
in deciding about the position of the sampling points. If nothing is known before
hand, then it may seem logical to select the sampling points at nodes or centers of a
suitable mesh over the study area. This is the most uniform sampling procedure as
shown in Fig. 2.4.
These sampling patterns can be divided into three different categories as the reg-
ular, random, and aggregated or clustered. Figure 2.4a and b are of regular sampling
procedures. In Fig. 2.4c and d, the randomness is in small scales and the random pat-
tern remains within the sub-areal regular grids. In Fig. 2.4b although each sub-area
is sampled, in Fig. 2.4d only randomly chosen sub-areas are sampled. In Fig. 2.4a
the maximum distance between the two neighboring points cannot be greater than
twice of the sub-area diagonal length. In Fig. 2.4c the distance is of several times
2.4 Sampling 29

the main diagonal length of the sub-area. Large-scale random sampling patters are
given in Fig. 2.4e, where there is no restriction on the distance between the sampling
points. In Fig. 2.4f there are three clusters of the spatial sampling, each with random
sampling patterns. Such categorized samplings are possible depending on the areal
occurrence of the phenomenon studied. For instance, if ore, water, or oil deposits are
intact from each other at three neighboring areas, then the cluster sampling patterns
arise.
Another feature of spatial sampling points is its uniformity concerning the fre-
quency of occurrence per area. If the density defined in this manner is equal in each
sub-area, then the spatial sampling is uniform; otherwise, it is non-uniform. This
definition implies that the regular sampling points in Fig. 2.4a and b, in addition to
small-scale random pattern in Fig. 2.4c, are all uniform because there is one point
per sub-area.
Uniformity gains significance if there are many sampling points within each sub-
area. In geological or meteorological studies, sub-areas are quadrangles between
two successive longitudes and latitudes. For instance, such a situation is shown
in Fig. 2.5 where each quadrangle has random number of random sampling
points. Hence, the question is whether the sampling point distribution is uniform
or not?
The pixels present a regular mesh over a region, which is conceptually similar
to numerical solution of analytical models. The difference is that in the case of
pixels the measurements (brightness values) are known, but this is not the case in
numerical analysis where the values either at the center of each cell or at each node
are necessary for calculation.
A set of point data from a region on interested event such as the elevations at
25 different sites can be presented in three columns, two of which are location
descriptions and the third column includes the regional variable, which is elevation
in this case (see Table 2.3). This table is an example for triple values as mentioned
before.

• • •
• • • • •
• • • •
•
• • • •
• • • • • • •
• •
• • •
• • • •
• • •
• •
• • • • •
• • • • • •
Northing • • •
• • •
• • • • • •
• • • •
• •
• •• • •
• • • • • •
• •
• • • •
• • • • •
• • • •
• •
• • •• • • •
• • •
• • • • •• •
Fig. 2.5 Quadrangle •
sampling Easting
30 2 Data Types and Logical Processing Methods

Table 2.3 Elevation data

Elevation Elevation
Points Easting Northing (m) Points Easting Northing (m)

1 534371.3 2449324 60 14 569732.8 2353515 750

2 589853.8 2362837 116 15 620966.6 2362115 570
3 539568.8 2432734 570 16 612983.4 2391577 110
4 605430.4 2359240 228 17 596556.5 2397926 700
5 626244 2350163 710 18 619828.7 2399006 60
6 619177.6 2370404 60 19 527607.4 2385670 710
7 570437.2 2454979 520 20 658855.3 2376279 1240
8 560157.4 2447557 280 21 520716.1 2377359 540
9 563857.7 2453414 11 22 655595.2 2355951 280
10 615788.4 2361155 660 23 646632.6 2390925 125
11 543309.8 2427824 720 24 650305.8 2366972 90
12 572452.4 2389806 910 25 641826.4 2348446 350
13 604572 2358312 11

In this table, the scatter of sample points can be obtained from the second and
third columns, which appear as in Fig. 2.6. The first appearance indicates that the
sampling points are irregularly distributed in the study area. However, it is not pos-
sible to say whether their distribution is random or there is a systematic correlation.
This question will be answered in Chapter 3.
This may not be a representative sampling set for the study area because there
are no sampling points in the upper right and lower left parts of the area.
Although the spatial data are understood as the sampling of a variable with
respect to longitudes and latitudes, it is also possible to consider another two
variables instead of longitude and latitude. For instance, similar to Table 2.3 in Table
2.4 calcium, magnesium, and chloride variables are given in three columns. These
values are taken at a single point in space at a set of time instances.

6
x 10
2.46

2.44

2.42
Northing

2.4

2.38

2.36

2.34
Fig. 2.6 Irregular sample 5.2 5.4 5.6 5.8 6 6.2 6.4 6.6
Easting x 105
points scatter
2.5 Number of Data 31

Table 2.4 Calcium, sodium, and chloride measurements

Calcium Sodium Chloride Calcium Sodium Chloride

Point (ppm) (ppm) (ppm) Point (ppm) (ppm) (ppm)

1 135.01 20.76 134.48 16 80.57 45.14 165.41

2 63.11 57.34 194.97 17 133.55 55.47 141.99
3 100.68 42.25 183.24 18 131.69 41.44 126.73
4 88.60 66.59 156.74 19 81.03 35.23 164.19
5 129.13 43.30 182.70 20 129.37 29.48 153.20
6 116.21 40.93 159.03 21 45.79 29.67 179.22
7 85.65 62.31 111.30 22 75.29 54.11 203.53
8 41.85 46.26 103.46 23 121.32 35.14 138.39
9 122.14 30.13 111.18 24 40.99 47.08 192.02
10 84.47 53.61 140.11 25 53.89 27.54 85.94
11 101.54 61.91 169.07 26 60.28 54.89 206.96
12 119.19 20.98 106.39 27 59.87 38.92 100.72
13 132.18 54.06 185.77 28 100.38 63.00 97.85
14 113.82 38.97 145.21 29 67.22 62.68 191.36
15 57.63 61.59 115.56 30 59.88 49.68 170.60

In this case there are three alternatives for data treatment purposes. The following
similar questions can be asked:

1) Does one want the change of chloride with respect to calcium and sodium? or
2) Does one want the change of calcium with respect to chloride and sodium? or
3) Does one want the change of sodium with respect to calcium and chloride?

Of course, each one of these cases presents different scatter diagrams as pre-
sented in Fig. 2.7. In each case, the third variable plays the role of elevation as in
the preparation of topographic maps.
In each case, the triplicate {e, n, z} has {e, n}, {e, z}, and (n, z} alternatives,
which are referred to as scatter diagrams.

2.5 Number of Data

In this book spatial data will imply at the minimum the existence of three points
within the study area, because they represent a spatial form as the simplest plain
trend. However, it must not be understood that only triplets are necessary for spa-
tial modeling. The practical question is how many data are necessary for spatial
data modeling. There is not an easy answer to such a question. The ready answers
as appearing in open literature such as 12 or 30 data points are not logical, but
they may provide vague and practical solutions depending on the circumstances.
If the purpose is to check whether the data confirms with the normal probability
distribution, then 12 data are enough. On the other hand, 30 is a number which is
based on empirical observations that in the representation of average meteorological
32 2 Data Types and Logical Processing Methods

70 220

65
200
60

55 180

Chloride (ppm)
Sodium (ppm)

50
160
45
140
40

35 120
30
100
25

20 80
40 50 60 70 80 90 100 110 120 130 140 40 50 60 70 80 90 100 110 120 130 140
Calcium (ppm) Calcium (ppm)
a b
70

65

60

55
Sodium (ppm)

50

45

40

35

30

25

20
80 100 120 140 160 180 200 220
Chloride (ppm)
c

Fig. 2.7 Sample points scatter diagrams (a) Calcium-sodium; (b) Calcium-chloride; (c) Chloride-
sodium

phenomenon it is adopted by the World Meteorological Organization (WMO) as a

normal period. Unfortunately, neither of these numbers can be adopted in practical
applications without looking at the behavior of the concerned phenomenon features,
properties, and environmental conditions. In order to have answer in a more realistic
manner, the following points must be taken into consideration:

1) If the phenomenon is spatially homogeneous in a deterministic manner, then the

number of sample will be as small as possible. In fact, in the case of perfectly
homogeneous spatial distribution only one sample point is enough. This is a very
theoretical situation and difficult to come across in earth sciences studies. Logi-
cally, in such a case there is no spatial variation and therefore the variance of the
spatial variation is equal to zero. This implies that the smaller the standard devi-
ation the smaller is the representative sample size. Numerical expression of this
situation can be achieved by standard deviation, similarity, or correlation coeffi-
cients (Şen, 2002). The more the correlation coefficient (close to either +1, i.e.,
completely positively proportional or –1 as the negatively inverse), the smaller
will be the data number. Finally, the more the similarity coefficient (equal to 1)
the smaller will be the data number.
2.5 Number of Data 33

2) If the regional phenomenon is randomly homogeneous (heterogeneous), it is

then necessary to take as much data sample (measurement) as possible. How-
ever, there must be an upper limit for the data number, which may be decided
not only on the basis of statistical criterion but also at times more significantly
on the cost of samples and on sampling occasions such as in rare events (floods,
droughts, etc.).
3) Areal coverage of the samples must also be decided by taking into considera-
tion the effective area of influence of each sample, which will be explained in
Chapter 3 in detail.

2.5.1 Small Sample Length of Independent Models

In general, statistical parameters from finite length records are biased estimations
of population (very long samples) counterparts. Estimates whose population values
(expectations) are not equal to the parameter they estimate are said to be biased.
This is due to the lack of complete information concerning the phenomenon and,
consequently, the parameter will be under- or over-estimated with respect to its true
population value. In practice, the true population parameters, which are independent
of sample length, are not known, but it is possible to estimate them from the avail-
able finite length records. As mentioned before, as a rule of thumb it is necessary to
have at least 30 values in order to have normal parameter estimations. However, 30-
year period cannot be equally valid for all random or regionalized variables (ReV).
It is very much a function of the correlation structure. In short, the more the persis-
tence, (correlation) the smaller the necessary data. In practical applications, it is of
great interest to know the number of data that is necessary to have a stable value on
the average. This is equivalent to saying that the variance of the average parameter
estimation must be constant. From this point, Cramer (1946) showed theoretically
for the normal independent ReVs that the variance, VI (x), of the arithmetic averages
for sample length n is

σ2
VI (x) = , (2.1)
n

where σ 2 is the unknown population variance and n is the number of data. This is
also due to the central limit theorem that the average of random samples accords in
normal pdf with mean equal to the average of the data, x, and the variance of the
averages is given as in Eq. (2.1). The square root of this expression is referred to as
the standard error, e, of estimate of arithmetic mean. If the original data come from
an independent ReV process with population mean, μ, and standard deviation, σ,
then the finite sample averages will have the same arithmetic mean with variance
as in Eq. (2.1). This means that the sample average is an unbiased estimate of the
arithmetic mean with a standard error as
34 2 Data Types and Logical Processing Methods

0.4

0.35

0.3

0.25
Probability

0.2

0.15

0.1

0.05 Risk region Risk region

Reliability
region
0
–4 –3 –2 –1 0 1 2 3 4
Standard variate, x

Fig. 2.8 Standard Gaussian pdf

σ
e= √ , (2.2)
n

which decreases with the square root of the sample length. ReVs are sampled over
various space (or time) intervals. It is necessary to take 10 (90)% standard error
(reliability, significance) level corresponding to standard deviate, x90 value from a
standard Gaussian distribution as shown in Fig. 2.8. This level separates the whole
area under the standard normal pdf into two parts as reliability and risk regions (Şen,
1998a).
It is possible to obtain a practical chart between the three variables in Eq. (2.2),
which gives the relationship between the data number depending on the reliability
level as in Fig. 2.9.

Example 2.4
In an extensive area there is an unconfined groundwater aquifer in rather homoge-
nous sandstone. The field tests indicated that the radius of influence of each well is
650 m. So how many samples must be drilled such that there will not be interference
between the adjacent wells in a total region of 5 km2 ? For the numerical answer, it
is necessary to first calculate the area of influence for each well as 3.14(650)2 =
1,326,650 m2 = 1.3266 km2 . Now the number of wells can be found as 5/1.3266 =
3.769 ∼ 4 wells.
2.5 Number of Data 35

100

90
5 10 25 50 100 n = 150
80

70
Standard error

60

50

40

30

20

10

0
0 200 400 600 800 1000 1200 1400
Standard deviation

Fig. 2.9 Sample number with the standard deviation and reliability level

Example 2.5
In a region n = 10 well samples of electric conductivity (EC) in micromhos per
centimeter are recorded as given in Table 2.5. What is the representative number of
data?

Table 2.5 EC samples

Data number 1 2 3 4 5 6 7 8 9 10

EC 770 1020 997 790 750 760 765 850 1029 900

The arithmetic average and the standard deviation of the data values are 863.1
μmhos/cm and 114.82 μmhos/cm, respectively. The samples are assumed to have
spatial independence. Find the number of data for e = 10%.
Since the standard deviation for the given example is 114.82 ppm then at 10%
standard error level one can find the number of representative data from Fig. 2.9 as
100.

2.5.2 Small Sample Length of Dependent Models

The smaller the sampling interval (distance) the more is the correlation between
nearby observations, and consequently the simple result in Eq. (2.2) cannot be
36 2 Data Types and Logical Processing Methods

applied directly to the averages of dependent series. The physical phenomena within
the earth sciences (as atmosphere) that give rise to such features are non-linear
dynamic systems with limited predictabilities. Therefore, it is not convenient to have
a very persistent process in their systematic modeling, but rather lower-order pro-
cesses seem more convenient without any mathematical complications. For instance,
Schubert et al. (1992) proposed a first-order Markov process to provide a general
picture of the persistence-based model behaviors compared to the General Circu-
lation Model (GCM) in atmospheric sciences. First-order Markov processes have
short memory of the correlation function and therefore they are not sufficient in the
GCM. The same authors then attempted to improve the situation upon the statistical
model by fitting low-order univariate Auto-Regressive Integrated Moving Average
(ARIMA) models to the control run of the GCM. Detailed information concerning
these models is available from Box and Jenkins (1976) as a natural extension of
the first-order Markov model, and they are useful in modeling atmospheric behav-
iors (Chu and Karz, 1985). In addition to finite sample length, the autocorrelation
structure of the process causes further source of bias (Kendall, 1954; Quenouille,
1956).
In order to model persistence within earth sciences ReVs, herein, the ARIMA
(1,0,1) model is considered. It is the mixture of separate stochastic processes includ-
ing autoregressive and moving average models. The numbers in the argument as
(1,0,1) imply that this type of ARIMA model is composed of first-order autore-
gressive (Markov) and first-order moving average processes with zero order differ-
ence between successive values. Generally, the model is written mathematically as
follows:

xi = φxi−1 − θεi−1 + εi , (2.3)

where φ and θ are the autoregressive and moving average parameters, respectively,
and εi is a zero-mean independent (white noise) random variable. The autocorrela-
tion structure of the same process is presented in terms of the model parameters as
(Box and Jenkins, 1976)

ρ = 0,
(φ − θ) (1 − φθ)
ρ1 = , (2.4)
(1 + θ2 − 2φθ)
ρi = φ ρi−1 ( i ≥ 2).

These expressions reduce to the white noise case when φ and θ are equal to zero;
to first-order Markov process case if θ = 0 and φ = ρ; and finally, to the moving
average process for φ = 0. For this model, the variance of time averages can be
calculated by using model parameters, which are the fundamental quantities related
to some statistical parameters that can be estimated from available data. In order to
illustrate this point, let us consider the data smoothened using a simple arithmetic
average of length m less than record length n. Hence, the arithmetic average for such
a subsample length n is
2.5 Number of Data 37

1
n
Xn = Xi . (2.5)
n
i=1

By taking first the square and then the expectation operator on both sides leads after
some algebra to
⎡ ⎤
1 ⎣ 
n n 
n
V(Xn ) = 2 E(X2i ) + E(Xi Xj )⎦ − μ2 . (2.6)
n
i=1 i=1 j=1

Of course, it is assumed E(Xi ) = μ. It is well known from mathematical statistics

that for identically distributed random variables E(X2i ) = E(S2 )+μ2 and E(Xi Xj ) =
E(S2 )ρ|i−j| + μ2 and their substitution into the last expression yields
 
E(S2 ) 
n
VA (Xn ) = n+2 (n − i) ρi . (2.7)
n2
i=1

The substitution of ARIMA (1,0,1) autocorrelation structure from Eq. (2.4) leads to


E(S2 ) 2 ρ1

VA (Xn ) = n + (1 − φ) − (1 − φn
) . (2.8)
n2 (1 − φ)2

This expression provides a common basis for the change calculations by giving the
variance of the pdf of the ReV means for finite lengths, i.e., sub-samples, from a
complete data set. For large samples (n > 30) the distribution of means converge to
a normal pdf due to the well-known central limit theorem in statistics. Square root
of Eq. (2.8) is equal to the standard deviation of such a normal pdf. Consequently, it
can be used for determining whether the small sample mean value in a given REV
is significantly different from its long-term mean value, μ, supposedly calculated
from the whole record. For this purpose, the executions of the following steps are
necessary:

1) Identify the underlying stochastic or ReV model for the given earth sciences
phenomenon.
2) Find the theoretical variance of the averages by substituting the necessary model
parameters into Eq. (2.8).
3) Consider the population pdf of the given data averages as a normal pdf with
mean, μ, and standard deviation σA = VA (Xn )
4) Find the standard deviate, t, for the means calculated from a given length, m, as

Xn − μ
t= . (2.9)
σA
38 2 Data Types and Logical Processing Methods

5) Accept a certain level of significance, αs , and find the corresponding confi-

dence limit, ± tc , from the standard normal pdf table given in any statistics book
(Davis, 2002).
6) If t ≤ |tc |, then the conclusion is that only sampling errors cause random vari-
ability and there is no long-term persistence variability (change); otherwise, the
earth sciences data are effected from systematic changes due to some atmo-
spheric and/or environmental reasons.

For the first-order Markov process, similar expressions to Eq. (2.8) can be
obtained provided that φ is substituted by ρ1 , which leads to

E(S2 ) 2 ρ1

VA (Xn ) = n+ (1 − ρ1 ) − (1 − ρ1 ) .
n
(2.10)
n2 (1 − ρ1 )2

This expression reduces to Eq. (2.1) for ρ1 = 0, which corresponds to the case of
independent model as explained in the previous section.
Generally, even for very small samples (n < 30) one can use Chebyshev inequal-
single value, say σA , selected at random to
ity, which states that the probability of

deviate from μ of the pdf more than ± VA (Xn ) is less than or equal to 1 K2 , and
it can be expressed in mathematical form as
   
P |σ A − μ| ≥ K VA (Xn ) ≤ 1 K .
2
(2.11)

This inequality yields upper and lower limits on the probability of deviation of a
given magnitude from the mean value. Hence, it is possible to find confidence inter-
vals on either sides of the average value.

Example 2.6
Let us calculate 95% confidence (reliability) limits for an ARIMA (1, 0, 1) model
with model parameters φ = 0.6 and for a finite length data, n = 15. First of all, from
Eq. (2.4) one can find that ρ1 = 0.34, the substitution of which with other relevant
values into Eq. (2.8) gives VA (Xn ) = 0.40 σ. Confidence
level of 95% implies that
 2
1 K = 0.95 or K = 1.026 and accordingly K VA (Xn ) = 0.41 σ. Consequently,
the confidence interval limits are. Finally, if the calculated square root of variance
of a finite data averages with n = 15 lies outside of these limits, then there is a trend
with the possibility of ReV change. Of course, for any desired small sample size
confidence interval limits can be calculated easily according to the aforementioned
procedure.
On theother hand, for the moving average case φ = 0, and hence from Eq. (2.4)
ρ1 = −θ (1 + θ2 ) and its substitution into Eq. (2.8) leads to
 
  σ2 2 (n − 1) θ
VA Xn = 2 n − . (2.12)
n 1 + θ2
2.5 Number of Data 39

It is important to notice that when θ = 0 this expression becomes identical to

Eq. (2.1). On the other hand, the well-known statistical estimate of small sample
variance, S2 , is given as

1 
n
S2 = (Xi − X)2 . (2.13)
n−1
i=1

The existence of n–1 rather than n in the denominator of Eq. (2.13) is because of
obtaining unbiased variance estimate for independent processes. Indeed, taking the
expectation of both sides leads to E(S2 ) = σ2 . This indicates with no hesitation
that it is possible to substitute σ2 in Eq. (2.1) by its sample estimate for indepen-
dent ReVs. However, as will be shown in the following sequel this is not true for
dependent processes. It has already been shown by Şen (1974) that for dependent
processes
 
2 
n
E(S ) = σ
2 2
1− (n − i)ρi , (2.14)
n(n − 1)
i=1

where ρi represents lag-i autocorrelation coefficient. This last expression shows that
there is bias effect not only due to the finite sample length, n, but also more sig-
nificantly due to autocorrelation structure of the ReV. Substitution of Eq. (2.4) into
Eq. (2.14) leads, after some algebraic manipulations, to
 
2 ρ1 n(1 − φ) − (1 − φn )
E(S ) = σ 1 −
2 2
. (2.15)
n(n − 1) (1 − φ)2

This expression proves that there is always an under-estimation of the population

variance from a given finite data set; the amount of bias depends on the finite sample
length and the model parameters that govern the dependence structure. In order to
obtain an unbiased estimate of the population variance, the sample estimate of the
variance must be divided by the curly bracket term in Eq. (2.15) as explained in
detail by Şen (1974).
A common point of the equations derived above is their asymptotic convergence
to a common value (σ2 /n) as the finite size of data n increases. In order to objectively
define a characteristic sample length for convergence, a certain level of relative error
percentage αr will be adopted and it is defined generally as

σ2 − E(S2 )
αr = 100 .
σ2
Hence, substitution of Eq. (2.15) into this expression leads to the most general form
of the relative error for the ARIMA (1,0,1) model as
 
2 ρ1 n(1 − φ) − (1 − φn )
αr = x100. (2.16)
n(n − 1) (1 − φ)2
40 2 Data Types and Logical Processing Methods

10

8
ρ =0.1
6
α
φ =0
4 .9

2 0.3 0.5 0.7

0.1

0
2 20 40 60 80
n

Fig. 2.10 Variance of average estimate relative error change with sample length (ρ1 = 0.1)

Based on this expression the change of relative error, αr , with the sample length n
for a set of ρ1 values are presented as charts in terms of φ values in Fig. 2.10.
A close inspection shows an exponential decrease in all the charts with sample
length. Increase in ρ1 implies increase in the relative error percentage for a given
sample length. In these figures the Markov model case corresponds to the curves
where ρ1 = φ. These charts are useful tools in finding equivalent independent
model length for a given dependent model length, provided that the relative error
percentage is given. For instance, when 10% error level is acceptable in the first-
order Markov model small data mean variance estimation with ρ1 = 0.3, then it is
possible to find from chart in Fig. 2.10 that after at least eight samples the ReV will
not have dependence structure. However, for the same error level and dependence
coefficient, an ARIMA (1,0,1) model requires at least 11, 16, and 48 samples for
φ = 0.5, 0.7, and 0.9, respectively.
These charts can also be used as indicators whether there is a trend in a given
ReV. For example, if the underlying generating mechanism is identified using the
procedures of Box and Jenkins (1976) as an ARIMA (1, 0, 1) model with ρ1 = 0.5
and φ = 0.7, then the error percentage is calculated from Eq. (2.16). If this error
percentage is greater than the theoretically calculated counterpart, then there is a
possibility of trend.
Another relative error for measuring any deviation from the independent model
with no long-term systematic features can be defined by considering Eqs. (2.1) and
(2.8) as

VA (Xn) − VI (Xn )
β = 100 .
VA (Xn )
This expression provides information about the necessary sample length concerning
the variance of climatic sub-sample averages. It is possible to write this expression
explicitly in terms of model parameters from Eqs. (2.1) and (2.8) as
2.6 Regional Representation 41
⎧ ⎫
⎪
⎪ ⎪
⎪
⎨ 1 ⎬
β = 100 1− ⎪ . (2.17)
⎪
⎪ 2ρ1
n) ⎪
⎩ n+ (1 − θ) − (1 − θ ⎭
(1 − θ)2

Finally, if the variance of short data and the variance of sub-sample averages are
calculated for the same sample length, then the relationship between these two types
of relative errors can be found from Eqs. (2.16) and (2.17) as
⎡ ⎤
⎢ 1 ⎥
β = 100 ⎣1 −
n−1 ⎦
, (2.18)
1+ αr
100
which implies that β is equal to zero only when αr approaches to zero. Otherwise,
for any given αr , there is an implied β-type error. The graphical representation of
Eq. (2.18) is given in Fig. 2.11. If any pair of αr , β, and n values are known, then the
third one can be determined from this chart easily.

Fig. 2.11 Variance of 30

climatic series average
estimate relative error change
with sample length (ρ1 = 0.3)
ρ =0.3
20

α
φ=
0.9
10
0.7
0.30.5
0.1

0
2 20 40 60 80
n

2.6 Regional Representation

Decision-making about which model to use for spatial estimation should be based on
some simple criteria as the number of data points, variability range, arithmetic aver-
age, standard deviation, geometrical configuration of sampling points, distances,
and in general distance-regional variability. Classical parameters such as the arith-
metic average, standard deviation, skewness coefficient in terms of summary statis-
tics can be readily known by the reader, and therefore they will not be elaborated in
this book (see Davis, 2002).
42 2 Data Types and Logical Processing Methods

2.6.1 Variability Range

In spatial analysis, the purpose is to estimate ReV at any desired point within the
study area using the measurements at a set of sampling points. Such a task will
depend not only on the internal variability at a fixed site of the variable but also more
significantly on the regional scatter of sampling points and the number of points.
Provided that the internal variability range does not differ more than 5% between
the maximum and the minimum data values, then one can depend on either the
arithmetic average or more specifically on the mode (the most frequently appearing
data value) as the sole representative for whole region. This method depends on
any number of data without data number specification. If the maximum and the
minimum ReV values are ZM and Zm , respectively, then the range of data variability
is defined as

RZ = ZM − Zm . (2.19)

The maximum percentage error can be defined as follows:

RZ
em = 100 . (2.20)
ZM

If em ≤ 5, then the regional representative value can be taken as the arithmetic

average, Z, at any point. This needs no detailed spatial modeling and it shows that
the phenomenon is not complex within the area and its behavior is more or less
homogeneous.
In order to be on a better side instead of arithmetic value, the mode, MZ , value
can be adopted. The mode is the most frequently occurring data vale within the
whole record.
Whatever the case may be, the regionalized variable is represented by a constant
value and, therefore, there will be deviations from this constant level, which are
the errors that could not be accounted by the constant value. Mathematically, it
is possible to suggest a procedure where the sum of square deviations (SSD) is
the minimum. Hence, if the constant level of n data values Zi (i = 1, 2, . . ., n) is
indicated by ZC , then the SSD becomes


n
SSD = (Zi − ZC )2 , (2.21)
i=1

which on expansion becomes


n 
n
SSD = Z2i − 2ZC Zi + n Z2C . (2.22)
i=1 i=1
2.6 Regional Representation 43

Since the minimum value is sought, the derivative of this last expression with respect
to ZC leads to

∂ (SSD) 
n
= −2 Zi + 2 n ZC . (2.23)
∂ ZC
i=1

If this is set equal to zero, then the result becomes

1
n
ZC = Zi , (2.24)
n
i=1

which is equal to the arithmetic average, Z. This proves that irrespective of any
practical requirement, the use of arithmetic average as a representative value for
the regionalized variable is a mathematical necessity. Further interpretation of
Eq. (2.24) can be given after its explicit form as

1 1 1 1
ZC = Z1 + Z2 + Z3 + . . . + Zn1 , (2.25)
n n n n

which means that the arithmetic average is the summation of point data value multi-
plied by a factor (1/n as weighting factor), which may have different interpretations.

1) The factors may be considered as weights for each data point, and therefore
they may represent some specific feature of the data point apart from the data
point value. For instance, weight may be the influence area of the point or some
important coefficient such as the relationship of this data value with the available
points. In this latter case, the factors may be the function of correlation between
data point pairs in addition to distances. In Eq. (2.25) the factors are the same,
which implies the isotropic behavior of the regionalized phenomenon, which
is not the case in natural events, and therefore each factor is expected to be
different than others.
2) In Eq. (2.25) equivalence of factors as 1/n reminds one that if there are n data
points, provided that they have equally likely future occurrence chances, the
probability of occurrence for each data value is p = 1/n. This corresponds to
random field case where the occurrences are completely independent from each
other, i.e., there is no regional correlation within the regionalized variable. How-
ever, in actual situations these probability of occurrences are not equal.
3) Another significant point from Eq. (2.25) is that the summation of the factors or
probabilities is always equal to 1. Furthermore, the factors are in percentages.
4) Finally, each factor is distance-independent whereas in any natural phenomenon
there are regional dependences, which may be simply expressed as inverse dis-
tance or inverse distance square weightings without natural consideration (see
Section 2.6.2).
44 2 Data Types and Logical Processing Methods

If the word “weight” is used instead of the aforementioned factor or probability,

then the arithmetic average expression in Eq. (2.25) can be written for more general
uses as the weighted average as


n
ZC = αi Zi , (2.26)
i=1

where αi ’s are the weightings. Consideration of percentage property of the weight-

ings, this last expression can be written in its most explicit form as
n 
 
Wi
ZC = Zi , (2.27)
WT
i=1

where Wi is the weight attached to i-th data point and WT represents the total weight
as the summation of all weights

WT = W1 + W2 + . . . + Wn , (2.28)

with n number of data points. The weights that are assigned to the estima-
tion points according to the weighting function are adjusted to sum to 1.0.
Therefore, the weighting function actually assigns proportional weights and
expresses the relative influence of each estimation point. A widely used ver-
sion of the weighting process assigns a function whose exact form depends
upon the distance from the location being estimated and the most distant point
used in the estimation procedure. This inverse distance-squared weighting func-
tion is then scaled so that it extends from 1 to 0 over this distance. Equation
(2.26) is in the form of linear multiple regression and it furnishes the basis
of all the linear estimation models in spatial analysis as will be explained in
this book.
Logically, the use of mode value is preferable because it is the most fre-
quently occurring data value within the given set of data points. The basis of
mode is probabilistic rather than mathematical, and it has the most likely prob-
ability of occurrence. A decision is to be made between the arithmetic aver-
age and the mode value in practical applications. If the pdf of the regional-
ized variable is symmetric, then the two concepts fall on each other. Otherwise,
the use of mode must be preferable, but mathematically its calculation is not
easy and therefore in any statistical or mathematical modeling arithmetic aver-
age is used invariably. The reader should keep in mind that although it yields
the minimum error, it does not abide with the most frequently occurring data
value. The arithmetic average value renders the mathematical procedures into a
tractable form.
2.6 Regional Representation 45

2.6.2 Inverse Distance Models

As mentioned in the previous section, the general regional estimation procedure is
given in Eqs. (2.26) or (2.27). In these expressions the most important part is the
attachment of weights. Once the weights are known, the remaining is simple calcu-
lation of the desired value at the point of estimation. Rational and logical procedures
are straightforward and one can conclude that as the distance, d, between two sites
increases, their effect on each other becomes meager; and hence, if only the distance
is considered without the regional variability feature of the phenomenon concerned,
then the first rational approach suggests that the weights can be taken as inverse dis-
tances, 1/d. However, in such an approach very small distances yield big weights and
at zero distance it is not defined. Another path of thought is the gravitational force
between two heavenly bodies where the attraction force is inversely proportional
with the square of distance, 1/d2 . It is even possible to generalize the inverse dis-
tance methodology as 1/dα , where α is a power parameter assuming values greater
than 1. The explicit expression used for the estimation, Ẑj at point j according to
IDP is given as
n 1
d α Zi
i=1 ij
Ẑj = n , (2.29)
1
d α
i=1 ij

where dij is the effective separation distance between grid node “j” and the neigh-
boring points “i,” Zi ’s are the neighboring point ReV values and α is the weighting
(smoothing) power. It is possible to derive special features from this general expres-
sion, which are well known in the literature. The last equation becomes equal to
inverse distance and inverse square distance for α = 1 and α = 2, respectively. The
slopes at the points used in the estimation procedure are weighted according to the
distances between the estimation node and the other points. Various inverse distance
functions are presented in Fig. 2.12.
Weighting is assigned to data through the use of a weighting power that con-
trols how the weighting factors drop off as distance from a grid node increases. The
greater the weighting power the less effect points far from the grid node has dur-
ing interpolation. As the power increases, the grid node value approaches the value
of the nearest point. For a smaller power, the weights are more evenly distributed
among the neighboring data points. Normally, inverse distance function behaves as
an exact interpolator, but still does not take into consideration the inherent ReV
variability in the phenomenon concerned. In calculating a grid node, the weights
assigned to the data points are fractions and the sum of all the weights is equal to
1.0 (Eq. 2.28). When a particular observation is coincident with a grid node, the
distance between that observation and the grid node is 0.0, and that observation is
given a weight of 1.0 while all other observations are given weights of 0.0. Thus, the
46 2 Data Types and Logical Processing Methods

10

9
α=3
8
α=2
7

6
Weight, w

α=1
5

3 α = 0.5

1 α = 0.3

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Distance, d

Fig. 2.12 Various inverse distance models

grid node is assigned the value of the coincident observation. Here α is a mechanism
for buffering this behavior. When one assigns a non-zero smoothing parameter, no
point is given an overwhelming weight so that no point is given a weighting factor
equal to 1.0.
One of the characteristics of inverse distance function is the generation of “bull’s-
eyes” surrounding the position of observations within the gridded area. One can
assign a smoothing parameter during inverse distance function to reduce the “bull’s-
eye” effect by smoothing the interpolated grid. Inverse distance function is a very
fast method for gridding. With less than 500 points, one can use all data search types
and gridding proceeds rapidly.

2.7 Sub-areal Partition

The whole study area may be partitioned into regular or irregular sub-areas each
with the same feature, hence the regional changes are considered as partially homo-
geneous within each sub-area. For instance, in Figs. 2.3 and 2.4 although the sub-
areas are in the form of regular quadrangles, sample points are located within these
sub-areas in a systematic or random manner. In this section, instead of points, sub-
areas are considered in the modeling of ReV variability.
2.7 Sub-areal Partition 47

2.7.1 Triangularization
The number of data points is also important in deciding which spatial prediction
model must be adopted. No need to say that spatial variability is describable by at
least three data points, which give the shape of a triangle; and, therefore, in the early
approaches before the advent of computers, triangularization method was preferred
due to its simplicity. If there are n data points, it is possible to obtain n–2 adjacent
triangles. For instance, five sampling points (A, B, C, D, and E) in Fig. 2.13 yield to
three triangles. Each corner is a measurement station. The ReV within each triangle
may be represented as the arithmetic average of the data values at three apices. In
this manner, rather than the use of arithmetic average over all the study area, it is
partitioned into triangular sub-areas where the arithmetic averages are used. In this
manner, the amount of the error in the global arithmetic average usage is reduced
significantly.
The problem in triangularization is that one point record enters more than one
sub-area. The question is, is it not better to define an influence area for each sub-
area separately? So that the whole study region is divided into a set of sub-areas
with influence of the data point.
In practice, given the scarcity of gauges and the spatial variability of ReV, for
instance, in the case of precipitation many storms completely miss several gauges
within a drainage area. Therefore, two basic tasks must be performed, namely the
assessment of the representativeness of point rainfall and picture derivations of spa-
tial patterns that reflect reality. Summer (1988) states, “In the ideal world, it should
be possible to track and then model, mathematically or statistically the passage of a
storm across an area with a great degree of accuracy and precision. In reality, this
is very difficult, and one must be content with generalized spatial models of storm
structure relating intensity or depth to storm area.” Kriging and stochastic methods
for the areal average estimation (AAE) based on the spatial correlation coefficient
are summarized by Bras and Rodriguez-Iturbe (1985). However, the use of these
methods needs recordings at many stations for the results to be reliable. Tabios and
Salas (1985) compared several AAE methods with rainfall variability and concluded
that a geostatistical method, ordinary and universal (Kriging, Chapter 4), with spa-
tial correlation structure is superior to Thiessen polygons, polynomial interpretation,
and inverse-distance weighting. Hevesi et al. (1992) suggested the use of multi-
variate geostatistical techniques for areal precipitation estimation in mountainous

A
B

E C

Fig. 2.13 Triangularization D

48 2 Data Types and Logical Processing Methods

terrain. Reliable estimates by these techniques are particularly difficult when the
areal coverage of stations is sparse or when precipitation characteristics vary greatly
with locations. Such situations frequently occur in arid regions due to sporadic
and haphazard meteorological occurrences. On the other hand, Kedem et al. (1990)
have shown, by considering satellite images and simple probability models, that the
higher the rainfall the smaller the affected area over large regions. All these methods
require high-speed computers and they are not as practical as conventional proce-
dures such as the arithmetic average, Thiessen polygons, or isohyetal map tech-
niques, which do not require much data (Chow, 1964).
An alternative AAE calculation method is presented by Akin (1971). It is
assumed that the precipitation over the sub-area varies linearly between the three-
corner (triangular) gauge points. Thus, at any point (x, y) interior to the n-th sub-area
the precipitation height Hn (x,y) is expressed as

Hn (x,y) = αn + βn x + γn y, (2.30)

where αn , βn , and γn are constants related to the gauge measurements at the

corners. It is possible to write three simultaneous equations one for each apex
(i, j, k) as

Hi (x,y) = αn + βn xi + γn yi .
Hj (x,y) = αn + βn xj + γn yj . (2.31)
Hk (x,y) = αn + βn xk + γn yk .

The solution of constants from these equations in terms of known quantities leads
to


αn = ai Hi + aj Hj + ak Hk /2An ,


βn = bi Hi + bj Hj + bk Hk /2An , (2.32)


γn = ci Hi + cj Hj + ck Hk /2An

where

ai = xj yk − xk yj .
bi = yj − yk . (2.33)
ci = xk − xj .

Following a cyclic permutation of (i, j, k) the sub-area An can be calculated as

follows:


An = ai + aj + ak /2. (2.34)
2.7 Sub-areal Partition 49

The differential volume of rainfall at any point within the sub-area is defined as

dQ = H(x,y)dA. (2.35)

So that the total volume of rainfall associated with the sub-area becomes theoreti-
cally as
! !


Qn = αn + βn xn + γn yn dxdy, (2.36)

where the substitution of the above relevant expressions leads after some algebra,
finally, to the volume of rainfall for the n-th sub-areas as


Qn = An αn + βn (xi + xj + xk )/3 + γn (yi + yj + yk )/3 . (2.37)

In the case of m triangular sub-areas, the total rainfall volume becomes


m
Q= Qn , (2.38)
n=1

and, corresponding total area is


m
A= An . (2.39)
n=1

Finally, the ratio of Eq. (2.38) to Eq. (2.39) gives the AAE height over m sub-areas
as
Q
H= .
A
By means of this procedure, the AAE area and volume are easily calculated if the
gauge locations and rainfall amounts are known.
The development is analogous to some elementary concepts used in finite ele-
ment analysis techniques. Consider a region of interest with station locations and
amounts that are plotted on a map, then a series of straight lines are drawn arbi-
trarily to connect every gauge points with the adjacent gauges. Straight lines are
drawn in anti-clockwise direction as shown in Fig. 2.14. These straight lines should
produce a series of triangles “not necessary to have the same shape.”
Each triangle area is known as sub-area and the corners of the triangles are shown
by [i, j, k]. Precipitation values at the corners are denoted as Hi , Hj , Hk (i, j, k = 1,
2, . . . , n), where n is the number of sub-areas. Triangularization of all the stations
for the northern part of Libya is shown in Fig. 2.15. On the other hand, Table 2.6
shows sub-areas for this study.
50 2 Data Types and Logical Processing Methods

1
H
K
Y H

(x,y) 1
K
i
5 X dA

2 Hn(x,y)

4
1
3 H

Fig. 2.14 Notations used for a triangular sub-area

2
14

16 13 3
MEDITERRANEAN SEA
17 15 12
18 32
15 14 13 4
23 11 2
11
33 12 10 10 1
19
27 31 22 9
20 30 24
N
35 28 9
16 34 29 8
21 8 3
7
GULF OF SIRT 5
26 36 37

23 6

22 24 27
6
26 4
28 7 5

25

1
18
19 Scale

0 100 200 300 400 500 km

29 - point number
37 - area number

Fig. 2.15 Triangular mesh for northern Libya

2.7 Sub-areal Partition 51

Table 2.6 Values of i, j, k for each sub-area

Sub-area Sub-area
number i j k number i j k

1 1 2 3 20 15 16 29
2 3 4 5 21 16 17 29
3 3 5 1 22 17 18 29
4 5 6 1 23 18 26 29
5 6 7 1 24 18 27 26
6 7 8 20 25 18 18 27
7 7 8 20 26 19 20 27
8 9 10 20 27 19 1 7
9 10 21 20 28 19 7 20
10 20 11 21 29 20 21 25
11 11 12 21 30 21 22 28
12 12 22 21 31 22 23 28
13 12 13 22 32 15 25 24
14 13 23 22 33 28 23 25
15 13 24 23 34 26 25 29
16 13 14 24 35 26 28 25
17 14 25 24 36 27 28 26
18 14 15 25 37 20 28 27
19 15 29 25

Convenient software is developed to calculate the sub-areas, associated AAR

depths, and volumes falling on each sub-area as well as the total area of interest.
The results are shown for monthly AARs in Table 2.7, while the seasonal AAR
depths and volumes are presented in Table 2.8.
These values are computed by developed software over the study area as a sum-
mation of 37 sub-areas already shown in Fig. 2.15 and Table 2.8.

Table 2.7 Monthly averages of rainfall depths and volumes

Months Average depth (mm) Average volume (×106 m3 )

Jan 28.2 97.400

Feb 12.7 64.889
Mar 16.2 47.684
Apr 4.4 16.564
May 5.5 19.980
Jun 0.41 1.600
Jul 0.03 0.160
Aug 0.17 0.800
Sep 2.9 9.060
Oct 7.13 23.230
Nov 15.7 39.600
Dec 28.9 112.10
52 2 Data Types and Logical Processing Methods

Table 2.8 Seasonal averages of rainfall depths and volumes

Average volume
Season Average depth (mm) (×106 m3 )

Winter 72.800 274.389

Spring 26.100 84.228
Summer 0.613 2.560
Autumn 25.700 71.890

2.8 Polygonizations
The idea is to surround each data point with a polygonal area of influence so that
the number of points will be equal to the number of sub-areas. In the case of n data
points, there will be n area of influence.

2.8.1 Delaney, Varoni, and Thiessen Polygons

These polygonal methods are all related to each other with simple differences but the
basic logic is the same. Different names are used in different disciplines for the same
method. According to these polygon methods, the study area can be partitioned into
a set of convex polygons, each containing only one measurement point such that
every point within a given polygon is closer to the measurement point than any
other measurement points. Each polygon defines the area of influence around the
measurement point (Fig. 2.16). Each one of these polygons is also called as Thiessen
polygon whose boundaries define the area that is closer to each point relative to all
other points.
They are geometrically defined by the perpendicular bisectors of the lines
between all points. A Voronoi diagram is sometimes also known as a Dirichlet
(1850) tessellation, with cells that are Dirichlet regions, Thiessen or Voronoi poly-
gons (Dirichlet, 1850; Voronoi, 1907; Thiessen, 1912). On the other hand, the
Delaunay triangulation and Voronoi diagram in 2D space are dual to each other in
the graph theoretical sense. Voronoi diagrams are named after Russian mathemati-
cian who defined and studied the general n-dimensional case in 1908. Voronoi dia-
grams that are used in geophysics and meteorology to analyze spatially distributed
data (such as rainfall measurements) are called Thiessen polygons. In climatology,
Voronoi diagrams are used to calculate the rainfall of an area, based on a series
of point measurements. In this usage, they are generally referred to as Thiessen
polygons.
Thiessen method is quick to apply because once the sub-polygons are fixed with
a set of observation points, they remain the same all the time. The only change
occurs artificially when additional observation points are added to the available set
of measurement points. It is based on the hypothesis that for each point in the area,
2.8 Polygonizations 53

•
• • •
• • •
• •

•
•
•
• •
• •
•
•
•
•• • • • •
•

• • •
•
• • • •
• ••
•
• • • •
Fig. 2.16 Polygonalization

the best estimate of ReV is the measurement physically closest to that point. This
concept is implemented by drawing perpendicular bisectors to the straight lines con-
necting each two-measurement stations, which yields with the consideration of the
watershed boundary a set of closed areas known as Thiessen polygons. Based on the
given measurement stations, the sub-polygons are obtained according to the follow-
ing steps:

1) Connect each station to each nearby station with a straight line. These lines
cannot cross and should connect only the nearest stations. The end product is
several triangles.
2) Each side of the triangles is then bisected with a perpendicular line, thus forming
polygons around each station.
3) Using an appropriate method calculate the total area, A, and sub-areas repre-
sented by each polygon (A1 , A2 , . . ., An ). The number of sub-areas is equal to
the number of measurement locations (Eq. 2.28).
4) Calculate the areal average estimation (AAE) of ReV as ZC according to the
weighted average formulation in Eq. (2.27). In this equation, Wi ’s correspond
to the sub-area at measurement location i with measurement Zi and WT is the
total area.
54 2 Data Types and Logical Processing Methods

Example 2.7
Consider the measurement locations A, B, C, and D as shown in Fig. 2.17. Let the
precipitation values be as 3.5, 2.9, 5.8, and 4.2 mm, respectively. Draw the Theissen
sub-polygons and calculate the areal average ReV value. The total area is given as
A = 50 km2 , and after the polygon partitioning each sub-polygon area are calculated
as AA = 12 km2 , BA = 17 km2 , CA = 13 km2 and DA = 8 km2 .
Using Eq. (2.27) the Thiessen areal average of ReV estimation for the entire area
becomes
12×3.5 + 17×2.9 + 13×5.8 + 8×4.2
ZC = = 4.174 mm .
50

In contrast, the simple arithmetic average of precipitations is 4.1 inches. Dif-

ferences between arithmetic and Thiessen averages increase for non-uniform ReVs
when the Thiessen areas differ widely. In the above given example since the ReV
measurements are close to each other, the AAE of ReV is close to the arithmetic
average. However, if the ReV values at the same measurement locations are 2.5,
15.3, 8.5, and 22.9 mm, then the AAE ReV value becomes

12×2.5 + 17×15.3 + 13×8.5 + 8×22.9

ZC = = 11.675 mm,
50
which indicates significant difference from the arithmetic average.
Unfortunately, the most commonly used Thiessen (1912) polygon method
for AAE calculations do not consider the areal precipitation amounts recorded
at individual stations in the partition of the whole catchment area into smaller
polygonal sub-areas. Therefore, once the polygons are obtained on the basis station

A
C

Fig. 2.17 Thiessen polygons

2.8 Polygonizations 55

location configuration, they remain the same as long as the measurement locations
do not change or there are no additional stations. However, it is logically plausible
to expect that the sub-areas should change in response to the spatial variation of the
phenomenon concerned. In other words, the partition should be based not only on
the measurement location network configuration but also on the ReV measurements
at each location.

2.8.2 Percentage-Weighted Polygon (PWP) Method

It is a new, simple, practical, and objective AAE method for determining the areal
average of the spatial event based on a sparse and/or irregular network of mea-
surement locations (Şen, 1998b). This method takes into account ReV measure-
ment percentage weightings for each station and also has geometrical advantages,
i.e., a better representation of the ReV on the study area compared to the conven-
tional Thiessen polygon method. For instance, ReV data such as precipitation show
a considerable spatial variation over any region, as explained by Huff and Neill
(1957), Stout (1960), Jackson (1972), and Summer (1988). Wilson and Atwater
(1972) suggested that this variation is due to differences in the type and scale of
precipitation-producing models, which are strongly influenced by local or regional
factors such as topography and wind direction. The AAE of the ReV over an area is
most conveniently determined from a well-sited network of measurement locations,
which show the local variations of ReV. For most areas, each measurement station is
assumed to represent a very considerable area around it. However, this is a restrictive
and frequently invalid assumption because of the spatial variation of ReV, especially
for short distances (durations), such as during severe storms (Summer, 1988). There
is no guarantee that point measurements provide reliable estimation for immediate
surrounding areas. Relief often leads to large variations in ReV over relatively short
distances. In the case of, say, precipitation if upward motion of air occurs uniformly
over thousands of square kilometers, the associated precipitation has usually light
or moderate intensity and may continue for a long time. On the other hand, con-
vective storms accompanied by compensating downdrafts (as best illustrated by the
thunderstorm) may be extremely intense, but their areal extent and local duration
are comparatively limited. In practice, given the scarcity of gauges and the spatial
variability of ReV, many stations may miss complete measurement. Therefore, two
basic tasks must be performed:

1) Assess the representativeness of point measurement.

2) Derive pictures of spatial patterns, which reflect reality.

After deciding on the triangles, the following procedure is necessary for dividing
the study area into polygons, leading to the percentage weighting method. If the
precipitation values at three apices of a triangle are A, B, and C, then their respective
percentages are
56 2 Data Types and Logical Processing Methods

PA = 100A/(A + B + C), (2.40)

PB = 100B/(A + B + C), (2.41)

and

PC = 100C/(A + B + C), (2.42)

respectively. Hence, it is possible to find the three-variable percentage data of con-

stant sums for each triangle. A two-dimensional plot of three variables can be shown
on a triangular graph paper as one point (see Fig. 2.18). Such papers are very com-
mon tools in earth sciences (Koch and Link, 1971).
In order to demonstrate the method more explicitly, the following step-by-step
algorithm is provided, which can be computerized:

1) Draw lines between each adjacent pair of precipitation stations. Hence, a set of
triangles is obtained, which cover the study area.
2) For each triangle calculate the precipitation percentage at its apices according
to Eqs. (2.40, 2.41, and 2.42). Consider in each station that each apex has the
value of 100% with 0% on the opposite side.
3) Consider bi-sector, which connects an apex to the midpoint of the opposite
side, and graduate it into 100 equal pieces.
4) By making use of one precipitation percentage calculated in step 2, mark it
along the convenient bi-sector starting from the opposite side toward the apex.
5) Draw a parallel line from this marked point in step 4 to the side opposite to the
apex considered with its precipitation percentage.
6) Repeat steps 4 and 5 for the next precipitation percentage and find similar
parallel line this time to another opposite side.
7) The intersection of these two lines defines the key point for the triangle
considered.
8) In order to check the correctness of this key point, repeat steps 4 and 5 for the
remaining third precipitation percentage value. If the parallel line to the side

100
0
0

10

Fig. 2.18 The triangular

100

coordinate paper 0
2.8 Polygonizations 57

crosses through the aforementioned key point, then the procedure of finding
the key point for the triangle is complete. Otherwise, there is a mistake either
in the precipitation percentage calculations or in the location of marked points
along the bi-sectors in steps 3 through 6 inclusive.
9) Return to step 2 and repeat the process for the triangles constructed in step 1.
In this manner each triangle will have its key point. The location of this
point within the triangle depends on the percentages of recorded precipitation
amounts at the three adjacent apices. The greater the precipitation percentage
for an apex the closer the point will lie to this apex. It is not necessary that the
triangles resulting from a given set of stations in a network should be exactly
equilateral. However, in the Thiessen method, for an obtuse-angle triangle the
intersection of the three perpendicular bi-sectors occurs outside the triangular
area.
10) Key points at adjacent triangles are connected with each other to form poly-
gons, each including a single precipitation station.
11) The boundaries of polygons around the basin perimeter are defined by drawing
a perpendicular to the sides of triangles from the key points. Now, the division
of the whole basin area into sub areas is complete.

The triangular or polygon methods are preferred in practice when there are sev-
eral data points about eight to ten.

Example 2.8
A simple example of plotting on a triangular coordinate paper is presented in
Fig. 2.19 for an obtuse-angle triangle.
If the precipitation amounts at three apices are A = 12.5 cm, B = 20.1 cm,
and C = 7.4 cm, then the corresponding percentages from Eqs. (2.40, 2.41,
and 2.42) are pA = 31, pB = 50, and pC = 19. In Fig. 2.19, point D on the
A–A bi-sector corresponds to 31% of the A–A length starting from A , which
lies on the B–C side of the triangle representing 0%. A parallel line to side
B–C is drawn from point D. Likewise, the next percentage is considered for
the precipitation amount at apex B. The bi-sector B–B is drawn starting from
point B on the side A–C toward B. On the bi-sector, point E corresponding
to 50% is depicted. A parallel line from this point to the side A–C is drawn.
Finally, the intersection of these two parallels at point F defines the “key point’’
for triangle ABC. The following steps are necessary for the implementation of
the PWP:

1) Three adjacent stations are considered such as in Fig. 2.19, where each apex
is coded by its longitude (XA , XB , and XC ), latitude (YA , YB , and YC ) and
precipitation value (ZA , ZB , and ZC ).
2) The slopes (mAB , mBC , and mCA ) of triangle sides are calculated by making
use of the apices coordinates.
58 2 Data Types and Logical Processing Methods

A Bisectors

B'
F Parallel to BC
D
E
B
A' C
Pa
ra
lle
l to
AC

Fig. 2.19 Subdivision of triangular area

3) Determine the straight line equations perpendicular to each of the sides, but
crossing from the opposite apex by analytical formulations. First of all, the
coordinates of the projection point such as A and B of each apex on the oppo-
site side must be calculated. For instance, the coordinates XA and YA of point
A can be expressed in terms of known quantities as

1 " #
XA = m2BC (XB ) + mBC (YA − YB ) + XA , (2.43)
m2BC + 1

1  
YA = XA − XA + YA , (2.44)
mBC

where

YB − YC
mBC = . (2.45)
XB − XC

Similarly, the coordinates XB and YB of point B on A–C side are

1 " #
XB = m2
(X
CA C ) + mCA (YB − YC ) + XB , (2.46)
m2CA + 1

1  
YB = XB − XB + YB , (2.47)
mCA
2.8 Polygonizations 59

where
YC − YA
mCA = . (2.48)
XC − XA

4) The precipitation amounts (ZA , ZB , and ZC ) are used in order to find points
along the perpendiculars starting from the side toward the apex, which divide
each one in proportions λ1 , λ2 , and λ3 . By use of these ratios and the previous
known quantities, the coordinates of points A , B , and C are determined. For
instance, the coordinates XA and YA are defined as

XA = xA + λ1 (XA − xA ). (2.49)

YA = YA + λ2 (YA − YA ). (2.50)

5) On the other hand, from the precipitation measurements at the three apices of
any triangle, the total precipitation T value becomes

T = ZA + ZB + ZC ,

which can be written in terms of percentages as

ZA /T + ZB /T + ZC /T = 1.0,

or
λ1 + λ2 + λ3 = 1.0

Hence, lambda values are defined as the percentage of precipitation amounts

at the three apices of a triangle. Similarly, the coordinates of point B’ can be
written with the relevant notations as

XB = xB + λ2 (XB − xB ), (2.51)

YB = YB + λ2 (YB − YB ), (2.52)

where
ZB
λ2 = . (2.53)
ZA + ZB + ZC

6) The straight lines are drawn, which pass through point A and parallel to the
B–C side of the triangle and similarly parallel to A–C crossing through B .
Let these parallel lines be D1 and D2 . The equations of these lines are given
as follows:
60 2 Data Types and Logical Processing Methods

D1 = mBC X + (YA − mBC X A ). (2.54)

D2 = mCA X + (YB − mCA X B ). (2.55)

7) The intersection point of lines D1 and D2 is referred to as the “key point,” MK ,

for triangle ABC. The coordinates of this key point (XK , YK ) can be found by
simultaneous solution of Eqs. (2.54) and (2.55).
8) Previous steps are repeated for each triangle, representing sub-area within the
whole catchment.
9) Adjacent key points are connected with each other, hence constituting poly-
gons each enclosing a single meteorology station. The key point coordinates
become
1
   
Xk = mCA X B − mBC X A − Y B − Y A . (2.56)
mCA − mBC

 
Yk = mBC Xk − X A + Y A (2.57)

10) The area of each polygon is calculated by one of the available convenient math-
ematical methods.
11) Multiplication of each polygon area by its representative station precipitation
value gives the volume of water that is gathered over this polygonal sub-area.
12) Summation of these volumes for all the relevant polygons that cover the whole
catchment area gives the total water volume which falls over the catchment
area.
13) Division of this total water volume by the total catchment area yields the AAE
value for the catchment.

In general, the comparison of the Thiessen and PWP methods indicates the fol-
lowing significant points:

1) If triangulation of stations produces obtuse triangles, then the Thiessen method

gives bi-sector intersections that fall outside the triangle area. However, in the
PWP method all the key points lie within the triangle itself. In this way the com-
mon effect of precipitation amounts at triangle apices falls within the triangle
domain.
2) In the Thiessen method the station lies within the sub-areal polygon at almost
equal distances from the sides. This implies that almost equal weights are given
to precipitation amounts in the separation of polygons. On the contrary, the PWP
method produces polygon sides closer to the station, depending on the relative
magnitude of precipitation amounts between adjacent stations. This implies that
in the PWP method the center of recorded precipitation amounts at three neigh-
boring stations (not the geometrical center of station locations) plays the most
significant role in the areal partition calculations.
2.8 Polygonizations 61

3) The size of any closed polygon for a station in the PWP method is always
smaller than the corresponding Thiessen method. Hence, more refined partition
of the catchment area into sub-areas is obtained. This implies that more refined
AAE values result by use of the PWP method. In fact, in the PWP method the
biggest precipitation record station results in the smallest sub-area.
4) Variations in the precipitation amounts, due to some different storms, change
the polygonal partition in the PWP method, whereas Thiessen polygons remain
the same.

Example 2.9
The implementation of PWP method was applied first to spatial precipitation data
given by Wiesner (1970) and Bruce and Clark (1966) as in Fig. 2.20, where the geo-
metric configuration of precipitation stations are given for the same data according
to the Thiessen and PWP methods. In order to assess the reliability of these meth-
ods, the average sum of square deviations (ASSD) from the AAE value is calculated
for each method as
1 
n
2
ASSD = Pi − AAE , (2.58)
n
i=1

A(0.91)

B(1.62)
C(2.54)

D(4.23)
E(2.8)
B(5.8)
A(3.1)

F(0.83)
G(3.86) C(2.54) D(3.3)

H(2.10)

E(1.5)
F(1.0)
I(1.84)
Catchment boundary Catchment boundary
Inter-station connections Inter-station connections
Theissen polygon Theissen polygon
Percentage weighted polygon Percentage weighted polygon
(a) (b)
Fig. 2.20 Precipitation data (a) Wiesner, (b) Bruce, and Clark
62 2 Data Types and Logical Processing Methods

Table 2.9 AAE of precipitation

Wiesner data Bruce and Clark data

Method AAE (inches) ASSD (inches2 ) AAE (inches) ASSD (inches2 )

Thiessen 2.67 1.34 3.10 2.54

PWP 2.45 1.15 2.50 2.37

where n is the number of stations. The results for various methods are shown in
Table 2.9.
The PWP method yields smaller AAE and ASSD values than the Thiessen
method.
As mentioned earlier, with the Thiessen method, in the obtuse triangle BDF a per-
pendicular bi-sector’s intersection appears to the left of BF line (see Fig. 2.20a). By
considering precipitation amounts at stations B and F, it is not possible to expect a
value of about 4 inches. On the other hand, in the PWP method there is no such con-
fusion. Wiesner (1970) has shown, after completion of an isohyetal map for the same
precipitation data, that the storm peak occurs within a small area around station D in
Fig. 2.20a. This is also confirmed by the PWP method as shown in Fig. 2.20a. Again,
in the same figure Thiessen method yields only one closed polygon, whereas PWP
provides three such polygons. Extensions of Thiessen boundary polygons between
the pairs of stations F and I and E and I suffer from the same drawbacks in that it
is not possible to have precipitation values greater than 1.8 inches between F and I,
and more than 2.83 inches between E and I. Similar general interpretations are valid
also for Fig. 2.20b where there are two odd bi-sectors’ intersection points that lie
outside the triangles, resulting from the application of Thiessen method. These lie
to the left of line A–F and below line F–E. Comparatively, even in this figure, the
closed polygon from the PWP method is smaller in size than the Thiessen case. The
PWP method can be applied to local or regional scales due to the following reasons:

1) The assumption that convective, intense precipitation is comparatively limited in

areal extent and duration is valid for small or regional scales. Thinking in larger
space and time scales (e.g., global and monthly), there are also broader areas
with intense precipitation, such as the large-scale convergence zones (ITCZ and
SPCZ) and the monsoon regions.
2) The limited areal extent and duration of intense convective precipitation com-
pared to stratiform, light precipitation is generally compensated by the fact that
many storms might be missed by the rain gauge network (see the “Introduction”
section). So if the rain gauge network is nearly equally distributed, there is not
any reason to compensate for this a second time by reducing the areal weight
of heavier precipitation, unless there is not really an indication that intense pre-
cipitation is over-represented due to the spatial distribution of the rain gauges in
the network. This might be the case in arid regions, because there are almost no
stations in desert regions. However, in mountainous areas precipitation is often
2.8 Polygonizations 63

underestimated, because most rain gauge stations are located in valleys and do
not catch very well the heavy precipitation on the slopes or mountain peaks.
3) The proposed PWP method is very convenient in areas where the precipitation
stations are irregularly located.

Example 2.10
In order to determine the AAE of rainfall from ten different meteorological sta-
tions, the PWP method is applied together with the other conventional methods
(arithmetic mean, Thiessen polygon, isohyetal map technique) to the Southeastern
Anatolia Region of Turkey (Bayraktar et al., 2005). Total monthly rainfall values of
these stations in 1993 are used and presented in Table 2.10.
About 1/500,000 scaled maps are used to draw the sub-areas, which are measured
with a planimetry. For each method, AAR values are calculated with the help of
Eq. (2.27).
Application map of Thiessen method and areal values of Thiessen polygons are
merely given in Fig. 2.21, because the key points remain the same as long as the
meteorological stations do not change. Monthly isohyetal maps are prepared for
January, April, July, and October as in Fig. 2.22.
For the same months rainfall values and sub-areas are given in Fig. 2.23a,b,c,d.
In PWP method, which is the main purpose of this study, values of rainfall and per-
centage weightings are calculated for each of the three adjacent stations constituting
sub-triangles.
During the application, PWP rainfall values are calculated by considering Eqs.
(2.40, 2.41, and 2.42) used for the determination of sub-areas. PWP method cal-
culations and sub-areal values are given in Fig. 2.23a,b,c,d. For the comparison of
different methods all results are presented collectively in Table 2.11.
As can be seen from this application, the PWP method yields more reliable
results and smaller AAE values depending on the regional variability of the rainfall
amounts over the catchment area. For instance, in July, rainfall values have many
regional variations. This is due to the semi-arid characteristics of the study area and

Table 2.10 Total monthly rainfall values of stations (mm)

Stations Jan Feb Mar Apr May Jun Jul Aug Sept Oct Nov Dec

Silvan 104 116 116 84 61 6 0 1 1 19 69 116

Siirt 115 107 111 106 66 9 1 0 5 48 86 102
Ahlat 64 67 79 84 62 23 9 4 15 50 68 53
Bitlis 180 156 141 107 65 19 5 3 14 58 97 130
Van 42 35 46 58 41 17 6 3 12 44 49 83
Şırnak 136 121 143 141 64 5 2 1 6 33 83 123
Başkale 47 43 69 102 100 42 20 9 16 39 43 38
Nusaybin 92 62 63 63 40 1 1 0 1 16 43 79
Şemdinli 106 155 151 195 56 9 3 4 5 77 119 97
Hakkari 102 105 125 146 57 15 3 1 10 26 76 91
64 2 Data Types and Logical Processing Methods

AHLAT
N

1695 km2
VAN
BITLIS
4415 km2 4550 km2

SILVAN
BASKALE
2390 km2 1545 km2
SIIRT

5770 km2 4540 km2 3670 km2

HAKKARI 575 km2

SIRNAK
2250 km2 THIESSEN POLYGONS SEMDINLI
Scale : 1: 5,000,000

NUSAYBIN

Fig. 2.21 Application map of Thiessen method (January, April, July, October 1993)

frequent occurrence of convective rainfalls, which are not expected to cover large
areas. It is noted that in July, Başkale station surrounding appears as the most rain-
fall reception area, which is represented by a sub-area of 1,545 km2 in the Thiessen
method, whereas by the sub-areas of 3,140 km2 and 519 km2 in the arithmetic mean
and the isohyetal map technique, respectively. On the other hand, it is represented
by a sub-area of 153 km2 when PWP method is considered. Hence, the represen-
tation of the most rainfall intercepting meteorology station with the smallest sub-
area gives rise to a smaller value of AAE for the catchment area during convective
storms. The monthly AAEs of the catchment area in this month according to iso-
hyetal map, Thiessen Polygon, and arithmetic mean techniques are estimated as 44,
41, and 54%, respectively. They are smaller than the PWP method AAE value. The
areal rainfall variability in October is comparatively less than July because of the
north polar maritime air mass penetration into the study area, which gives rise to
frontal rainfall occurrences. It is well known that the frontal rainfalls have more
extensive areal coverage than the convective rainfall types and consequently the
percentages are smaller in the PWP method calculations. This is tantamount to say-
ing that percentage calculations from the three adjacent stations are not very dif-
ferent as they are in the convective rainfall events. It is timely to state that similar
effects can be expected from the orogrophic rainfall occurrences to the convective
rainfall types in mountainous areas. Such distinctions cannot be considered in the
Thiessen approach, where the sub-areas are determined according to the geometric
configuration of the station locations, without the consideration of actually recorded
rainfall amounts at these stations. For example, in October, the AAE calculations
by using the PWP method yielded 13, 15, and 18% smaller values than other three
 2.8
Polygonizations

Fig. 2.22 Application map of isohyetal technique (a) January, (b) April, (c) July, (d) October
65
 66

Fig. 2.23 Application map of PWP technique (a) January, (b) April, (c) July, (d) October
2 Data Types and Logical Processing Methods
2.9 Areal Coverage Probability 67

Table 2.11 AAE value for each method (mm)

Method January April July October Annual

Percentage weighting polygon 89.22 92.70 2.32 33.75 54.49

Isohyetal map 102.72 102.72 4.12 38.96 62.13
Thiessen polygon 106.30 104.30 3.94 39.74 63.57
The arithmetic average 98.76 108.49 5.00 40.97 63.31

conventional methods. This is due to the lesser areal variability of rainfall in this
month than in July. Furthermore, the AAE amounts by use of the PWP method and
the annual average rainfall amounts are found 12, 14, 14% and on the average 13.5%
lesser than the isohyetal map, Thiessen polygon, and arithmetic average methods,
respectively. Arithmetic average method yields in January smaller AAE than all
other methods because this is the month where the effect of frontal rainfall is the
most dominant with high rainfalls in the area except low rainfalls at few stations,
such as at Van and Başkale stations as 42 and 47 mm, respectively. Last but not the
least, the arithmetic average method is affected from extreme rainfall values more
than any other method.
As the frontal and more homogeneous type of rainfalls are recorded during a
specific period, then it is expected that the results of the PWP and Thiessen poly-
gon methodologies will approach to each other, provided that there are approximate
rainfall records at each meteorological station. However, in practice the use of PWP
method is recommended. In fact, whatever the circumstances are, it is preferable to
adopt PWP method since it yields almost the same result as the Thiessen method if
the conditions are satisfied as the homogeneous type of rainfalls.

2.9 Areal Coverage Probability

A rigorous probabilistic methodology is developed on the basis of random field

concept for determining various probability of precipitation (PoP) areal coverage
probability (ACP) over a region. The necessary analytical formulations are derived
with the restrictive assumption that the precipitation occurrences at different sites
and times are independent from each other. However, the regional heterogeneity in
precipitation occurrence is exemplified by either non-identical probability of precip-
itation values or non-uniform threshold levels. Finally, extreme value probabilities
of precipitation coverage area are formulated in their most general forms. The prob-
ability statements derived are helpful in predicting the regional probable potential
future precipitation or drought coverage areas. The drought basic time unit can be
adopted as hour, day, week, month, or year in addition to unit area for spatial analy-
sis. If the basic unit is day, then each day can be labeled as “rainy” and “non-rainy”
and the calculations are performed accordingly (Şen, 2008). Similarly, in the case
of regular division of an area into a set of sub-areas as in Fig. 2.24b, each sub-unit
may have either a “dry” or a “wet” spell. For instance, in Fig. 2.24a and b there are
six and nine time and sub-area units, respectively.
68 2 Data Types and Logical Processing Methods

Fig. 2.24 Time and spatial y

units Space

1 2 3

4 5 6

1 2 3 4 5 6 7 8 9
Time x
a b

y
Space

Time x
a b

Fig. 2.25 Dry time and sub-area units

If “dry” units are shown by black and “wet” units by white colors, then in
Fig. 2.25 there are four and five time and sub-areal dry units, respectively,
As the number of basic unit increases, then the “dry,” D, and “wet,” W, time and
spatial features start to become clearer as in Fig. 2.26.
In this section precipitation phenomenon is modeled as a random field where
time-wise probabilities at a fixed site are referred to as PoP, and spatial probabilities
for a fixed time instant are ACP. The areal probability is, in fact, the fraction of the
area hit by rainfall. It does not provide a means for depicting which of the sub-areas
is going to be hit with precipitation event. However, it simply represents the esti-
mate of what the fractional coverage will be. Furthermore, it is clearly a conditional
probability being conditioned by whether at least one precipitation event actually
occurs in the sub-area during a certain time interval.

y
Space

Time
K S K S
x
b
Fig. 2.26 Dry durations and areas
2.9 Areal Coverage Probability 69

The PoP at any desired threshold value, x0 , such as standard 0.01 inches is equiv-
alent to the exceedence probability of this value. If the pdf of precipitation at site i
is denoted by fi (X), then the PoP, pi , at this site is

!∞
pi = fi (X) dX.
x0

If the forecast area is thought in terms of n sample sites, then the average areal
probability p can be defined formally as

1
n
p= pi . (2.59)
n
i=1

In the special case where the point probability is uniform over the forecast area, p
is equal to the common value of PoP. Note that since p is a lumped value it does
not, in general, contain detailed information provided by the set of individual point
probabilities. However, if for some reason some of the sub-areas vary in size, then
each probability value must be percentage area weighted. It should be noted that the
areas associated with sub-areas need to be small enough that a single probability
value can be applied to each one.

2.9.1 Theoretical Treatment

In the most general case, none of the sites have equal PoPs, which implies that
the random field is heterogeneous. It is quite likely that probabilities might vary
from place to place even within a single area. In practice, in addition to the spatial
correlation variations, the following three situations give rise to heterogeneous and
anisotropic random fields. These are:

1) Identical pdfs of precipitation at different sites, but non-uniform threshold

levels.
2) Non-identical pdf of precipitation at sites, but uniform threshold level.
3) Non-identical pdfs at sites and non-uniform threshold levels.

Let the PoP and its complementary probability at j-th site within a region of n
sites be given by pj and qj (j = 1, 2, . . ., n), respectively. They are point-wise mutu-
ally exclusive and complementary, i.e., pj + qj = 1. The ACP, P(A = i) including i
sites can be evaluated through enumeration technique after the application of sum-
mation and multiplication rules of probability theory. Hence
n n n $
n
P(A = i) = 1
i! pj1 pj2 . . pji qk
j1=1 j2=1 ji k=1 (2.60)
j2 = j1 ji = ji−1 k = j1 ,
70 2 Data Types and Logical Processing Methods

or succinctly
n n n $
n
P(A = i) = 1
i! pj1 pj2 . . . pji qk
j1 =1 j2 =j1 ji k=1 (2.61)
j2 = j1 ji = ji−1 k = ji ,

% &
$
n n $
n
P(A = i) = 1
i! pjk1 qk2
k1 =1 jk1 =1 k2 =1 (2.62)
jk1 = jm k2 = jm
m = 1, 2, . . . jkm−1 ,

where the multiplication of i summations in the brackets includes all the possible
combinations of i precipitation occurrences at n sites, whereas the last multiplication
term corresponds to possible no-precipitation combinations. For identical PoPs the
term in brackets simplifies to n(n–1). . .(n–i+1)pi and the last multiplication to qn—I ,
hence Eq. (2.62) reduces to
 
n
P(A = i) = pi qn−i , (2.63)
i

which is actually the Binomial distribution with two-stage Bernoulli trials (Feller,
1967).
This is the well-known Binomial distribution with mean np and variance npq.
The probability, pA , that all the sites, hence area, are covered by precipitation at an
instant can be found from Eq. (2.62) as

'
n
pA = pi . (2.64)
i=1

Comparison of Eqs. (2.59) and (2.64) yields

p > pA , (2.65)

and accordingly

min (p1 ,p2 , . . . ,pn ) < p < max (p1 ,p2 , . . . ,pn ). (2.66)

The conditional probability of partial area, given a group of randomly or system-

atically selected k precipitation occurrences or non-occurrences, is of great interest
in practical applications. The joint probability that any k1 of n sites have signifi-
cant precipitation occurrences can be defined by virtue of independence and hence
multiplication rule of the probability theory as
2.9 Areal Coverage Probability 71

P(X1 > x0 , X2 > x0 , . . . ,Xk1 > x0 ) = pk1 . (2.67)

Assuming that any one of the sites is equally likely to experience a precipitation, the
conditional probability that only a certain group of i sites has precipitation can be
written as
 k1
i
P(X1 > x0 , X2 > x0 , . . . ,Xk1 > x0 /A = i) = .
n
By definition the joint probability becomes
 k1
i
P(X1 > x0 , X2 > x0 , . . . ,Xk1 > x0 ,A = i) = P(A = i).
n
The conditional ACP of precipitation can then be obtained after dividing this last
expression by Eq. (2.67), which leads to
 k1
i
P(A = i/X1 > x0 , X2 > x0 , . . . ,Xk1 > x0 ) = P(A = i). (2.68)
np

Hence, in order to obtain the most general case of heterogeneous PoP’s conditional
areal coverage, it is necessary to substitute Eq. (2.62) into Eq. (2.68). On the other
hand, for homogeneously distributed PoPs, Eq. (2.68) takes its explicit form by the
substitution of Eq. (2.63)
 k1  
  i n
P A = i/X1 > x0 , X2 > x0 , . . . ,Xk1 > x0 = pi qn−1 . (2.69)
np i

Since some initial information is given with certainty, the conditional probability in
Eq. (2.69) has more information content than the original unconditional distribution
in Eq. (2.63). This fact can be objectively shown by considering the variances of
pdfs in Eqs. (2.63) and (2.69), which are npq and (n–1)pq, respectively. Hence,
the conditional variance is smaller, and therefore conditional ACP is more certain.
However, the unconditional areal precipitation coverage expectation from Eq. (2.63)
is equal to np, whereas the conditional coverage area expectation is greater and can
be obtained from Eq. (2.68) as np + q.
Similarly, the conditional ACP of precipitation, given that a group of k2 sites
have no-precipitation, can be found for homogeneous PoPs as
  
n−i n
P(A = i/X1 < x0 , X2 < x0 , . . . , Xk2 < x0 ,) = qi pn−1 . (2.70)
nq i

Finally, the conditional ACP of precipitation, given that a group of k1 sites has
precipitation and another group of k2 sites has no-precipitation, is obtained as
72 2 Data Types and Logical Processing Methods

P(A = i/X1 > x0 ,X2>  x0 , . . . ,Xk1 > x0 ,Xk1+1 > x0 ,Xk1 +2 < x0 , . . .Xk2 < x0 ,)
( )( ) n
= nqi n−i
np qi pn−1 .
i
(2.71)
The probability expressions in Eqs. (2.68, 2.69, 2.70, and 2.71) can be effectively
employed to study regional precipitation occurrence patterns.

2.9.2 Extreme Value Probabilities

In practice, it is also important to know the probability of maximum and minimum

areal coverage of precipitation. The probability of maximum areal coverage, AM , of
precipitation can be found provided that the precipitation coverage area evolution
along the time axis is assumed to be independent from each other. In general, for m
time instances the probability of maximum areal coverage of precipitation to be less
than or equal to an integer value, i, can be written as

P(AM ≤ i) = [P(A ≤ i)]m , (2.72)

where P(A ≤ i) can be evaluated, in general, from Eq. (2.62) for heterogeneous PoPs
by employing


i
P(A ≤ i) = P(A = j).
j=0

The substitution of which into Eq. (2.72) leads to

⎡ ⎤m

i
P(AM ≤ i) = ⎣ P(A = j)⎦ .
j=1

However, it is a well-known fact that P(AM = i) = P(AM ≤ i) – P(AM ≤ i–1), and

therefore one can obtain consequently that
⎡ ⎤m ⎡ ⎤m
i  
 i−1  

n n j n−j ⎦
P(AM = i) = ⎣ pj qn−j ⎦ − ⎣ pq . (2.73)
j j
j=0 j=0

It is then possible to rewrite Eq. (2.73) for homogeneous PoPs from Eq. (2.63)
explicitly as
⎡ ⎤m ⎡ ⎤m
i  
 i−1  

n n
P(AM = i) = ⎣ pj qn−j ⎦ − ⎣ pj qn−j ⎦ .
j j
j=0 j=0
2.10 Spatio-Temporal Drought Theory and Analysis 73

Hence, the probability P(AM = n) that the whole area is covered by precipitation
can be obtained from Eq. (2.70) as

P(AM = n) = 1 − (1 − pn )m . (2.74)

One can interpret from this expression that for small regions the number of sites
is also small, in which case the probability in Eq. (2.74) is not zero and there is a
chance for the whole area to be covered by precipitation. Similarly, the probability
of minimum areal coverage, Am , of precipitation can be written for homogeneous
PoPs as
⎡ ⎤m ⎡ ⎤m
n   n  
n n
P(Am = i) = ⎣ pj qn−j ⎦ − ⎣ pj qn−j ⎦ . (2.75)
j j
j=i j=i+1

2.10 Spatio-Temporal Drought Theory and Analysis

Let an agricultural land be divided into m mutually exclusive sub-areas, each with
the same spatial and temporal drought chance. The Bernoulli distribution theory can
be employed to find the extent of drought area, Ad , during a time interval, t. The
probability of n1 sub-areas stricken by drought can be written according to Bernoulli
distribution as (Feller, 1967)
 
m
Pt (Ad = n1 ) = pnr 1 qm−n
r
1 pr + qr = 1.0. (2.76)
n1

This implies that out of m possible drought-prone sub-areas, n1 have deficit and
hence the areal coverage of drought is equal to n1 or in percentages n1 /m. For the
subsequent time interval t, there are (m–n1 ) drought-prone sub-areas. Assuming
that the evolution of possible deficit and surplus spells along time axis is indepen-
dent over mutually exclusive sub-areas, similar to the concept in Eq. (2.76), it is
possible to write for the second time interval that
n2 
  
m m − n1
P2t (Ad = n2 ) = pnr 1 pnr 2 −n1 qm−n 1 qm−n2 , (2.77)
n1 n2 − n1 r r
n1 =0

where n2 is the total number of drought-affected sub-areas during the second time
interval. By considering Eq. (2.76), this expression can be rewritten succinctly in
the form of recurrence relationship as


n2
P2t (Ad = n2 ) = Pt (Ad = n1 ) P (Ad = n2 − n1 ), (2.78)
n1 =0
74 2 Data Types and Logical Processing Methods

where P(Ad = n2 − n1 ) is the probability of additional (n2 – n1 ) sub-areas to be

effected by deficit during the second time interval out of the remaining (m–n1 )
potential sub-areas from the previous time interval. With the same logic, extension
of Eq. (2.78) for any successive time interval, i, furnishes all the required drought-
area probabilities as


ni
Pit (Ad = ni ) = P(i−1)t (Ad = ni−1 ) P (Ad = ni − ni−1 ) . (2.79)
ni−1 =0

For i = 1 this expression reduces to its simplest case, which does not consider time
variability of drought occurrences as presented by Şen (1980) and experimentally
on digital computers by Tase (1976). Furthermore, the probability of agricultural
drought area to be equal to or less than a specific number of sub-areas j can be
evaluated from Eq. (2.79) according to


j
Pit (Ad ≤ j) = Pit (Ad = k). (2.80)
k=0

The probability of having, n1 , deficit sub-areas, given that there are n1 deficit sub-
areas at the beginning of the same time interval within the whole region, can be
expressed as
  
  m n1 n1 −n1 m−n n1
Pt Ad = n1 |Ad = n1 = pnr 1 pt qr qt ,
n1 n1 − n1
or shortly,
 
  n1 n1 −n1 n1
Pt Ad = n1 |Ad = n1 = Pt (Ad = n1 ) pt qt . (2.81)
n1 − n1

It should be noted that always n1 ≥ n1 and the difference j = n1 − n1 gives the
number of transitions. On the basis of Eq. (2.81), a general equation for the marginal
probability of observing n1 deficit spells at the end of the same time interval, after
simple algebra, becomes

m=ni
  
 
  
 k = n n
Pt Ad = n1 = Pit Ad = k + ni i pkt qt i . (2.82)
k
k=0

Hence, the regional agricultural drought occurrences during the second time interval
follow similarly to this last expression, and generally, for the i-th step, it takes the
following form:
2.10 Spatio-Temporal Drought Theory and Analysis 75

Fig. 2.27 Probability of 0.4

drought area for

PROBABILITY
0.3 t=1 2 3 4 5
multi-seasonal model m=10;
pr =0.3; pt =0.2)
0.2

0.1

0.0
0 1 2 3 4 5 6 7 8 9 10
AREA (m)

m=ni


 
 
  
 k + n k ni
Pit Ad = ni = Pit Ad = k + ni i pt qt . (2.83)
k
k=0

Its validity has been verified using digital computers. The pdfs of areal agricultural
droughts for this model are shown in Fig. 2.27, with parameters m = 10, pr =
0.3, pt = 0.2 and i = 1, 2, 3, 4, and 5.
The probability functions exhibit almost symmetrical forms irrespective of time
intervals although they have very small positive skewness.
Another version of the multi-seasonal model is interesting when the number of
continuously deficit sub-areas appear along the whole observation period. In such a
case, the probability of drought area in the first time interval can be calculated from
Eq. (2.76). At the end of the second time interval, the probability of j sub-areas
with two successive deficits, given that already n1 sub-areas had SMC deficit in the
previous interval, can be expressed as
 
n1 j n −j
P2t (Ad = j |Ad = n1 ) = Pt (Ad = n1 ) pt qt 1 . (2.84)
j

This expression yields the probability of having n1 sub-areas to have deficit out of
which j sub-areas are hit by two deficits, i.e., there are (n1 – j) sub-areas with one
deficit. Hence, the marginal probability of continuous deficit sub-area numbers is


m−j  
k+j j
P2t (Ad = j) = Pt (Ad = k + j) pt qkt .
j
k=0
In general, for the i-th time interval it is possible to write


m−j  
k+j j
Pit (Ad = j) = P(i−1)t (Ad = k + j) pt qkt . (2.85)
j
k=0

The numerical solutions of this expression are presented in Fig. 2.28 for
m = 10, pr = 0.3 and pt = 0.5. The probability distribution function is positively
skewed.
76 2 Data Types and Logical Processing Methods

Fig. 2.28 Drought area 0.5

probability for multi-seasonal i=1
model (m=10; pr =0.3; 0.4

PROBABILITY
i=2
pt =0.2) i=3
0.3

0.2

0.1 i=4
0.0
0 1 2 3 4 5 6 7 8 9 10
NUMBER OF SUBAREAS

2.10.1 Drought Parameters

The global assessment of model performances can be achieved on the basis of

drought parameters such as averages, i.e., expectations and variances, but for
drought predictions the pdf expressions as derived above are significant. The
expected, i.e., average number of deficits, Ei (Ad ), over a region of m sub-areas
during time interval, it is defined as


m
Ei (Ad ) = kPit (Ad = k). (2.86)
k=0

Similarly, the variance, Vi (Ad ), of drought-affected area is given by definition as


m
Vi (Ad ) = k2 Pit (Ad = k) − E2i (Ad ). (2.87)
k=0

Substitution of Eq. (2.79) into Eq. (2.86) leads to drought-stricken average area
within the whole region as


i−1
Ei (Ad ) = mpr qkr , (2.88)
k=0

or succinctly
( )
Ei (Ad ) = m 1 − qir . (2.89)

Furthermore, the percentage of agricultural drought-stricken area, Pi A , can be

calculated by dividing both sides by the total number of sub-areas, m, leading to

PiA = (1 − qir ). (2.90)

2.10 Spatio-Temporal Drought Theory and Analysis 77

1
qr = 0.2
0.9 0.4

0.8 0.6
Average drought coverage area

0.7
0.8
0.6

0.5
0.9
0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7 8 9 10
Number of sub areas

Fig. 2.29 Drought percentage areal coverage

Figure 2.29 shows the change of drought-stricken area percentage with the num-
ber of deficit sub-areas, i, for given deficit probability, qr .
For regional drought variations in the first time interval (i = 1), from Eq. (2.90),
P1 A = pr . On the other hand, for the whole area to be covered by drought, theoret-
ically i →c ∞ and therefore P∞ A = 1. It is obvious that the temporal agricultural
drought percentage for a region of m sub-areas at any time i, is pr ≤ Pi A ≤ 1.
As the probability of deficit spell in a region increases, the average drought
area attains its maximum value in a relatively shorter time as can be written from
Eq. (2.90) as

i = Ln(1 − PiA )/Ln(1 − ρr ). (2.91)

Hence, this expression provides the opportunity to estimate the average time period
that is required for a certain percentage of the region to be covered by drought.
Figure 2.30 indicates the change of i with pr , which is the surplus probability.
Furthermore, in practical applications the probability of deficit can be approxi-
mated empirically as 1/m or preferably as 1/(m+1). The substitution of these condi-
tions into Eq. (2.91) gives

i = Ln(1 − PiA )/[Ln(m/(m + 1)]. (2.92)

This expression confirms that regional drought durations are affected mainly by its
size rather than its shape, as was claimed by Tase and Yevjevich (1978).
78 2 Data Types and Logical Processing Methods

7
Average areal drought coverage area

m = 10
6

5
qr = 0.8

1 0.6
0.4
0.2
0
1 2 3 4 5 6 7 8 9 10
Time

Fig. 2.30 Drought time change by pr

The next significant regional drought parameter is the variance, which is a mea-
sure of drought variability. In general, the smaller the variance the smaller is the
areal drought coverage percentage. The variance of the regional persistence model
can be found from Eqs. (2.76) and (2.79) after some algebra as
 
Vi (Ad ) = m 1 − qtr qtr . (2.93)

Similarly, it is better in practice to standardize this variance by dividing both sides

by the total area, m, which gives percentage variance, PV i,T in the case of i sub-area
coverage after T time duration,

PV i,T = (1 − qir )qTr . (2.94)

Furthermore, consideration of Eq. (2.90) together with this last expression yields
the relationship between percentages of average and variance drought coverage as

PV i,T = PiA qTr . (2.95)

Figure 2.31 shows the change of regional drought variance percentage with i
deficit affected number of sub-areas at different times for a given deficit probability,
qr = 0.7.
2.10 Spatio-Temporal Drought Theory and Analysis 79

0.8 1

0.9
0.7 qr=0.2
qr=0.2
0.8
Variance of areal droughr coverage

0.6

Variance of drought coverage

0.7

0.5 i = 1 subarea i = 3 subareas

0.6

0.4 0.5 0.4

0.4 0.4 0.6

0.3

0.3
0.8
0.2 0.6
0.2

0.1 0.8 0.9

0.1
0.9
0 0
0 5 10 15 0 5 10 15
Time Time

0.9

0.8
Variance of areal drought coverage

i = 10 subareas
0.7

0.6

0.5

0.4
0.9
0.3 0.8

0.2 0.6
0.4

0.1
qr=0.2

0
0 5 10 15
Time

Fig. 2.31 Areal drought variance percentage variation

Another application of spatio-temporal agricultural drought occurrence is possi-

ble by considering spatial deficit probability, qr , and temporal surplus probability,
pt . It is rather cumbersome to find a concise expression for the expectation of this
case at all times. However, for the first time interval,

E1 (Ad ) = m (1 − qr ) (1 − pt ) , (2.96)

which exposes explicitly the contribution of the regional and temporal dry spell
effects on the average areal drought. Figure 2.32 shows drought spatial and temporal
average variations for the given probabilities.
Finally, from the variance of the drought area coverage by simple spatio-temporal
model considerations, it is possible to derive for the first time interval that
80 2 Data Types and Logical Processing Methods

Average areal drought coverage area

5
qr = 0.4

4 m = 10

qr = 0.6
3

2
qr = 0.8

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
SMC deficit time probability

Fig. 2.32 Average areal drought coverage by considering spatio-temporal variations during the
first time interval

Fig. 2.33 Average variance 2.5

of drought area for various Pt = 0.3
basic probability values 0.2

2.0

0.1
1.5
Vi (Ad)

1.0
0.0

0.5

0.0
0 1 2 3 4 5 6 7 8 9 10
Time (i)
References 81

V1 (Ad ) = E1 (Ad ) (qr + pt pr ) . (2.97)

The numerical solution of this expression is given for various combinations of pr

and pt in Fig. 2.33.
As a result the pdfs of regional and temporal agricultural droughts are derived
for independent dry and wet spell occurrences. Two basically different probabilistic
models are proposed for regional drought modeling. Regional drought parameter
variations are assessed graphically. The following conclusions are valid for regional
and temporal drought occurrences:

1) Drought occurrences are dependent on the regional and temporal dry and wet
spell probabilities, as well as size of the region considered.
2) Drought area distribution within a region without considering temporal probabil-
ities becomes negatively skewed as its size increases. Initially, it can be approx-
imated by a normal distribution. For multi-seasonal model the same distribution
has an approximate normal pdf, provided that continuous drought duration is not
considered. Otherwise, it is positively skewed.
3) Drought probabilities over a region are more affected by its size.

References
Akin, J. E., 1971. Calculation of areal depth of precipitation. J. Hydrol. 12, 363–376.
Bayraktar, H., Turalioglu, F. S., and Şen, Z., 2005. The estimation of average areal precipitation by
percentage weighting polygon method in Southeastern Anatolia Region, Turkey. Atmos. Res.
73, 149–160.
Box, G. E. P., and Jenkins, G. M., 1976. Time Series Analysis, Control and Forecasting. Holden
Day, San Francisco.
Bras, R. L., and Rodriguez-Iturbe, I., 1985. Random Functions and Hydrology. Addison-Wesley
Publishing Co., Reading, MA, 599 pp.
Bruce, J. P., and Clark, R. H., 1966. Hydrometeorology. Pergamon Press, New York, 319 pp.
Chow, V. T., 1964. Handbook of Applied Hydrology. McGraw-Hill, New York.
Chu, P. S., and Karz, R.W., 1985. A time-domain approach. Mon. Wea. Rev. 113, 1876–1888.
Cramer, H., 1946. Mathematical Methods of Statistics. Princeton University Press, New Jersey,
p. 213.
Davis, A., 2002. Statistics Data Analysis Geology. John Wiley and Sons, London, 638 pp.
Dirichlet, G. L., 1850. Über die Reduktion der positiven quadratischen Formen mit drei unbes-
timmten ganzen Zahlen. Journal für die Reine und Angewandte Mathematik 40, 209–227.
Feller, W., 1967. An Introduction to Probability Theory and its Application. John Wiley and Sons,
New York, 509 pp.
Fisher, R. A., 1912. On an absolute criterion for filtering frequency curves. Messenger Math. 41,
155.
Gauss, K. F., 1809. Theory of the motion of the heavenly bodies about the sun in conic section.
Dover, New York (1993).
Hevesi, J. A., Istok, J. D., and Flint, A. I., 1992. Precipitation estimation in mountainous terrain
using multi-variate geostatistics, Part I: Structural analysis. J. Appl. Meteorol. 31, 661–676.
Huff, F. A., and Neill, J. C., 1957. Areal representativeness of point rainfall. Trans. Amer. Geophys.
Union 38(3), 341–351.
82 2 Data Types and Logical Processing Methods

Jackson, I. J., 1972. Mean daily rainfall intensity and number of rainy days over Tanzania. Geogr.
Ann. A. 54, 369–375.
Kedem, B., Chiu. L. S., and Karni, Z., 1990. An analysis of the threshold method for measuring
area-average rainfall. J. App. Meteor. 29, 3–20.
Kendall, M. G., 1954. Note on the bias in the estimation of autocorrelation. Biometrika 42,
403–404.
Koch, S., and Link, R. F., 1971. Statistical Analysis of Geological Data. Dover Publications, New
York, 375 pp.
Kolmogorov, A. N., 1941. Interpolation and extrapolation von stationaren zufalligen folgen. Bull.
Acad. Sci. USSR, Ser. Math. 5, 3–14.
Quenouille, M. H., 1956. Notes on bias in estimation. Biometrika 43, 353–360.
Schubert, S. D., Saurez, M. J., and Schemm, J. K., 1992. Persistence and predictability in a perfect
model: J. Atmos. Sci. 49(5), 256–269.
Şen, Z., 1974. Small sample properties of stationary stochastic processe4s and Hurst phenomenon
in hydrology. Unpublished Ph. D. Thesis, Imperial College of Science and Technology, Uni-
versity of London, 256 pp.
Şen, Z., 1980. Regional drought and flood frequency analysis: theoretical considerations. J. Hydrol.
46, 265–279.
Şen, Z., 1998a. Small sample estimation of the variance of time-averages in climatic time series.
Int. J. Climatol. 18, 1725–1732.
Şen, Z., 1998b. Average areal precipitation by percentage weighted polygon method. ASCE J.
Hydrol. Eng. 3(1), 69–76.
Şen, Z., 2002. İstatistik Veri İşleme Yöntemleri (Hidroloji ve Meteoroloji) (Statistical Data Treat-
ment Methods – Hydrology and Meteorology). Su Vakfı yayınları, 243 sayfa (in Turkish).
Şen, Z., 2008. Kuraklik Afet ve Modern Modelleme Yontemleri (Drought Disaster and Modern
Modeling Methods). Turkish Water Foundation (in Turkish).
Stout, G. E., 1960. Studies of severe rainstorms in Illinois. J. Hydraul. Div. Proc. ASCE, HY4,
129–146.
Summer, G., 1988. Precipitation Process and Analysis. John Wiley and Sons, New York, 455 pp.
Tabios, G. O. III, and Salas, J. D., 1985. A comparative analysis of techniques for spatial interpo-
lation of precipitation. Water Resour. Bull. 21, 365–380.
Tase, N., 1976. Area-deficit-intensity characteristics of droughts. Hydrology Paper 87, Colorado
State University, Fort Collins.
Tase, N., and Yevjevich, Y., 1978. Effects of size and shape of a region on drought coverage.
Hydrol. Sci. Bull. 23(2), 203–213.
Thiessen, A. H., 1912. Precipitation averages for large areas. Mon. Wea. Rev. 39, 1082–1084.
Voronoi, G., 1907. Nouvelles applications des paramètres continus à la théorie des formes quadra-
tiques. Journal für die Reine und Angewandte Mathematik 133, 97–178.
Wiesner, C. J., 1970. Hydrometeorology. Chapman and Hall Ltd., London, 232 pp.
Wiener, N., 1942. The extrapolation, interpolation and smoothing of stationary time series. OSRD
370, Report to the Services 19, Research Project DIC-6037, MIT.
Wiener, N., 1949. The Extrapolation, Interpolation and Smoothing of Stationary Time Series. John
Wiley & Sons, Inc., New York.
Wilson, J. W., and Atwater, M. A., 1972. Storm rainfall variability over Connecticut. J. Geophys.
Res. 77(21), 3950–3956.
Zadeh, L. A., 1965. Fuzzy sets. Information Control 8, 338–353.
Zadeh, L.A., 1973. Outline of a new approach to the analysis of complex systems and decision
processes. IEEE Trans. Syst. Man. Cybern. 3, 28–44.
Chapter 3
Classical Spatial Variation Models

Abstract Men have tried to model spatial behavior of the natural phenomena since
a long time, with initiative simple models such as the weighting functions, which
are supposed to represent regional dependence structure of the phenomenon con-
cerned. Unfortunately, commonly employed weighting functions are not actually
data-dependent and hence they are applicable invariably in each spatial prediction,
which is not convenient since each spatial phenomenon will have its own spatial
dependence function. Spatial data distribution can be uniform, randomly uniform,
homogeneous, isotropic, clustering, etc., which should be tested by a convenient test
as described in the text. Besides statistically it is also possible to depict the spatial
variation through trend surface fit methods by using least squares technique. In this
chapter, finally, adaptive least squares techniques is suggested in the form of Kalman
filter for spatial estimation.

Keywords Cluster sampling · Data group · Geometric weights · Kalman

filter · Nearest neighbor · Random field · Spatial pattern · Trend surface ·
Uniformity test

3.1 General

The spatial nature of earth sciences phenomena expelled the researchers to explore
spatial statistical procedures whereas the classical statistics remained at the service
as usual. In general, any phenomenon with spatial variations is referred to as the
ReV by Matheron (1963). ReVs fall between the random fields where the spatial
variations have independence and deterministic variability depending on the spa-
tial correlation value. The most significant methodologies of the spatial analysis is
Kriging, which is discussed in Chapter 5, but prior to their development earth scien-
tists were using different empirical and geometrical rational approaches in assessing
ReVs. Hence, in this chapter these rather preliminary and simple but effective meth-
ods will be discussed with their drawbacks. The trend surface analysis is an offshoot
of statistical regression; kriging is related to time series analysis, and contouring is
an extension of interpolation procedures.

Z. Şen, Spatial Modeling Principles in Earth Sciences, 83

DOI 10.1007/978-1-4020-9672-3_3,  C Springer Science+Business Media B.V. 2009
84 3 Classical Spatial Variation Models

3.2 Spatio-Temporal Characteristics

Our minds are preconditioned on the Euclidean geometry and, consequently, ordi-
nary human beings are bound to think in 3D space as length, width, and depth
in addition to the time as the fourth dimension. Hence, all the scientific formula-
tions including differential equations and others include space and time variability.
Earth, environmental, and atmospheric variables vary along these four dimensions.
If their changes along the time are not considered, then it is frozen in time; and
therefore, a steady situation remains along the time axis, but ReV in concern
has variations in the space. A good example for such a change can be consid-
ered as geological events, which do not change appreciably during human life-
time. For instance, iron content of a rock mass varies rather randomly from a
point to another within the rock mass and hence spatial variation is considered.
Another example is the measurement of rainfall amounts at many irregularly located
meteorological stations spread over an area, i.e., simultaneous measurement of
rainfall amounts; again the time is kept constant and the spatial variations are
sought.
On the other hand, there are time variations that are referred to as temporal vari-
ations in the natural phenomenon. For such a variation, it suffices to measure the
event at a given location, which is the case in any meteorology station or in ground-
water wells. Depending on the time evolution of the event whether it is continuous
or not, time series records can be obtained. A time series is the systematic measure-
ment of any natural event along the time axis at regular time intervals. Depending
on this time interval, time series is called as hourly, daily, weekly, monthly, or yearly
time series.
Contrary to time variations, it is not possible to consider space series where
the records are kept at regular distances, except in very specific cases. For exam-
ple, if water samples along a river are taken at every 1 km, then the measure-
ments provide a distance series in regularity sense. In fact, distance series are
very limited as if there are no such data sequences in practical studies. On the
other hand, depending on the interest of event, they appear as time series data
but they are not time series due to irregularity or randomness in the time inter-
vals between successive occurrences of the same event. Flood and drought occur-
rences in hydrology correspond to such cases. One cannot know the duration of
floods or droughts. Likewise, in meteorology the precipitation occurrence or any
dangerous levels of air pollutants concentrations do not have time series character-
istics.
Any natural event evolves in the four-dimensional human visualization domain
and consequently its records should involve the characteristics of both time and
space variability. Record with this property is said to have spatio-temporal variation.
It is the main purpose of this chapter to present the classical methodologies for
spatial data treatment regional estimations and interpretations, based on simple but
effective methodologies.
3.3 Spatial Pattern Search 85

3.3 Spatial Pattern Search

Human is biased toward regular pattern geometries, appearances, and presentation
rather than irregular or random features. Although at the first instance he/she can dis-
tinguish visually between the regular and the irregular patterns in space, he/she may
not know their quantitative assessment methodologies at the time. For instances,
everyday the pattern of clouds is different than previous days and the most regular
ones are either overcast situation or cloudless patterns. However, in between, there
are millions of different irregular patterns that cannot be distinguished simply from
each other in fine scales. It is, therefore, necessary to sample such situations at a
set of irregular measurement sites and then use quantitative methodologies for the
identification of a set of parameters or a spatial map for each case, which can then
be compared and classified into different categories.
Usually the pattern is invisible, such as oil, gold, and groundwater resources,
sub-surface geological features, but based on the sampling and thereafter the data
treatment with objective methodologies as well as personal expertise and intuition,
the general spatial features may be depicted in the form of maps.
The worth of data in earth sciences and geology is very high since most of the
interpretations and decisions are based on their qualitative and quantitative infor-
mation contents. This information is hidden in representative field samples, which
are analyzed for the extraction of numerical or descriptive characteristics, which
are referred to as data. Data collection in earth sciences is difficult, expensive, and
requires special care for accurately representing the geological phenomenon. After
all, various parameters necessary for the description and modeling of the geological
event, such as bearing capacity, fracture frequency, aperture, orientation, effective
strength, porosity, hydraulic conductivity, chemical contents, etc., are hidden within
each sample, but they represent a specific point in space and time. Hence, it is pos-
sible to attach temporal and three spatial reference systems with data, as shown in
Fig. 3.1.
In geological sciences and applications, the concerned phenomenon can be exam-
ined and assessed through the collection of field data and accordingly meaningful
solutions can be proposed. It is, therefore, necessary to make the best use of avail-
able data from different points. Geological data are collected either directly in the
field or field samples are transferred to laboratories in order to make necessary anal-
yses and measurements. For instance, in hydrogeology domain, among field mea-
surements are the groundwater table elevations, pH, and total dissolved solution
(TDS) readings whereas some of the laboratory measurements are chemical ele-
ments in parts per million (ppm). There are also office calculations that yield hydro-
geological data such as hydraulic conductivity, transmissivity, and storage coeffi-
cients. In the mean time, many other data sources from soil surveys, topographic
measurements, geological prospection, remote sensing evaluations, and others may
also support data for further investigation. The common property of these mea-
surements and calculations of laboratory analysis is that they include uncertainty
86 3 Classical Spatial Variation Models

Fig. 3.1 Data variation by Data

(a) time, (b) space

Time
a

*A
*B

*C

*F *D
0 10 km
*E

attached at a particular time and sampling point. Hence, the first question is how to
deal with rather uncertainly (randomly) varying data? At times the data are random,
sometimes chaotic, and still in other cases irregular or very regular. These changes
can be categorized into two broad classes as systematic and unsystematic. System-
atic data yield mathematically depictable variations with time, space, or both. For
instance, as the depth increases so does the temperature and this is an example for
systematic variation. Especially, if there is only one type of geological formation,
then systematic variation becomes more pronounced. Otherwise, on the basis of
rather systematic variation on the average, there are unsystematic deviations, which
might be irregular or random. Systematic and unsystematic data components are
shown in Fig. 3.2.
In many studies systematic changes are referred to as the deterministic compo-
nents, which are due to systematic natural (geography, astronomy, climatology) fac-
tors that are explainable, and unsystematic variations are unexplainable or random
parts that need special treatment.

Data Systematic

Random
Fig. 3.2 Systematic and
unsystematic components Time
3.4 Spatial Data Analysis Needs 87

3.4 Spatial Data Analysis Needs

It is obvious that the spatial patterns or arrangements of individual sampling loca-
tions in any study area have different patterns. The problem is to visualize quali-
tatively and quantitatively the change of an ReV over these points and then over
the area. Appreciation of some differences between the set records of concerned
variable at a set of measurement stations is the first step prior to any formal model
application. It helps in the final modeling stage if the variation feature of the ReV
variable is deduced first with simple but effective methods. In order to achieve such
a preliminary work, the following steps become helpful:
1. Data Collection: Since the sampling locations are considered on a plane,
their locations can be described by abscissas, X, and ordinates, Y. This procedure
is referred to finding the easting and northing from the longitudes and latitudes.
Although it is not possible to meet always in the practical studies, it is preferable
that the spatial distribution of sampling sites should be rather evenly distributed over
the study area, and if possible constitutes points of a regular grid. The first descrip-
tive exposition of sampling sites may be their location distribution, but without its
relationship to the ReV it does not mean much in spatial analysis. The sampling
locations may already be fixed for researchers, such as the existing well locations
(water or oil), meteorology stations, urban areas. If the distribution of a variable
Z with respect to easting and northing is given as in Table 3.1, then its frequency
distribution can be calculated for various classes and presented in Fig. 3.3.
It is obvious from this figure that as the number of classes increase, the categories
change both in class limits and in frequency values.
On the other hand, it is possible to see the scatter of these data as in Fig. 3.4.
If ReV is denoted in general as Z, then the points in this scatter diagram can be
grouped depending on some criterion as follows with respective symbols.

◦ Z < 189
+ 189 < Z < 536
∗ 536 < Z < 936
x 936 < Z < 1310
1310 < Z < 1632
♦ Z > 1632
These symbols help to make two types of interpretations, namely interclass and
smaller or less than a given class limits, two mutually exclusive but collectively
exhaustive groups. According to the attached symbols, one can deduce the following
interpretations:

a) The majority of small ReVs are gathered at high easting but low northing regions.
Two exceptional clusters of the same type appear at low easting and northing
regions and also in high easting but medium and high northing regions. In short,
there are three clusters of small ReVs. Each cluster has its special dependence
88 3 Classical Spatial Variation Models

Table 3.1 Easting, northing, and ReV

Easting Northing ReV (m) Easting Northing ReV (m)

40.78 30.42 31 41.25 29.02 30

40.52 30.3 100 41.73 27.23 232
39.72 40.05 1631 41.10 29.06 114
37.00 35.33 20 40.60 43.08 1,775
41.17 29.04 130 41.40 27.35 46
40.73 31.60 742 37.93 41.95 896
39.62 27.92 120 41.18 29.62 31
40.18 29.07 100 42.03 35.17 32
40.32 27.97 58 40.98 27.48 4
40.15 29.98 539 40.65 29.27 4
38.40 42.12 1578 39.83 34.82 1,298
40.13 26.40 3 38.50 43.5 1,671
41.17 27.8 183 38.75 30.53 1,034
40.55 34.97 776 38.78 35.48 1,053
37.88 40.18 677 37.75 30.55 997
39.58 28.63 639 37.97 32.55 1,032
39.60 27.02 21 37.20 28.35 646
39.92 41.27 1869 41.28 36.33 4
41.67 26.57 51 36.70 30.73 50
39.78 30.57 789 40.92 38.4 38
40.98 28.80 36 41.37 33.77 799
40.20 25.90 72 39.75 37.02 1,285
40.97 29.08 33 39.73 39.5 1,156
40.93 26.40 10 41.17 29.05 56
40.78 29.93 76 39.93 44.03 858
40.90 29.18 28 38.43 38.08 862

feature different than others. The frequency of low ReV occurrences fall within
this type of data.
b) The ReV variability with values between 189 and 536 takes place also at high
easting and low northing region, with the most concentrated cluster of ReVs less
than 189. In fact, these two groups can be assumed as one cluster in larger scale
as ReV values less than 536.
c) The ReV values that lie between 536 and 936 are scattered almost all over the
sampling area, but they do not fall within the major sampling area where easting
is high and northing is low. This group of sampling points has the maximum ReV
variability over the region in an independent manner. In other words, although
ReVs less than 189 have high regional variability, but at least they are fathered
in three clusters.
d) Those ReV values between 536 and 1,310 show a single cluster within the
medium easting and northing region, without intermixing with other classes.
They show a directional extension along the northeast southwest trend in
Fig. 3.4.
e) The high values of ReV more than 1,310 are gathered at medium easting but
high northing region of the sampling area.
3.4 Spatial Data Analysis Needs 89

30 30

25 25

20 20
Frequency

Frequency
15 15

10 10

5 5

0 0
0 200 400 600 800 100012001400 16001800 2000 0 200 400 600 800 100012001400 16001800 2000
ReV in meters Rev in meters
a b
30 30

25 25

20 20
Frequency

Frequency

15 15

10 10

5 5

0 0
0 200 400 600 800 100012001400 16001800 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000
Rev in meters Rev in meters
c d
Fig. 3.3 Relative frequency diagram of ReV with (a) 5-class; (b) 6-class; (c) 7-class; (d) 8-class

45

40
Northing (m)

35

30

25
Fig. 3.4 Spatial pattern of 36 37 38 39 40 41 42 43
each class Easting (m)
90 3 Classical Spatial Variation Models

f) If the symbols are followed from the smallest to moderate and then to the biggest
ReV values in Fig. 3.4, it is clear that there is an increasing trend from low nor-
thing sampling locations toward high northing direction. In the middle northing
region along the easting direction, there is a hill with low values at low and high
easting regions.

The above interpretations of the ReV data scatter based on the sampling locations
yield clues that become very useful in actual modeling scheme. So far only linguistic
interpretations are deduced and they are the basic ingredients or rules in fuzzy logic
system modeling (Zadeh, 1968; Şen, 2004).
It is important to notice from the qualitative information that distinctive features
are depicted by considering the differences or dissimilarities between the scatter
points and the attached ReV values at each point. The first difference implies the
distances between the sampling points and the second is the difference between
the ReV values at two sites. Hence, in any quantitative study these two differences
should be taken into consideration. These differences are considered independently
from each other, but there may be regional dependence between them. For instance,
visualization of a trend in step f implies such a regional dependence. As will be
explained in Chapter 5, the relationship between the distance of two sampling points
and the difference in their ReV values can be depicted by semivariogram (SV)
concept.
2. Univariate data description: It is necessary and helpful to explore the uni-, bi-,
and multivariate statistical properties of the ReV data irrespective of sampling point
locations. The second step in geostatistics (or, for that matter, any statistical analy-
sis) is to describe the data using univariate (and if possible, bivariate or multivariate)
descriptive statistics. Among the univariate descriptive statistics are the arithmetic
mean, median, mode, standard deviation, skewness, kurtosis, etc.
These parameters can be obtained either from the frequency diagram (as in Fig.
3.3) or through the classical mathematical expression of the parameters (Koch and
Link, 1971). Often used parameters are presented in Table 3.2 for the ReV data in
Table 3.1.
In spatial analysis one should note that there is only one univariate parameter
value for any given ReV variable. As will be explained later in more detail, the

Table 3.2 Univariate statistical

parameters Parameter Value

Minimum 3
Mode 4
Median 125
Arithmetic mean 506.05
Maximum 1,869
Standard deviation 571.11
Skewness 0.89
Kurtosis 0.41
3.4 Spatial Data Analysis Needs 91

Table 3.3 Representative

ReV values Easting (m) Northing (m) ReV (m)

40.78 30.42 31
40.52 30.3 100
41.17 29.04 130
39.62 27.92 120
40.18 29.07 100
40.32 27.97 58
Parameters Value
Minimum 31
Mode 100
Median 100
Arithmetic average 89.83333
Maximum 130
Standard deviation 37.96007
Skewness 37.96007
Kurtosis 0.363526

univariate parameter values do not change with direction in the sampling area. What-
ever the direction, although, the projection of sampling point sequence and distances
will change but the univariate statistical parameters will remain the same. This
explains why the univariate parameters cannot be useful in the spatial dependence
search directly. The spatial variation can be measured by considering comparatively
the properties of at least two ReV values at two distinctive sampling locations,
3. Bivariate data description: This can de achieved by considering at least two
sampling points simultaneously by comparing their ReV values. It can be achieved
simply by considering either a direction or without any direction. In order to famil-
iarize the reader with each one of these approaches, six ReV values with their easting
and northing values are considered from Table 3.1 and presented in Table 3.3. The
scatter of sampling points in this case is given in Fig. 3.5.

30.5

30

29.5
Northing (m)

29

28.5

28

27.5
39.6 39.8 40 40.2 40.4 40.6 40.8 41 41.2 41.4
Fig. 3.5 ReV sampling
locations Easting (m)
92 3 Classical Spatial Variation Models

30.5 30.5
5 5
4 4
30 30
A
29.5 6 29.5

Northing (m)
Northing (m)

5 B 1
29 3 4 6 29 3 6
2
2 3 4 5
3 B
6
28.5 28.5
1
A
28 28
2 1 2
1

27.5 27.5
39.6 39.8 40 40.2 40.4 40.6 40.8 41 41.2 41.4 39.6 39.8 40 40.2 40.4 40.6 40.8 41 41.2 41.4
Easting (m) Easting (m)
a b
C 30.5
30.5 5
5 4
4
4 30
30
5 d34 d35
29.5
29.5
Northing (m)
Northing (m)

d36
29 3 6
29 3 3 6
d32
d31
6 28.5
28.5

1 28
28 2
1 2 2 1

27.5 27.5
39.6 39.8 40 40.2 40.4 40.6 40.8 41 41.2 41.4 39.6 39.8 40 40.2 40.4 40.6 40.8 41 41.2 41.4
C Easting (m)
Easting (m)
c d

30.5 30.5
5 5
4 4
30 30
d64 d64
d65 d65
29.5 29.5
Northing (m)
Northing (m)

d63 d63
6 29 3 6
29 3

28.5 28.5 d61 d62

d61 d62

28 28
2 2
1 1

27.5 27.5
39.6 39.8 40 40.2 40.4 40.6 40.8 41 41.2 41.4 39.6 39.8 40 40.2 40.4 40.6 40.8 41 41.2 41.4

Easting (m) Easting (m)

e f

Fig. 3.6 (a) A–A direction; (b) B–B direction; (c) C–C direction; (d) Partial directional
(Point 3) configuration; (e) Partial directional (Point 6) configuration; (f) Unidirectional (Global)
configuration

On the basis of the data in this table, the spatial relationship between the distance
and the ReV values can be thought in three categories, namely punctual, directional,
and global assessments.
In the directional case all the sampling points are projected onto desired num-
ber of directions and the ReV values are considered not at the sampling points but
3.5 Simple Uniformity Test 93

Fig. 3.7 Directional ReV (m)

variations (a) A–A direction;
(b) B–B direction; (c) C–C * *
100 * *
direction
50
* Distance (m)
1 2 3 4 5 6
a
ReV (m)

* *
100 * *
50 *
* Distance (m)
1 3 24 5 6
b
ReV (m)
* *
100 *
50 * *
Distance (m)
4,5 3 6 1 2
c

at the projection points. In this manner, two-dimensional sampling point scatter is

transformed into one-dimensional directional samplings as shown in Fig. 3.6.
A close inspection of these figures indicates that the sequence of sampling points
and the distances between two successive points differ from each case. This is the
first indication that depending on the sequential properties, the spatial variations
show different appearances. The directions correspond to cross sections of three-
dimensional ReV along these directions. In order to confirm this situation, Fig. 3.7
presents the variation of ReV values with distance along the direction. It is obvious
that although the univariate statistical parameters remain as they are in Table 3.3,
whatever the direction, they expose different patterns.
Comparison of these three cross sections shows different patterns. Especially, in
section C–C points 4 and 5 fall on the same directional point and in this case the
representative ReV value is taken as the average of the two values. The directional
graphs show the change of ReV with distance, but the distances among the points
are irregular. Had it been that they are regular, then the classical time series analysis
could be applied for their statistical analysis. In practice, it is time consuming to
rely on such directions and the question is whether there is an easy statistical way to
inherent the common features of all the directions and then to express the regional
variability by a mathematical formulation, which helps to make ready calculations
provided that the spatial data of ReV are available. This can be achieved by SV, or
covariogram analysis (Chapter 5).

3.5 Simple Uniformity Test

In the case of sampling point scatter as in Fig. 3.4, it is necessary to decide whether
the points in each sub-quadrangle (sub-area) arranged in a manner have more or
94 3 Classical Spatial Variation Models

less the same uniformity? In the case of uniformity, there is no superiority among
sub-areas and each sub-area has the same likelihood of sampling point occurrences.
If there are n sampling points over the whole area and the number of sub-areas is k,
then the expected average uniform sampling number, UE , for each sub-area is

n
UE = . (3.1)
k

This indicates the average number of sampling points per sub-area. However, the
actual sampling point count, Ci , (i = 1, 2, . . ., k) in i-th sub-area is different from
each other and in general from the uniform sampling number. It is possible to check
the uniformity by the Chi-square test as


k
(Ci − UE )2
χ2 = . (3.2)
UE
i=1

The Chi-square distribution has ν = k − 2 degrees of freedom, and one can find
from the Chi-square distribution tables in any statistical text book (Benjamin and
Cornell, 1970) the critical Chi-square value, χcr2 , that corresponds to this degree of
freedom. If χ 2 ≤ χcr2 , then the distribution of the points in each sub-area over the
whole area is uniform.

Example 3.1
The positions of many earthquake measurement stations are given in Fig. 3.8, and
it has been divided into nine equal sub-areas. The question is whether the spatial
scatter of station locations is uniformly distributed in the area or not?

90

80

70 1 2 3
60
Northing (km)

50

40 4 5 6

30

20

10 7 8 9
0
Fig. 3.8 Earthquake station 0 10 20 30 40 50 60 70 80 90
location scatter Easting (km)
3.6 Random Field 95

Table 3.4 Sub-areal stations

Sub-area No. Station No.

1 7
2 15
3 8
4 7
5 6
6 12
7 4
8 5
9 12
Total 75

Table 3.4 indicates the number of stations in each sub-area where the station
locations on the boundary are considered as belonging to the right and upper sub-
areas.
According to Eq. (3.1), the expected average uniform number of stations in each
sub-area is 75/9 = 8.33, which can be taken as a round number equal to 8. The
application of Eq. (3.4) leads to χ 2 = 14. Since the degree of freedom is ν = 9 – 2
= 7, the critical χcr2 value at 5% significance level appears as 14.2 (i.e., χ 2 ≤ χcr2 )
and hence the distribution of the points is almost uniform.

3.6 Random Field

If the distribution of sampling points in the study area is not uniform, then one starts
to suspect whether they constitute a random field or not? The random field is the
collection of many randomly scattered points in an area where there is no spatial
correlation between the points. If the points are completely random without any
spatial (regional) dependence, then the only way of their spatial treatment is the
probability axioms.
Let us consider the whole study area, A, with m sampling locations. If they are
random, then the probability of one sampling point occurrence can be expressed as
a percentage,

m
λ= , (3.3)
A

which lies between 0 and 1 exclusive. Now the total area is considered in terms
of very small sub-areas where the number of sub-areas, n, is much larger than the
number, m, sampling point (n >> m). This means that m/n goes to zero as the number
of sub-areas increase. Each one of the n sub-areas can be considered as a pixel (see
Fig. 2.2), which is very small, compared to total area. This makes it possible to
consider that each pixel is almost equal to the influence area of one sampling point.
It is, therefore, impossible to have two sampling points within each pixel. Hence,
the area of influence for each sampling point can be calculated as
96 3 Classical Spatial Variation Models

A
a= , (3.4)
n

which is the pixel area. Since each sampling point has the occurrence probability
of λ per area, then in the influence area the probability of sampling occurrence
becomes
A
po = λa = λ . (3.5)
n
Accordingly, the probability of non-occurrence of the sampling point in a pixel can
be obtained simply as
A
pn = 1 − p0 = 1 − λ . (3.6)
n
If the question is to find k sampling sites in n pixels, then in the remaining n−k
pixels there will not be any sampling point. Since the spatial sampling point occur-
rences and non-occurrences are independent from each other, the production oper-
ation of the probability theory gives the probability of k sampling point occurrence
according to Binomial pdf as
      
n n A k A n−k
pk = (p0 )k (pn )n−k = λ 1−λ . (3.7)
k k n n

After this line, the remaining is the application of pure mathematical principles; and
as n goes to very large numbers (mathematically positive infinity), then Eq. (3.7)
takes its new shape as

pk = e−λA [(λA)k /K!], (3.8)

which has the name of Poisson distribution in common statistics. This equation
requires only the rate of sampling points, λ (Eq. 3.3), sampling point number, k, and
the area of the concerned region. The combined value λA indicates the mean number
of stations per quadrant. All these values are practically observable or calculatable,
and therefore it is possible to calculate pk from Eq. (3.8). If for a given confidence
interval (90 or 95%) the critical probability value, pcr is found from the Poisson
distribution tables in any textbook on statistics (Benjamin and Cornell, 1970) and if
pk < pcr , then the sampling points are randomly distributed in the area.
However, in earth sciences the distribution of sampling points is exactly neither
random nor uniform.

Example 3.2
It is now time to search for the spatial random character of the same example given
in the previous section. As shown in Fig. 3.9 the whole area is divided into 9×9 =
81 quadrants.
3.6 Random Field 97

Fig. 3.9 Quadrants and 90

station locations
80

70

60

Northing (km)
50

40

30

20

10

0
0 10 20 30 40 50 60 70 80 90
Easting (km)

The mean number of stations in each quadrant is λA = 75/81 = 0.926. The first
column in Table 3.5 shows the number of quadrants with 1, 2, 3, etc., stations. The
probability of occurrence calculations is given in the second column.
According to Eq. (3.2) one can calculate that χ 2 = 1.064. Since there are five
categories, the degree of freedom is ν = 5 – 2 = 3. The critical value of χ 2 for ν = 3
and the significance level of 5% is 7.81. The test statistics is less than this critical
value and so the hypothesis of randomness is acceptable.
The mean number of stations per quadrant λA and its variance can be esti-
mated as

T
(r1 − m/T)2
i=1
s2 = , (3.9)
T−1

where ri is the number of stations per quadrant and T is the number of quadrantsis. It
is well known that in Poisson distribution the arithmetic average is equal to variance,
and by using this rule it is possible to make further interpretations. For instance, if

Table 3.5 Applied and theoretical numbers of quadrants

Number of quadrants
Number of Equation (3.8)
stations calculations Theoretical Actual

0 0.3961 32 31
1 0.3668 30 29
2 0.1698 14 14
3 0.0524 4 6
4 0.0121 1 1
Total 0.9972 81 81
98 3 Classical Spatial Variation Models

the arithmetic average is greater (smaller) than the variance, the scatter of stations
is more uniform (clustered) than random. If the two parameters are equal to each
other, the scatter of stations accords with a complete random behavior. However,
at this stage it must be kept in mind that some sampling differences occur between
these two parameters and in practice it is not possible to have them equal in an exact
manner.

3.7 Cluster Sampling

Due to stratification of many earth sciences events the sampling procedure also
shows grouping pattern in different non-overlapping clusters (classes) as in Fig. 3.3.
In the modeling of cluster sampling there are two stages. First is the modeling of
cluster centers and then the second is the modeling of sampling points within each
cluster relative to its center. The first is inter-cluster modeling and the second is
within cluster modeling. As stated in many publications, the negative exponential
distribution can be used to model the occurrence of clustered points in space in a
manner equivalent to the use of the Poisson model to represent random field points.
The cluster point distribution is modeled according to the Poisson distribution and
the points in each cluster by the logarithmic distribution. Hence, the negative expo-
nential distribution function provides the occurrence of k sampling points probabil-
ity as (Davis, 2002)

  k  n
n+k−1 po 1
Pk = , (3.10)
k 1 + po 1 + po

where k is the number of sampling points, po is the probability of event occurrence at

a sampling site, and n is the degree of clustering of the occurrences. If n is large, the
clustering is less spelled and the sampling point distribution is close to random field
distribution of the points. However, as n approaches to zero the clustering pattern
becomes more pronounced over the area. The density of sampling points is

λ = npo . (3.11)

The following mathematical approximation is considered for solving Eq. (3.10):

1
P0 = , (3.12)
(1 + po )n

and the following recursive formula,

( )
(n + k − 1) 1
po + po
Pk = Pk−1 , (3.13)
k
3.7 Cluster Sampling 99

helps to calculate the subsequent probabilities. The clustering parameter can be cal-
culated as
 m 2
n= T
, (3.14)
s2 − m
T

where s2 is the variance in the number of occurrences per tract, which is defined as

m
λA = ,
T
or
m
T= . (3.15)
λA

Likewise, the probability of occurrence can be calculated as

m
λ
p0 = = T.
n n

Example 3.3
The same example given in the previous section can be adopted for the application of
cluster test. It is already calculated that λA = 0.926. The variance, s2 , in the number
of occurrences per quadrangle (sub-area) is 0.897. With these values at hand, the
clustering effect from Eq. (3.14) becomes n = 27.74. Hence, from Eq. (3.12) p0 =
0.0323, and it is possible to calculate from Eq. (3.13) the probability that a given
quadrant will have 1, 2, 3, etc., stations as shown in Table 3.6.
In order to test the theoretical values with the actual correspondences, it is nec-
essary to apply chi-square test, which yields χ 2 = 1.249. Since there are five cate-
gories, the degree of freedom is ν = 5 – 2 = 3. The critical value of χ 2 for ν = 3
and the significance level of 5% is 7.81. The test statistics is less than this critical
value and so the hypothesis of randomness is acceptable.

Table 3.6 Negative exponential distribution expected number of quadrants with r stations

Number of quadrants
Number of stations Probability Theoretical Actual

1 0.4140 33.5 31
2 0.3593 29.1 29
3 0.1616 13.1 14
4 0.0501 4.0 6
5 0.0120 1.0 1
Total 1.0020 81.0 81
100 3 Classical Spatial Variation Models

3.8 Nearest Neighbor Analysis

Rather than considering quadrangles in the subdivision of the total area into sub-
areas and then to make calculations on the basis of the sampling points within the
sub-area, it is preferable to consider the neighbor points next to each other. Log-
ically, if there are n sampling points within a total area, A, then the size of each
equal sub-area, a, is simply calculated according to Eq. (3.4). This area is supposed
to include two nearest points on the average. It is considered as a square and there-
fore the side length, L, of the square is the square root of the area,
√
L= a. (3.16)

Hence, the mean distance between two points is

1
dT = L. (3.17)
2

The variance of the average distance between two points can be calculated ratio-
nally as

(4 − π )
σT2 = A. (3.18)
4π n2

In the derivations of Eqs. (3.17) and (3.18), the area is assumed without boundary,
i.e., very extensive. However, this is not the situation in practice, and therefore these
statistical parameters are without areal extent restrictions and hence they provide
under-estimations. If the constants are worked and the standard error of estimate is
calculated, the result becomes

0.26136
sTe =  . (3.19)
a
n2

If the number of sampling points is more than six, then the distribution of the average
distance between nearest neighbors comply with a normal pdf. The mean and the
variance of this pdf are given by Eqs. (3.17) and (3.19), respectively. Hence, for
decision the standardized normal pdf value, x, can be obtained as

d − dT
x= . (3.20)
sTe

As mentioned earlier the theoretical Eqs. (3.17) and (3.18) are in under-estimations,
and therefore a correction factor less than 1 must be imported. Many researchers
suggested different adjustments, but the one given by Donnelly (1978) found
frequent use. According to his extensive numerical simulation studies, the theoreti-
cal mean distance and its variance can be expressed as (Davis, 2002)
3.8 Nearest Neighbor Analysis 101

• • • • • • • • • • • • •
• • • • • • • • • • • • • •
• • • • • • • • • • • • •
• • • • • • • • • • • • • •
• • • • • • • • • • • • •
• • • • • • • • • • • • • •
• • • • • • • • • • • • •
• • • • • • • • • • • • • •

a b
• • • • • • • • • • •
• • • • •• • • • • • • •
•
• • • • • • • •
•
• • • • • • • • • • •
•
• • • • • • • •
• • • •
• • • •• • • • • • •
• • • • • • •
• • • • • • • •
•
•
c d
• • •••
• •
•

• • •
• ••• ••
• • •• •
•
••• ••• •• •
• • • • •
•••• •
• •
• •

e f
Fig. 3.10 Nearest-neighbor ratio statistics

*  
1 A 0.412 P
dT ≈ + 0.514 + √ , (3.21)
2 n n n

and
√
A A
σT2 ≈ 0.0070 2 + 0.035P 2.5 , (3.22)
n n

where P is the perimeter of the regular map. The ratio, Rd , of the expected and
observed mean nearest-neighbor distances can be used to indicate the spatial pat-
tern as
102 3 Classical Spatial Variation Models

d
Rd = . (3.23)
dT

This ratio approaches to zero where all the sampling points are very close to each
other with almost negligible average distance. Another extreme point appears as Rd
approaches 1, where the random field scatter of sampling points takes place. When
the mean distance to the nearest neighbor is maximized, then Rd takes its maximum
value as 2.15. Figure 3.10 indicates the sampling point distribution with different
distance ratio indices.
The distance ratio indices for Fig. 3.8a to f are equal to 2.15, 1.95, 1.20, 0.89,
0.31, and 0.11, respectively.

3.9 Search Algorithms

One of the critical differences between various local-fit algorithms for interpolating
to a regular grid is the way in which “nearest neighbors” are defined and found.
Some may depend on the number of nearest points and others adopt a radius of
influence and depend on the points within this area of influence (Chapter 2). Some
search techniques (such as the optimum interpolation technique) may be superior
to others in certain situations, especially concerning the data points arrangement.
There are a variety of search procedures that may be selected by the user. The sim-
plest method finds the n nearest neighboring data points in a Euclidean distance
sense, regardless of their irregular distribution around the estimation node, which is
fast and satisfactory if the observations are distributed in a comparatively uniform
manner, but provides poor estimates if the data are closely spaced along widely
separated traverses.
One of the objections to a simple nearest neighbor search is that all the near-
est points may lie, for instance, in a narrow wedge on one side of the grid node
(Fig. 3.11). Hence, the estimation is essentially unconstrained, except in one direc-
tion. It is possible to avoid such a situation by restricting the search in some way,
which ensures that the points are equitably distributed about the estimation point.
The simplest method introduces a measure of radial search as a quadrant. Some
minimum number of points must be taken from each of the four quadrants around
the estimation point. An elaboration on the quadrant search is an octant search,
which introduces a further constraint on the radial distribution of the points used
in the estimation procedure. A specified number of control points must be found in
each of the 45◦ segments surrounding the estimation point. These constrained search
procedures require finding and testing more neighboring points than in a simple
search, which increases time consumption. It is also possible to consider all points
within a radius R of the estimation point. Estimates made using the other search
procedures are based on a fixed number of points collected at variable distances
from the grid node. This search algorithm uses a variable number of points within a
fixed distance of the estimation point.
3.9 Search Algorithms 103

Fig. 3.11 Regular directional

nearest point systems, (a) x
Quadrant; (b) Octant x x
x x
x x
x x
x
x
x
x
x x
x

a) Quadrant search

x
x x
x x
x x
x x
x
x
x x
x
x x

b) Octant search

Any constraints on the search for the nearest points, such as a quadrant or octant
requirement, will obviously expand the size of the neighborhood around the estima-
tion point. This occurs because some nearby control points are likely to be passed
over in favor of more distant points, in order to satisfy the requirement that only a
few points may be taken from a single sector. Unfortunately, the autocorrelation of
a surface decreases with increasing distance, so more remote points are less closely
related to the estimation point. This means the estimate may be poorer than if a
simple nearest neighbor search procedure is used.

3.9.1 Geometric Weighting Functions

Operational objective analysis procedures are suggested for interpolating in a prac-

tical way the information from unevenly distributed observations to a uniformly
distributed set of grid points.
The earliest studies were by Gilchrist and Cressman (1954), who reduced the
domain of polynomial fitting to small areas surrounding each node with a parabola.
Bergthorsson and Döös (1955) proposed the basis of successive correction methods,
which did not rely only on interpretation to obtain grid point values but also a pre-
liminary guess field is initially specified at the grid points. Cressman (1959) devel-
oped a number of further correction versions based on reported data falling within a
specified distance R from each grid point. The value of R is decreased with succes-
sive scans and the resulting field of the latest scan is taken as the new approximation.
104 3 Classical Spatial Variation Models

Barnes (1964) summarized the development of a convergent weighted-averaging

analysis scheme, which can be used to obtain any desired amount of detail in the
analysis of a set of randomly spaced data. The scheme is based on the supposi-
tion that the two-dimensional distribution of an atmospheric variable can be repre-
sented by the summation of an infinite number of independent waves, i.e., Fourier
integral representation. A comparison of existing objective methods up to 1979 for
sparse data is provided by Goodin et al.(1979). Their study indicated that fitting a
second-degree polynomial to each subregion triangular in the plane, with each data
point weighted according to its distance from the subregion, provides a compromise
between accuracy and computational cost. Koch et al. (1983) presented an extension
of the Barnes method, which is designed for an interactive computer scheme. Such
a scheme allows real-time assessment both of the quality of the resulting analyses
and of the impact of satellite-derived data upon various meteorological data sets.
Both Cressman and Barnes methods include power parameter and radius of influ-
ence values, which are rather subjectively determined in the current meteorological
practice.
In any optimum analysis technique, the main idea is that the estimation at any
point is considered as a weighted average of the measured values at irregular sites.
Hence, if there are i = 1, 2, . . . , n measurement sites with records Zi , then the esti-
mated site solar irradiation, ZC , can be calculated according to Eq. (2.27), which is
most commonly used in different disciplines because of its explicit expression as
the weighted average. Weighting functions proposed by Cressman (1959), Gandin
(1963), and Barnes (1964) also appear as sole functions of the distances between
the sites only. Unfortunately, none of the weighting functions are event-dependent,
but they are suggested on the basis of the logical and geometrical conceptualiza-
tions of site configuration. Various geometrical weighting functions are shown in
Fig. 3.12.
The weighting functions that are prepared on a rational and logical basis without
consideration of regional data have the following major drawbacks.

(a) They do not take into consideration the natural variability of the regional vari-
ability features. For instance, in meteorology, Cressman (1959) weightings are
given as
⎧
⎨ R2 −r2i,E
for ri,E ≤ R
W(ri,E ) = R2 +r2i,E , (3.24)
⎩0 for r ≥ R
i,E

where W(ri,E ) corresponds to Wi in Eq. (2.27); di,E is the distance between esti-
mation point and other points; R is the radius of influence, which is determined
subjectively by personal experience.
(b) Although weighting functions are considered universally applicable all over
the world, they may show unreliable variability for small areas. For instance,
within the same study area, neighboring sites may have quite different weight-
ing functions.
3.9 Search Algorithms 105

Fig. 3.12 Weighting w=1

functions 1.0
2 2
w= R –r
R + r2
2

w = exp ⎡–4 ( r ) ⎡
2
⎣ R ⎣

2 2 4
w= R –r
R + r2
2

0
0 1
r
R

(c) Geometric weighting functions cannot reflect the morphology, i.e., the regional
variability of the phenomenon. They can only be considered as practical first
approximation tools.

A generalized form of the Cressman model with an extra exponent parameter α

is suggested as
⎧ α
⎨ R2 −r2i,E
for ri,E ≤ R
W(ri,E ) = R2 +r2i,E . (3.25)
⎩
0 for ri,E ≥ R

The inclusion of α has alleviated the aforesaid drawbacks to some extent, but
its determination still presents difficulties in practical applications. Another form
of geometrical weighting function was proposed by Sasaki (1960) and Barnes
(1964) as
 ( ) 
ri,E 2
W(ri,E ) = exp −4 . (3.26)
R

In reality, it is expected that weighting functions should reflect the spatial depen-
dence behavior of the phenomenon. To this end, regional covariance and SV func-
tions are among the early alternatives for the weighting functions that take into
account the spatial correlation of the phenomenon considered. The former method
requires a set of assumptions such as the Gaussian distribution of the regionalized
106 3 Classical Spatial Variation Models

variable. The latter technique, semivariogram (SV), does not always yield a clear
pattern of regional correlation structure (Şen, 1989).

3.10 Trend Surface Analysis

The main purpose of trend surface analysis is to summarize data that are numerous
to grasp readily by eye from the list of numbers. A spatial trend is a systematic
change usually in a geographical (longitude and latitude) direction in the ReV. The
trend surface analysis is also called a “global interpolator” because it uses all data.
Most interpolators (called “local interpolators”) use only a sample of points to esti-
mate the value at each grid node location. It fits the entire surface with a polyno-
mial equation by the ordinary least squares method. Once trend surfaces were the
only basic tool as maps for communication in earth sciences. In trend analysis, the
statistical model coefficients are estimated and the equation is presented as a con-
tour map. Trend surface presentation in the form of contour maps is the same as
the preparation of topographic maps, where the spatial dependence is not consid-
ered explicitly. The mathematical model of the trend surface is ready in the form
of polynomial equation, which relates the geographic variables x and y to the ReV,
z as

z = a0 + a1 x + a2 y + a3 x2 + a4 xy + a5 y2 , (3.27)

where ai ’s are the model coefficients to be estimated from the available data recorded
at a set of geographical points within the study area. Although there are many sta-
tistical models, this model is appropriate for a smooth concave and convex surface.
The locations of local maximum and minimum can be obtained from Eq. (3.27) by
taking the partial derivative with respect to x and y, which yields

∂z
= a1 + 2a3 x + a4 y,
∂x
and
∂z
= a2 + a4 y + 2a5 y.
∂y
When these derivatives are set equal to zero, the solution for the extreme points
become
a2 a4 − 2 a1 a5
xe = ,
4 a3 a5 − a24
and
a1 a4 − 2 a2 a3
ye = .
4 a3 a5 − a24
3.10 Trend Surface Analysis 107

Different interpretations can be made of an extreme case depending on the analysis.

Among the possibilities are that the extreme point may be a structural high, low, or
inflection point.
In any trend surface analysis the only limit on the number of coefficients is that
one cannot have more than the number of sample points. It is good for approximat-
ing the general form of the surface. Estimates of individual values are not very good
(including at sample points). The interpolated surface will always have the same
value as the sample value where there are sample points, and assigns each point to
be interpolated the same value as the nearest sample point. It gives a stair-step shape
to the surface discontinuities.
Trend surface analysis is a popular numerical technique in trying to explore
the three-dimensional features of any phenomena based on the geographical coor-
dinates; but due to a set of restrictive assumptions they cannot be regarded as
successful at the time. Hence, it is frequently misused and, consequently, the con-
clusions based on such uses may also be in great error. Majority of problems arise
from the measurement site distribution, lack of real data fit, extensive computational
requirements, and at times inappropriate applications. In general, trend surface fit-
ting methodologies are multivariate statistical model fitting to a set of regional
measurement data with restrictive assumptions. A uniform density of data points
is important in many types of analysis, including trend surface methods.
Trend analysis separates the ReV into two complementary components, namely
regional nature of deterministic variations and local fluctuations around the regional
component. The regional and local components are dependent on the scale of the
ReV. In any trend analysis there are three variables.

1) The basic variables are the geographical coordinates, which give the exact loca-
tions of the sampling points in the study area. For trend analysis the longitude
and latitude are converted to easting and northing values with respect to a com-
mon reference point. In trend analysis the choice of the reference point does not
make any difference in further calculations.
2) The trend component is the regional representation of the event in rather deter-
ministic manner, which has the form
∧
z = a0 + a1 x + a2 y + a3 x2 + a4 xy + a5 y2 , (3.28)

where x and y are some function or combination of the geographical coordinates

of the sample locations. This expression yields the trend component as at the i-th
location of the ReV. Although there are non-linear terms, this equation is called
linear because the terms are added together.
3) The trend coefficients are chosen such that the squared deviations of measure-
ments from the trend surface are minimized. The sum of the squares from the
trend surface is equivalent to the variance of the ReV. The multiple linear regres-
sions are also defined similarly, and therefore the trend analysis is similar to the
multivariate analysis in the classical statistics.
108 3 Classical Spatial Variation Models

3.10.1 Trend Model Parameter Estimations

It is possible to obtain the coefficient estimations in Eq. (3.27) through the appli-
cation of regression methodology with the criterion of least squares analysis. This
is a very formal way of deriving the coefficient estimation expressions in terms of
the data values. However, in this book a more practical expression is obtained by
successive independent variable term multiplication and average taking procedure.
The procedure says the following steps.

1) Consider the main Eq. (3.27) and then take the arithmetic average of both sides,
which leads to

z = a0 + a1 x + a2 y + a3 x2 + a4 xy + a5 y2 . (3.29)

This is the first estimation expression where all the arithmetic averages can be
obtained from an available data set. Since there are six coefficients as unknowns,
it is necessary to obtain five more expressions.
2) Multiply both sides of the main equation by the first term variable on the right-
hand side and then take the arithmetic averages. This procedure yields finally,

zx = a0 x + a1 x2 + a2 yx + a3 x3 + a4 x2 y + a5 xy2 . (3.30)

3) Apply the same procedure as in the previous step, but this time multiplying both
sides by the second independent variable, i.e., y, which gives

zy = a0 y + a1 xy + a2 y2 x + a3 yx2 + a4 xy2 + a5 y3 . (3.31)

4) Multiply both sides by x2 and then take the arithmetic average of both sides,
which leads to

zx2 = a0 x2 + a1 x3 + a2 x2 y + a3 x4 + a4 x3 y + a5 x2 y2 . (3.32)

5) This time the independent variable is xy and accordingly its use under the light
of aforementioned procedure results in

xyz = a0 xy + a1 x2 y + a2 x y2 + a3 x3 y + a4 x2 y2 + a5 xy3 . (3.33)

Finally, considering the last independent term variable as y2 , the same rule yields
the following expression

y2 z = a0 y2 + a1 y2 x + a2 y3 + a3 y2 x2 + a4 xy3 + a5 y4 . (3.34)

Hence, the necessary equations are obtained for the trend surface model coefficients
estimation. In matrix notation these equations can be written explicitly as
3.10 Trend Surface Analysis 109

⎡ ⎤−1
⎡ ⎤ 1 x y x2 xy y2 ⎡ ⎤
a0 z
⎢ ⎥
⎢ a1 ⎥ ⎢ x x2 xy x3 x2 y xy2 ⎥ ⎢xz ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ a2 ⎥ ⎢ y xy y2 x2 y xy2 y3 ⎥ ⎢yz ⎥
⎢ ⎥=⎢ ⎥ ⎢ ⎥
⎢ a3 ⎥ ⎢ x2 x3 x2 y x4 x3 y x2 y2 ⎥ ⎢ x2 z ⎥ .
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ a4 ⎦ ⎢ ⎥ ⎣ xyz ⎦
⎣ xy x2 y xy2 x3 y x2 y2 xy3 ⎦
a5 y2 z
y2 xy2 y3 x2 y2 xy3 y4
The right-hand side of this last expression includes various averages that are calcula-
ble from a given set of data, and therefore the parameter estimation is possible after
the inverse of square matrix. This matrix is symmetrical around the main diagonal.
It is also possible to construct the titles of calculation table by considering these
averages as in Table 3.7 the following table titles.

Table 3.7 Trend surface calculation table

x y z x2 xz xy yz y2 x3 x2 y x2 z xyz y2 z xy2 y3 x4 x3 y x2 y2 xy3 y4

The arithmetic average of each column gives the elements of each matrix on the
right-hand side of Eqs. (3.31, 3.32, 3.33, 3.34, 3.35, 3.36, and 3.37).

Example 3.4
A set of earthquake records are given for some part of Turkey, and the magnitude is
required to be related to easting (x) and northing (y) coordinates for the prediction of
Richter magnitude. The first three columns of Table 3.8 give the easting, northing,
and earthquake magnitude values, respectively. For the sake of argument, simple
linear trend surface equation is adopted as

z = a + bx + cy, (3.35)

where a, b, and c are the trend surface parameters. According to the steps given
above, the rest of the table is prepared accordingly.
There are three unknowns and it is necessary to obtain three equations in the light
of practical trend surface calculations. By considering the last row of averages from
Table 3.8, one can write the necessary equations simply as

6.38 = a + 39.12b + 33.30c

249.64 = 39.12a + 1533.62b + 1306.02c
211.97 = 33.30a + 1306.02b + 1150.11c.

The simultaneous solution of these equations yields a = 5.63, b = 0.0314, and c =

−0.0143, and hence the final linear trend surface expression is given as

z = 5.56 + 0.0314x − 0.0143y.

110 3 Classical Spatial Variation Models

Table 3.8 Earthquake data and trend calculations

Number x y z xz X2 xy yz y2

1 40.3 38.4 5.1 205.53 1624.09 1547.52 195.84 1474.56

2 38 27 6 228 1444 1026 162 729
3 39 39 6.3 245.7 1521 1521 245.7 1521
4 40.5 42.7 6.2 251.1 1640.25 1729.35 264.74 1823.29
5 38 26.5 6 228 1444 1007 159 702.25
6 40.18 38.1 6.75 271.215 1614.432 1530.858 257.175 1451.61
7 42.5 26.4 5.9 250.75 1806.25 1122 155.76 696.96
8 41 34 6.2 254.2 1681 1394 210.8 1156
9 36 30 6.25 225 1296 1080 187.5 900
10 40.65 27.2 7.75 315.0375 1652.423 1105.68 210.8 739.84
11 40.6 27.1 6.4 259.84 1648.36 1100.26 173.44 734.41
12 40.1 26.8 6.9 276.69 1608.01 1074.68 184.92 718.24
13 38 30 6.9 262.2 1444 1140 207 900
14 40.27 36.38 7.1 285.917 1621.673 1465.023 258.298 1323.504
15 39.26 26.71 7 274.82 1541.348 1048.635 186.97 713.4241
16 35.5 34 5.8 205.9 1260.25 1207 197.2 1156
17 39.7 42.8 5.3 210.41 1576.09 1699.16 226.84 1831.84
18 39.96 41.94 6.8 271.728 1596.802 1675.922 285.192 1758.964
19 38.55 30.78 5.9 227.445 1486.103 1186.569 181.602 947.4084
20 41.33 43.41 6 247.98 1708.169 1794.135 260.46 1884.428
21 38 30.5 5.9 224.2 1444 1159 179.95 930.25
22 37.03 29.43 6.1 225.883 1371.221 1089.793 179.523 866.1249
23 35.84 29.5 7 250.88 1284.506 1057.28 206.5 870.25
24 36.54 27.33 7.3 266.742 1335.172 998.6382 199.509 746.9289
25 40.94 43.88 6 245.64 1676.084 1796.447 263.28 1925.454
26 38.1 27.1 6.5 247.65 1451.61 1032.51 176.15 734.41
27 40.5 26.5 6.1 247.05 1640.25 1073.25 161.65 702.25
28 40.2 37.9 6.1 245.22 1616.04 1523.58 231.19 1436.41
29 37.98 44.48 7.6 288.648 1442.48 1689.35 338.048 1978.47
Averages 39.12172 33.30483 6.384483 249.6336 1533.642 1306.022 211.9668 1150.113

3.11 Multisite Kalman Filter Methodology

Trend surface analysis depends on continuous and regular mathematical surfaces,
which does not consider individual features in any site separately, but globally fits
the model to the available spatial data at fixed time. It is necessary to repeat the same
procedure for any subsequent time interval independently from the previous step. It
is, therefore, necessary to adopt a spatial model, which digests the spatial data at a
given time and then updated the model output with the new recordings at the same
locations in the region. It is possible to achieve such a task through Kalman Fil-
ter (KF) procedure. KF is an optimal state estimation process applied to a dynamic
system that involves random perturbations. The atmosphere can be regarded as an
uncertain dynamical system (Kalman, 1960; Dee, 1991). More precisely, KF gives a
linear, equitable, and minimum error variance recursive algorithm to optimally esti-
mate the unknown state of a dynamic system from the noisy data taken at continuous
3.11 Multisite Kalman Filter Methodology 111

or discrete real-time intervals. KF supplies a measure of accuracy of an analysis in

the form of its error covariance (Kalman and Bucy, 1961). Therefore, it allows the
impact of different observation sets to be compared.
KF is the most general approach to statistical estimation and prediction. It has
been shown by Harrison and Stevens (1975) that all forecasting methods are special
cases of KF. This filter can deal with changes in the model, parameters, and vari-
ances. The difficulty with KF is that many technical questions have not yet been
answered satisfactorily. The approach itself has grown out of engineering practices.
Consequently, many statisticians and operation researchers know little about it, or
find it difficult to understand, because it is most often described in state space nota-
tion. Furthermore, many practical difficulties still exist as the initial estimates for
parameters, variances, covariance, and transition matrix. As shown in Fig. 3.13,
the discrete KF is an iterative procedure containing several elements, which are
described in the next sections.
The filter is supplied with initial information, including the prior estimate of
initial parameters, which is based on all the knowledge about the process, and the
error covariance associated with it, and these are used to calculate a Kalman gain.
The error between the parameter estimation and the measured data is determined
and multiplied by Kalman gain to update the parameter estimate and estimation
error. The updated error and parameters are used as inputs to a model, in order
to predict the projected error and parameters at the next time instance. The first
derivation of the KF recursive equations is the “filter” equation. The equations used
in the discrete KF are given in detail by Brown and Hwang (1992). Latif (1999)
has performed multisite KF development and application for Turkish precipitation
in detail.

Enter prior estimate and the error

covariance matrix associated with it.

Compute Kalman
gain.

Project ahead the

prior estimate and
the error covariance Update estimate
matrix associated with measurement
with it.

Compute error
covariance for
updated estimate.
Fig. 3.13 KF iterative
procedure
112 3 Classical Spatial Variation Models

The objective of this section is to investigate and develop a KF model approach

to multisite precipitation modeling, and prediction in addition to the assessment of
associated errors. In order to have an online prediction operation, it is desirable to
be able to deal with a multitude of rainfall events. The precipitation predictor should
not be fixed with time and space, but should adapt itself to the evolving meteoro-
logical conditions. Any stochastic model is associated with various uncertainties.
KF consists of combining two independent estimates to form a weighted estimate
or prediction. One estimate can be a prior prediction or an estimate based on prior
knowledge, and the other a prediction based on new information (new data). The
purpose of KF is to combine these two pieces of information to obtain an improved
estimate.

3.11.1 One-Dimensional Kalman Filter

Suppose that there is a random variable, Xk , whose values should be estimated at a
set of certain times t0 , t1 , t2 , . . . ,tn and that Xk−1 satisfies the following dynamic
system equation,
Xk = ϕXk−1 + Wk−1 . (3.36)

In this expression, φ is a known parameter relating Xk to Xk−1 , and Wk−1 is

a pure independent random number with zero mean Wk−1 = 0 and the variance
equals to Q. Let the measurement, Zk , at time k is given as

Zk = Hk Xk + Vk , (3.37)

where Vk is another white noisy with zero mean and variance, R, and H is the
measurement parameter.
In order to improve the prior estimate X̂k/k−1 , the noisy measurement at time k,
Zk is used as

X̂k/k = X̂k/k−1 + Kk (Zk − Hk X̂k/k−1 ), (3.38)

where X̂k/k is the updated estimate and Kk is the Kalman gain. Notice that (Zk −
Hk X̂k/k−1 ) is just the error in estimating Zk . For deciding on the value of K, the
variance of the error be computed as
" # ( )2 
E (Xk − X̂k/k )2 = E Xk − X̂k/k−1 − Kk (Zk − Hk X̂k/k−1 )
" #2 (3.39)
= E (1 − Kk Hk )(Xk − X̂k/k−1 ) + Kk Vk
= Pk/k−1 (1 − Kk Hk )2 + RK2k ,

where the cross-product terms drop out because Vk is assumed to be uncorrelated

with Xk and X̂k/k−1 . So, the variance of updated estimation error is given by
3.11 Multisite Kalman Filter Methodology 113

Pk/k = Pk/k−1 (1 − Kk Hk )2 + RK2k . (3.40)

If it is necessary to minimize the estimation error, then minimization of Pk/k is

required by differentiating Pk/k with respect to Kk and setting the derivatives equal
to zero. A little algebra shows that the optimal Kk is obtained as
" #−1
K = HPk/k−1 Pk/k−1 (Kk )2 + R . (3.41)

In these calculations Xk is a column matrix with many components. Then Eqs. (3.35,
3.36, 3.37, 3.38, 3.39, 3.40 and 3.41) become matrix equations, and the simplicity
as well as the intuitive logic of the KF become obscured. The covariance matrices
for Wk−1 and Vk vectors are given by


 Qk , i = k
E Wk WTi = ,
0, i = k


 Rk , i = k
E Vk VTi = ,
0, i = k


E Wk VTi = 0, for all k and i.,

where Subscript T denotes transpose of the vector.

These equations fall into two groups, time and measurement update equations.
The time update equations are responsible for projecting forward (in time) the cur-
rent state and error covariance estimates to obtain the a priori estimates for the next
time step. The measurement update equations are responsible for incorporating a
new measurement into the a priori estimate to obtain an improved estimate. The
time update equations can also be thought of as predictor equations, while the mea-
surement update equations can be thought of as corrector equations. Indeed, the final
estimation algorithm resembles that of a predictor algorithm for solving numerical
problems, as shown in Fig. 3.14.
The specific equations for the time updates are as follows:

X̂k+1/k = k+1 k X̂k/k , (3.42)

and

Pk+1/k = k+1 k Pk/k Tk+1 k + Qk . (3.43)

On the other hand, the measurement update (corrector) equations are as follows:

Kk = Pk/k−1 HTk (Hk Pk/k−1 HTk + Rk )−1 , (3.44)

114 3 Classical Spatial Variation Models

Fig. 3.14 The ongoing

discrete KF cycle

Time Update Measurement Update

("Predictor") ("Corrector")

X̂k/k = X̂k/k−1 + Kk (Zk − Hk X̂k/k−1 ), (3.45)

and

Pk/k = Pk/k−1 − Kk Hk Pk/k−1 . (3.46)

It is possible to summarize KF estimation execution steps under the light of pre-

vious derivations as follows:

1) Enter prior estimate X̂k/k−1 , which is based on all our knowledge about the pro-
cess prior to time tk−1 , and also suggest the error covariance matrix associated
with it Pk/k−1 .
2) Compute the Kalman gain as Kk = Pk/k−1 HTk (Hk Pk/k−1 HTk + Rk )−1 .
3) Update the estimate with measurement Zk as X̂k/k = X̂k/k−1 + Kk (Zk −
Hk X̂k/k−1 ).
4) Compute error covariance for updated estimate as Pk/k = Pk/k−1 − Kk Hk Pk/k−1 .
5) Project ahead the updated estimate X̂k/k and the error covariance matrix associ-
ated with it Pk/k , to use it as a prior estimation for the next time step X̂k+1/k =
k+1 k X̂k/k and finally Pk+1/k = k+1 k Pk/k Tk+1 k + Qk .

Once the loop is entered as shown in Fig. 3.12, then it can be continued as much
as necessary. Initially, when the model parameters are only rough estimates, the
gain matrix ensures that the measurement data are highly influential in estimating
the state parameters. Then, as confidence in the accuracy of the parameters grows
with each iteration, the gain matrix values decrease, causing the influence of the
measurement data in updating the parameters and associated error to reduce.
3.11 Multisite Kalman Filter Methodology 115

3.11.2 Kalman Filter Application

To investigate and develop a KF model approach to multisite precipitation model-
ing, 30-year records (1956–1985) of annual rainfall for the 52 stations are used. As
shown in Fig. 3.15 these stations are distributed over an area approximately covering
all of Turkey, with more concentration in the northwestern part.
The study area lies between latitudes 36◦ : 70 N and 42◦ : 03 N and longitude
25 : 90 E and 44◦ : 03 E, extended over an area of about 780,576 square kilometers.
◦

The geographic locations (latitude, longitude, and elevation in meter above mean sea
level) of precipitation stations are represented in Table 3.9.
Precipitation is characterized by variability in space and time. In addition, there
are many factors affecting the magnitude and distribution of precipitation, elevation
of station above mean sea level, various air mass movement, moisture, temperature,
pressure, and topography. The magnitude and distribution of precipitation vary from
place to place and from time to time, even in small areas. The application of multisite
KF model as developed in this chapter approach to multisite precipitation modeling,
which illustrates some interesting points in the annual precipitation pattern.
Figure 3.16 provides the observed and estimated annual rainfall values at Ada-
pazarı from 1956 to 1984.
It is to be noticed from this figure that the observed and estimated values follow
each other closely, which indicates that KF provides an efficient method for model-
ing of annual rainfall. Some statistical parameters of annual observed and estimated
rainfall values during the time period (1956–1985) are summarized in Table 3.10.
From another point of view, Fig. 3.16 provides the observed and estimated annual
rainfall values at 52 selected stations in Turkey for 1985. From Fig. 3.17 and Table
3.10, again as noticed before, in the case of one station (Adapazarı), the observed
and estimated values follow each other closely, which indicates that KF provides an
efficient method for modeling of annual rainfall in both time and space dimension.

800

700

600
Distance (km)

500

400

300

200

100
0 km 100 km 200 km

0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
Distance (km)

Fig. 3.15 Distribution of rainfall stations over Turkey

116 3 Classical Spatial Variation Models

Table 3.9 Station locations and elevation above mean sea level

No. Station No. Station name Latitude (N) Longitude (E) Elevation (m)

01 068 Adapazarı 40◦ :78 30◦ :42 31

02 662 Ali Fuat Paşa – Adapazarı 40◦ :52 30◦ :30 100
03 102 Ağrı 39◦ :72 40◦ :05 1,631
04 350 Adana 37◦ :00 35◦ :33 20
05 620 Bahçeköy – İstanbul 41◦ :17 29◦ :05 130
06 070 Bolu 40◦ :73 31◦ :60 742
07 150 Balıkesir 39◦ :62 27◦ :92 120
08 116 Bursa 40◦ :18 29◦ :07 100
09 115 Bandırma – Balıksir 40◦ :32 27◦ :97 58
10 122 Bilecık 40◦ :15 29◦ :98 539
11 164 Bitlis 38◦ :40 42◦ :12 1,578
12 112 Çanakkale 40◦ :13 26◦ :40 3
13 054 Çorlu – Tekirdağ 41◦ :17 27◦ :80 183
14 084 Çorum 40◦ :55 34◦ :97 776
15 280 Diyarbakır 37◦ :88 40◦ :18 677
16 651 Dursun bey – Balıkesir 39◦ :58 28◦ :63 639
17 653 Edremit – Balıkesir 39◦ :60 27◦ :02 21
18 096 Erzurum 39◦ :92 41◦ :27 1,869
19 050 Edirne 41◦ :67 26◦ :57 51
20 124 Eskişehir 39◦ :78 30◦ :57 789
21 058 Florya – İstanbul 40◦ :98 28◦ :80 36
22 673 Gökçeada – Çanakkale 40◦ :20 25◦ :90 72
23 062 Göztepe – İstanbul 40◦ :97 29◦ :08 33
24 674 İpsala – Edirne 40◦ :93 26◦ :40 10
25 010 İzmit 40◦ :78 29◦ :93 76
26 063 Kartal – İstanbul 40◦ :90 29◦ :18 28
27 059 Kumköy – İstanbul 41◦ :25 29◦ :02 30
28 601 Kırklareli 41◦ :73 27◦ :23 232
29 011 Kandilli 41◦ :10 29◦ :06 114
30 098 Kars 40◦ :60 43◦ :08 1,775
31 052 Luleburgaz – Kirklareli 41◦ :40 27◦ :35 46
32 210 Siirt 37◦ :93 41◦ :95 896
33 020 Şile – İstanbul 41◦ :18 29◦ :62 31
34 026 Sinop 42◦ :03 35◦ :17 32
35 056 Tekirdağ 40◦ :98 27◦ :48 4
36 118 Yalova – İstanbul 40◦ :65 29◦ :27 4
37 132 Yozgat 39◦ :83 34◦ :82 1,298
38 170 Van 38◦ :50 43◦ :50 1,671
39 190 Afyon 38◦ :75 30◦ :53 1,034
40 195 Kayseri – Erkilet 38◦ :78 35◦ :48 1,053
41 240 Isparta 37◦ :75 30◦ :55 997
42 244 Konya 37◦ :97 32◦ :55 1,032
43 292 Muğla 37◦ :20 28◦ :35 646
44 030 Samsun 41◦ :28 36◦ :33 4
45 300 Antalya 36◦ :70 30◦ :73 50
46 034 Giresun 40◦ :92 38◦ :40 38
47 074 Kastamonu 41◦ :37 33◦ :77 799
48 090 Sivas 39◦ :75 37◦ :02 1,285
49 092 Erzincan 39◦ :73 39◦ :50 1,156
50 061 Sarıyer – İstanbul 41◦ :17 29◦ :05 56
51 100 Iğdır 39◦ :93 44◦ :03 858
52 200 Malatya – Erhav 38◦ :43 38◦ :08 0862
3.11 Multisite Kalman Filter Methodology 117

Fig. 3.16 Observed and 1600

estimated annual rainfall
Observed
values at Adapazarı from Estimated
1956 to 1984 1400

Rainfall (mm)
1200

1000

800

600
1955 1960 1965 1970 1975 1980 1985
Time (Year)

Contour maps of observed and estimated annual rainfall for 1956–1985 and per-
centage errors of estimated annual rainfall are presented in Figs. 3.18 and 3.19,
respectively. In Fig. 3.18, dashed lines indicate the estimated annual average precip-
itation contours.
According to the areal values of observed and estimated annual rainfalls, the mul-
tisite KF method has a slight tendency toward under-estimation. Standard deviation
of estimated value is smaller than that of observed one (Fig. 3.20). Its mean is less
variable and, therefore, more smoothed than the observed values.
Furthermore, Fig. 3.17 proves that the estimated values of annual rainfall at most
of the sites in the study area are close to the observed values, especially in the part
where more stations are available such as in the northwestern part of Turkey. The
percentage error of estimated values vary from –6 in station number 52 (underes-
timation) to 6 in station number 49 (overestimation), with overall average about
0.121%.
The magnitude and distribution of precipitation vary from place to place and
from time to time even in small areas. Describing and predicting the precipitation
variability in space and/or time are the fundamental requirements for a wide variety
of human activities and water project designs. In this chapter, a KF technique has
been developed for the prediction of annual precipitation amounts at a multiple site.
In this manner, the precipitation amount at any year and stations is predicted by
considering all the available stations interrelationships.
Once a model has been selected, then KF processing requires specification
of initial state vector, error covariance matrix associated with this initial state
vector, system noise covariance, measurement noise covariance, state transition
matrix, and connection matrix. Most of this information should be based on
 Table 3.10 Statistics parameters of observed and estimated annual rainfall values (1956–1985)
118

No. Station name Mean (mm) St. Dev. (mm) Minimum (mm) Maximum (mm) Range (mm) Skewness
Obs. Est. Obs. Est. Obs. Est. Obs. Est. Obs. Est. Obs. Est.

1 Adapazarı 835.5 834.6 183.4 135.3 606.3 611.7 1,564 1,332 957.7 720.3 2.18615 1.50665
2 Ali Fuat Paşa – Adapazarı 660.6 661 107.7 85.1 484.9 519.1 922.5 851.1 437.6 332 0.36170 0.13996
3 Ağrı 506.7 504.7 91.9 66.6 352 389.1 693.1 612.7 341.1 223.6 0.13271 −0.0978
4 Adana 690.4 688.7 221.3 184.4 319.4 360 1,265 1,149 945.6 789 0.86862 0.66023
5 Bahçeköy – İstanbul 1,287 1,284 478.6 354.7 848.1 866.2 2,698 2,325 1,850 1,459 1.73873 1.56783
6 Bolu 549.2 549.2 83.3 59.3 377.4 429.1 716.8 659.7 339.4 230.6 0.14283 0.16372
7 Balıkesir 606.7 606.1 163.6 126.3 362.6 385.6 1,192 1,036 829.4 650.4 1.59256 1.25458
8 Bursa 680.6 680.1 109.1 82.7 447.8 521.1 871 835 423.2 313.9 0.16262 −0.0062
9 Bandırma – Balıkesir 719 718.9 149.3 115 456.2 494.7 1,086 988.2 629.8 493.5 0.71735 0.55033
10 Bilecik 445.1 444.5 75.4 57.5 320.6 332.8 624.7 563.1 304.1 230.3 0.62528 0.39746
11 Bitlis 1,164 1,160 287.5 235.8 564.5 734.2 1,866 1,738 1,301 1,004 0.27270 0.51462
12 Çanakkale 636.6 640.4 139.9 108.9 414.8 459.8 977.7 868.5 562.9 408.7 0.64583 0.45769
13 Çorlu – Tekirdağ 570.6 571.1 104.1 81.5 431.2 456.4 801.3 759.9 370.1 303.5 0.58876 0.64617
14 Çorum 432.1 430.7 79.7 59.1 285.5 296.8 606.7 542.9 321.2 246.1 −0.05457 −0.2570
15 Diyarbakır 490 489 134.5 99.3 146.3 276 748.8 671 602.5 395 −0.10185 0.24572
16 Dursun bey – Balıkesir 779.1 782.2 171.6 130.6 479.3 575 1,199 1,079 719.7 504 0.36409 0.15470
17 Edremit – Balikesir 696.2 697.2 144.1 111.6 512.9 548.4 1,175 1,061 662.1 512.6 1.36238 1.2798
18 Erzurum 410.1 409.8 85.3 61.3 291.1 308.1 638.5 551.8 347.4 243.7 0.92707 0.49036
19 Edirne 593.2 594.4 102.2 72.1 430.5 496.6 863.6 792.2 433.1 295.6 0.85735 0.87714
20 Eskişehir 393.1 391.8 94.8 51.7 215.8 220.1 524 486.8 326.2 266.7 −0.18598 −0.9474
21 Florya – Istanbul 656.4 655.5 117.3 89.3 500.8 540 1,026 942.6 525.2 402.6 1.02088 1.07883
22 Gökçeada – Çanakkale 792.5 796.4 212.2 162.8 483 543.8 1,451 1,235 968 691.2 1.28517 0.92635
23 Göztepe – Istanbul 698.7 696.7 127.4 98.3 538.8 559 1,047 990.9 508.2 431.9 0.87054 0.94386
24 İpsala – Edirne 612.5 616.3 98.7 77.5 386.5 434.7 808.9 763.9 422.4 329.2 0.15228 0.11054
25 İzmit 766.6 764.9 150.2 119.9 554.9 557.2 1,088 1,012 533.1 454.8 0.37969 0.12170
26 Kartal – İstanbul 651.1 649.5 118.5 90.9 475.6 496.9 871.9 832.4 396.3 335.5 0.16123 0.05385
27 Kumköy – İstanbul 796.1 793.2 185.8 143.7 519.5 581.9 1,278 1,145 758.5 563.1 1.02989 0.88983
3 Classical Spatial Variation Models
 3.11

Table 3.10 (continued)

No. Station name Mean (mm) St. Dev. (mm) Minimum (mm) Maximum (mm) Range (mm) Skewness
Obs. Est. Obs. Est. Obs. Est. Obs. Est. Obs. Est. Obs. Est.

28 Kirklareli 589.4 592.5 133.9 97.5 371.5 417.5 1,000 876.9 628.5 459.4 1.17951 0.99290
29 Kandilli 827.7 825.5 148.7 117.5 600.5 631.5 1,230 1,143 629.5 511.5 0.54675 0.42092
30 Kars 470.5 472.2 106.1 78.2 298.5 361.7 718.8 661.1 420.3 299.4 0.75373 0.81958
31 Luleburgaz – Kırklareli 652.3 655.4 182.2 138.8 399.7 437.4 1,360 1,159 960.3 721.6 2.00788 1.57797
32 Siirt 684.5 682.4 183.4 141.5 420.8 474.5 1,229 1,060 808.2 585.5 1.02095 0.91065
33 Şile – İstanbul 800.8 796.2 251.5 211.6 454.6 491.5 1,697 1,537 1,242 1,045 1.74568 1.66792
34 Sinop 640.2 639.7 134.2 109.6 414.7 455.2 990.4 917.1 575.7 461.9 0.83072 0.77820
35 Tekirdağ 604.2 606.6 192.6 141.2 405.2 434.8 1,464 1,192 1,059 757.2 3.1807 2.50102
36 Yalova – İstanbul 770.4 769.2 266.9 202.7 473.2 484.7 1,959 1,617 1,486 1,132 3.216 2.60607
Multisite Kalman Filter Methodology

37 Yozgat 565.8 562.6 109.2 86 391 412.9 858.2 771.3 467.2 358.4 0.72179 0.44662
38 Van 377.9 377.9 62.7 47.6 267.9 292.5 486 461.1 218.1 168.6 −0.1035 0.04482
39 Afyon 408.6 407.7 89.5 69 239 259.3 618 576.6 379 317.3 0.14329 −0.0192
40 Kayseri – Erkilet 364.8 365.2 60 45.3 263 278.6 535 475.3 272 196.7 0.75670 0.2949
41 Isparta 557.6 557.8 160 122.7 332 371.4 968 891.5 636 520.1 0.70164 0.85167
42 Konya 326.1 325 77.6 63.7 193 225 545 483.7 352 258.7 0.95732 0.66484
43 Muğla 1,165 1,164 298.3 229.3 658 767.4 1,805 1,672 1,147 904.6 0.53785 0.44423
44 Samsun 693.7 694.2 129.4 103.2 442.2 503.5 1,011 955 568.8 451.5 0.45975 0.72549
45 Antalya 1,074 1,072 293.4 228.3 554 608.7 1,914 1,747 1,360 1,138 0.29616 0.28473
46 Giresun 1,240 1,244 153.2 125.5 1,057 1,090 1,679 1,668 622 578 1.17704 1.63857
47 Kastamonu 457.6 455 82.2 67.3 308 312.1 617 574.9 309 262.8 0.17054 −0.1840
48 Sivas 411.8 410.1 79.2 56.4 286 298.8 575 519.8 289 221 0.03905 0.09282
49 Erzincan 368.8 367.6 71.7 56.8 257 277.9 563 504.1 306 226.2 0.86580 0.62331
50 Sarıyer – İstanbul 796.6 794.3 150.8 130.6 574.9 597.8 1,171 1,102 596.1 504.2 0.91685 0.72847
51 Iğdır 251.9 253.2 77.3 55.5 114.5 154.9 501.2 422.9 386.7 268 1.0145 1.00863
52 Malatya – Erhav 402.7 403.1 100 72.1 248 292.1 593 533.9 345 241.8 0.42067 0.33824
119
120 3 Classical Spatial Variation Models

1600
Observed
Estimated

1200
Rainfall (mm)

800

400

0
0 5 10 15 20 25 30 35 40 45 50 55
Station Number

Fig. 3.17 Observed and estimated annual rainfall values at selected stations in Turkey for 1985

physical understanding and on all the previous knowledge about the process prior
to tk-1 . If little historical information is available to specify the above matrices, then
KF may be started with very little objective information and adapted as the data
become available. However, the less the initial information, greater diagonal ele-
ments should be selected in the covariance matrices. In this manner, the algorithm
will have flexibility to adjust itself to sensible values in a relatively short space of
time.
The average amount of rainfall values at the selected stations are used as the ele-
ments of initial state vector. Sufficiently great diagonal elements of error covariance
matrix are needed with initial state vector provided in the initial moment. Then, the
prediction error covariance steadily decreases with time and arrives at a stable value
after some steps, indicating the efficiency of the prediction algorithm.
After the initial-state descriptions are read as first step, then the Kalman gain
matrix for the one-step prediction can be computed, with necessary assumptions, as
the connection matrix is unity, i.e., all stations are reporting their observations. The
diagonal elements of measurement noise covariance matrix is taken smaller than
those of the system covariance matrix, because the observed values are relatively
noise-free compared with the errors which result from the system.
Initially, when the model parameters are only rough estimates, with little objec-
tive information, the Kalman gain matrix ensures that the measurement data are
highly influential in estimating the state parameters. Then, as confidence in the accu-
racy of the parameters grows with each iteration, the gain matrix values decrease,
causing the influence of the measurement data in updating the parameters and asso-
ciated error.
3.11 Multisite Kalman Filter Methodology 121

800

700

600
DISTANCE IN (KM)

500

400 1960

300

200

100
0 km 100 km 200 km
0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
DISTANCE IN (KM)
800

700

600
DISTANCE IN (KM)

500
1965
400

300

200

100
0 km 100 km 200 km
0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
DISTANCE IN (KM)
800

700

600
DISTANCE IN (KM)

500

400 1970

300

200

100
0 km 100 km 200 km
0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
DISTANCE IN (KM)

Fig. 3.18 Contour map of observed and estimated annual precipitation 1960–1985
122 3 Classical Spatial Variation Models

800

700

600
DISTANCE IN (KM)

500

400
1975
300

200

100
0 km 100 km 200 km
0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
DISTANCE IN (KM)
800

700

600
DISTANCE IN (KM)

500

400
1980
300

200

100
0 km 100 km 200 km
0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
DISTANCE IN (KM)
800

700

600
DISTANCE IN (KM)

500

400 1985

300

200

100
0 km 100 km 200 km
0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
DISTANCE IN (KM)

Fig. 3.18 (continued)

3.11 Multisite Kalman Filter Methodology 123

800

700

600
Distance (km)

500

400 1960
300

200

100
0 km 100 km 200 km

0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
Distance (km)
800

700

600
Distance (km)

500

400 1965

300

200

100
0 km 100 km 200 km

0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
Distance (km)
800

700

600
Distance (km)

500

400 1970

300

200

100
0 km 100 km 200 km

0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
Distance (km)

Fig. 3.19 Contour map of percentage error of estimated annual rainfall 1960–1985
124 3 Classical Spatial Variation Models

800

700

600
Distance (km)

500

400
1975
300

200

100
0 km 100 km 200 km

0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
Distance (km)
800

700

600
Distance (km)

500

400 1980

300

200

100
0 km 100 km 200 km

0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
Distance (km)
800

700

600
Distance (km)

500

400 1985

300

200

100
0 km 100 km 200 km

0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
Distance (km)

Fig. 3.19 (continued)

3.11 Multisite Kalman Filter Methodology 125

Fig. 3.20 Standard deviation 400

of observed and estimated Observed
areal rainfall values for
Predicted
1956–1985
350

Standard Deviation (mm)

300

250

200

150
1955 1960 1965 1970 1975 1980 1985
Time (Year)

With the assumption of an initial estimate, the measurement is used to improve

the initial estimate by a linear blending of the noisy measurement and the prior
estimate. The covariance matrix associated with the updated estimate is computed
by using the error covariance matrix associated with initial state vector, connection
and Kalman gain matrices.
The updated estimates can be projected ahead via the transition matrix, where
the contribution of system error can be ignored, because it has zero mean. However,
the estimate of transition matrix may be difficult, but the system is quite robust to
the transition matrix values, and that they can, therefore, be set to fixed values. They
have a minimal effect on the results. However, for simplicity transition matrix is
assumed to be unity.
Projecting ahead the error covariance matrix, where the updated error covariance
matrix can be computed, and for the system noise covariance matrix, one can use
his/her judgment to select appropriate values. He/she can then examine the actual
operation of the filter and adjust these values online if the situation changes at a later
time. Similar to the covariance of initial values, KF is started with large diagonal
elements in the system noise covariance matrix.
The projected ahead estimation and the error covariance matrix are used as initial
estimation for the next time step. Once the multisite KF loop is entered, then it can
be continued as much as necessary.
It is to be noticed that for one station the observed and estimated values follow
each other closely, which indicates that KF provides an efficient method for mod-
eling of annual rainfall. The observed and estimated annual rainfall values at 52
selected stations in Turkey, for 1985, follow each other closely, which indicates that
126 3 Classical Spatial Variation Models

KF provides an efficient method for modeling of annual rainfall in both time and
space dimensions.
The estimated values of annual rainfall at most of the sites in the study area are
close to the observed values, specially in the part where more stations are available,
such as in the northwestern part of Turkey. The percentage error of estimated values
vary from –6 (underestimation) to 6 (overestimation), with overall average about
0.121%.

References
Barnes, S. L., 1964. A technique for maximizing details in numerical weather map analysis. J.
Appl. Meteor. 3, 396–409.
Benjamin, J. R., and Cornell, C. A., 1970. Probability Statistics and Decision Making in Civil
Engineering. McGraw-Hill Book Inc., New York.
Bergthorsson, P., and Döös, B. R., 1955. Numerical weather map analysis. Tellus 7, 329–340.
Brown, R. G., and Hwang Y. C., 1992. Introduction to Random Signals and Applied Kalman Fil-
tering (Second edition). John Wiley & Sons, New York.
Cressman, G. P., 1959. An operational objective analysis system. Mon. Wea. Rev. 87, 367–374.
Davis, A., 2002. Statistics Data Analysis Geology. John Wiley and Sons, New York, 638 pp.
Dee, D. P., 1991. Simplification of Kalman filter for meteorological data assimilation. Q. J. R.
Meteorol. Soc. 117, 365–384.
Donnelly, K. P., 1978. Simulations to determine the variance and edge effect of total nearest neigh-
bor distance. In: Hodder, I. (Ed.), Simulation Studies in Archaeology. Cambridge University
Press, Cambridge, UK, pp. 91–95.
Gandin, L. S., 1963. Objective Analysis of Meteorological Fields. Hydromet Press, Leningrad,
242 pp.
Gilchrist, B., and Cressman, G. P., 1954. An experiment in objective analysis. Tellus 6, 309–318.
Goodin, W. R., McRea, G. J., and Seinfeld, J. H., 1979. A comparison of interpolation methods for
sparse data: application to wind and concentration fields. J. Appl. Meteor. 18, 761–771.
Harrison, P. J., and Stevens, C. F., 1975. Bayesian forecasting. University of Warwick, Working
Paper No. 13.
Kalman, R. E., 1960. A new approach to linear filtering and prediction problems. Trans ASME,
Ser. D, J. Basic Eng. 82, 35–45.
Kalman, R. E., and Bucy, R., 1961. New result in linear filtering and prediction theory. Trans
ASME, Ser. D, J. Basic Eng. 83, 95–108.
Koch, S., and Link, R. F., 1971. Statistical Analysis of Geological Data. Dover Publications, New
York, 375 pp.
Koch, S. E., DesJardins, M., and Kocin, P. J., 1983. An iterative Barnes objective map analysis
scheme for use with satellite and conventional data. J. Appl. Meteor. 22, 1487–1503.
Latif, A. M., 1999. A Kalman filter approach to multisite precipitation modeling in meteorology.
Unpublished Ph.D. thesis, Istanbul Technical University, 125 pp.
Matheron, G., 1963. Principles of geostatistics. Econ. Geol. 58, 1246–1266.
Matheron, G., 1965. Les Variables Regionalisees et leur Estimation. Masson, Paris, 306 pp.
Sasaki, Y., 1960. An objective analysis for determining initial conditions for the primitive equa-
tions. Tech. Rep. 60-16T.
Şen, Z., 1989. Cumulative semivariogram models of regionalized variables. Math. Geol. 21,
891–903.
Şen, Z., 2004. Fuzzy Logic and System Models in Water Sciences. Turkish Water Foundation
Publication, Istanbul, Turkey, 315 pp.
Zadeh, L. A., 1968. Fuzzy algorithms. Inform. Control 12, 94–102.
Chapter 4
Spatial Dependence Measures

Abstract Rather than geometrical weighting functions as in Chapter 2, it is

preferable to obtain spatial dependence function from a set of measurements points.
Prior to such a functional derivation, it is necessary to examine the isotropy and
homogeneity of the spatial data directionally and the point-wise features of the
regionalized variable (ReV). The basics of semivariogram (SV), with its different
components such as sill, nugget, and radius of influence, are presented in descrip-
tive and application manners. Similar to SV cumulative SV (CSV) and point CSV
(PCSV), concepts are explained with applications to groundwater quality data. It
is emphasized that PCS helps to depict the spatial behavior features around any
sampling point by taking into consideration the contribution from the surrounding
measurement points. It is shown that for each location of measurement it is possible
to obtain the radius of influence, if necessary along any direction, and their regional
contour maps provide the radius of influence at non-measurement locations. Once
the radius of influence is known, then it is possible to depict which nearby measure-
ment locations should be taken into consideration in the calculation of unknown data
value. The validity of any method can be decided on the basis of cross-validation
error minimization. A new concept of spatial dependence function (SDF) is devel-
oped without the need that regional data have normal (Gaussian) probability distri-
bution. The application of SDF is presented for earthquake and precipitation data
from Turkey.

Keywords Cumulative semivariogram · Homogeneity · Isotropy · Nugget · Point

cumulative semivariogram · Range · Spatial correlation · Sample
semivariogram · Sill · Spatial dependence function

4.1 General

Uncertainty within earth sciences data and geologic maps is rarely analyzed or dis-
cussed. This is undoubtedly due, in part, to the difficulty of analyzing large sets of
paper files. The increasing use of computers for storing, retrieving, and serving large
data sets has made it easier to systematically analyze these data and preparation of
maps and models from them.

Z. Şen, Spatial Modeling Principles in Earth Sciences, 127

DOI 10.1007/978-1-4020-9672-3_4,  C Springer Science+Business Media B.V. 2009
128 4 Spatial Dependence Measures

General error theory can be a useful method for characterizing the uncertainty
within spatial data. For most data, errors can be due to inaccuracies in record keep-
ing, description, identification of behavior and location, generalization, and corre-
lation. While analysis by error theory can be employed to evaluate the uncertainty
within the data used in mapping, other methods can be used to more fully evalu-
ate the uncertainty within specific maps. As already mentioned in Chapter 2, the
area of influence method was developed to use three map characteristics (longi-
tude, latitude, elevation) for estimating uncertainty within the maps of earth sci-
ences. This method uses the spatial distribution of data points, the probability of
mis-identification of the targeted unit, and the size of the targeted geologic feature
to calculate the probability that additional, unidentified targets can exist. Insight
gained from the use of error theory and the area of influence methods can be used
to describe the uncertainty included in spatial maps.
Earth scientists need maps based even on few samples for extensive decisions
in executing certain projects. It is, therefore, necessary for him/her to be able
to visualize the subject through limited amount of data preliminarily, which can
then be expanded to more effective directions with the coming of additional data.
Although there is ready software for mapping if the user is not familiar with some
basic principles of the methodology implemented, then the conclusions may be
at bias at the minimum. At the early stages of any study, maps are needed for
efficient spatial relationships without any involved mathematical formulations. In
general, maps relate three variables (triplets as mentioned in Chapter 2), most of
the time two are the longitude and latitude or geographical coordinates of the
sample points and therefore they show the spatial variability of a single variable.
However, in this chapter imaginary maps of three different earth science vari-
ables are also presented for better logical arguments, interpretations, and conclu-
sions. Maps help to perceive large-scale spatial relationships easily on a small piece
of paper.
Maps are based on the point data measurements, the distances between the points,
and the density of points. Since most natural phenomenon is of continuous type,
maps are representations of finite number of measurement sites and their continuous
surface expressions. Hence, the more and better scattered the measurement points
within the study area the better is the map representation of the natural phenomenon.
The most common ones are the topographic maps with continuous contour lines of
elevations, which are derived from a set of discrete location surveys. In the drawing
of maps not only the measurements but also the artistic talent and skills of the expert
are also taken into consideration. Hence, in many early maps subjective biases have
entered the domain of mapping, but recently more objective mapping methodologies
are developed and the objectivity is more enhanced.
The reliability of contour maps is directly dependent on the total density of the
sampling points as well as on their uniform distribution (Chapter 3). However, in
practice the uniformity of the sampling points is seldom encountered and the maps
are prepared by avoiding this point. For instance, numerical, statistical, or prob-
abilistic versions of weather prediction methods are based on the data available
from irregularly distributed sites within synoptic regions. The very success of such
4.2 Isotropy, Anisotropy, and Homogeneity 129

prediction methods significantly depends on the production of regularly gridded data

at nodes of a mesh laid over the region concerned. Hence, in any spatial prediction
study the following two steps are important:
1) that measured data at practically convenient but irregular sites must be trans-
ferred to regularly selected grid points, and then,
2) the use of these grid point estimates in a convenient modeling scheme through
digital computers.

4.2 Isotropy, Anisotropy, and Homogeneity

In operational regional interpolation systems, many simplifying assumptions are

made about the nature of the correlation, and it is represented by an analytic func-
tion of the distance. This assumption implies that the ReV is statistically homo-
geneous and isotropic. Gustafsson (1981) has defined the statistical homogeneity
properties and isotropy of meteorological fields in reference to the covariance. A
ReV is homogeneous and second-order stationary, if the covariance is independent
of a translation of the two positions. This means that the covariance depends only
on the difference between the positional vectors, i.e., distance. Homogeneity in ref-
erence to the covariance implies that the variance is constant. A ReV is isotropic in
reference to the covariance, if the covariance is independent of a rotation in the field
around the center point on the line between two positions.
For any practical application of the spatial modeling, it is necessary to construct
a model for the spatial correlation to be used in the scheme. To simplify the con-
struction of these correlation models, it is an advantage if the homogeneity and
isotropy conditions are fulfilled. Much work has gone into proper choice of cor-
relation functions of ReV. One simplifying assumption is that the correlations are
assumed to be isotropic and homogeneous, which make them dependent only on dis-
tance as stated by Thiebaux and Pedder (1987) and Daley (1991). For instance, the
real atmosphere is anything but homogeneous and isotropic, and hence it is assumed
that the deviation from the first guess (climatologic mean) field are homogeneous
and isotropic. These assumptions are made in most, if not all, interpolation anal-
ysis schemes (Bergman, 1979; Lorenc, 1981; Rutherford, 1976; Schlatter, 1975).
Also homogeneity and isotropy were assumed by Gandin (1963), Eddy (1964) and
Kruger (1969a,b) for most meteorological fields.
Natural phenomena physical processes have preferred orientations. For exam-
ple, at the mouth of a river the coarse material settles out fastest, while the finer
material takes longer to settle. Thus, the closer one is to the shoreline the coarser
the sediments while the further from the shoreline the finer the sediments. When
interpolating at a point, an observation 100 m away, but in a direction paral-
lel to the shoreline, is more likely to be similar to the value at the interpolation
point than is an equidistant observation in a direction perpendicular to the shore-
line. Anisotropy takes these trends in the data into account during the gridding
process.
130 4 Spatial Dependence Measures

Usually, points closer to the grid node are given more weight than points farther
from the grid node. If, as in the example above, the points in one direction have
more similarity than the points in another direction, it is advantageous to give points
in a specific direction more weight in determining the value of a grid node. The rel-
ative weighting is defined by the anisotropy ratio. The underlying physical process
producing the data as well as the sample spacing of the data are important in the
decision of whether or not to reset the default anisotropy settings.
Anisotropy is also useful when data sets have fundamentally different units along
different dimensions. For example, consider plotting a flood profile along a river.
The x coordinates are locations, measured in km along the river channel. The t coor-
dinates are time, measured in days. The Z(x, t) values are river depth as a function of
location and time. Clearly in this case, the x and t coordinates would not be plotted
on a common scale, because one is distance and the other is time (Fig. 4.1). One
unit of x does not equal one unit of t. While the resulting map can be displayed with
changes in scaling, it may be necessary to apply anisotropy as well.
Another example of anisotropy might be employed for an isotherm map (equal
temperature lines, contour map) of average daily temperature over a region.
Although the X and Y coordinates (as Easting, say X and Northing, say Y) are
measured using the same units, the temperature tends to be very similar. Along
north–south lines (Y lines) the temperature tends to change more quickly (getting
colder as one heads towards the north) (see Fig. 4.2). In this case, in gridding the
data, it would be advantageous to give more weights to data along the east–west
axis than along the north–south axis. When interpolating a grid node, observations
that lie in an east–west direction are given greater weight than observations lying an
equivalent distance in the north–south direction.
In the most general case, anisotropy can be visualized as an ellipse. The ellipse
is specified by the lengths of its two orthogonal axes (major and minor) and by
an orientation angle, θ. The orientation angle is defined as the counter clockwise
angle between the positive X and, for instance, minor axis (see Fig. 4.3). Since the
ellipse is defined in this manner, an ellipse can be defined with more than one set of
parameters.

Z (x, t)

Fig. 4.1 Spatio-temporal

depth variations t
4.2 Isotropy, Anisotropy, and Homogeneity 131

Istanbul Black Sea

Aegean Sea

Mediterranean Sea

Fig. 4.2 Average annual temperature of Turkey

For most of the gridding methods, the relative lengths of the axes are more impor-
tant than the actual length of the axes. The relative lengths are expressed as a ratio
in the anisotropy group. The ratio is defined as major axis divided by minor axis. If
it is equal to 1, then the ellipse takes the form of a circle. The angle is the counter
clockwise angle between the positive X axes and minor axis. The small picture in
the anisotropy group displays a graphic of the ellipse to help describe the ellipse.
An anisotropy ratio less than 2 is considered mild, while an anisotropy ratio greater
than 4 is considered severe. Typically, when the anisotropy ratio is greater than 3,
its effect is clearly visible on grid-based maps. The angle is the preferred orientation
(direction) of the major axis in degrees.
An example where an anisotropy ratio is appropriate is an oceanographic survey
to determine water temperature at varying depths. Assume the data are collected
every 1,000 m along a survey line, and temperatures are taken every 10 m in depth
at each sample location. With this type of data set in mind, consider the problem of
creating a grid file. When computing the weights to assign to the data points, closer
data points get greater weights than points farther away. A temperature at 10 m in

Y, Northing Minor axis

θ X, Easting

Fig. 4.3 Anisotropy ratio

and angle Major axis
132 4 Spatial Dependence Measures

depth at one location is similar to a sample at 10 m in depth at another location,

although the sample locations are 1,000s of meters apart. Temperatures might vary
greatly with depth, but not as much between sample locations.

4.3 Spatial Dependence Function

The first step is referred to as the objective analysis and the second one is the spa-
tial modeling phase. For sure, a sound objective analysis is the primary prerequisite
of successful modeling. For instance, meteorologists strive for effective interpola-
tion in order to enhance their mesoscale analysis and forecasts. Objective analysis
studies of meteorological variables started with the work by Panofsky (1949). He
attempted to produce contour lines of upper-wind movements by fitting third-order
polynomials and employing least squares method to the observations at irregular
sites. The least squares method leads to predicted field variables, which depend
strongly on the distribution of data points when a suitable polynomial is fitted to
full grid. Optimum analysis procedures are introduced to meteorology by Eliassen
(1954) and Gandin (1963). These techniques employ historical data about the struc-
ture of the atmosphere to determine the weights to be applied to the observations.
Here, the implied assumption is that the observations are spatially correlated. Con-
sequently, observations that are close to each other are highly correlated; hence, as
the observations get farther apart the spatial dependence decreases. It is a logical
consequence to expect regional dependence function as in Fig. 4.4, assuming that
at zero distance the dependence is equal to 1 and then onward there is a continuous
decrease or decreasing fluctuations depending on the ReV behavior.
In this figure there are three spatial dependence functions (SDFs) as A, B, and
C. Logically, A and B indicate rather homogeneous and isotropic regional behavior
of ReV whereas C has local differences at various distances. However, all of them
decrease down to zero SDF value. The distance between the origin and the point
where the SDF is almost equal to zero shows the radius of influence as R1 or R2 .
Provided that the ReV behavior is isotropic (independent of direction), the radius
of area can be calculated as a circle around each station, with radius equal to the
radius of influence. These are subjective and expert views about the spatial depen-
dence structure of any ReV. Their objective counterparts can be obtained from a
Spatial dependence

A B C
0 Distance
R1
Fig. 4.4 Spatial dependence
functions R2
4.3 Spatial Dependence Function 133

set of spatial data as will be explained later in this chapter. The spatial predictions
are then made by considering a spatial model with a domain equal to the radius of
area. For instance, Gilchrist and Cressman (1954) reduced the domain of polyno-
mial fitting to small areas surrounding each node with a parabola. Bergthorsson and
Döös (1955) proposed the basis of successive correction methods, which did not
rely only on interpretation to obtain grid point values but also a preliminary guess
field is initially specified at the grid points (Chapter 5). Cressman (1959) developed
a number of further correction versions based on reported data falling within a spec-
ified distance R from each grid point. The value of R is decreased with successive
scans (1,500, 750, 500 km, etc.) and the resulting field of the latest scan is taken
as the new approximation. Barnes (1964) summarized the development of a conver-
gent weighted-averaging analysis scheme, which can be used to obtain any desired
amount of detail in the analysis of a set of randomly spaced data. The scheme is
based on the supposition that the two-dimensional distribution of a ReV can be rep-
resented by the summation of an infinite number of independent waves, i.e., Fourier
integral representation. A comparison of existing objective methods up to 1979 for
sparse data is provided by Goodin et al. (1979). Their study indicated that fitting a
second-degree polynomial to each sub-region triangular in the plane with each data
point weighted according to its distance from the sub-region provides a compromise
between accuracy and computational cost. Koch et al. (1983) presented an extension
of the Barnes method, which is designed for an interactive computer scheme. Such
a scheme allows real-time assessment both of the quality of the resulting analyses
and of the impact of satellite-derived data upon various earth sciences data sets.
However, all of the aforementioned objective methods have the following common
drawbacks.
1) They are rather mechanical without any physical foundation, but rely on the
regional configuration of irregular sites. Any change in site configuration leads
to different results although the same ReV is sampled.
2) They do not take into consideration the spatial covariance or correlation struc-
ture within the ReV concerned.
3) They have constant radius of influence without any directional variations.
Hence, spatial anisotropy of observed fields is ignored. Although some
anisotropic distance function formulations have been proposed by Inman (1970) and
Shenfield and Bayer (1974), all of them are developed with no explicit quantitative
reference to the anisotropy of observed field structure of the ReV.
According to Thiebaux and Pedder’s (1987) assessment of the work done by
Bergthorsson and Döös, “the most obvious disadvantage of simple inverse distance-
weighing schemes is that they fail to take into account the spatial distribution of
observations relative to each other.” Two observations at equidistant from a grid
point are given the same weight regardless of relative values at measurement sites.
This may lead to large operational biases in grid point data when some observations
are much closer together than others within the area of influence.
Especially after 1980s, many researchers are concentrated on the spatial covari-
ance and correlation structures of the ReVs. Lorenc (1981) has developed a
134 4 Spatial Dependence Measures

methodology whereby, first of all, the grid points in a sub-region are analyzed simul-
taneously using the same set of observations and then sub-areas are combined to
produce the whole study area analysis. Some papers are concerned with the deter-
mination of unknown parameters of the other covariance functions or SDFs, which
provide required weightings for ReV data assimilation. Along this line, the idea
proposed by Bratseth (1986) depends on the interpretation of the ReV covariances
into the objective analysis. His analysis caused a recent resurgence of the succes-
sive correction method in which the optimal analysis solution is approached. His
method uses the correlation function for the forecast errors to derive weights that
are reduced in regions of higher data density. Later, Sashegyi et al. (1993) employed
his methodology for the numerical analysis of the data collected during the Genesis
of Atlantic Lows Experiment, (GALE). Practical conclusions of Bratseth’s approach
have been reported by Franke (1988) and Seaman (1988).
On the other hand, Buzzi et al. (1991) described a simple and economic method
for reducing the errors that can result from the irregular distribution of data points
during an objective analysis. They have demonstrated that a simple iterative method
for correcting the analysis generated by an isotropic distance-weighting scheme
applied to a heterogeneous spatial distribution of observations cannot improve anal-
ysis accuracy, but also results in an actual frequency response that approximates
closely the theoretical response of the predicted weight-generating function. They
have shown that in the case of heterogeneous spatial sampling, a Barnes analysis
could produce an unrealistic interpolation of the sampled field even when this is
reasonably well resolved by error-free observations. Iteration of a single correction
algorithm led to the method of successive correction (Daley, 1991). The method of
successive correction has been applied as a means to tune adaptively the a posteriori
weights. Objective analysis schemes are practical attempts to minimize the variance
estimation (Thiebaux and Pedder, 1987).
Pedder (1993) provided a suitable formation for successive correction scheme
based on a multiple iteration using a constant influence scale that provides more
effective approach to estimate ReVs from scattered observations than the more con-
ventional Barnes method which usually involves varying the influence scale between
the iterations. Recently, Dee (1995) has presented a simple scheme for online esti-
mation of covariance parameters in statistical data assimilation systems. The basis of
the methodology is a maximum-likelihood approach in which estimates are obtained
through a single batch of simultaneous observations. Simple and adaptive Kalman
filtering techniques are used for explicit calculation of forecast error covariance
(Chapter 3). However, the computational cost of the scheme is rather high.
Field measurements of ReVs such as ore grades, chemical constitutions in ground
water, fracture spacing, porosity, permeability, aquifer thickness, dip and strike of
a structure are dependent on the relative positions of measurement points within
the study area. Measurements of a given variable at a set of points provide some
insight into the spatial variability. This variability determines the ReV behavior as
well as its predictability. In general, the larger the variability, the more heteroge-
neous is the ReV environment and as a result the number of measurements required
to model, simulate, estimate, and predict the ReV is expected to be large. Large
variability implies also that the degree of dependence might be rather small even
4.4 Spatial Correlation Function 135

for data whose locations are close to each other. A logical interpretation of such a
situation may be that either the region was subjected to natural phenomena, such as
tectonics, volcanism, deposition, erosion, recharge, climate change, or later to some
other human activities as pollution, ground water abstraction, mining, etc.
However, many types of ReVs are known to be spatially related in that the closer
their positions, the greater is their dependence. For instance, spatial dependence is
especially pronounced in hydrogeological data due to groundwater flow as a result of
the hydrological cycle, which homogenizes the distribution of chemical constituents
within the heterogeneous mineral distribution in geological formations.
The factors of ReVs are sampled at irregular measurement points within an area
at regular or irregular time intervals. No doubt, these factors show continuous vari-
ations with respect to other variables such as temperature, distance. Furthermore,
temporal and spatial ReV evolutions are controlled by temporal and spatial correla-
tion structures within the ReV itself. As long as the factors are sampled at regular
time intervals, the whole theory of time series is sufficient in their temporal mod-
eling, simulation, and prediction. The problem is with their spatial constructions
and the transfer of information available at irregular sites to regular grid nodes or
to any desired point. Provided that the structure of spatial dependence of the ReV
concerned is depicted effectively, then any future study such as the numerical pre-
dictions based on these sites will be successful. In order to achieve such a task it is
necessary and sufficient to derive the change of spatial correlation for the ReV data
with distance.
In order to quantify the degree of variability within spatial data, variance tech-
niques can be used in addition to classical autocorrelation methods (Box and Jenk-
ins, 1976). However, these methods are not helpful directly to account for the spa-
tial dependence or for the variability in terms of sample positions. The drawbacks
are due to either non-normal (asymmetric) distribution of data and/or irregularity
of sampling positions. However, the semivariogram (SV) technique, developed by
Matheron (1963, 1971) and used by many researchers (Clark, 1979; Cooley, 1979;
David, 1977; Myers et al., 1982; Journel, 1985; Aboufirassi and Marino, 1984;
Hoeksema and Kitandis, 1984; Carr et al., 1985) in diverse fields such as geology,
mining, hydrology, earthquake prediction, groundwater, can be used to character-
ize spatial variability and hence the SDF. The SV is a prerequisite for best linear
unbiased prediction of ReVs through the use of Kriging techniques (Krige, 1982;
Journel and Huijbregts, 1978; David, 1977).

4.4 Spatial Correlation Function

By definition, SCF, ρij between i and j, takes values between −1 and +1 and can be
calculated from available historical data as

(Zoi − Zi )(Zoj − Zj )
ρij =  , (4.1)
(Zoi − Zi )2 (Zoj − Zj )2
136 4 Spatial Dependence Measures

where over bars indicate time averages over a long sequence of past observations,
Zoi and Zoj represent observed precipitation amounts at these stations, and finally, Zi
and Zj are the climatological mean of precipitations. Furthermore, ρij is thought as
attached with the horizontal distance Di,j between stations i and j. Consequently, if
there are n stations, then there will be m = n(n–1)/2 pairs of distances and the cor-
responding correlation coefficients. Their plot results in a scatter diagram that indi-
cates the SCF pattern for the regional rainfall amounts considered as a random field.
Figure 4.5 presents such scatter diagrams of empirical SCFs concerning monthly
rainfall amounts (Şen and Habib, 2001). At a first glance, it is obvious from this fig-
ure that there are great scatters at any given distance in the correlation coefficients,
and unfortunately, one cannot easily identify a functional trend. The scatter can be
averaged out by computing mean correlation coefficient over relatively short dis-
tance intervals (Thiebaux and Pedder, 1987). The following significant points can
be deduced from these SCFs.

1) Each monthly average SCF shows a monotonically decreasing trend.

2) Due to averaging procedure within the first 15 km interval, it may appear in
Figs. 4.5 and 4.6 that the correlation coefficient at lag zero is not equal to +1 as
expected. Monthly average SCFs for the data considered are given in Fig. 4.6.
Herein, averaging is taken over successive 30 km intervals. This discrepancy is
entirely due to the averaging scheme rather than a physical reality. Hence, this is
not a physically plausible result but unavoidable consequence of the averaging
procedure.
3) The more the averaging correlation coefficient within the first 30 km the more
strongly related spatial correlation appears between the measurement sites.

4.4.1 Correlation Coefficient Drawback

Although the cross-correlation function definition can give a direct indication of

the dependence of variations from the mean at any two sites, it suffers from the
following drawbacks.

1) Auto-correlation and cross-correlation formulations require symmetrical (nor-

mal, Gaussian) pdf of data for reliable calculations. It is well established in
literature that most of the earth sciences data pdfs accord rarely with normal
(Gaussian) pdf but a Weibull, Gamma or logarithmic pdfs, which are all skewed
(Benjamin and Cornell, 1970; Şen, 2002).
2) Since the cross-correlation (as the auto-correlation) is valid for symmetrically
distributed ReVs (and random variables, RVs), the available data must be trans-
formed into normal pdf prior to application of these methods.
3) In the spatial calculation of the cross-correlation, it is necessary to have a
sequence of measurements with time at each site, which is not the case in many
earth sciences problems where there are only one measurement at each site.
 1.0 1.0 1.0
January February March
4.4

0.8 0.8
0.8
0.6 0.6
0.4 0.6
0.4
0.2 0.4
0.2

Correlation
0.0 0.2

Correlation
Correlation
0.0
–0.2 0.0 –0.2
0 200 400 600 800 1000120014001600
–0.2 –0.4
Distance (km) 0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
1.0 1.0 1.0
April May June
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2
Spatial Correlation Function

0.2 0.2 0.0

Correlation
Correlation
Correlation
0.0 0.0 –0.2
–0.2 –0.2 –0.4
–0.4 –0.4 –0.6
0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
1.0 1.0 August 1.0 September
July
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2
0.2 0.2 0.0

Correlation
Correlation
Correlation

0.0 0.0 –0.2

–0.2 –0.2 –0.4
–0.4 –0.4 –0.6
0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
1.0 1.0 1.0
October November December
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2

Correlation
Correlation
Correlation

0.0 0.0 0.0

–0.2 –0.2 –0.2
–0.4 –0.4 –0.4
0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
137

Fig. 4.5 Empirical spatial correlation functions

 1.0
1.0 1.0
0.8
January March
February
0.8 0.8
138

0.6
0.6 0.6
0.4

0.2 0.4 0.4

Correlation
0.0

Correlation

Correlation
0.2 0.2

–0.2
0.0 0.0
0 200 400 600 800 1000 1200 1400 1600
0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
Distance (km) Distance (km) Distance (km)
1.0 1.0 1.0

0.8
April 0.8
May 0.8
June
0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Correlation

Correlation
Correlation
0.0 0.0 0.0

–0.2 –0.2 –0.2

0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600

Distance (km) Distance (km) Distance (km)

1.0 1.0 1.0

0.8
July 0.8
August 0.8
September
0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Correlation

Correlation
Correlation
0.0 0.0 0.0

–0.2 –0.2 –0.2

4

0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
Distance (km) Distance (km) Distance (km)
1.0 1.0 1.0

0.8
October 0.8
November 0.8
December
0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Correlation

Correlation
Correlation
0.0 0.0 0.0

–0.2 –0.2 –0.2

0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600

Distance (km) Distance (km) Distance (km)

Spatial Dependence Measures

Fig. 4.6 Average and theoretical spatial correlation coefficient

4.4 Spatial Correlation Function 139

4) The correlation function measures the variation around the arithmetic average
values of the measurements at individual sites. However, in the spatial variability
calculations, a measure of relative variability between two sites is necessary.
5) Especially for the last two points, the SV (Matheron, 1963) or cumulative SV
(CSV) (Şen, 1989) concepts are developed and their modifications as the point
CSV (PCSV) are presented and used for the regional assessment of earth sci-
ences data.
For instance, Barros and Estevan (1983) presented a method for evaluating wind
power potential from a 3-month long wind record at a site and data from a regional
network of wind systems. Their key assumption was that “wind speed has some
degree of spatial correlation,” which is a logical conclusion, but they failed to
present an effective method for the objective calculation of the spatial variability,
except by employing cross- and auto-correlation techniques. Their statement does
not provide an objective measure of spatial correlation. Skibin (1984) raised the
following questions.
1) What is “a reasonable spatial correlation?” Are the correlation coefficients
between the weekly averages of wind speed a good measure of it? Answers
to these questions are necessary by any objective method. For instance, PCSV
technique can be employed for this purpose.
2) Do the weekly averages represent the actual ones?
3) How applicable to the siting of wind generators the results obtained by the use
of spatial correlation coefficients?
In deciding about the effectiveness of the wind speed measurement at a site
around its vicinity, the topographic and climatic conditions must be taken into con-
sideration. The smaller the area of influence the more homogeneous orographic,
weather, and climatologic features are and, consequently, the simplest is the model.
However, large areas more than 1,000 km in radius around any site may contain
different climates with different through and ridges, high and low pressure areas
with varying intensities. Furthermore, in heterogeneous regions with varying surface
properties (such as land–sea–lake–river interfaces) and variable roughness param-
eters, the local wind profile and wind potential can be affected significantly. The
wind energy potential and the groundwater availability are highly sensitive to height
variations of hills, valleys, and plains (Şen, 1995). The reasons for wind speed vari-
ations are not only of orographical origin but also of different flow regimes (i.e.,
anabatic–catabatic influences compared with hill top conditions, upwind compared
with leeside sites, flow separation effects). All these effects will lose their influence
further away from the siting point. It can be expected that smaller distance from
the site corresponds to a larger correlation. Here, again it is obvious that the spa-
tial dependence decreases with distance as in Fig. 4.5 (correlation property). Barros
and Estevan (1983) noticed that a small region had higher correlation coefficients
between the sites. On the contrary, the spatial independence increases with the dis-
tance (SV property).
Barchet and Davis (1983) have stated that better estimates are obtained when
the radius of influence is about 200 km from the site. However, this information is
140 4 Spatial Dependence Measures

region-dependent and there is a need to develop an objective technique whereby the

radius of influence can be estimated from a given set of sites.

4.5 Semivariogram Regional Dependence Measure

The regional correlation coefficient calculation requires a set of assumptions, which

are not taken into consideration in practical applications (Şen, 2008). First of all,
calculation of spatial correlation coefficient requires time series records at each site.
This is not possible in many earth sciences studies such as in ore grading, soil prop-
erties, hydrogeological parameters. Rather than a time series availability at each site,
there is only one record, say ore grade record at a set of sites, and therefore it is not
possible to calculate spatial correlation function. However, the only way to depict
the spatial correlation from a set of single records at a set of locations is through the
SV methodology.

4.5.1 SV Philosophy

The very basic definition of the SV says that it is the half squared-difference varia-
tion of the ReV by distance. ReV theory does not use the autocorrelation, but instead
uses a related property called the SV to express the degree of relationship between
measurement points in a region. The SV is defined simply as half-square (variance)
of the differences between all possible point pairs spaced a constant distance, d,
apart. The SV at a distance d = 0 should be zero, because there are no differences
(variance) between points that are compared to themselves. The magnitude of the
SV between points depends on the distance between the points. The smaller the
distance the smaller is the SV, and at larger distances SV value is larger. The SV
is a practical measure of average spatial changes. The underlying principle is that,
on the average, two observations closer together are more similar than two obser-
vations farther apart. This is a general statement where the directional changes are
not considered. The plot of the SV values as a function of distance from a point is
referred to as an SV. However, as points are compared to increasingly distant points,
the SV increases.
The simplest and most common form of ReV is a triplet, and therefore it is illu-
minating first to consider the surface in 3D and then according to the SV definition
it is possible to infer its shape intuitively by mental experiment.
1) Continuously deterministic uniform spatial data: If the ReV is a deterministic
horizontal surface of homogeneous, isotropic, and uniform data as in Fig. 4.7,
then the average half-difference square of such data is zero at every distance as
in Fig. 4.8.
2) Discontinuously deterministic partially uniform spatial data: The continuity in
Fig. 4.7 is disrupted by a discontinuous feature (cliff, fault, facies change,
boundary, etc.) as in Fig. 4.9.
4.5 Semivariogram Regional Dependence Measure 141

Fig. 4.7 Homogeneous, Z (x, y)

isotropic, and uniform ReV

x
0

Fig. 4.8 Perfectly γ (d)

homogeneous and isotropic
ReV SV

d
0

Z(x, y)

Z H (x, y)

ZL(x, y)

[ZH (x, y) –Z L(x, y)]

x
0

Fig. 4.9 Discontinuous surface

The average square-difference at various distances leads to an SV with a dis-

continuity at the origin (see Fig. 4.10), the amount of which is equal to the
square-difference between higher, ZH (x, y), and lower, ZL (x, y), data values as

γ (d) = [ZH (x,y)− ZL (x,y)]2 . (4.2)

The resulting SV is expected to take the shape as in Fig. 4.10, where there is a
non-zero value at the origin. Such a jump at the origin indicates discontinuity
142 4 Spatial Dependence Measures

γ (d)

2
[ZH (x, y) – Z L(x, y)]

d
0

Fig. 4.10 Completely random ReV SV

Fig. 4.11 Continuous linear

trend
Z(x, y)

x
0

embeddings in the spatial event and it is referred to as “sill” in geostatistics

literature.
3) Continuously deterministic spatially linear trend data: If the ReV is a linear sur-
face along the X axis as in Fig. 4.11, then the SV along the X axis by definition
has a quadratic form without any decrease (Fig. 4.12).
This SV does not have any horizontal portion and at large distances the slope
increases in an extreme manner.
4) Discontinuously deterministic spatially linear trend data: If the trend surface in
Fig. 4.11 has a discontinuity (Fig. 4.13), then the SV shape appears as in Fig.
4.14, where there is a jump at the origin, which is referred to as nugget effect in
SV terminology.

γ (d)

Fig. 4.12 Linear trend d

surface SV in X direction 0
4.5 Semivariogram Regional Dependence Measure 143

Fig. 4.13 Discontinuous Z(x, y)

trend surface

x
0

Fig. 4.14 Discontinuous γ (d)

trend surface SV in X
direction

nugget
d
0

5) Completely independent spatial data: If the ReV is completely random with no

spatial correlation as in Fig. 4.15, then the SV will be equal to the variance,
σ2 , of the ReV at all distances as in Fig. 4.16. A decision can be made about
the continuity (or discontinuity) and smoothness of the ReV by visual inspec-
tion from the sample SV. If at small distances the sample SV does not indicate
passage from the origin (nugget effect), then the ReV includes discontinuities,
where there is no regional dependence in the ReV at all. Its SV appears as a
horizontal straight line similar to SV in Fig. 4.10.

Fig. 4.15 Independent spatial data

144 4 Spatial Dependence Measures

Fig. 4.16 Completely γ (d)

random ReV SV
σ2

d
0

The SV in this spatial random event case is equivalent to the expectation of Eq.
(4.2), which after expansion and expectation E(.) operation application on both sides
leads to

 " # " #
E γ (d) = e Z2H (x,y) − 2E[ZH (x,y)ZL (x,y)] + e Z2L (x,y) .

Since the ReV is assumed as spatially independent with zero mean (expectation), the
second term of this expression is equal to zero and the other terms are
equal
 to the
variance, σ2 , of the spatial event. Finally, this last expression yields E γ (d) = 2σ 2 .
In order to have the SV expectation equal to the variance in practical applications,
it is defined as the half-square difference instead of squared-difference as in Eq.
(4.2). Consequently, the SV of an independent ReV appears as having a sill value
similar to Fig. 4.10, but this time the sill value is equal to the spatial variance of
the ReV.

4.5.2 SV Definition

The SV is the basic geostatistical tool for visualizing, interpreting, modeling, and
exploiting the regional dependence in an ReV. It is well known that even though the
measurement sites are irregularly distributed, one can find central statistical param-
eters such as mean, median, mode, variance, skewness, but they do not yield any
detailed information about the phenomenon concerned. The greater the variance the
greater is the variability, but unfortunately this is a global interpretation without
detailed useful information. The structural variability in any phenomenon within an
area can best be measured by comparing the relative change between two sites. For
instance, if any two sites, distant d apart, have measured concentration values Zi
and Zi+d, then the relative variability can simply be written as (Zi – Zi+d ). How-
ever, similar to Taylor (1915) theory concerning turbulence, the squared-difference,
(Zi − Zi+d )2 , represents this relative change in the best possible way. This squared-
difference has appeared first in the Russian literature as the “structure function”
of ReV. It subsumes the assumption that the smaller the distance, d, the smaller
will be the structure function. Large variability implies that the degree of depen-
dence among earth sciences records might be rather small even for sites close to
each other.
4.5 Semivariogram Regional Dependence Measure 145

In order to quantify the degree of spatial variability, variance and correlation tech-
niques have been frequently used in the literature. However, these methods cannot
account correctly for the spatial dependence due to either non-normal pdfs and/or
irregularity of sampling positions.
The classical SV technique has been proposed by Matheron (1963) to eliminate
the aforementioned drawbacks. Mathematically, it is defined as a version of Eq.
(4.28) by considering all of the available sites within the study area as (Matheron,
1963; Clark, 1979)

1 
nd
γ (d) = (Zi − Zi+d )2 , (4.3)
2nd
k=1

where k is the counter of the distance, which can be expanded by considering the
regional arithmetic average, Z, of the ReV as follows:

1
nd

   2
γ (d) = Zi − Z − Zi+d − Z
2
k=1
" 2     2 #
= Zi − Z − 2 Zi − Z Zi+d − Z + Zi+d − Z .

The elegancy of this formulation is that the ReV pdf is not important in obtaining the
SV, and furthermore, it is effective for regular data points. It is to be recalled, herein,
that the classical variogram, autocorrelation, and autorun techniques (Şen, 1978)
all require equally spaced data values. Due to the irregularly spaced point sources,
the use of classical techniques is highly questionable, except that these techniques
might provide biased approximate results only. The SV technique, although suitable
for irregularly spaced data, has practical difficulties as summarized by Şen (1989).
Among such difficulties is the grouping of distance data into classes of equal or
variable lengths for SV construction, but the result appears in an inconsistent pattern
and does not have a non-decreasing form as expected in theory. As the name implies
a SV, γ (d), is a measure of spatial dependence of an ReV.
Due to independence any cross-multiplication of Zi and Zj will be equal to zero
on the average, and hence this is equivalent to regional variance, σ2 , as explained in
the previous section. Figure 4.16 shows this mental experiment SV as a horizontal
straight line. Hence, at every distance the SV is dominated by sill value only. Expert
reasoning of SV models in the previous figures help to elaborate some fundamental
and further points as follows.

1) If the ReV is continuous without any discontinuity, then the SV should start
from the origin, which means that at zero distance SV is also zero (Figs. 4.8 and
4.12).
2) If there is any discontinuity within the ReV, then at zero distance a non-zero
value of the SV appears as in Figs. 4.10, 4.14, and 4.16.
3) If there is an extensive spatial dependence, then the SV has increasing values at
large distances (Figs. 4.12 and 4.14).
146 4 Spatial Dependence Measures

4) When the spatial dependence is not existent, then the SV has a constant non-
zero value equal to the regional variance of the ReV at all distances as in
Fig. 4.16.
5) Under the light of all what have been explained so far, it is logically and ratio-
nally obvious that in the case of spatial dependence structure in ReV the SV
should start from zero at zero distance and then will reach the regional variance
value as a constant at large distances. The SV increases as the distance increases
until at a certain distance away from a point it equals the variance around the
average value of the ReV, and will therefore no longer increase, causing a flat
(stabilization) region to occur on the SV, which is called as a sill (Fig. 4.17). The
horizontal stabilization level of sample SV is referred to as its sill. The distance
at which the horizontal SV portion starts is named as the range, R, radius of
influence or dependence length after which there is no spatial (regional) depen-
dence between data points. Only within this range, locations are related to each
other and hence all measurement locations in this region are the nearest neigh-
bors that must be considered in the estimation process. This implies that the
ReV has a limited areal extend over which the spatial dependence decreases or
independence increases in the SV sense as in Fig. 4.17.
The classical SV is used to quantify and model spatial correlations. It reflects
the idea that closer points have more regional dependence than distant points. In
general, spatial prediction is a methodology that imbeds the spatial dependence
in the model structure.
6) At some distance, called the range, the SV will become approximately equal
to the variance of the ReV itself (see Fig. 4.17). This is the greatest distance
over which the value at a point on the surface is related to the value at another
point. The range defines the maximum neighborhood over which control points
should be selected to estimate a grid node, to take advantage of the statistical
correlation among the observations. In the circumstance where the grid node
and the observations are spaced so that all distances exceed the range, Kriging
produces the same estimate as classical statistics, which is equal to the mean
value.
7) However, most often natural data may have preferred orientations and as a result
ReV values may change more along the same distance in one direction than
another (Fig. 4.3). Hence, in addition to distance, the SV becomes a function of
direction (Fig. 4.18).

γ (d) Range (radius of influence)

Scale Sill (regional variance)

Nugget effect
Fig. 4.17 Classical global d
SV and elements 0
4.5 Semivariogram Regional Dependence Measure 147

γ (d) γ (d)

d d
0 0
a b
Fig. 4.18 Classical directional SV (a) major axis, (b) minor axis

It is possible to view the general of SV as a three-dimensional function as the

change of SV value, γ(θ,d), with respect to direction, θ, and separation distance, d.
Of course, θ and d are the independent variables. In general, specification of any SV
requires the following information:
a) Sill (regional variance)
b) Range, (radius of influence)
c) Nugget, (zero distance jump)
d) Directional values of these parameters.

The last point is helpful for the identification of regional isotropy or anisotropy.
For the Kriging application the convenient composition of these parameters must be
identified through a theoretical SV. Whether a given sample SV is stationary or not
can be decided from its behavior at large distances. If the large distance portion of
the SV approaches a horizontal line, then it is stationary, which means intuitively
that there are rather small fluctuations with almost the same variance at every corner
of the region.
If the SV is generated from paired points selected just based on distance (with
no directional component), then it is called isotropic (iso means the same; tropic
refers to direction) or omni-directional. In this case, the lag-distance measure is a
scalar and the SV represents the average of all pairs of data without regard to their
orientation or direction. A standardized SV is created by dividing each SV value by
the overall sample variance, which allows SVs from different data sets on the same
entity for facilitating mutual comparison.
On the other hand, SVs from points that are paired based on direction and dis-
tance are called anisotropic (meaning not isotropic). In this case, the lag measure is
a vector. The SVs in this case are calculated for data that are in a particular direction
as explained in Section 4.3. The regularity and continuity of the ReV of a natural
phenomenon are represented by the behavior of SV near the origin. In SV mod-
els with sill (Fig. 4.17), the horizontal distance between the origin and the end of
SV reflects the zone where the spatial dependence and the influence of one value
on the other occur, and beyond this distance the ReVs Z(x) and Z(x+d) are inde-
pendent from each other. Furthermore, SVs, which increase at least as rapidly as
d2 for large distances d indicate the presence of drift (trend), i.e., non-stationary
mathematical expectation. Plot of SV graphs for different directions gives valuable
148 4 Spatial Dependence Measures

information about continuity and homogeneity. If SV depends on distance d only,

it is said to be isotropic; but if it depends on distance as well as direction, it is said
to be anisotropic. A properly fitted theoretical SV model allows linear estimation
calculations that reflect the spatial extent and orientation of spatial dependence in
the ReV to be mapped. Details on these points can be found in standard text books
on geostatistics (Davis, 1986; Clark, 1979).
There are also indicator SVs, which are calculated from data that have been trans-
formed to a binary form (1 or 0), indicating the presence or absence of some vari-
able, or values that are above some threshold. In the calculation of sample SVs the
following rules of thumb must be considered.

1) Each distance lag (d) class must be represented by at least 30–50 pairs of points.
2) The SV should only be plotted out to about half the width of the sampling space
in any direction.

Characterizing spatial correlation across the site through experimental SV can

often be the most time-consuming step in a geostatistical analysis. This is particu-
larly true if the data are heterogeneous or limited in number. Without a rationale for
identifying the major direction of anisotropy, the following steps might be useful in
narrowing the focus of the exercise.

1) Begin with an omni-directional SV with a bandwidth large enough to encompass

all data points on the site. In practice, maximum lag distance can be taken as 1/3
of the maximum distance between the data points.
2) Select the number of lags and lag distances sufficient to span a significant por-
tion of the entire site, and choose the lag tolerance to be very close in value to
the lag distance itself.
3) Calculate the SV. In most cases, data become less correlated as the distance
between them increases. Under these circumstances, the SV values should pro-
duce a monotonic increasing function, which approaches a maximal value called
the sill. In practice, this may not be the case with SV values that may begin high
or jump around as distance increases.
4) Adjust the number of lags and lag tolerances until, generally, a monotonic
increasing trend is seen in the SV values. If this cannot be achieved, it may
be that a geostatistical approach is not viable or that more complicated trends
are occurring than can be modeled. If a visual inspection of the data or knowl-
edge about the dispersion of contamination indicates a direction of correlation,
it may be more appropriate to first test this direction.
5) Assuming the omni-directional SV is reasonable, add another direction to the
plot with a smaller tolerance. You may have to adjust the bandwidth and angle
tolerance to produce a reasonable SV plot.
6) If the second direction rises slower to the sill or rises to a lower sill, then this is
the major direction of anisotropy.
7) If neither direction produces significantly lower spatial correlation, it may be
reasonable to assume an isotropic correlation structure.
4.5 Semivariogram Regional Dependence Measure 149

8) Add a cone structure with direction equal to the major direction plus 90◦ , and
model the SV results in this direction.
9) If the data are isotropic, choose the omni-directional SV as the major direction.

4.5.3 SV Limitations

The SV model mathematically specifies the spatial variability of the data set and
after its identification the spatial interpolation weights, which are applied to data
points during the grid node calculations, are direct functions of the Kriging model
(Chapter 5). In order to determine the estimation value all measurements within
the SV range are assigned weights depending on the distance of neighboring point
using the SV. These weights and measurements are then used to calculate the esti-
mation value through Kriging modeling. Useful and definite discussions on the
practicalities and limitations of the classical SV have been given by Şen (1989) as
follows.
1) The classical SV, γ(d), for any distance, d, is defined as the half-squared-
difference of two measurements separated by this distance. As d varies from
zero to the maximum possible distance within the study area, the relationship
of the half-square-difference to the separation distance emerges as a theoretical
function, which is called the SV. The sample SV is an estimate of this theoretical
function calculated from a finite number, n, of samples. The sample SV can be
estimated reliably for small distances when the distribution of sampling points
within the region is regular. As the distance increases, the number of data pairs
for calculation of SV decreases, which implies less reliable estimation at large
distances.
2) In various disciplines of the earth sciences, the sampling positions are irregu-
larly distributed in the region, and therefore, an unbiased estimate of SV is not
possible. Some distances occur more frequently than others, and accordingly
their SV estimates are more reliable than others. Hence, a heterogeneous reli-
ability dominates the sample SV. Consequently, the sample SV may have ups
and downs even at small distances. Such a situation gives rise to inconsistencies
and/or experimental fluctuations with the classical SV models, which are, by
definition, non-decreasing functions, i.e., a continuous increase with distance is
their main property. In order to give a consistent form to the sample SV, different
researchers have used different subjective procedures.
a) Journel and Huijbregts (1978) advised grouping of data into distance classes
of equal length in order to construct a sample SV. However, the grouping
of data pairs into classes causes a smoothing of the sample SV relative to
the underlying theoretical SV. If a number of distances fall within a certain
class, then the average of half-squared-differences within this class is taken
as the representative half-squared-difference for the mid-class point. The
effect of outliers is partially damped, but not completely smoothed out by
the averaging operation.
150 4 Spatial Dependence Measures

b) To reduce the variability in the sample SV, Myers et al. (1982) grouped the
observed distances between samples into variable length classes. The class
size is determined such that a constant number of sample pairs fall in each
class. The mean values of distances and half-squared-differences are used
for the classes as a representative point of sample SV. Even this procedure
resulted in an inconsistent pattern of sample SV (Myers et al., 1982), for
some choices of the number, m, of pairs falling within each class. However,
it was observed by Myers et al. that choosing m = 1,000 gave a discernible
shape. The choice of constant number of pairs is subjective and, in addition,
the averaging procedures smooth out the variability within the experimental
SV. As a result, the sample SV provides a distorted view of the variable in
that it does not provide, for instance, higher-frequency (short wave length)
variations. However, such short-wavelength variations, if they exist, are so
small that they can be safely ignored.

The above procedures have two basic common properties, namely predetermi-
nation of a constant number of pairs or distinctive class lengths and the arithmetic
averaging procedure for half-squared-differences as well as the distances. The for-
mer needs a decision, which in most cases is subjective, whereas the latter can lead
to unrepresentative SV values. In classical statistics, only in the case of symmetri-
cally distributed data the mean value is the best estimation, otherwise the median
becomes superior. Moreover, the mean value is sensitive to outliers. The following
points are important in the interpretation of any sample SV.
1) The SV has the lowest value at the smallest lag distances (d) and increases with
distance, leveling off at the sill, which is equivalent to the overall regional vari-
ance of the available sample data. It is the total vertical scale of the SV (Nugget
effect + sum of all component scales). However, linear, logarithmic, and power
SVs do not have a sill.
2) The range is the average distance (lag) within which the samples remain spa-
tially dependent, and it corresponds to the distance at which the SV values level
off. Some SV models do not have a length parameter; e.g., the linear model has
a slope instead.
3) The nugget is the SV value at which the model appears to intercept the ordinate.
It quantifies the sampling and assaying errors and the short-scale variability (i.e.,
spatial variation that occurs at distance closer than the sample spacing). It rep-
resents two often co-occurring sources of variability.
a) All unaccounted for spatial variability at distances smaller than the smallest
sampling distance.
b) Experimental error is often referred to as human nugget. According to
Liebhold et al. (1993), interpretations made from SVs depend on the size
of the nugget because the difference between the nugget and the sill (if there
is one) represents the proportion of the total sample variance that can be
modeled as spatial variability.
4.6 Sample SV 151

4.6 Sample SV
In practice, one is unlikely to get SVs that look like the one shown in Fig. 4.17.
Instead, patterns such as those in Fig. 4.19 are more common.
Important practical information in the interpretation and application of any sam-
ple SV is to consider only about d/3 of the horizontal distance axis values from the
origin as reliable.
A digression is taken in this book as for the calculation of sample SVs. Instead of
easting- and northing-based SVs, it is also possible to construct SVs based on triple
variables. In the following different triple values are assessed for the SV shapes and
interpretations. For instance, in Fig. 4.20 the chloride change with respect to calcium
and sodium are shown in 3D and various sample SVs along different directions are
presented in Fig. 4.21.
This figure indicates that the change of chloride data with respective independent
variables (magnesium and calcium) are of clumped type without leveling effect.
It is possible to consider Fig. 4.20 as having two parts, namely an almost linear
trend and fluctuations (drift) around it. In such a case a neighborhood definition
and weights assignments become impossible. Therefore, the ReV is divided into
two parts, the residual and the drift. The drift is the weighted average of points
within the neighborhood around the estimation value. The residual is the difference
between the ReV and the drift.
The residual is a stationary ReV in itself and hence allows construction of an
SV. However, once again the problem of not being able to define a neighborhood
arises. Therefore, an arbitrary neighborhood is chosen from which a drift can be
calculated. The calculation includes the points within the assumed neighborhood

γ (d) γ (d)

d d
0 0
a b

γ (d) γ (d)

d d
0 0
c d
Fig. 4.19 Common forms of sample SVs (a) random; (b) uniform; (c) clumped with leveling; (d)
clumped without leveling
152 4 Spatial Dependence Measures

Chloride (ppm)

m)
(pp
um
lci
Ca
Ma
gn
es
ium
(p
pm
)

Fig. 4.20 Triple surface of chloride

and a corresponding coefficient for each point, which will be explained in more
detail in the Kriging section. The only variable left in the calculation is the SV;
however, no SVs exist from which to obtain the SV. Therefore, a reasonable SV is
assumed and compared to the resultant residual SV. If the two are the same, then
the assumptions made about the neighborhood and SV are correct, and regional
estimation can be made. If they differ, then another SV must be used until they
become the same. It is possible to identify from an SV the optimum distance after
which regional dependence is zero or constant. By definition, the SV is half the
variance of the difference between all possible points at a constant distance apart.
The existence of underlying trend implies the expectation of sample SV similar to
ideal case as in Fig. 4.12.
Zero and 30◦ directional SV in Fig. 4.21 a and b have such a trend, whereas other
directional sample SVs along 60◦ and 90◦ have more or less random type, provided
that only d/3 of the distance axis variables (0–15) are considered.
It is obvious that practically there is no nugget effect in these sample SVs, which
is rather obvious from Fig. 4.20 where the trend surface does not have discontinuity
as in Fig. 4.12. Similarly, 3D representation of bicarbonate variation with calcium
and magnesium is presented in Fig. 4.22, with a global (without different directions)
sample SV in Fig. 4.23.
4.7 Theoretical SV 153

Column C:Cl
Direction:30.0 Tolerance:10.0
Column C: Cl
Direction:0.0 Tolerance:10.0
8000
4000
7000
3500
6000

Semivariogram
3000
Semivariogam

5000
2500

4000
2000

3000
1500

2000
1000

500 1000

0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40

Distance Distance

Column C: Cl
Column C:Cl
Direction: 60.0 Tolerance: 10.0
Direction: 90.0 Tolerance:10.0

7000 9000

8000
6000
7000
Semivariogram
Semivariogram

5000
6000

4000 5000

4000
3000

3000
2000
2000

1000
1000

0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40

Distance Distance

Fig. 4.21 Chloride sample SVs

Visual inspection of Fig. 4.22 indicates that a trend surface is embedded in the
ReV. This trend surface is almost horizontal with a very small angle, and accordingly
its global sample SV in Fig. 4.23 has almost horizontal sector within d/3.
Total dissolved solids (TDS) measurement is an indicator of the water quality
variation with respect to calcium and magnesium, and its variation is presented in
Fig. 4.24 with corresponding global SV in Fig. 4.25.

4.7 Theoretical SV

Useful discussion on the computation of the SV in one or two dimensions has been
given by Clark (1979) and Hohn (1988). The SV as computed from the data will
tend to be rather lumpy and the more irregular the data less regular it will appear.
Whatever the extend of lumpiness, the graph of γ (d) may often be linked to one or
154 4 Spatial Dependence Measures

Bicarbonate (ppm)

m)
( pp
m
lciu
Ca
Ma
gn es i
um
( pp
m)

Fig. 4.22 Triple surface of chloride

Column C: HCO3
800 Direction: 0.0 Tolerance: 90.0

700

600
Semivariogram

500

400
Reliability range

300
Nugget = 345; Slope = 10.3; Anisotropy = 1.0
200

100

0
0 5 10 15 20 25 30 35 40
Distance

Fig. 4.23 Bicarbonate sample SV

other of a small number of theoretical (ideal) and simple curves that relate γ (d) to
d, which are referred to as the theoretical SVs. These theoretical curves are mod-
els of SV that have been defined on theoretical and sample basis. The fitting of
the theoretical SV curve from a set of functions to the experimentally available
one derived from real data has been developed into the art of “structural analy-
sis” discussed in detail by a number of researchers (Journel and Huijbregts, 1978;
Myers et al., 1982; Hohn, 1988). The main theoretical SV types are linear, spherical,
exponential, Gaussian, or cubic types as will be explained later in detail. These
4.7 Theoretical SV 155

Total disssolved solids (ppm)

m)
( pp
m
siu
g ne
Ca Ma
lc ium
(pp
m)

Fig. 4.24 Triple surface of total dissolved solids

Column C: TDS
Direction: 0.0 Tolerance: 90.0

60000

50000
Semivariogram

40000

30000
Reliability range

20000

10000

0
0 5 10 15 20 25 30 35 40
Distance

Fig. 4.25 Total dissolved solids SV

156 4 Spatial Dependence Measures

functions express rather well qualitatively the characteristics of ReV and act as a
quantified summary of the structural information, which is then channeled into the
estimation procedures of the natural (geologic, hydrologic, meteorological, atmo-
spheric, etc.) phenomena.
In order to apply Kriging modeling to an ReV, the first step is to obtain sample SV
from the available data and then to match this sample SV to a theoretically suitable
mathematical function. It describes the relationship between the difference of values
and distance with a mathematical function. Several different types of functions can
be used, each with a different form to the distance function. In addition to its shape
or form, the SV model is described by three parameters, namely nugget, sill, and
range. Nugget shows how much variance is observed at a distance of zero. It shows
up because there may be variation at distances shorter than the sample spacing or
because there are errors in the measurements. It corresponds to discontinuity feature
existence in the ReV. The sill shows how much variation is observed when the SV
levels off at large distances. Most SVs become constant at large distances, provided
that the ReV has spatial dependence without any systematic trend component. Once
it is far enough away, there is no relationship in the ReV between the distances of
two points. The range shows how far one has to cover distance before the SV levels
off to the sill. At distances less than the range an ReV will be said to be spatially
dependent and beyond the range distance the ReV has no effect and there is no
spatial dependence. Finally, if the directional SVs are very different, one may need
to specify an anisotropy parameter. For instance, the N-S SV may be different from
the E-W SV (there is a different range, sill, nugget value, and rather different shape).
The development of an appropriate SV model for a data set requires the under-
standing and application of advanced statistical concepts and tools, which is the
science of SV modeling. In addition, the development of an appropriate SV model
for a data set requires knowledge of the tricks, traps, pitfalls, and approximations
inherent in fitting a theoretical SV model to real world data, which is the art of SV
modeling. Skill with the science and art are both necessary for success.
The development of an appropriate SV model requires numerous correct deci-
sions. These decisions can only be properly addressed with an intimate knowledge
of the data at hand, and a competent understanding of the data genesis (i.e., the
underlying processes from which the data are drawn).
Several SV models have been developed to describe the various underlying spa-
tial patterns in data. Examples of isotropic models including spherical, exponential,
linear, power, and Gaussian models are used as input in the Kriging ReV estimation
process (Chapter 5).

4.7.1 Simple Nugget SV

This corresponds to a random field with no regional dependence structure at all. In
such a case the ReV is represented by its variance, σ2 , only. Its mathematical form
is given as
4.7 Theoretical SV 157

σ2 for d>1
γ (d) = . (4.4)
0 for d=0

The random field that this SV represents is not continuous (Fig. 4.15). No matter
what the distance is (small or large), each ReV is completely independent and dif-
ferent from others. In this case the spatial analysis methodology is the probability
principles only (Fig. 4.16). The special case of this nugget SV occurs when there is
not any spatial variability but a uniform ReV value at each point (see Figs. 4.7 and
4.8). The mathematical model is then γ(d) = 0 for all d values.

4.7.2 Linear SV
If the ReV does not have any discontinuity, then its simplest model is given as,

γ (d) = βd, (4.5)

where β is the only model parameter with its meaning of the slope of the straight
line (see Fig. 4.26).
A more general form of theoretical linear SV is the mixture of nugget and the
linear SVs. It postulates a linear relationship between the cumulative half-squared-
difference and the distance as

γc (d) = α + βd. (4.6)

in which α corresponds to nugget effect and β is the slope of the theoretical SV as

in Fig. 4.27.

50
45
40 β=5
Semivariogram, γ(d)

35
30
25 β=3

20
15
10 β=1
5 β = 0.5
β = 0.3
0
0 1 2 3 4 5 6 7 8 9 10
Fig. 4.26 Linear SV model Distance, d
158 4 Spatial Dependence Measures

60

50
β=5
Semivariogram, γ(d)
40

β=3
30

20 β=1
β = 0.5
10 β = 0.3

0
0 1 2 3 4 5 6 7 8 9 10
Distance, d

Fig. 4.27 General linear SV

The unknowns α and β can be solved from a sample SV either by substitution

of two most characteristic points that should be preserved in the theoretical SV or
through a least squares regression line fitting. The first approach is rather subjective
but may represent expert in a better way. On the other hand, according to theory, the
expected value of the sample variance is the average value of the SV between all
possible pairs of sample locations, which yields one equation (Barnes, 1991). The
second equation is generated by equating the experimental SV for nearest neighbors
to the modeled SV. Thus, it is possible to write

γs (d) = α + βd,

γn (d) = α + βdn ,

where dn = average distance to the nearest neighbor, d = average inter-sample sep-

aration distance, γn (d) = one-half the averaged squared-difference between nearest
neighbors, and γs (d) = sample variance. By solving the two equations for the two
unknown parameters, and checking for unreasonable values, one can get the final
formulae used in the Kriging technology:
 
dγn (d) − dn γs (d)
α = max ,0 ,
d − dn
and
 
γs (d) − γn (d)
β = max ,0 .
d − dn
4.7 Theoretical SV 159

4.7.3 Exponential SV
The mathematical expression of theoretical exponential SV model is given as
(Cressie, 1993)
( )
γ (d) = α 1 − e−βd , (4.7)

where the two model parameters are α > 0 and β > 0. This model is used com-
monly in hydrological studies. Figure 4.28 shows exponential models with different
variance values and β = 1. It does not have a nugget value and hence represents
continuous ReVs, which are stationary, because at large distances exponential SV
approaches a horizontal asymptote.

4.7.4 Gaussian SV

Its general expression is given in the following mathematical form (Pannatier, 1996):

( )
γ (d) = α 1 − e−βd ,
2
(4.8)

where α > 0 and β > 0 are the two parameters and according to their values it takes
different forms. Figure 4.29 shows a set of Gaussian SV with different α values and
β = 1. Since at large distances there is a horizontal portion, the representative ReV
is also stationary.
At small distances the Gaussian model appears as proportional to d2 , which
implies that the ReV is smooth enough to be differentiable, i.e., the slope of SV
tends to a well-defined limit as the distance between the two points.

5
4.5
α=5
4
Semivariogram, γ(d)

3.5
α=3
3
2.5
2
1.5
α=1
1
0.5
0
Fig. 4.28 Exponential SV 0 1 2 3 4 5 6 7 8 9 10
model Distance, d
160 4 Spatial Dependence Measures

5
4.5 α=5
4

Semivariogram, γ(d) 3.5

α=3
3
2.5
2
1.5
α=1
1
0.5
0
0 1 2 3 4 5 6 7 8 9 10
Distance, d

Fig. 4.29 Gaussian SV model

4.7.5 Quadratic SV

Its mathematical expression is given by Alfaro (1980) as

αd (2 − d) d<1
γ (d) = , (4.9)
α d≥1

where the single model parameter α > 0. Its shape is given in Fig. 4.30 for different
α values. As it increases the sill value decreases, which means that the fluctuations
become smaller. It represents stationary ReVs with non-differentiable properties.

4.7.6 Rational Quadratic SV

Cressie (1993) provided this SV for use with the following expression:

 
d2
γ (d) = α , (4.10)
1 + d2

where α > 0 is the model parameter. It has similar behavior to Gaussian model
at small distances, which implies that the ReV is differentiable and rather contin-
uous. Increase in α parameter value causes increase in the sill level, which fur-
ther shows that the fluctuations in the ReV from its average level become bigger
(see Fig. 4.31).
4.7 Theoretical SV 161

Fig. 4.30 Quadratic SV

model

5
4.5
4 α=5
Semivariogram, γ(d)

3.5
3
2.5
α=3
2
1.5
1
0.5 α=1

0 1 2 3 4 5 6 7 8 9 10
Distance, d

Fig. 4.31 Rational quadratic SV model

4.7.7 Power SV
Its general form is presented by Pannatier (1996) as
+ +
γ (d) = α +dm + , (4.11)

where α > 0 and 0 < m < 2 are the model parameters. Convex and concave forms
result, respectively, for 0 < m < 1, and 1 < m < 2 as in Fig. 4.32. Besides, it reduces
 162 4 Spatial Dependence Measures

3.5 2.5

α = 1, m = 1.9
3 = 0.9
,m 2 α = 1, m = 1.7
1
α= .7 α = 1, m = 1.5
=0

Semivariogram, γ(d)
Semivariogram, γ(d)

2.5
1, m α = 1, m = 1.3
α=
1.5 α = 1, m = 1.1
2 ,m = 0.5
α=1
1.5 α = 1, m = 0.3 1

1 α = 1, m = 0.1
0.5
0.5

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5
Distance, d Distance, d
a b

Fig. 4.32 Power SV model, (a) concave, (b) convex

to a linear model (Eq. 4.5) for m =1. Depending on whether m > 1 (m < 1), the
SV represents a differentiable (non-differentiable) ReVs. Additionally, none of the
theoretical SVs approach to a sill value, and therefore the represented ReV does not
have stationary property. These SVs describe the same appearance of realizations at
every scale, and therefore they are referred to as self-similar ReVs. This is because
they appear as a straight line on a double-logarithmic paper with different slopes.
The main difference of this model from the others is that it has a non-zero value
for zero distance, i.e., it has a nugget effect. The forms of different cumulative SVs
resulting from Eq. (4.11) are shown in Fig. 4.32.

4.7.8 Wave (Hole Effect) SV

Cressie (1993) gave the general mathematical expression of this model as
 
sin d
γ (d) = α 1 − , (4.12)
d

where again α > 0 is the model parameter (see Fig. 4.33). This model describes
ReVs with excursions above the mean tend to be compensated by excursions below
the mean. It exhibits linear behavior at small distances, which implies that the cor-
responding ReVs are continuous but not differentiable. They are less smooth than
the realization of a random field

4.7.9 Spherical SV

Pannatier (1996) presented the general expression as

  
αd 1.5 − 0.5d2 d<1
γ (d) = , (4.13)
α d≥1
4.7 Theoretical SV 163

Fig. 4.33 Wave (hole effect) 7

SV model
6
α=5

Semivariogram, γ (d)
5

4
α=3
3

2
α=1
1 α = 0.5

0
0 1 2 3 4 5 6 7 8 9 10
Distance, d

with model parameter α > 0. At small distances it is proportional with distance,

and therefore its representative ReV is non-differentiable but continuous. At large
distance the existence of horizontal level (sill) implies that the ReV is stationary. Its
shapes are presented in Fig. 4.34.

Fig. 4.34 Spherical SV 5

model 4.5
4
Semivariogram, γ(d)

α=5
3.5
3
2.5
α=3
2
1.5
1
α=1
0.5 α = 0.5
α = 0.3
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Distance, d

4.7.10 Logarithmic SV

It is a single parameter model with mathematical expression as

γ (d) = α ln d d > 0. (4.14)

164 4 Spatial Dependence Measures

Fig. 4.35 Logarithmic SV 10

model
5 α=5
α=3
α=1

Semivariogram, γ(d)
0 α = 0.5

–5

–10

–15

–20

–25
0 0.5 1 1.5 2 2.5 3
Distance, d

It appears as a straight line on a semi-logarithmic paper (logarithmic distances)

and α represents the slope of this line. Various logarithmic SVs are presented in
Fig. 4.35. Since there is no sill, the corresponding ReV is non-stationary. It can be
used only for integrals over finite volumes and cannot be used directly with point
data values of the ReV (Kitanidis, 1997).

4.8 Cumulative Semivariogram

The CSV method proposed by Şen (1989) as an alternative to the classical SV tech-
nique of Matheron (1965) has various advantages over any conventional procedure
in depicting the regional variability, and hence spatial dependence structure CSV is
defined similarly to the SV, the only difference is that successive cumulative sum-
mations are adopted. CSV has all of the advantages claimed for SV; besides, it pro-
vides more objective way in deriving theoretical models to the regional dependence
behavior of the regionalized variable. Furthermore, standardization of CSV provides
a basis in identifying regional stationary stochastic models (Şen, 1992). The CSV is
a graph that shows the variation of successive half-squared difference summations
with distance. Hence, a non-decreasing CSV function is obtained, which exhibits
various significant clues about the regional behavior of the ReV. The CSV provides
a measure of spatial dependence. The CSV can be obtained from a given set of ReV
data by executing the following steps.

1) Calculate the distance di,j (i = j = 1, 2, . . . , m) between every possible pair of

sparse measurement sites. For instance, if the number of sample sites is n, then
there are m = n(n–1)/2 distance values.
4.8 Cumulative Semivariogram 165

2) For each distance, di,j calculate the corresponding half-squared differences, Di,j ,
of the ReV data. For instance, if the ReV has values of Zi and Zj at two distinct
sites at distance di,j apart, then the half-squared difference is

1 2
Di,j = Zi − Zj . (4.15)
2

3) Take the successive summation of the half-squared differences starting from

the smallest distance to the largest in order. This procedure will yield a non-
decreasing function as


m 
m
γ(di,j ) = Di,j . (4.16)
i=1 i=1

where γ(di,j ) represents CSV value at distance di,j .

4) Plot γ(di,j ) values versus the corresponding distance di,j . The result will appear
similar to the representative CSV functions as in Fig. 4.36. The sample CSV
functions are free of subjectivity because no a priori selection of distance classes
is involved in contrast to the analysis as suggested by Perrie and Toulany (1989)
in which the distance axis is divided into subjective intervals, and subsequently
averages are taken within individual intervals, which are regarded as the repre-
sentative value for this interval.

700 VM

600

500

400
γc(d)

300

200

100
RM
0
0 200 400 600 800 1000 1200 1400 1600 1800
Distance (km)

Fig. 4.36 Representative CSV

166 4 Spatial Dependence Measures

In his discussion on the practical difficulties of classical SV, Şen (1989) has
noticed the reliability of SV estimations at small distances, with regular distribution
of sampling points within the region. However, in various disciplines of geological
sciences, the sampling positions are irregularly distributed in the region (such as in
the earthquake epicenters) where some distances occur more frequently than others;
thus, a heterogeneous reliability dominates the sample SV. In order to overcome
these shortcomings, he developed CSV technique on the basis of the ReVs theory.
This technique possesses all of the objective properties of the classical SVs and, in
addition, it helps to identify the hidden and local dependencies within a region. It is
defined as the successive summation of half-squared differences, which are ranked
according to the ascending order of distances extracted from all possible pairs of
sample locations within a region. Mathematically, CSV is expressed as


k " #
γc (dk ) = d di (k = 1,2,...,m), (4.17)
i=1

where γc (dk ) is the value of the k-th ordered distance CSV value and superscript i
indicates the rank. The procedure of sample CSV calculations as well as its model
form and equations have been given by Şen (1989) as in Fig. II.12. These models
are counterpart to those of classical SV models, but with different interpretations of
the model parameters. The attributes and advantages of CSV can be summarized as
follows.

1) The CSV is a non-decreasing function; however, there may be local flat portions,
implying constancy of the regionalized variables at certain distance, i.e., the
same values have been observed at two locations h apart.
2) The slope of the theoretical CSV at any distance is an indicator of the depen-
dence between pairs of regionalized variables separated by that distance.
3) The sample CSV reflects even smaller dependencies between data pairs,
which are not possible to detect with classical SV due to the averaging
procedure.
4) The sample CSV is straight forward in applications and free of subjectivity,
because there is no need for a priori selection of distance classes. In fact, the
real distances are employed in the construction of the sample CSV rather than
class mid-point distance.
5) The CSV model may be used for irregularly distributed sample positions within
the study region.
6) The underlying model for any regionalized variable can be detected by plotting
the cumulative half-squared differences versus distances on arithmetic-semi-
logarithmic or double logarithmic. Appearance of sample CSV points on any
one of these papers as a straight line confirms the type of model. Such an oppor-
tunity is missing in the samples of classical SV.
7) Model parameter estimates are obtained from the slope and intercept values of
the straight line.
4.8 Cumulative Semivariogram 167

8) Any classical SV model has a theoretical CSV counterpart, which can be

obtained through an integration operation.
9) These characteristics and advantages make the CSV attractive for practical
applications.

4.8.1 Sample CSV

The CSV proposed in the previous section is applied to the transmissivity, total dis-
solved solids, and piezometric level records in the Wasia sandstone aquifer in the
eastern part of the Kingdom of Saudi Arabia (Şen, 1989). A complete hydrogeolog-
ical study of this area has been performed recently by Subyani (1987).
GAMA3 software developed for computing the classical SV by Journel and
Huijbregts (1978) has been applied to groundwater variables such as transmissiv-
ity, piezometric level, and total dissolved solids from the Wasia sandstone aquifer.
The resulting sample SV and sample CSV plots are presented in Figs. 4.37, 4.38,
and 4.39. It is clear from these figures that the half-squared-difference points are
scattered in such a way that it is not possible to distinguish a clear pattern in the
sample SVs, which suffer from fluctuations even at small distances. Comparisons
of the sample SVs in these figures with the sample CSVs indicate that the latter are
more orderly and have distinctive non-decreasing patterns.

2.8

CSV
2.4

2.0
γ c (d) (×10–3 m4/s2)

1.6
3.0
SV
γ (d) (×10–3)

1.2
2.0

0.8
1.0

0.4 0
0 40 80 120 140 180
d (km)
0
0 20 40 60 80 100 120 140 160 180 200
d (km)

Fig. 4.37 Sample CSV for Wasia sandstone transmissivity

168 4 Spatial Dependence Measures

140

CSV
120

100
12 SV
2
γcd (mg/l)

2
80

γ cd (mg/l)
8

60
4

40 0
0 40 80 120 140 180
d (km)
20

0
0 20 40 60 80 100 120 140 160 200
d (km)

Fig. 4.38 Sample CSV for Wasia sandstone dissolved solids

280

240
30 SV
200
γc(d) (m2)

20
γ c(d) (m2)

160
10

120 0
0 40 80 120 140 180
d (km)
80

40
CSV

0
0 10 20 30 40 50 60 70 80 90 100
d (km)

Fig. 4.39 Sample CSV for Wasia sandstone piezometric level

4.8 Cumulative Semivariogram 169

A sample CSV often yields more or less a straight line for large distances, which
corresponds to the sill concept in the classical SV. Furthermore, the sample CSV
starts as a curve with different curvatures over some distance domain before it
becomes almost a straight line. The length of the distance domain over which the
sample CSV occurs as a curve is a counterpart of the range in the classical SV.
Hence, it is straightforward to determine the range from the sample CSV. The piezo-
metric level sample CSV in Fig. 4.39 shows an initial range portion, which has zero
half-squared-differences for about 10 km. Such a portion implies physically that the
piezometric level does not change significantly within distances less than 60 km.
In fact, the Wasia aquifer has remained free of any tectonic movements, it is exten-
sive, and the recharge is negligible, but it is discharged by local well groups that are
situated at large distances from each other (Powers et al., 1966).

4.8.2 Theoretical CSV Models

In order to be able to apply Kriging estimation techniques to an ReV, a functional
relationship must be established between the distance and the measure of regional
dependence, which, herein, is the CSV. These models must be non-decreasing func-
tions. Although numerous functions have this property, in practice restricting them
to a few simple ones is desirable. By considering basic definitions of both the clas-
sical and the CSVs, they may be related through an integration as

!d
γc (d) = γ (u) du, (4.18)
0

or through differentiation as
+
dγc (u) ++
γ (d) = u=d . (4.19)
du +

Therefore, a CSV counterpart may be found for any given classical SV using Eq.
(4.18). Furthermore, Eq. (4.19) indicates that the theoretical classical SV value at
any distance is equal to the slope of the theoretical CSV at the same distance. In the
following, models which have been used previously for SVs by many researchers
will be assessed from the CSV point of view.

4.8.2.1 Linear Model

The linear CSV model signifies that the phenomenon has regional independence and
it evolves in accords with an independent (white noise) process. Hence, in an objec-
tive analysis of such a phenomenon one needs to know only the mean and standard
deviation, and consequently the prediction of any grid point value can be withdrawn
from an independent process because the radius of influence is equal to zero. This is
170 4 Spatial Dependence Measures

the most extreme case, which rarely happens in earth sciences domains. This model
postulates a linear relationship between the cumulative half-squared difference and
the distance as

γc (d) = α + βd, (4.20)

in which α and β are the model parameters (Fig. 4.40a). The sample CSV of the
regionalized variable that abides by this model will appear as a straight line on
arithmetic paper.
In fact, α is the intercept on the CSV axis and β is the slope of this straight
line. This slope corresponds to the sill value in the classical SV, which represents a
pure nugget effect (Şen, 1989). Furthermore, β represents exactly the variance of the
underlying random field. Hence, the smaller the slope of the straight line, the smaller
the random fluctuation in the ReV. If the slope is equal to zero, theoretically, this
indicates a complete deterministic uniform variation in the ReV. The sample CSV
scatter diagram and the fitted regression line to pH values measured at 71 sample
locations within the Umm Er Radhuma Limestone aquifer in the Eastern Provinces,
Saudi Arabia (Fig. 4.41), has the form

γc (d) = −0.213 + 1.144d,

1
12 a 12 b β=
β>1

8 8
γc(d)

γc(d)

4 Linear 4 Power

0 0
0 4 8 12 16 20 0 4 8 12 16 20
d d

c 12 d
00

12
10
β=1

β=

8 8
γc(d)
γc(d)

1
β=

4 4
Exponential
Logaritmic
0 0
0 4 8 12 16 20 0 4 8 12 16 20
d d
Fig. 4.40 Theoretical CSVs
4.8 Cumulative Semivariogram 171

15

data
model
12

9
γc(d)

0
0 1 2 3 4 5 6 7 8 9 10 11 12
d (km)

Fig. 4.41 Sample CSV for pH in Umm Er Radhuma aquifer

from which the parameter estimates are α = −0.213 and β = 1.144. The hydro-
chemical data were presented by Şen and A1-Dakheel (1985) for major anions and
cations.

4.8.2.2 Power Model

This is a two-parameter model which yields a set of different shapes for the theoret-
ical CSV (Fig. 4.40b). The mathematical expression for this model is

γc (d) = αdβ , (4.21)

in which α is the scale parameter and β is the shape parameter. Because 0 < β < 2
for a theoretical SV from a power family (Journel and Huijbregts, 1978, p. 165),
parameter β for the theoretical CSV in Eq. (4.21) is restricted to the range 1 < β < 3.
The derivative of Eq. (4.21) yields also a power form for the classical SV. Obviously,
use of a double logarithmic paper facilitates parameter estimation. Sulfate concen-
trations in the Umm Er Radhuma aquifer groundwater show on double logarithmic
paper a more or less straight line pattern (Fig. 4.42).
The mathematical expression of this straight line by the regression technique can be
found as

log γc (d) = 0.46 + 0.841d,

172 4 Spatial Dependence Measures

30
data
model
101
γc(d)

100

10–1
101 100 101 20
d (km)

Fig. 4.42 Sample CSV for sulfate in Umm Er Radhuma aquifer

hence, parameter estimates are log α = 0.46 or α = −2.88 and β = 0.84. The original
form of this model prior to transformation can be written as γc (d) = 2.88d0.84 .

4.8.2.3 Exponential CSV

The general form of this CSV can be written as

γc (d) = αeβd , (4.22)

where α and β are scale and shape parameters, respectively. The main difference of
this model from the others is that it has a non-zero value for zero distance, i.e., it has
a nugget effect. Forms of different CSVs resulting from Eq. (4.22) are shown (Fig.
4.40c). The sample CSV can be checked for concordance with this model by plotting
log γc (d) vs. d on semi-logarithmic paper. If the sample points appear as a straight
line, the exponential model is the generating mechanism of the regional variability
within the regionalized variable. The slope of this line directly yields an estimate of
β, whereas the intercept on the γc (d) axis leads to an estimate of β. This model does
not have a unique classical SV, which has appeared in the geostatistical literature.
The sample CSV for bicarbonate concentrations in the Umm Er Radhuma aquifer
appears as a straight line on semi-logarithmic paper (Fig. 4.43). The appearance of
this straight line implies that the convenient model for bicarbonate concentrations
for this aquifer is of exponential type. The regression line of this scatter diagram is
4.8 Cumulative Semivariogram 173

100

data
model
γc(d)

10–1
0 1 2 3 4 5 6 7 8 9 10 11 12
d (km)

Fig. 4.43 Sample CSV for bicarbonate in Umm Er Radhuma aquifer

log γc (d) = −0.86 + 0.079d,

and, correspondingly, the model parameter estimates are log γc (d) = −0.86 or α
= 0.14 and β = 0.079. Hence, the original form of the model can be written as
γc (d) = 0.14e0.079d .

4.8.2.4 Logarithmic CSV

The mathematical expression of this model can be written as

α + β log d for d > 1
γc (d) = , (4.23)
0 for d < 1

in which α and β are the two model parameters. This model differs from the expo-
nential one in that it has an intercept on the distance axis similar to the sample
CSV for piezometric level (Fig. 4.39). Different forms of the logarithmic models
are presented (Fig. 4.40d). The model can be depicted from a sample CSV plotted
on semi-logarithmic paper as γc (d) vs. log d. If the sample points appear as a straight
line, the validity of the logarithmic model is confirmed. The slope of this straight
line is equal to β, and the cumulative half-squared difference corresponding to d = 1
174 4 Spatial Dependence Measures

yields the estimate of a. Such a model is similar to what is referred to in the classical
SV terminology as the De Wijsian model (DeWijs, 1972).
Other models for the CSV can be constructed from classical CSV models through
Eq. (4.18). For instance, the exponential model of the classical SV, which is
" #
γ (d) = α 1 − e−βd , (4.24)

corresponds to a CSV model, which is

 
1 −βd
γc (d) = α d − e , (4.25)
β

in which α and β are model parameters. A close inspection of Eq. (4.25) indicates
that for large distances, (1/β)exp(−βd) ≈ 0; consequently, at large distances this
model appears as a straight line (on arithmetic paper) whose slope is an estimate
of α. In addition, this model has an intercept value, γc (0), which is equal to α/β.
Provided that α is known from the slope at large distances, this ratio yields the
estimate of β. These α and β values are the parameters of the classical exponential
SV model. This last example shows that the CSV method may help to estimate the
parameters of the classical SV by simple graphical procedures.

4.8.2.5 Gaussian CSV

The Gaussian classical SV corresponds to the CSV model,
 , 
2π
γc (d) = α d − φ (d, β) , (4.26)
β

where φ(d, β) is the area under the normal probability density function (with zero
mean and variance 1/β) from 0 to d. Obviously, α can be estimated as the slope of
this straight line. That is

* !d  2 
β d β
φ(d, β) = exp − dd. (4.27)
2π 2
0

Last but not the least, it is also possible to have different types of CSV mod-
els along different directions for the same earth sciences phenomenon. In such a
situation, there is structural heterogeneity within the phenomenon.
Various theoretical CSV models are fitted to the sample CSV by simple least
squares technique. A weighted or generalized least-squares approach would proba-
bly be preferable because the sample CSV values are correlated and do not have
equal variance. Future researches should be directed toward how to implement
4.9 Point Cumulative Semivariogram 175

a weighted or generalized least squares approach in particular, what should the

weights be, and how strong are the correlations between neighboring CSV values?

4.9 Point Cumulative Semivariogram

The PCSV function is proposed by Şen and Habib (1998) as the CSV calculated
for a single point (site). The PCSV identifies the spatial variability of earth sciences
ReV around a single site rather than the whole region. It presents the regional effect
of all the sites within the study area on a particular site. Consequently, the number
of PCSVs is equal to the number of available sites. Each PCSV provides a basis for
nearby station variability interpretations, and their mutual comparisons at different
sites lead to invaluable information for describing the heterogeneity of the ReV in
an area. The treatment of available spatial data at n sites according to the following
steps leads to sample PCSV for a particular site.

1) Calculate the arithmetic average, Z, and standard deviation, Zs , from the avail-
able data. Standardize the data according to the following formulation,

Zi − Z
zi = . (4.28)
ZS

2) Calculate the distances between the desired site and the remaining sites. If there
are m sites, the number of distances is n−1, di (i = 1, 2, . . . , n−1).
3) For each pair calculate the squared differences as (zc –zi )2 where zc and zi are
the ReVs at the concerned and i-th sites, respectively. Consequently, there are
(n–1) squared differences.
4) Rank the distances in ascending order and plot distances di versus correspond-
ing successive cumulative sums of half-squared differences. Hence, a non-
decreasing function is obtained similar to Fig. 4.36, which is named as the sam-
ple PCSV for the desired site.
All these steps imply that the PCSV, γ(dc ) for site C can be expressed as

1
n−1
γ (dc ) = (zD − zi )2 . (4.29)
2
i=1

5) Application of these steps in turn for each site leads to n sample PCSVs of ReV.

These sample PCSVs are the potential information sources in describing the ReV
characteristics around each site. Among these characteristics are the radius of influ-
ence, spatial dependence, and structural behavior of the regionalized variable near
the site such as the nugget (sudden changes) and sill effects, and heterogeneity as
will be explained in the following section.
176 4 Spatial Dependence Measures

Example 4.1
The PCSV proposed in the previous section is applied to seismic data from Turkey
(Erdik et al., 1985). Conventional probabilistic procedures are used to construct
the seismic hazard maps for Turkey. Seismic hazard is defined as the probability
occurrence of ground motion due to an earthquake of a particular site capable of
causing significant loss of value through damage or destruction at a given site within
a definite time period. The PCSV technique is applied to all sites; some of the sample
PCSVs are presented in Fig. 4.44, and after grouping, the major sample trends of
PCSV appeared as shown in Fig. 4.45. Consideration of these figures individually
or comparatively leads to the following significant interpretations, which can be
implemented in any future seismic hazard mapping models of Turkey (Şen, 1997).

1) Individual seismic PCSVs have rather different appearance from each other,
which indicates that the regional seismic distribution in Turkey is heteroge-
neous. It is a necessary and sufficient requirement for areal heterogeneity that
the sample PCSV should exhibit different patterns within the limits of sam-
pling error at different sites. A first glance through the whole sample PCSVs
gave the impression that, in general, there are ten distinctive categories of them
within Turkey. These categories are labeled alphabetically from A to J as shown
in Fig. 4.45. Each of these categories has different features, which reflect the
seismic record behavior around the station concerned. For instance, those sam-
ple PCSVs, which are in category F (see Fig. 4.45), have an initial part with
convex curvature (at small distances) and then either single or multiple broken
straight lines follow at large distances. On the other hand, in category D sample
PCSVs do not have any curvature but many broken straight lines and have an
intercept on the horizontal distance axis. An abundance of broken straight lines
indicates the heterogeneity involved around the station concerned at different
distances.
In category B the sample PCSVs expose single straight line, indicating that
there are rather homogeneous areas of influence around these stations.
The stations within each category can be regarded as homogeneous collec-
tively, and hence it is clear that the whole study area has about ten distinct
homogeneity regions of seismic variation. Further useful interpretations about
the sample PCSVs can be listed as follows:
2) Some of the sample PCSVs does not pass through the origin. This is tanta-
mount to saying that the seismic occurrences at these sites cannot be consid-
ered as regionally smooth processes, but rather the seismic event is under the
control of some local and/or regional geological factors. This further implies
that in the spatial seismic occurrences, uniform conditions do not prevail but
rather complex combination of multitude tectonic events. Last but not the
least, if the sample PCSV has an intercept on the vertical axis, it implies the
existence of a nugget effect in the regional variability at the site concerned
(Fig. 4.45 J).
3) Some of the sample PCSVs have intercepts, R0, on the horizontal (distance)
axis (see categories C, F, and I). This means that at distances less than R0 the
4.9 Point Cumulative Semivariogram 177

40

40
Point CSV

Point CSV
20

20
0

0
0 500 1000 1500 0 500 1000 1500
Distance (km) Distance (km)

40
40
Point CSV

Point CSV
20

20
0

0 500 1000 1500 0 500 1000 1500

Distance (km) Distance (km)
40
40

Point CSV
Point CSV

20
20 0

0 500 1000 1500 0 500 1000 1500

Distance (km) Distance (km)

Fig. 4.44 PCSVs of seismic events in Turkey

sample PCSV value is equal to zero, hence from Eq. (4.29) Zc ∼ = Zi imply-
ing structural control within the regional seismic event. Furthermore, in gen-
eral, big seismic values follow big seismic values and small ones follow small
seismic values, i.e., there are isolated islands of high or low seismic locations
around the site.
178 4 Spatial Dependence Measures

Point CSV

Point CSV
A B
Distance Distance
Point CSV

Point CSV

C D
Distance Distance
Point CSV

Point CSV

E F
Distance Distance
Point CSV

Point CSV

G H
Distance Distance
Point CSV

Point CSV

I J
Distance Distance

Fig. 4.45 Seismic PCSV categories in Turkey

4.9 Point Cumulative Semivariogram 179

4) That each sample PCSV fluctuates about a straight line for large distances.
The existence of straight line portions in the PCSV implies that seismic activ-
ities at large distances are independent from each other. This is equivalent to
the regional sill property in the classical SV, but PCSV provides information
about the sill at individual sites (see Fig. 4.45 J). These portions correspond to
horizontal segments at large distances in classical SV as defined by Matheron
(1963). Furthermore, this is the only range where the classical formulations
keep their validity.
5) A group of sample PCSVs passes through the origin (see category in Fig. A, B,
D, E, G, and H). Such a property on the PCSV diagram implies the continuity
of seismic effects from the site outward. Continuity means that there are no
nugget effects or discontinuities within the regional seismic variable at the site
concerned.
6) Some of the sample PCSVs has “curvature portions” for a moderate range of
distances (categories E and G). In fact, such a range corresponds to the distance
scale as defined in the turbulent flow by Taylor (1915). After this range, the
PCSVs converge to straight lines (Şen, 1989). The initial curvature implies
that the seismic at these sites has regional dependencies, which weaken toward
the end of curvature distance range, (Şen, 1989). Since curvatures are convex,
there are positive regional structural dependence (Şen, 1992). Furthermore, the
curvature implies that the seismic has areal structural dependence.
7) In some of the sample PCSVs, there is no curvature part at all (categories B and
F). Such a situation is valid in cases where the regional seismic distributions
arise predominantly due to the activities of external factors only. Furthermore,
there is no structural correlation, i.e., seismic phenomena evolve randomly over
the region concerned.
8) As suggested by Şen (1992), the sample CSVs help to identify the underlying
generating mechanism of regional phenomenon. Likewise, the sample PCSVs
provide clues about the seismicity-generating mechanism around the site con-
cerned. For instance, if the sample PCSV passes through the origin and has
straight line portions only, then the regional phenomenon concerned complies
with independent (white noise) process with no regional dependence at all.
However, when the sample PCSV is in the form of straight line but does not
pass through the origin, then a moving average process is the underlying gen-
erating mechanism of the regional variability (Şen, 1992). The PCSVs in cate-
gory F have such a property, and therefore it is possible to conclude that moving
average mechanisms are dominant at these sites.
9) In the case of a single straight line following a curved initial portion, the slope
of long-distance straight line portion is related to the regional standard devia-
tion of the underlying precipitation generating mechanism.
10) As mentioned above, the PCSV is an indicator of cumulative similarity of
the seismic varıatıon at a station with other stations. Practically, if the two
PCSV at different sites follow the same pattern within the limits of sampling
errors, then they are said to be similar. Such a similarity implies heterogene-
ity to exist between two sites. They may be partially similar to each other at
some distances if the PCSV values at the same distances are close to each other
180 4 Spatial Dependence Measures

again within the limits of an acceptable sampling error. In practice, in order to

appreciate the distances over which the seismic records measured at different
sites are similar to each other, it is helpful to look at the regional map for fixed
PCSV levels. Herein, for the sake of discussion two PCSV levels at 10 and
20 are considered for construction of such a regional similarity map. Table 4.1
shows the distances obtained from the PCSV graphs at 10 and 20 PCSV values.
Obviously, it is possible to prepare similar maps for any desired study area.

Table 4.1 Similarity measures at fixed level of PCSV

Fixed PCSV Levels

Station Number 10 20

1 180 305
2 300 1110
3 440 940
4 405 1295
5 300 1110
6 295 150
7 500 800
8 540 760
9 595 1120
12 370 600
13 180 435
14 280 435
20 470 1310
21 325 520
22 460 1065
23 220 500
24 200 280
30 580 1460
31 250 500
32 400 1070
33 305 880
34 350 1035
40 360 975
41 95 115
42 450 735
43 405 900
50 370 1165
51 220 470
52 430 650
53 280 420
60 115 295
61 300 420
62 320 1025
63 400 1205
70 705 1050
71 210 395
72 300 750
4.9 Point Cumulative Semivariogram 181

Table 4.1 (continued)

Fixed PCSV Levels

Station Number 10 20

73 275 700
80 100 205
81 405 1110
82 620 800
83 215 500
91 200 645
92 200 500
93 320 1135
100 200 280
101 325 580
102 220 620
103 230 680
104 120 195
105 110 290
106 500 645

The relevant similarity maps are shown in Figs. 4.46 and 4.47. For a fixed PCSV
level the smaller is the distance the more the seismic activity effect at the site and the
smaller is the regional dependence, i.e., the location of the site has relatively intense
seismic occurrences than the other sites or the regions. For instance, in Fig. 4.46 the
map shows intense seismic occurrences in the eastern part of Turkey, surrounded by
the 120 contour line. The next most intensive seismic varıatıons are observed in the
central, northeast, and western portions of the country where contour lines of 280
occur. However, the least sensitive locations are in the southern parts, with similarity
contours of about 600.
On the other hand, at the higher level of similarity as presented in Fig. 4.47 the
study area seems more heterogeneous, but major distinctive regional zones as in
Fig. 4.46 remain the same.
Principles of PCSV have been explained and applied to seismic data in Turkey.
This type of PCSVs provides detailed information about a regional variable at and
near the measurement sites as well as among the sites. The main purpose of the CSV
technique is to check the heterogeneity of the regionalized variable. If the empirical
point CSVs at different sites have similar patterns within a certain error band such as
5–10%, then the regionalized variable is homogeneous and otherwise heterogene-
ity exists. The point CS concept brings an additional new concept, which provides
an opportunity to make spatial variability interpretations at each site rather than
regionally. Interpretations of relevant PCSV at any site provide useful information
concerning the smoothness, structural control, regional dependence, continuity, and
the radius of influence. The PCSV methodology proposed in this paper is applied
to the scattered seismic varıatıon over Turkey. Finally, a similarity map is obtained,
which provides a basis for the regional heterogeneity assessments.
182 4 Spatial Dependence Measures

35° N

35° N
25° E 45° E

Fig. 4.46 Similarity map at 10 PCSV value

35∞ N

35∞ N
25∞ E 45∞ E

Fig. 4.47 Similarity map at 20 PCSV value

4.10 Spatial Dependence Function 183

4.10 Spatial Dependence Function

As already mentioned, calculation of a sample CSV (or PCSV) leads to a non-
decreasing function with distance. It is said in Section 2.2 that all the classical
weighting functions appear as a non-increasing function with distance. It is, there-
fore, logical to execute the following steps in order to obtain a valid and stan-
dard weighting function from the sample CSV similar to the classical weighting
functions.

1) Depict on the sample CSV the maximum distance, RM , and corresponding

sample CSV value, VM . RM corresponds to the distance between the two far-
thest station locations in any study area (see Fig. 4.36).
2) Divide all the distances (CSV values) by RM (by VM ) and the result appears as
a scaled form of the sample CSV within limits of 0 and 1 on both axes. This
shows the change of dimensionless CSV with dimensionless distance.
3) Subtraction of the dimensionless CSV values from the maximum value of one
appears as a non-increasing function of the dimensionless distance as shown in
Fig. 4.48, which has similar pattern to all the classical weighting functions as
explained in the previous section. This function is referred to as the standard
dependence function (SDF).

Example 4.2
In order to explain the experimental CSV and thereof derived SDFs, the monthly
rainfall amounts at a set of stations are considered, each with at least 30 years of
records in the northwestern part of Turkey (Şen, 1997). Some of the CSVs are
presented in Fig. 4.49 with corresponding classical and calculated SDF values in
Fig. 4.50.

1.0

0.8

0.6
Weighting

0.4

0.2

0.0 0 0.2 0.4 0.6 0.8 1.0

Fig. 4.48 A representative
SDF Standard distance
184 4 Spatial Dependence Measures

4000 1500
January February

1000
CSV (mm2)

CSV (mm2)
2000
500

0 0
0 100 200 0 100 200
Distance (km) Distance (km)

8000 3000
March
April
CSV (mm2)

CSV (mm2)
4000 2000

2000 1000

0 0
0 100 200 0 100 200
Distance (km) Distance (km)

1000 1500
May June
CSV (mm2)

CSV (mm2)

1000
500
500

0 0
0 100 200 0 100 200
Distance (km) Distance (km)

1500 2000
July August
CSV (mm2)

CSV (mm2)

1000
1000
500

0 0
0 100 200 0 100 200
Distance (km) Distance (km)

Fig. 4.49 Monthly CSVs

4.10 Spatial Dependence Function 185

3000 8000
September October

2000
CSV (mm2)

4000

1000

0 0
0 100 200 0 100 200
Distance (km) Distance (km)

3000 1000
November December
CSV (mm2)

CSV (mm2)
2000
500

1000

0 0
0 100 200 0 100 200
Distance (km) Distance (km)

Fig. 4.49 (continued)

It is possible to conclude from Fig. 4.49 that

1) They all reach almost horizontal CSV value at large distances, which mean that
after a certain distance there is no regional effect of one station on other stations’
rainfall amount. This distance corresponds to R in Eq. (3.24, 3.25, and 3.26).
2) Initially, all the CSV’s have an intercept on the horizontal distance axis at about
5 km. This corresponds to almost the smallest distance between the stations.
3) The smallest and the greatest CSV values at large distances occur during July
and October, respectively, which are the transition months in this region from
the Black Sea to the Mediterranean Sea climate in July and vice versa in
October.

On the other hand, Fig. 4.50 includes the geometric weighting functions already
given in Fig. 3.31 for the sake of comparison. The following interpretations and
conclusions can be drawn from these figures for each month.

1) In January, the experimental CSV weighting function does not conform by any
of the classical models. Initially, at small distances it is above all the models and
then becomes closer to exponential model, but only up to about 0.30 dimension-
less distances, deviating from it thereafter.
186 4 Spatial Dependence Measures

January February
1.0 1.0
Eq. (3.21) Eq. (3.21)

Eq. (3.21) Eq. (3.21)

CSV
CSV

0.5 0.5

Eq. (3.25) Eq. (3.25)

0 0
0 0.5 1.0 0 0.5 1.0
Distance Distance
March April
1.0 1.0
Eq. (3.21) Eq. (3.21)

Eq. (3.21) Eq. (3.21)

CSV
CSV

0.5 0.5

Eq. (3.25) Eq. (3.25)

0 0
0 0.5 1.0 0 0.5 1.0
Distance Distance
May June
1.0 1.0
Eq. (3.21) Eq. (3.21)

Eq. (3.21) Eq. (3.21)

CSV

CSV

0.5 0.5

Eq. (3.25) Eq. (3.25)

0 0
0 0.5 1.0 0 0.5 1.0
Distance Distance
July Agust
1.0 1.0
Eq. (3.21) Eq. (3.21)

Eq. (3.21) Eq. (3.21)

CSV

0.5
CSV

0.5

Eq. (3.25) Eq. (3.25)

0
0 0 0.5 1.0
0 0.5 1.0
Distance Distance

Fig. 4.50 Monthly weighting functions

4.10 Spatial Dependence Function 187

September October
1.0 1.0
Eq. (3.21) Eq. (3.21)

Eq. (3.21) Eq. (3.21)

CSV
CSV

0.5 0.5

Eq. (3.25)
Eq. (3.25)
0 0
0 0.5 1.0 0 0.5 1.0
Distance
Distance
November November
1.0 1.0
Eq. (3.21) Eq. (3.21)

Eq. (3.21)
Eq. (3.21) CSV
0.5
CSV

0.5

Eq. (3.25) Eq. (3.25)

0 0
0 0.5 1.0 0 0.5 1.0
Distance Distance

Fig. 4.50 (continued)

2) In February, perhaps initial portion confirms with the exponential model, but
later becomes closer to the power model.
3) March experimental CSV follows the power law of the weighting functions at
almost every distance.
4) Similar pattern to March repeats itself in April, but in May although initially
there is a portion abiding with the power model, then it converts to exponential
model. In June, the experimental CSV weighting function comes even closer to
the exponential weighting function.
5) During September very small and big distances come closer to the ratio and
power models, respectively, but for moderate distance it has an exponential
form. So the mixture of three models appears as a hybrid model in represent-
ing the regional variability. Such a phenomenon cannot be estimated only by
considering one of the classically available models. In the remaining months,
similar interpretations are valid.

These discussions indicate that classical geometric weighting models do not have
full justification for the whole of the meteorological phenomenon, but they are good
first approximations. They cannot be valid for the whole regional variability in any
study area.
188 4 Spatial Dependence Measures

Example 4.3 For the implementation of the SDF and spatial estimation method-
ology, the iron (Fe) percentage (%) concentration dataset used by Clark (1979, p.
95, Table 4.5) is adopted. Although there are 50 sampling points, in order to make
comparison with fuzzy clustering results by Pham (1997) only 21 sites are given
in Table 4.2, because he considered only 21 sampling points from Clark’s table.
Figure 4.51 shows the locations of the sites within the study area.
It is obvious that there is a rather uniform distribution of these sites in a repre-
sentative manner over the whole study area.
The CSV is given in Fig. 4.52. The maximum CSV corresponding to 350
m is 4,250, and accordingly the SDF is obtained and the results are presented
in Fig. 4.53.
Similar to all the regional estimation procedures, weighted average formulation
as in Eq. (2.27) is used together with the weights obtained from SDF in Fig. 4.53.
In order to assess the validity of the proposed weighted average procedure, a
cross-validation technique is used. According to this a data value at one site is sup-
posed to be unknown and is removed from the data set. This removed value is then
estimated with the remaining set of data and by using the SDF together with Eq.
(2.27). This procedure is repeated for all the sites, knowing that a data removed for
estimation at its location is put again in the set for the estimation at another location.

Table 4.2 Measured and estimated Fe % concentrations and errors

Sample No. Measured Estimate Relative error (%)

(1) (2) (3) (4)

1 35.5 32.85 7.45

2 29.4 32.78 11.51
3 36.8 33.72 8.36
4 33.7 34.15 1.32
5 35.3 33.33 5.58
6 32.5 32.56 0.19
7 30.6 33.88 10.73
8 30.1 33.52 11.37
9 40.1 33.76 15.81
10 31.6 33.84 7.10
11 34.8 33.32 4.24
12 28.6 32.76 14.55
13 41.5 33.48 19.32
14 33.2 33.51 0.92
15 34.3 33.56 2.31
16 31.0 33.82 9.09
17 29.6 32.50 9.78
18 40.4 33.70 16.59
19 28.5 34.08 19.59
20 24.4 34.55 41.62
21 39.5 32.91 16.69
Average 33.40 33.46 11.15
4.10 Spatial Dependence Function 189

Location of the sites

180

150
Northing (km)

120

90

60

30

0
0 100 200 300 400 500
Easting (km)

Fig. 4.51 Fe % concentration locations

4500

4000

3500
SEMIVARIOGRAM

3000
CUMULATIVE

2500

2000

1500

1000

500

0
0 50 100 150 200 250 300 350 400 450
DISTANCE

Fig. 4.52 Fe % concentration sample CSV

There are two procedures in the estimation of the site Fe % concentrations. In the
first one, all the other sites are considered for their simultaneous contributions, and
therefore in the estimation of any site’s Fe % concentration, all the distances from
this site to others (n−1 sites) are measured from the map in Fig. 4.51. Subsequently,
these distances are entered into Fig. 4.53 on the horizontal axis and the correspond-
ing RDF weights are found from the vertical axis for each distance. In this manner,
all the sites are treated equally by the same procedure and hence, instead of mea-
sured values, their estimations through the SDF and the cross-validation procedure
are calculated. Column (3) in Table 4.2 shows estimated Fe % concentrations and
their corresponding relative errors are calculated as the ratio of the absolute differ-
ence between the measured and the estimated values divided by the measured value
multiplied by 100, which is shown in column (4).
190 4 Spatial Dependence Measures

1,2

1
1- DIMENSIONLESS CSV

0,8

0,6

0,4

0,2

0
0 50 100 150 200 250 300 350 400 450
DISTANCE (KM)

Fig. 4.53 Fe % concentration (a) dimensionless CSV, (b) SDF

|measured – estimated|
Relative error = 100 ∗ .
measured

It is obvious that for almost half of the sites, the relative error is more than 10%,
which indicates the unsuitability of the so far proposed procedure. For extreme Fe
% concentration sites, the relative errors are very high. However, the averages of
measured and estimated values are very close together with 2% relative error. On
this basis, it may be concluded that the proposed procedure yields reasonable values
on the average, but fails to estimate individual site values. This procedure takes into
account the contribution of all the sites in the estimation and disregards the concept
of the radius of influence.
In order to assess visually, the correspondence between the measured and the
estimated values, Fe % concentrations are presented in Fig. 4.54a against the site
number sequency along the horizontal axis. Unfortunately, consideration of all the
sites in this regional estimation procedure by using Eq. (2.27) and SDF function
for weights calculation yields to average estimations, which do not represent the
high and low points satisfactorily as obvious from Fig. 4.54a. On the other hand,
Table 4.3 shows also the results by the inverse distance square and four fuzzy clustral
procedures as suggested by Pham (1997). The comparisons of average relative errors
in the last row of this table indicate that the SDF method has the least relative error
percentage. It is even better than the five-cluster case, which has been stated as the
best solution by Pham (1997).
In order to improve the representativeness of the Fe % concentrations region-
alized variable estimations at sites, herein, an adaptive new technique is suggested,
which does not only estimate the regional value at a site but also provides the number
of the nearest sites that should be considered in the best possible regional estimation.
4.10 Spatial Dependence Function 191

Table 4.3 Fe % concentration estimations by different techniques

Inverse square
Sample No. RDF distance Fuzzy clustering
(1) (2) (3) (4)

Three Four Five Six

1 32.85 33.44 34.16 34.19 34.83 33.10
2 32.78 31.92 33.28 32.39 31.76 33.20
3 33.72 30.87 34.04 30.50 31.28 31.00
4 34.15 30.77 33.96 33.60 30.23 31.15
5 33.33 36.67 33.11 33.53 35.75 35.23
6 32.56 35.61 33.46 33.11 34.10 34.69
7 33.88 33.24 33.97 33.60 33.00 33.22
8 33.52 35.47 33.63 33.22 33.95 37.46
9 33.76 32.62 33.88 32.83 34.23 34.44
10 33.84 33.24 33.38 33.28 33.12 32.38
11 33.32 36.66 33.84 33.04 34.02 34.29
12 32.76 30.50 33.61 32.19 31.63 31.69
13 33.48 29.85 33.93 33.93 31.32 30.17
14 33.51 34.60 34.11 33.82 30.95 32.20
15 33.51 37.35 33.23 32.96 33.54 34.68
16 33.82 33.24 34.34 33.78 33.78 34.06
17 32.50 34.67 33.05 33.25 38.10 35.17
18 33.70 34.04 34.93 34.37 34.47 35.73
19 34.08 34.12 33.78 33.21 34.68 34.98
20 34.55 32.29 31.87 32.06 32.48 32.13
21 32.91 34.06 33.16 33.18 32.74 31.98
Averages 33.45 33.58 33.65 33.14 33.33 33.49
Average relative 11.15 12.66 12.55 12.65 12.01 12.61
errors (%)

Accordingly, the radius of influence is defined as the distance between the estima-
tion site and the far distant site within the adjacent sites that are considered in the
regional estimation procedure. The following steps are necessary for the application
of this adaptive procedure.

(a) Take any site for cross-validation and apply Eq. (2.27) by considering the near-
est site only. Such a selection is redundant and corresponds to the assumption
that, if only the nearest site measurement is considered, then the regional esti-
mation will be equal to the same value. This means that in such a calculation
the radius of influence is the minimum and equal to the distance between the
estimation and the nearest sites.
(b) Consider now two of the nearest sites to the estimation site and apply the
RDF weighting method according to Eq. (2.27). Consideration of two sites will
increase the radius of influence as the distance between the estimation and the
next nearest site and the estimation value will assume weighted value of the
two nearest sites. Since the weights and measurements are positive numbers,
192 4 Spatial Dependence Measures

Measured
43
Fuzzy estimated
IRON-ORE PERCENTAGE

40

37

34

31

28

25

22
0 5 10 15 20 25

SITE NUMBER
a

43 Measured
SDF estimated
40
IRON-ORE PERCENTAGES

37

34

31

28

25

22
0 5 10 15 20 25

SITE NUMBER
b

Fig. 4.54 Measured and estimated Fe % concentration, (a) whole points, (b) adaptive method

the estimated value will be between the measurements at the two nearest sites.
There will be a squared estimation error as the square of the difference between
measured and estimated values.
(c) Repeat the same procedure now with three nearest stations and calculate the
square error likewise. Subsequently, it is possible to continue with 4, 5, . . . ,
(n−1) nearest sites consequently, and for each case to calculate corresponding
4.10 Spatial Dependence Function 193

Table 4.4 Detailed calculation steps for site-14 regional estimation

Number of
nearest sites Estimation Square error
(1) (2) (3)

2 31.94 1.58
3 29.46 13.96
4 30.75 6.00
5 32.42 0.60
6 32.21 0.97
7 32.38 0.67
8 32.82 0.14
9 32.59 0.37
10 32.72 0.23
11 33.21 0.00
12 33.12 0.01
13 33.48 0.08
14 33.80 0.36
15 33.69 0.24
16 33.63 0.18
17 33.57 0.13
18 33.55 0.12
19 33.48 0.08
20 33.51 0.09

square error. The first one with the least square error yields the number of near-
est sites for the best regional Fe % concentration estimation. The distance of
the farthest site in such a situation corresponds to the radius of influence. As an
example, herein, only site-14 calculations are presented in detail in Table 4.4. It
is obvious that when Eq. (2.27) is applied by considering 11 nearest sites, the
estimation error square becomes the least with the radius of influence equal to
127.47 m.

Application of the above adaptive procedure to Fe % concentrations results in

the estimation values, number of the nearest sites with the minimum squared error,
and the radius of influence, which are presented in Table 4.5. Notice that this table
includes the complete list of samples as given by Clark (1979, p. 95, Table 4.5). This
table yields the following points:

1) The adaptive estimation procedure gives average iron ore concentration value
similar to average measurements with 2.31% error. Hence, similar to all
the previous methods, adaptive estimation also gives reasonable average
values.
2) Comparison of average relative error in Table 4.4 with average relative errors in
Table 4.4 shows clearly that the adaptive method with 4.74% error is the best
among all approaches, and the reduction in the relative error implies that devia-
tions from average level are taken into account effectively. Figure 4.49b presents
the adaptive estimations together with the measured values for Pham (1997) data
194 4 Spatial Dependence Measures

set. If Fig. 4.49a and b are compared, it is then obvious that deviations are better
accounted by the adaptive method.
3) The adaptive approach provides the radius of influence for each station as shown
in the last column of Table 4.5. The average radius of influence is about 88.5 m
with maximum and minimum values of about 223 and 13 m, respectively.

By making use of the radius of influence from Table 4.5, it is possible to construct
equal radii regional map as shown in Fig. 4.55.

Table 4.5 Estimations and radius of influence

Relative error Number of Radius of

Easting Northing % Fe Estimation (%) nearest sites influence (m)
(1) (2) (3) (4) (5) (6) (7)

0 170 34.30 32.15 6.27 3 13.25

10 40 35.50 33.10 6.76 10 199.06
15 135 28.60 30.50 6.64 3 41.23
55 145 29.40 29.10 1.02 2 41.23
125 20 41.50 38.65 6.87 2 58.31
175 50 36.80 36.91 0.30 5 69.46
120 180 33.40 31.52 5.63 6 73.16
160 175 36.00 33.27 7.58 5 97.84
240 184 30.20 31.30 3.64 4 86.17
260 115 33.20 33.21 0.03 11 127.47
235 15 33.70 33.84 0.42 3 69.46
365 60 34.30 34.26 0.12 4 94.34
285 110 35.30 35.83 1.50 3 47.17
345 115 31.00 33.17 7.00 8 127.47
335 170 27.40 28.12 2.63 6 65.13
325 195 33.90 34.75 2.51 6 123.10
350 235 37.60 38.02 1.12 5 73.10
290 230 39.90 41.15 3.13 4 68.17
10 390 27.20 29.12 7.06 3 55.64
85 380 34.20 32.75 4.24 3 112.51
50 270 30.20 31.60 4.64 6 78.32
200 280 30.40 31.06 2.17 4 85.76
400 355 39.90 41.73 4.59 5 67.23
360 335 40.00 42.16 5.40 5 72.15
335 310 40.60 41.18 1.43 4 83.72
5 195 33.90 34.73 2.45 3 198.12
20 105 32.50 32.50 0.00 9 164.50
25 155 29.60 30.14 1.82 3 50.25
50 40 30.60 33.82 10.52 14 222.99
155 15 40.40 39.50 2.23 3 52.50
145 125 30.10 31.63 5.08 3 92.19
130 185 35.30 33.18 6.01 3 187.15
175 185 41.40 39.17 5.39 3 59.37
220 90 28.50 28.82 1.12 2 51.48
205 0 40.10 37.96 5.34 4 82.46
265 65 24.40 32.34 32.54 3 51.48
390 65 31.60 32.73 3.58 2 67.29
4.10 Spatial Dependence Function 195

Table 4.5 (continued)

Relative error Number of Radius of

Easting Northing % Fe Estimation (%) nearest sites influence (m)
(1) (2) (3) (4) (5) (6) (7)

325 105 39.50 33.89 14.20 4 60.21

0 170 34.30 32.15 6.27 3 13.25
310 150 34.80 34.77 0.09 4 61.03
385 165 29.90 28.16 5.82 5 97.25
325 220 37.80 36.08 4.55 6 75.34
375 215 29.80 26.18 12.15 5 56.76
200 230 37.40 35.83 4.20 4 82.63
55 375 27.40 28.17 2.81 3 62.75
395 245 36.50 35.08 3.89 5 192.50
165 355 40.80 39.15 4.04 5 55.25
270 285 32.90 34.75 5.62 4 155.50
365 340 40.00 40.80 2.00 5 86.13
330 320 44.10 45.83 3.92 5 69.12
330 290 41.40 37.16 10.24 4 60.73
Averages 34.50 34.42 4.74 5 88.51

350

300

250
Northing

200

150

100

50

0
0 50 100 150 200 250 300 350 400
Easting

0 50 100 150 200

Fig. 4.55 Iron ore radius of influence map (m)

196 4 Spatial Dependence Measures

Table 4.6 Regional interpolation

Radius of Estimation Fe
Site No. Easting Northing influence (m) (%)
(1) (2) (3) (4) (5)

1 36 136 63 30.02
2 27 90 172 32.71
3 86 116 97 30.13
4 45 29 200 33.81
5 95 47 118 34.88
6 186 23 68 38.49
7 218 58 50 32.63
8 222 129 95 31.77
9 268 98 80 32.35
10 327 134 93 33.47
11 272 43 58 28.99
12 340 40 80 32.30
13 331 65 77 32.78
14 327 87 70 33.08
15 368 96 103 34.37

From this map one can know the relevant radius of influence for any desired point
within the study area. Once this radius is determined, then a circle with the center
at the prediction point is drawn. The measurement sites within this circle are taken
into consideration in the application of Eq. (2.27) for regional estimation through
the SDF weights.

15 sites
1,2
21 sites

0,8
RDF

0,6

0,4

0,2

0
0 0,2 0,4 0,6 0,8 1 1,2
Dimensionles distance

Fig. 4.56 Measured and estimated SDFs

4.10 Spatial Dependence Function 197

After having completed the cross-validation procedure and the map of the radius
of influence, it is now time to present spatial interpolation procedure with RDF
usage as follows.

1) Select any number, say 15, of spatially scattered points within the study area as
shown in Fig. 4.51. These sites are the locations without measurements. For the
sake of argument, they are selected rather arbitrarily with easting and northing
coordinates as shown in Table 4.6.
2) Radius of influence of each site is determined from the map in Fig. 4.55 and
written in the forth column of Table 4.6.
3) Consideration of the radius of influence for each site defines the number of
measurement sites within this radius, which are basis for the Fe% concentration

42
SDF without radius of Inverse distance square
42
influence
38
38
Observed
Observed

34
34

30 30

26 Measurements 26 Measurements

26 30 34 38 42 26 30 34 38 42
42
42
Fuzzy clustering with 3 Fuzzy clustering with 4
38 groups 38 groups
Observed
Observed

34 34

30 30

26 Measurements 26 Measurements
26 30 34 38 26 30 34 38 42

42 Fuzzy clustering with 5 Fuzzy clustering with 6

42
groups groups
38 38
Observed

Observed

34 34

30 30

26 Measurements 26 Measurements

26 30 34 38 42 26 30 34 38 42

Fig. 4.57 Comparative model verifications

198 4 Spatial Dependence Measures

estimation through Eq. (2.27) at this site. Hence, the measurement sites that will
be considered in the spatial interpretation of the Fe % concentration at the site
are identified.
4) Subsequently, distances between the interpolation site and the effective mea-
surement sites are calculated.
5) The entrance of these distances on the horizontal axis in the SDF (Fig. 4.49b)
yields the weights on the vertical axis.
6) Substitution of all the relevant values into Eq. (2.27) provides the Fe % con-
centration value estimations at each site, which are shown in the last column of
Table 4.6.
7) In order to check the reliability of the estimations, the question is now whether
these spatial estimations will yield almost the same SDF or not. For this purpose,
the SDF calculation steps are applied to data in Table 5.6.
8) Figure 4.56 indicates the resulting SDF for the measured and estimated Fe %
concentrations. The maximum relative difference between these two SDFs is
less than 5%, which confirms the practical validity of the SDF adaptive estima-
tion procedure methodology as proposed in this paper.

It is now time to compare all the methods with the measured data on a Cartesian
coordinate system with the measured values on the horizontal axes and the model
outputs on the vertical axes as in Figs. 4.57 and 4.58.

42

40

38
RDF with radius of influence

36

34

32

30

28

26

24
24 26 28 30 32 34 36 38 40 42
Measured

Fig. 4.58 SDF model verification

References 199

References
Aboufirassi, M., and Marino, M. A., 1984. A geostatistically based approach to the iden-
tification of aquifer transmissivities in Yolo Basin, California. Math. Geol. 16(26),
125–137.
Alfano, M., 1984. Statistical inference of the semivariogram and the quadratic model. In: G. Verly,
and others (Eds.), Geostatistics for Natural Resources Characterization. Reidel, Dordrecht,
Holland, pp. 45–53.
Barchet, W. R., and Davis, W. E., 1983. Estimating long-term mean winds from short-term wind
data. Technical Report Number PNL-4785, Pacific Northwest Lab. Richland, WA (USA),
31 pp.
Barnes, R. J., 1991. The variogram sill and the sample variance. Math. Geol. 23(4), 673–678.
Barnes, S. L., 1964. A technique for maximizing details in numerical weather map analysis. J.
Appl. Meteor. 3, 396–409.
Barros, V. R., and Estevan, E. A., 1983. On the evaluation of wind power from short wind records.
J. Appl. Meteor. 22, 1116–1123.
Benjamin, J. R., and Cornell, C. A., 1970. Probability Statistics and Decision Making in Civil
Engineering. McGraw-Hill Book Inc., New York.
Bergman, K., 1979. Multivariate analysis of temperatures and winds using optimum interpolation.
Monthly Weather Rev. 107, 1432–1444.
Bergthorsson, P., and Döös, B. R., 1955. Numerical weather map analysis. Tellus 7, 329–340.
Box, G. E. P., and Jenkins, G. M., 1976. Time Series Analysis, Control and Forecasting. Holden
Day, San Francisco.
Bratseth, A. M., 1986. Statistical interpretation by means of successive corrections. Tellus 38A,
438–447.
Buzzi, A., Gomis, D., Pedder, M. A., and Alonsa, S., 1991. A method to reduce the adverse impact
that inhomogeneous station distributions have on spatial interpolations. Mon. Wea. Rev. 119,
2465–2491.
Carr, J. R., Bailey, R. E., and Deng, E. D., 1985. Use of indicator variograms for enhanced spatial
analysis. Math. Geol. 17(8), 797–812.
Clark, I., 1979. The semivariogram – Part 1. Eng. Min. J. 180(7), 90–94.
Cooley, R. L., 1979. A method of estimating parameters and assessing reliability for models of
steady state groundwater flow, 2, Applications of statistical analysis. Water Resour. Res. 15,
603–617.
Cressie, N. A. C., 1993. Statistics for Spatial Data (revised edition). Wiley, New York, 900 pp.
Cressman, G. P., 1959. An operational objective analysis system. Mon. Wea. Rev. 87, 367–374.
Daley, R., 1991. Atmospheric Data Analysis. Cambridge University Press, Cambridge, 475 pp.
David, M., 1977. Geostatistical Ore Reserve Estimation. Elsevier, New York, 340 pp.
Davis, J., 1986. Statistic and data analysis in geology. John Wiley & Sons, Inc., New York, 560 pp.
Dee, D. P., 1995. On-line estimation of error covariance parameters for atmospheric data assimila-
tion. Mon. Wea. Rev. 123, 1128–1145.
DeWijs, H. J., 1972. Method of successive differences applied to mine sampling: Trans. Inst. Min.
Metal. Sect. A, Min. Industry 81, 78–81.
Eddy, A., 1964. The objective analysis of horizontal wind divergence fields. Quart. J. Roy. Meteo-
rol. Soc. 90, 424–440.
Eliassen, A., 1954. Provisional report on calculation of spatial covariance and autocorrelation of the
pressure field, Report No. 5 (Oslo: Videnskaps-Akademiet Institut for Vaer og Klimaforskning).
Erdik, M., Doyuran, V., Akkaş, N., and Gülkan, P., 1985. A probabilistic assessment of the seismic
hazard in Turkey. Tectonophysics 117(3/4), 295–330.
Franke, R., 1988. Statistical interpretation by iteration. Mon. Wea. Rev. 116, 961–963.
Gandin, L. S., 1963. Objective Analysis of Meteorological Fields. Leningrad, Gidromet; Translated
from Russian, Jerusalem, Israel Program for Scientific Translations, 1965, 242 pp.
Gilchrist, B., and Cressman, G. P., 1954. An experiment in objective analysis. Tellus 6, 309–318.
200 4 Spatial Dependence Measures

Goodin, W. R., McRea, G. J., and Seinfeld, J. H., 1979. A comparison of interpolation meth-
ods for sparse data: application to wind and concentration fields. J. Appl. Meteor. 18,
761–771.
Gustafsson, N., 1981. A review of methods for objective analysis. In: L. Bengtsson, M. Ghil, and E.
Källén (Eds.), Dynamic Meteorology: Data Assimilation Methods. Springer-Verlag, pp. 17–76.
Hoeksema, R. J., and Kitandis, P. K., 1984. An application of the geostatistical approach to
the inverse problem in two dimensional groundwater modeling. Water Resour. Res. 20(7),
1003–1020.
Hohn, M. E., 1988. Geostatistics and Petroleum Geology. Van Nostrand Reinhold, New York, NY.
Inman, R. L., 1970. Operational objective analysis schemes at the National Severe Storms Forecast
Center. U.S. National Severe Storms Laboratory Tech. Circular 10, Norman, OK, 50 pp.
Journel, A. G., and Huijbregts, C. I., 1978. Mining geostatistics. Academic Press, London, 710 pp.
Journel, A. G., 1985. The deterministic side of geostatistics. Math. Geol. 17(1), 1–15.
Kitanidis, P. K. 1997. Introduction to Geostatistics: Applications in Hydrogeology. Cambridge
University Press, Cambridge, 249 pp.
Koch, S. E., DesJardins, M., and Kocin, P. J. 1983. An iterative Barnes objective map analysis
scheme for use with satellite and conventional data. J. Appl. Meteor. 22, 1487–1503.
Krige, D. G., 1982. Geostatistical case studies of the advantages of log-normal, De Wijsian Kriging
with mean for a base metal mine and a gold mine. Math. Geol. 14(6), 547–555.
Kruger, H. B., 1969a. General and special approaches to the problem of objective analysis of
meteorological variable. Quart. J. Roy. Meteorol. Soc. 95(403), 21–39, January.
Kruger, H. B., 1969b. Objective analysis of pressure height data for the tropics. Monthly Weather
Rev. 94(4), 237–257.
Liebhold, A. M., Halverson, J., and Elmes, G., 1993. Quantitative analysis of the invasion of gypsy
moth in North America. J. Biogeogr. 19, 513–520.
Lorenc, A. C., 1981. A global three-dimensional multivariate statistical interpolation scheme.
Monthly Weather Rev. 109, 701–721.
Matheron, G., 1963. Principles of geostatistics. Econ. Geol. 58, 1246–1266.
Matheron, G., 1965. Les Variables Regionalisees et leur Estimation. Masson, Paris, 306 pp.
Matheron, G., 1971. The theory of regionalized variables and its applications. Ecole de Mines,
Fontainbleau, France.
Myers, D. E., Begovich, C. L., Butz, T. R., and Kane, V. E., 1982. Variogram models for regional
groundwater geochemical data: Math. Geol. 14(6), 629–644.
Pannatier, Y., 1996. VARIOWIN: Software for Spatial Data Analysis in 2D. Springer-Verlag, New
York, NY.
Panofsky, H. A., 1949. Objective weather map analysis. J. Meteor. 6, 386–392.
Pedder, M. A., 1993. Interpolation and filtering of spatial observations using successive correla-
tions and Gaussian filters. Mon. Wea. Rev. 121, 2889–2902.
Perrie, W., and Toulany, B., 1989. Correlations of sea level pressure field for objective analysis.
Mon. Wea. Rev. 117, 1965–1974.
Pham, T. D., 1997. Grade estimation using fuzzy-set algorithms. Math. Geol. 29(2), 291–305.
Powers, R. W., Ramirez, L. F., Redmond, C. D., and Elberg, E. L., 1966. Geology of the Arabian
Peninsula. Sedimentary Geology of Saudi Arabia. U.S. Geol. Survey, Prof Paper 560-D, 1–47,
New York.
Rutherford, I. D., 1976. An operational three-dimensional multivariate statistical objective analysis
scheme. Proceedings of the JOC Study Group, Conference on four-dimensional Data Assimi-
lation, Paris, November 17–21,1975. The GARP Programme on Numerical Experimentation,
Report No. 11, January 1976.
Sashegyi, K. D., Harms, D. E., Madala, R. V., and Raman, S., 1993. Application of the Brat-
seth scheme for the analysis of GALE data using a mesoscale model. Mon. Wea. Rev. 121,
2331–2350.
Schlatter, T. W., 1975. Some experiments with a multivariate statistical objective analysis scheme.
Monthly Weather Rev. 103, 246–257.
References 201

Seaman, R. S., 1988. Some real data tests of the interpolation accuracy of Bratseth’s successive
correction method. Tellus 40A, 173–176.
Şen, Z., 1978. Autorun analysis of hydrologic time series. J. Hydrol. 36, 7585.
Şen, Z., 1989. Cumulative semivariogram models of regionalized variables. Int. J. Math. Geol.
21(3), 891–903.
Şen, Z., 1992. Standard cumulative semivariograms of stationary stochastic processes and regional
correlation: Math. Geol. 24, 417–435.
Şen, Z., 1995. Applied Hydrogeology for Scientists and Engineers. CRC Lewis Publishers, Boca
Raton, FL, 496 pp.
Şen, Z., 1997. Objective Analysis by cumulative semivariogram technique and its application in
Turkey. J.Appl. Meteorol. 36(12), 1712–1724.
Şen, Z., 2002. İstatistik Veri İşleme Yöntemleri (Hidroloji ve Meteoroloji). Su Vakfı yayınları, 243
sayfa.
Şen, Z., and Habib, Z. Z., 1998. Point cumulative semivariogram of areal precipitation in moun-
tainous regions. J. Hydrol. 205, 81–91.
Şen, Z., and Habib Z., 2001. Monthly spatial rainfall correlation functions and interpretations for
Turkey. Hydrol. Sci. J. 46(5), 829–829.
Şen, Z., 2008. Solar Energy Fundamentals and Modeling Techniques. Atmosphere, Environment,
Climate Change and Renewable Energy. Springer Verlag, 276 pp.
Shenfield, L., and Bayer, A. E., 1974. The utilization of an urban air pollution model in air man-
agement. Proc. 15th Meeting of the Expert Panel on Air Pollution Modeling, NATO Committee
on the Challenges to Modern Society, Brussels, Belgium, 35 pp.
Skibin, D., 1984. A simple method for determining the standard deviation of wind direction,
J. Atmos. Oceanic Tech., 1, 101–102.
Subyani, A. M., 1987. Hydrogeology of the Wasia aquifer and its geostatistical modeling. Unpub-
lished M.Sc. Thesis, Faculty of Earth Sciences, King Abdulaziz University, 170 pp.
Taylor, G. I., 1915. Eddy motion in the atmosphere. Phil. Trans. Roy. Soc. A 215, 1.
Thiebaux, H. J., and Pedder, M. A., 1987. Spatial Objective Analysis, Academic Press, London,
299 pp.
Chapter 5
Spatial Modeling

Abstract In general, spatial variability is concerned with different values for any
property, which is measured at a set of irregularly distributed geographic locations
in an area. The aim is to construct a regional model on the basis of measurement
locations with records and then to use this model for regional estimations at any
desired point within the area.
Earth science phenomenon varies both in time and in space; and its sampling is
based on measurement stations’ configuration. In many practical applications mea-
sured data are seldom available at the point of interest, and consequently the only
way to transfer the solar irradiation data from the measurement sites to the esti-
mation point is through regional interpolation techniques using powerful models.
The spatial variability is measured in the most common way through the recorded
solar irradiation time series at individual points. The relative variability between
the stations (the difference of simultaneous values between each pair of stations) is
treated commonly by spatial autocorrelation function, which is used for inter-station
dependence based on a set of restrictions.
Spatial variability is the main feature of regionalized variables, which are very
common in physical sciences. In practical applications, the spatial variation rates of
the phenomenon concerned is of great significance in fields such as solar engineer-
ing, agriculture, remote sensing, and other earth and planetary sciences. A set of
measurement stations during a fixed time interval (hour, day, month, etc.) provides
records of the regionalized variable at irregular sites, and there are few methodolo-
gies to deal with this type of scattered data. There are various difficulties in making
spatial estimations originating not only from the regionalized random behavior but
also from the irregular site configuration.
Optimum interpolation modeling technique is presented for spatio-temporal pre-
diction of regionalized variable (ReV) with application to precipitation data from
Turkey. Kriging method is explained with simple basic concepts and graphics and
then various Kriging application methodologies are explained for the spatial data
modeling. The distinctions between simple, ordinary, universal, and block Krig-
ing alternatives are presented in detail. Finally, triple diagram mapping method-
ology for spatial modeling is presented with applications at five different climate
locations.

Z. Şen, Spatial Modeling Principles in Earth Sciences, 203

DOI 10.1007/978-1-4020-9672-3_5,  C Springer Science+Business Media B.V. 2009
204 5 Spatial Modeling

Keywords Spatial estimation · Optimum interpolation · Spatial correlation

function · Cross-validation · Geostatistics · Kriging · Intrinsic property · Simple
Kriging · Ordinary Kriging · Universal Kriging · Block Kriging · Triple
diagram · Regional rainfall pattern

5.1 General

Modeling is a procedure that helps researchers, planners, politicians, and many

experts alike to make future predictions in time or spatial estimations in a region. It
is an interactive procedure where the natural event records (data), modeler’s expert
view, and prediction (estimation) stages work parallel at times and places in a serial
manner for the best conclusions and numerical result. During a modeling proce-
dure three different worlds are entered in a sequential manner. These are real-world
affairs, interpretation in conceptual world, and mathematical formulation stage in
symbolic world. The real world can be grasped by visual observations, experiments,
and numerical sampling. It is, therefore, very useful to make field trips to the study
area and have verbal (linguistic) information directly from the local authorities and
settlers about the phenomenon and its ReV behaviors. This very first stage includes
verbal and numerical data collection. Verbal information can be in the form of log-
ical premises, which relate the concerned output variable to various input effective
factors. The second stage consists of simple assumptions for the simplification of
phenomenon investigation, which is very complex especially in the earth sciences.
Among such simplifications are the scale of the model, behavior of different input
and output variables such as their isotropy, homogeneity, uniformity, geometrical
shape (rectangle, cylindrical, circular, etc.), geological or atmospheric setup sim-
plifications. Depending on the type of the problem, it is possible first to try and
apply conventional methods such as scatter plots between input and output variables
so as to acquire insight about the internal behavior of the phenomenon, basic his-
tograms, simple regressions, which may provide some further illuminations about
the internal generating mechanism of the ReV functionality. After such basic and
simple information, more representative formulations for the modeling are sought
and questioned by updating basic assumptions and simplifications. Once a satis-
faction is felt, then the formulation is applied with the data available and then the
prediction (estimation) is achieved, which must be checked against the observation
values (cross-validation). Whatever is the complexity, suitability, and verifiability of
the model, there is always a difference between the model output and the measure-
ments. This is the model error and then the model must be trained according to the
desired error level. For instance, if the relative error is not less than desired level, say
5%, then the existing model must be updated either by releasing the basic assump-
tions or by changing the model formulation structure. Hence, the validation of the
model is reiterated according to the previous sequence and this procedure is repeated
until the desired level of error or less is reached. This is tantamount to saying that
the model is verified with available numerical and verbal information, and finally it
5.2 Spatial Estimation of ReV 205

is ready for temporal and/or spatial predictions (estimations) at locations where the
data measurements are not available.
The first spatial model is performed by Student (1907), who summed up the
number of the particles per unit area instead of the analysis of the spatial positions
of particles in a liquid. He divided the 1 mm2 area of a hemacytometer into 400 equal
squares and counted the yeast cells. Later, Fisher (1935) used the spatial analysis in
agricultural area. Yates (1938) searched about the influence of spatial correlation
at the randomization process, where completely random regionalized variables are
considered for modeling purposes.
This chapter is concerned with the spatial prediction of the ReV and two of the
most important procedures, namely the optimum interpolation and Kriging models,
are presented in detail.

5.2 Spatial Estimation of ReV

The spatial interpolation of earth sciences data aims at estimating the ReV value at
a given site based on the nearby observations (measurements) sites (Fig. 5.1). Most
often the input variables are longitudes and latitudes (easting and northing) and the
ReV is one of the earth sciences variables, i.e., triplicate as mentioned in Chapter 2.
This is a problem of operational earth sciences that is regularly encountered in
the spatial estimation procedures. The 2D statistically optimum interpolation mod-
els are useful in the analysis and modeling of ReVs, where spatial dependence func-
tions (SDF) are used for depicting the radius of influence in the case of isotropic
ReVs or area of influence for anisotropic ReVs variability (Chapters 2 and 4). One
of the fundamental, effective, and objective analysis is the transformation of mea-
surements from irregularly distributed sites to regular gridded networks for use in
the numerical prediction schemes or mapping methodology through spatial mod-
els such as the Kriging methodology, which assigns a spatial estimation value to
any point without measurement through a suitable estimation procedure. Hence, in
practice, there are twofold purposes in describing the spatial variability in earth sci-
ence ReVs by objective analysis.

Measurement sites

Grid sites

Radius of influence

Estimation site

Fig. 5.1 Elements in a

spatial estimation
206 5 Spatial Modeling

1) A regular grid of nodes serves as the initial data for numerical forecast models
or mapping methodologies.
2) Coordinates of a rather great number of iso-line points help to construct equal-
value lines (contour lines) such as isohyetal maps.

In any spatial prediction, the problems are with spatial (areal and elevation) mod-
eling and with the transfer of information from available irregular measurement
sites to regular grid nodes or to any desired point of interest. In general, ReV at any
location bears some relationship to nearby locations. The strength of the relation-
ship normally decreases as the distance between the location increases (Chapter 4,
Section 4.10). There are various spatial data modeling techniques for data interpo-
lation from the measurement sites to any desired point as follows (Schlatter, 1988).

1) Surface fitting methods: The first objective analysis method in meteorology was
the surface fitting type devised by Panofsky (1949). In this method, the analysis
value is represented as a continuous mathematical function, which fits irregularly
scattered measurements. Among these methods are the polynomial interpolation
(Panofsky, 1949), orthogonal polynomials (Dixon, 1969, 1976); splines (Fritsch,
1971), and finally spectral approaches (Flattery, 1970).
2) Empirical linear interpolations: The value of any variable at a particular location
is estimated as a weighted sum of observations. Among such interpolation tech-
niques are the iterative successive correction methods (Cressman, 1959; Barnes,
1964), which are already explained in Chapters 2 and 4.
3) Statistical objective analysis: These are estimation methods at any desired point
where spatial correlation (dependence) structure determines the weights appli-
cable to each observation. The major approaches in this category are the optimal
interpolation (Gandin, 1963), the covariance models for atmospheric variable
(Buell, 1958), adaptive filtering (Kalman, 1960; Şen, 1980, 1983), and recently
the CSV method (Şen, 1997).
4) Variational techniques: These include more mathematical abstractions than other
methods and two of them are the incorporation of dynamic constraints (Sasaki,
1958) and the fitting models to data at different times (Ledimet and Talagrand,
1986).
5) Geostatistical approaches: These models are based on the classical SV and Krig-
ing techniques (Matheron, 1963; Clark, 1979; Journel and Huijbregts, 1978).
Various alternatives of SV are presented by different authors (see Chapter 4),
6) Objective modeling: The analysis of ReV produces the optimum solution in the
sense that interpolation error is minimized on the average. This method allows
for the extraction of as much useful information as possible from the measure-
ments. The problems associated with optimum interpolation analysis can be
summarized as follows.

a) It requires knowledge of covariance, which is often not known and thus its
estimate is necessary from the available data. Establishing such an estimate
is often fraught with difficulty as a host of local factors are involved.
5.3 Optimum Interpolation Model 207

b) Essentially, one must determine the priority about which station measure-
ments are significantly correlated with the value at the point of estimation
(interpolation), i.e., one must determine a region of influence around the
interpolation point as in Fig. 5.1.

5.3 Optimum Interpolation Model

A technique commonly used for meteorological analysis is the “optimal interpola-

tion model” (OIM), which is very similar to Kriging method except that some prior
knowledge of climatological (obtained from a 6 or 12 hour numerical forecast) back-
ground fields are assumed. Apparently, the optimum interpolation method was sug-
gested first by Eliassen (1954). However, Gandin (1963) developed the method over
a number of years and his name is associated with it in the literature. Eddy (1964)
was also one of the early developers of this method. Detailed literature review on
this subject is given by Şen and Habib (2000).
In any numerical analysis technique the main idea is that estimation at any point
is considered as a weighted average of the measured values at irregular sites. Hence,
if there are Zoi (i = 1, . . ., n) observation stations with the number of prediction
site k, then the general prediction in the form of weighted average formula becomes
similar to Eq. (2.27) as

n
Wi Zoi
i=1
Zak = n , (5.1)
Wi
i=1

where Zak is the estimate value and Wi ’s show the weightings between the i-th site
and the grid point k (Fig. 5.1). In the literature, all weighting functions that are
proposed by various researchers appear as functions of the distance between the
sites only. The major drawbacks of such weighting functions are already presented
in Chapter 2.
OIM assumes that analyses of data are represented by first guess data plus cor-
rections, which are linear combinations of differences between the first guess and
the observed data. The coefficients of linear combination are determined statistically
so as to minimize the expected square error of data. The coefficients of linear com-
bination are expressed by error covariance matrices of observed and forecast data
(estimated data), when forecast data are used as “first guess.” Thus the covariance
matrix has a great influence on the final estimation. It is assumed that measurements
are spatially correlated and such correlations are calculated as in Chapter 4 (Figs. 4.5
and 4.6). This implies that measurements are close together in clusters, i.e., highly
correlated, and that as they get farther apart they become independent similar to
what has been explained in the SV definition in Chapter 4. Although the method
presented here is a fully 2D version of optimum interpolation, similar approach is
208 5 Spatial Modeling

applicable to 3D or to multivariate problems. In all OIM the following points should

be kept in mind.

1) One of the significant advantages in the use of OIM is the ability to estimate a
variability at any site from the observations at adjacent sites.
2) The interpolation weights depend on the statistical structure of the ReVs mea-
sured as a sequence of time series and not on the individual measurement values
at a given time.
3) The expected analysis errors are produced as a function of the distribution and
accuracy of the data.
4) It is more expensive computationally than other commonly used methods. The
method is computer intensive as the number of computations and amount of
computer storage are concerned.
5) The correlation (dependence) models require a long history of first guess field
for he accurate determination of empirical coefficients.
6) The statistical error functions are estimates, not exact values.

Furthermore, the basic assumptions that are embedded in any OIM can be sum-
marized as follows.

1) The measurements have a spatial dependence, which implies that as long as they
are close together the spatial dependence is high, otherwise they become more
independent.
2) There is no dependence (no correlation) between the measurements and the first
guess field errors.
3) OIM as described here may be applied to any scalar field when the correlation
and error patterns for that field are known.
4) The field to be analyzed is statistically homogeneous and isotropic.

Consider any grid point, k, with n observations, Zoi (i = 1, . . ., n) around it, which
will be used to calculate the analyzed value, Zak , at this grid point. The main idea
g
is that Zak is determined as the sum of a first guess value, Zk , at grid point k plus
a linear combination of the deviations of the observed values from the first guess
g
values (Zoi − Zi ) (see Fig. 5.2). This is to say that

g

n
g
Zak = Zk + Wi (Zoi − Zi ), (5.2)
i=1

where n is the number of stations (in other words, it may also be considered as
the number of influencing stations). The Wi ’s are the interpolation weights; as intu-
itively mentioned above, one would expect the Wi ’s to be positive and decrease
monotonically with increasing distance from the grid point, k.
Assume that the true value, ZTk , can be subtracted from both sides of the above
equation. Usually, the true value is not known, but some knowledge may be known
5.3 Optimum Interpolation Model 209

Fig. 5.2 Location of grid

point and measurement site

T o
Zi Zi
ΔX
g
Zi
Grid point Measurement site
ΔY

g
Zk
a T
Zk Zk

about its statistical parameters, and that will prove useful in the following method.
With these considerations, Eq. (5.2) becomes

g

n
g
(ZTk − Zak ) = (ZTk − Zk ) − Wi (Zoi − Zi ). (5.3)
i=1

The values of the guess fields at each observation location may be determined
by using bilinear interpolation. The difference between each observation and corre-
sponding guess values may be computed as
g g
Zoi − Zi = (Zoi − ZTi ) + (ZTi − Zi ) = Z̃i + Ẑi , (5.4)

where the deviation of the true value ZTi from the first guess value at observed station
is Ẑi (guess error at station i); the deviation of true value from the first guess at grid
point is Ẑk (guess error at grid k); and, finally, Z̃i denotes the deviation of true value
from the observed value at station i (the observational errors at station i), which can
be written as
g
⎫
Ẑi = (ZTi − Zi ) ⎪
⎪
⎬
g
Ẑk = (ZTk − Zk ) (5.5)
⎪
⎪
Z̃ = (Zo − ZT ) ⎭
i i i

By substituting Eqs. (5.4) and (5.5) into Eq. (5.3), one can obtain


n
(ZTk − Zak ) = Ẑk − Wi (Z̃i + Ẑi ). (5.6)
i=1

Furthermore, the interpolation weights Wi are obtained by condition that the mean
square error (MSE) of interpolation is the minimum, which corresponds to the
method of least squares. This condition can be expressed as
210 5 Spatial Modeling

E = (ZTk − Zak )2 = Minimum, (5.7)

where the over bar denotes an ensemble average in the case of a large grid point
number. Insertion of Eq. (5.6) into Eq. (5.7) and evaluation of the square term
leads to
- .2
n n n
E = Ẑk − Wi (Z̃i + Ẑi ) = Ẑ2k + Wi Wj (Z̃i + Ẑi )(Z̃j + Ẑj )
i=1 i=1 j=1
n
−2 Wi (Ẑk Z̃i + Ẑk Ẑi )
i=1 (5.8)
n n
= Ẑ2k + Wi Wj (Z̃i Z̃j + Z̃i Ẑj + Ẑi Z̃j + Ẑi Ẑj )
i=1 j=1
n
−2 Wi (Ẑk Z̃i + Ẑk Ẑi ),
i=1

where i = 1, . . ., n and j = 1, . . ., n. By assuming that there is no dependence (no

correlation) between the observational and the first guess field errors, one can obtain

Z̃i Ẑj = Ẑi Z̃j = Ẑk Z̃i = 0. (5.9)

Consideration of these equations renders Eq. (5.8) to a simplified form as


n 
n 
n
E = Ẑ2k + Wi Wj (Z̃i Z̃j + Ẑi Ẑj ) − 2 Wi (Ẑk Ẑi ), (5.10)
i=1 j=1 i=1

which can be rewritten as follows:


n 
n 
n
E = γkk + Wi Wj (γij + γ̃ij ) − 2 Wi γik . (5.11)
i=1 j=1 i=1

The elements in this last expression are defined as

g
γkk = (ZTk − Zk )2 , (5.12)

which represents the variance of the guess error at the grid point k, and

g g
γij = (ZTi − Zi )(ZTj − Zj ) (5.13)

is the covariance of the guess error at location i and with guess error variance at
location j as

g g
γki = (ZTk − Zk )(ZTi − Zi ), (5.14)
5.3 Optimum Interpolation Model 211

which is the guess error covariance at the grid point k with the guess error at
location i,

γ̃ij = (Zoi − ZTi )(Zoj − ZTj ). (5.15)

This is the covariance of the observation error at the location i with the observation
error at station j.
The statistical interpolation is a minimum variance estimation procedure and
attempts to minimize the expected analysis error variance. The problem is to find
the weights Wi that minimize the variance. Hence, differentiating Eq. (5.11) with
respect to each of the weightsWi and equating the result to zero after necessary
algebraic manipulations lead succinctly to


n
Wi (γij + γ̃ij ) = γkj . (5.16)
i=1

Interpolation of Eq. (5.2) using weights from Eq. (5.16) is called “optimum interpo-
lation.” However, these weights are optimal only if the observation and first guess
error covariance are correct. On the other hand, if the assumed observation and first
guess error covariance are not correct, then Eq. (5.7) is not minimized strictly. In this
case, the interpolation weights in Eq. (5.10) are not optimal and it is called statistical
interpolation.
Multiplication of Eq. (5.16) by Wj and summation for j =1, . . ., n, leads to


n 
n 
n
(γij + γ̃ij )Wi Wj = Wj γkj . (5.17)
i=1 j=1 j=1

Finally, subtraction of this expression from Eq. (5.11) gives the minimum interpo-
lation error as


n
E = γkk − Wi γik . (5.18)
i=1

5.3.1 Data and Application

The region of application is Turkey, with stations as shown in Fig. 5.3. The grid for
this analysis is on a polar stereographic projection oriented along the 35E longitude
line (Habib, 1999). There are 16 columns and 8 rows in the array, for a total of 128
grid points.
The data employed in this study are the mean monthly precipitation records
collected from the statistics published by Turkish State Meteorological Office.
212 5 Spatial Modeling

778
BLACK SEA
667

556 400 Km

300 Km
AEGEAN SEA

445 Sivas
200 Km
334
100 Km
222
0 Km
111

0 MEDITERRANEAN SEA
0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668

Fig. 5.3 Grid points and measurement sites in Turkey

Fifty-two stations for 30-year period from 1960 to 1990 are selected with monthly
rainfall measurements.
The spatial scatter distribution of these measurement sites is given in Fig. 5.4,
which shows an irregular pattern. In general, the transfer of information from the
measurement sites to grid points is necessary for any type of modeling as numeri-
cal solution or mapping the ReVs. Here each site monthly precipitation time series,
which may also be monthly average groundwater levels, groundwater quality val-
ues (calcium, magnesium, sodium, potassium, sulfate, bicarbonate, chloride, nitrate,
total dissolved solids, electric conductivity, etc.) and alike.
In practical applications of the optimum interpolation such as the analysis of rain-
fall, one uses the climatological mean as the first guess value. Hence, the following
expression is considered for the interpolation point value (Habib, 1999).

45
. Measurement
+ Grid points
40
Northing

35

30

25
Fig. 5.4 Spatial distributions 36 37 38 39 40 41 42 43
of the data points Easting
5.3 Optimum Interpolation Model 213


n
Zak = Zk + Wi (Zoi − Zi ), (5.19)
i=1

where Za k and Z0 i are the calculated and observed rainfall values corresponding
to the arithmetic averages of Zk and Zi, respectively, at interpolation point k and
observation stations i = 1, 2, . . ., n, and Wi ’s are the interpolation weights. In order
to calculate these weights, the interpolation formula can be obtained by multiplying
both sides of Eq. (5.19) by (Z0 j – Zj ) and taking the expectation leading to


n
Wi ρij = ρkj , (5.20)
i=1

where ρij is the spatial correlation coefficients between stations i and j, and ρkj
between stations j and k. This equation is valid in the case when E(Zk ) = E(Zk a ),
which implies unbiased estimator. In short, interpolation weights, Wi , are dependent
on the statistical structure of the spatial correlation function (SCF) of the rainfall
records at irregular sites. Once the SCFs are obtained from the available data, then
the value of ρkj can be read from this function depending on the distance between k
and j, and, consequently, the only unknowns in Eq. (5.20) are the weights that can be
calculated from the set of n linear equations. The correlation functions are presented
in the previous chapter (Figs. 4.5 and 4.6). The expected analysis error, εkj , at grid
point k, which results from the introduction by using information at location j, can
be expressed as (Habib, 1999)


n
εk = 1 − Wi ρki = 1 − ρkj . (5.21)
i=1

Most often in practice, 0 < ρkj < 1 and therefore 0 ≤ εkj ≤1. It is obvious that
the expected error does not depend directly on the observed values but again
on the spatial statistical structure of the rainfall amounts. Under the light of the
aforementioned discussion, the following OIM steps are necessary for practical
applications.

1) Specify the geographical locations (longitude and latitude) of interpolation

points.
2) Specify the estimation locations.
3) Compute a first guess field value (climatologic means, i.e., arithmetic averages)
for each station locations.
4) Compute the background error correlations, which correspond to the differences
between observed and average values.
5) Find a suitable model for the background error SCF.
6) Select the measurement sites that will influence the interpolation point.
7) Solve the system in Eq. (5.20) and obtain the interpolation weights.
214 5 Spatial Modeling

START

PROCESS OBSERVATION AND BACKGROUND DATA

COMPUTE AND MODEL BACKGROUND ERROR CORRELATION

CHOOSE THE OBSERVED STATION THAT

IS USED TO CALCULATE ERROR VALUES

SOLVE THE SYSTEM OF EQUATIONS (3)

TO OBTAIN THE INTERPOLATION WEIGHT
n
∑ Wi (ρ i = ρ k)
i=1

COMPUTE THE ANALYSIS VALUES AT

GRID POINT BY USING EQUATION (1)
n
– –
Z k = Z k +∑ Wi (Z io – Z i )
a

i=1

CALCULATE THE EXPECTED ANALYSIS

ERROR AT GRID POINT BY USING
EQUATION (7)
n
ε k = 1 – ∑ Wi ρki
i=1

NO COMUTATION COMPLETE
FOR ALL GRID?

YES

END

Fig. 5.5 OIM execution steps

5.3 Optimum Interpolation Model 215

8) Compute the interpolation point value by using Eq. (5.19).

9) Calculate the expected analysis error at the interpolation point by using
Eq. (5.21).
10) Repeat steps (5) to (8) for all desired interpolation points. Figure 5.5 shows the
necessary steps in the execution of a regional modeling procedure according to
OIM approach.

5.3.1.1 Spatial Correlation Function

Although there are several models to represent the correlation by continuous func-
tions, the basic modeling form is adapted as “the negative-exponential” for all
monthly spatial correlation types. The full correlation array is reduced to distance-
interval-averaged values (Fig. 4.6). Hence, most of the heterogeneity is averaged
out in the computed lag-correlation and the computation time is greatly reduced. In
general, the negative-exponential model has the following mathematical expression,

R(d) = a + be−d/c , (5.22)

where R(d) represents the theoretical SCF and d is considered as the distance
between i-th and j-th stations; a, b, and c are the model parameters. These parameters
are determined by fitting a mathematical function to the array of computed corre-
lation versus corresponding distances between measurement sites. Each one of the
monthly average SCF is fitted to this model by employing the ordinary least squares
regression approach. The resulting parameter values for each month are presented
in Table 5.1, in addition to the overall monthly average parameter values.
All the monthly spatial correlation functions (SCF), say for Sivas station (see
Fig. 5.3), are confined between January and July as one can see from Fig. 5.6. It

Table 5.1 Spatial correlation model parameters

Month a b C

January −0 .15553 0.90883 1352.15221

February 0.27266 0.55969 270.50807
March 0.00000 0.85671 575.8087
April 0.12362 0.71309 255.97296
May 0.06599 0.78690 220.55975
June −0.06992 0.62878 506.51055
July 0.11930 0.65853 131.18509
August −0.08718 0.61655 653.76575
September −0.29293 0.95016 909.51275
October 0.11815 0.65535 370.6699
November −0.33159 0.82909 1707.53125
December 0.03030 0.51511 636.63236
Average −0.01650 0.72650 623.2176
216 5 Spatial Modeling

1.00

0.80 1

0.60
Correlation

11

12
0.40 4 2

3
5
7 6

0.20 10

9 8

0.00
0 200 400 600 800 1000
Distance (Km)

Fig. 5.6 Spatial correlation function of Sivas city for different months

is also clear from this figure that summer months have less SCF values than winter
months.
This is due to the fact that in this region cyclonic rainfall types are dominant in
winter whereas convectional type of rather local rainfall occurs in summer season.
Of course, orographic rainfall effects play more effective role in winter season than
summer. However, it is not possible to identify this type of rainfall from the SCFs.
In Fig. 5.6 one can note that the maximum continuity appears in winter months.
This is due to the fact that in this region frontal rainfall types are dominant in winter
whereas conventional type of rather local rainfall occurs in summer season.
At large distances, February has strong persistent SCF, which implies that in this
month there are cyclonic rainfall types because they cover large areas. Similar trends
are also observed in April and May. On the other hand, comparatively steeper SCFs
indicate convectional rainfall types, which appear over rather smaller area during
summer months.
It is possible to show the variation of SDF for any measurement station as the
center of concentric contour. The extreme and average SCFs are shown for Sivas
city as given in Fig. 5.7.
The benefit from these figures is that at any given month of the year and a given
correlation level, say 0.050, it is possible to determine the influence area around the
5.3 Optimum Interpolation Model 217

1.00

0.80

Maximum
0.60
Correlation

Average

0.40

Minimum
0.20

0.00
0 200 400 600 800 1000
Distance (Km)

Fig. 5.7 Sivas correlation contour lines

station. On the average, the minimum correlation value for 200 km (other distances
have similar interpretations) appears in July and the maximum correlation value is
in January. It is also clear from this figure that summer months have less spatial
correlation values than winter months. Table 5.2 shows the values of the percentage
of the variance of observational error, which has a minimum of 14.32 in March and
a maximum of 50.46 in November. Similarly, the correlation of the observed values
varies between a minimum of 0.4975 in November and a maximum of 0.8567 in
March.
It is possible to find the correlation ρ(d) of the true (first guess) values of the
meteorological variables at distance d as

R(d)
ρ(d) = , (5.23)
R(0)

where R(d) and R(0) are the correlations of the observed values at d and at zero
distances, respectively. This equation is used to calculate the correlation of the first
guess (true) values of the ReVs. The negative-exponential model in Eq. (5.22) at
zero distance can be defined as

R(0) = a + b, (5.24)
218 5 Spatial Modeling

Table 5.2 Some pertinent statistics

Month R (0) Percentage of σe2

January 0.7533 24.67

February 0.8323 16.76
March 0.8567 14.32
April 0.8367 16.32
May 0.8528 14.71
June 0.5588 44.11
July 0.7775 22.24
August 0.5293 47.06
September 0.6572 34.27
October 0.7735 22.65
November 0.4975 50.25
December 0.5454 45.46
Average 0.7059 29.40

and, finally, by substituting Eqs. (5.22) and (3.24) into Eq. (5.23) it is possible to
obtain
 
a + be−d/c
ρ(d) = . (5.25)
a+b

This expression can be used for the indirect evaluation of the correlation of the
first guess (true) values of ReVs by means of extrapolating the correlation of the
observed values at zero distance d = 0 as shown in Fig. 5.8, or by using Eq. (5.23)
or Eq. (5.25) for finding the correlation of the true values.

5.3.1.2 Expected Error

The expected analysis error is a by-product of the OIM analysis procedure as already
stated in Eq. (5.18). It is computed at each analysis point and is a function of the
amount, quality, and distribution of the data near each point. If no data are present,
the expected error remains unchanged from its initial value. The calculation and
modeling of errors result in an analysis that is not necessarily optimum (Lorenc,
1981). The expected error is thus a measure of what the analyzer “thinks” the error
is (Schlatter and Branstator, 1987).
Figure 5.9 includes maps of the expected error of rainfall calculated from
Eq. (5.18) for whole months at Sivas station. It is obvious that there are a num-
ber of areas where expected error changes quite rapidly. For instance, there is a
bull’s-eye feature in all months, which is enclosed by the minimum expected error.
There is also a similar area of less circular contours surrounding the region at the
north center of the study area.
At the first glance, it is possible to depict that contours are less densely occur-
ring over terrestrial area during September–February winter period. However, the
5.3 Optimum Interpolation Model 219

ρ (0) 1 ρ (0) = 1

R(0) This line is extrapolated from nonzero (15 km)

distance to zero distance
Correlation

R(D)

0 15
Distance (km)

Fig. 5.8 Estimation of zero intercept value

contours are very dense from March to October. The reason for such a digression
is due to the fact that the rainfall occurrences are comparatively more sporadic (i.e.,
regionally random) during March–October duration. This is also because the con-
vective rainfall occurrences appear almost independently from each other. However,
areal continuity of winter rainfall is a signature of cyclonic weather movements.
Such continuity results in comparatively very small expected error amounts, say
for instance in January (Fig. 5.9), where errors vary between 0.05 and 0.15 over
the Anatolia within Turkey. On the contrary, in July (Fig. 5.9) error band varies
between 0.20 and 0.70. Table 5.3 shows the relation between expected mean square
error (MSE) and months.
One can note that the maximum expected error appears in summer months (for
example, in July 0.3480), because the rainfall has more discontinuous ReV as
explained above. On the other hand, the minimum expected error is in winter months
(for example, in January 0.0193), because the rainfall in this season is areally exten-
sive and more continuous.

5.3.1.3 Data Search and Selection Procedure

In order to find the weights Wi , the inversion of the correlation (covariance) matrix
must be computed at each grid point. The size of the correlation matrix is directly
 778 778 778

667 667 667

220

400Km 400Km 400Km

556 556 556

300Km 300Km 300Km

445 445 445
200Km 200Km 200Km
334 334 334
100Km 100Km 100Km
222 222 222
0Km 0Km 0Km
111 111 111

0 0 0
0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668 0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668 0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668

778 778 778

667 667 667

400Km 400Km 400Km

556 556 556

300Km 300Km 300Km

445 445 445
200Km 200Km 200Km
334 334 334
100Km 100Km 100Km
222 222 222
0Km 0Km 0Km
111 111 111

0 0 0
0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668 0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668 0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668

778 778 778

667 667 667

400Km 556 400Km 400Km

556 556

300Km 300Km 300Km

445 445 445
200Km 200Km 200Km
334 334 334
100Km 100Km 100Km
222 222 222
0Km 0Km 0Km

111 111 111

0 0
0
0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668 0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668 0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668

778 778 778

667 667 667

400Km 400Km 400Km

556 556 556

300Km 300Km 300Km

445 445 445
200Km 200Km 200Km
334 334 334
100Km 100Km 100Km
222 222 222
0Km 0Km 0Km
111 111 111

0 0 0
0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668 0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668 0 111 222 334 445 556 667 778 890 1001 1112 1223 1334 1446 1557 1668
5 Spatial Modeling

Fig. 5.9 Equal correlation lines between Sivas city and other stations
5.3 Optimum Interpolation Model 221

Table 5.3 Monthly expected errors

Month MSE

January 0.0193
February 0.0935
March 0.0834
April 0.1700
May 0.2500
June 0.1000
July 0.3480
August 0.1200
September 0.0657
October 0.0994
November 0.0238
December 0.0463
Average 0.1141

related to the number of observations permitted to influence each grid point. The
choice of a search strategy that controls the stations which are included in the inter-
polation procedure is an important consideration in any approach to OIMs.
The most common approach in choosing the stations that contribute to the inter-
polation is to define a search neighborhood within which all available stations will
be used. Herein, a simple search strategy is adapted using all station within a circular
search neighborhood with a limited radius of influence.
Meleschko and Prigodich (1964) have shown that the interpolation error reaches
a minimum at about six to eight measurements and shows no further improvement
with the inclusion of more sites. In order to fix this idea with the data at hand, the
change of expected MSE is plotted versus the number of neighboring station for
each month. However, it appeared that such graphs are very similar to each other.
It is objectively seen from Fig. 5.10 that on the average the number of influencing

0.032
Mean square of expected error

0.030

0.028

0.026

0.024

0.022

0.020

0.018
0 2 4 6 8 10 12 14 16 18 20
Number of influencing stations

Fig. 5.10 Relationship between the number of influencing stations

222 5 Spatial Modeling

stations does not change significantly above four sites. For more station numbers,
the expected error mean square remains on almost the same minimum level.

5.3.1.4 Cross-Validation of the Model

The accuracy of OIM is investigated with the help of numerical experiments from
the surrounding sites and in comparison to the interpolated values with the mea-
surements, (Gandin, 1965). Similarly, cross-validation techniques allow comparison
of estimation and true values using only the information available in the measure-
ments (Isaaks and Srivastava, 1989). In a cross-validation technique, the interpo-
lation method is tested at the measurement site, and the measurement value at a
particular station is discarded temporarily from the surrounding sites. Once the esti-
mate is calculated, it is possible to compare it with the measurement value that is
initially removed from the station. This procedure is repeated for all available sites,
with the comparison of observed, interpolated, and actual values by using simple
statistical parameters such as means and standard deviations. The statistical param-
eters of the comparison between various estimations methods and of the true values
are summarized in Table 5.4 for January 1984.
In general, the difference between the average estimate from the OIM and the
true average reflects a tendency toward over-estimation or under-estimation. How-
ever, in this case the analysis method has a slight tendency toward over-estimation,
as seen from Table 5.4. The estimations have less variability than measurements.
However, this reduced variability of analysis values is often referred to as smooth-
ing. Therefore, the measurements are less smoothed than the analysis values. A
scatter plot of measurements versus estimations provides additional evidence on
how well an estimation method has been performed. Figure 5.11 shows the scat-
ter plot of measurements versus estimations for January 1984, which indicates that
there is a good relationship between estimations and measurements, especially for
less than 200 mm. This is also supported by the root MSE, which is 29.9, and the
mean percentage error is −9.07%. Furthermore, the square of the correlation, which
is a measure of the variance explained by the model, is equal to 0.76.
In order to see the correspondence matching between measurements and estima-
tions as well as monthly rainfall amounts, the monthly statistical parameter varia-
tions are presented in Figs. 5.12 and 5.13. It is seen that on the basis of averages,

Table 5.4 Univariate statistics of optimum interpolation analysis

Parameters Observation Estimation

n 52 52
Mean 76.6 74.70
Minimum 5 0
Maximum 361 266.9
Range 356 267
Variance 3407.60 2905.61
Std. dev. 58.37 53.90
5.3 Optimum Interpolation Model 223

Fig. 5.11 Scatter plot of 300

observed versus estimated
values for January 1984

Estimated values (mm)

200

100

0
0 100 200 300 400
Observe values (mm)

measurements and estimations are very close to each other with less than 1% error.
Although there are more discrepancies for monthly standard deviations, generally
2.5% but high errors appear in November and December.

100
OBSERVED VALUES
ESTIMATED VALUES
80
RAINFALL (MM)

60

40

20

Fig. 5.12 Arithmetic mean 0

of observed and predicted 1 2 3 4 5 6 7 8 9 10 11 12
monthly values MONTHS
224 5 Spatial Modeling

Fig. 5.13 Standard deviation 100

of observed and predicted OBSERVED VALUES
monthly values ESTIMATED VALUES
80

STANDARD DEVIATION
60

40

20

0
1 2 3 4 5 6 7 8 9 10 11 12
MONTHS

5.4 Geostatistical Analysis

It is a branch of applied statistics with the original purpose of spatial data processing
for estimating changes in ReVs. However, the principles have been applied to a vari-
ety of areas in geology and other scientific disciplines. A unique aspect of geostatis-
tics is the use of ReVs, which are variables that fall between random and completely
deterministic behaviors (Chapter 1). ReVs describe phenomena with geographical
distribution (e.g., elevation of ground surface) and exhibits spatial continuity, which
is not always possible to sample at every location. Therefore, unknown values must
be estimated from the data taken at specific locations. The size, shape, orientation,
and spatial arrangement of the sample locations are termed as the support and influ-
ences the capability to predict the unknown samples. If any of these characteristics
change, then the unknown values also change. The sampling and estimation of ReVs
are done so that a pattern of variation in the phenomenon under investigation can
be mapped, such as a contour map for a geographical region. In addition to spa-
tial and temporal variation in nature, spatial and temporal dependence or continuity
also occurs. Geostatistics include a set of statistical procedures that can be used to
analyze and model spatial relationships in nature. The spatial variation is somewhat
similar to the time variation in some properties; time observations are dependent and
occur in equally spaced time intervals in 1D, whereas the spatial observations are
dependent and occur in 2D or 3D and the distance between points are not necessarily
equal intervals as they are in time variations. In earth sciences, most of the natural
phenomena behave simultaneously in space and time, as soil moisture content, water
level, and precipitation data may consist of long time series at various locations. The
spatio-temporal variation of the natural phenomena implies a significant amount of
5.4 Geostatistical Analysis 225

uncertainty. Moreover, this variation is, in general, non-homogeneous in space and

non-stationary in time.
Matheron (1963) was the first to use the term “geostatistics,” which was defined
by him as “the application of the formalism of random function to the reconnais-
sance and estimation of natural phenomena.” The basis of the method is the ReV
theory and concepts of this theory can be split into two main successive steps.

1) Establish the theoretical basis for expressing the structural properties of a natural
phenomenon in a useful mathematical form as SV or CSV.
2) Provide a particular means for solving various problems of estimation such as
the Kriging methodology, which guarantees a solution to the estimation problem
(Kriging) and deals with the ReV by using the probabilistic theory of random
function.

By now, however, there are a number of excellent books on the subject, includ-
ing both introductory (Clark, 1979) and advanced aspects (David, 1977; Journel
and Huijbregts, 1978). The classical parametric and non-parametric geostatistics
are introduced recently for application in the field of earthquake ground motion
evaluation (Glass, 1978; Carr and Glass, 1984, 1985; Carr et al., 1985).
Almost all variables encountered in the earth sciences can be regarded as ReVs.
For seismic zonation this is an ideal one in describing earthquake ground motions.
Each observation of ground motion can be considered simply as a unique realization
of an Rev, which adequately represents local random behavior tempered by global
attenuation. Furthermore, provided that a valid SV can be developed, Kriging is well
suited for the estimation process to result in data regionalization.

5.4.1 Kriging Technique

Geostatistics, which is also referred as ReV, was basically developed for statistical
study of the natural phenomena in mining field, including other spatial phenomena
in earth sciences (Matheron, 1963; Davis, 1986). The natural phenomena that have
spatial distribution vary from one place to another with apparent continuity struc-
ture. The first step in geostatistics is to study and determine the spatial structure
and dependence from the measurement characteristics of the natural phenomena
(Chapter 4). The spatial structure can be established by SV, CSV, or PCSV as con-
venient. Using the underlining generation structure, it is possible to estimate the
characteristics at unsampled locations and extend findings of the regional behavior
in natural phenomena through Kriging techniques.
In any geostatistical modeling there are two stages as the estimation of SV
(CSV or PCSV) from a given set of regionalized data for the extraction of spatial
dependence and then Kriging methodology, which provides estimates for locations
where no data are available. The SV is the source of information used in Kriging
to achieve optimal weighting functions for mapping. Kriging uses theoretical SV
226 5 Spatial Modeling

in calculating estimates at a set of grid nodes. These Kriging estimates are best-
linear-unbiased-estimates (BLUE) of the ReV at a set of locations, provided that the
surface is stationary and the correct form of the theoretical SV has been determined
(Chapter 4).
In many disciplines such as petroleum exploration, mining, groundwater, and
water pollution analysis, data are available at a set of predetermined spatial locations
(water and oil wells, meteorology stations, etc.). The purpose is to make regional
estimation at any location based on the available data at these locations. It is nec-
essary to often have maps based on a regular grid and the estimations are used to
produce 2D contour maps or 3D surface plots. In theory, only Kriging method of
grid generation can produce better estimates (in the sense of being unbiased and
having minimum error). In practice, the effectiveness of the technique depends on
the correct specification of several parameters that describe the SV and the model
of the drift (regional trend). However, because Kriging is robust, even with a naive
selection of parameters it will do no worse than conventional grid estimation proce-
dures (Chapter 2).
The price that must be paid for optimality in estimation is computational com-
plexity compared to the techniques presented in Chapter 3. A large set of simultane-
ous equations must be solved for every grid node estimated by Kriging. Therefore,
computer run times will be significantly longer if a map is produced by Kriging
rather than by conventional gridding. Kriging can be computationally very intense
but increasingly available in software packages, and it is the best method for many
purposes. In addition, an extensive prior study of the data must be made to test for
stationarity, determine the form of the SV, set the neighborhood size, and select
the proper order of the drift, if it ever exists. These properties are not independent
and because the system is underdetermined, trial-and-error experimentation may be
necessary to determine the best combination. For this reason, to warrant the addi-
tional costs of analysis and processing, Kriging probably should be applied in those
instances where the best possible estimates of the surface are essential, the data are
of reasonably good quality, and the estimates of the error are needed.

5.4.1.1 Intrinsic Property

Mathematically speaking, let x be a point in space and Z(x) the value of the function
at point x. It is usually highly variable and non-continuous, and therefore cannot be
estimated directly. Thus the structure with variation will be studied by examining
its increment. The basic idea of the theory is to consider the function Z(x) as one
realization of random function from the exhaustive ensemble. This is tantamount to
assuming a well-defined and unique sequence of numerical values into a realization
of random process. Only one realization of that process is available and the problem
is to find the characteristics of the ReV in order to make the estimation of unknown
points possible. Obviously, it is not rigorously possible to infer the probability law of
any random function from its single realization, because it has limited finite number
of sample points. Thus, many realizations, Z1 (x), Z2 (x), . . ., Zk (x) of the ReV are
required for identifying the probability law. In order to render the situation into a
5.4 Geostatistical Analysis 227

practically tractable one, the assumptions of spatial homogeneity and stationarity

are needed. An ReV is said to be strictly stationary if any conventional statistical
properties of the medium such as the mean, variance, covariance do not change with
distance. This assumption is seldom encountered in natural phenomena. Therefore,
an ReV is to be first-order stationary when the random variable has the same mean
value irrespective of location. However, in the linear geostatistics, it will be enough
to assume first of all that the weak or second-order stationarity exists. The second-
order stationarity assumption consists of two following conditions.

1) The arithmetic average, m (expected value for any term), value of the ReV is the
same all over the area,

E [Z(x)] = m (5.26)

or equivalently,

E [Z(x) − Z(x + d)] = 0. (5.27)

2) The spatial covariance of the ReV is the same all over the field of interest.

Cov {[Z (x) − m] [Z (x + d) − m]} = Cov (d) , (5.28)

where d is the distance separating between two sites at locations x and x + d.

If it is assumed that for every d this spatial covariance is independent of the
location, then the second-order stationarity is valid. By holding the full second-
order stationarity, the covariance Cov(d) approaches the variance, σZ2 of the random
variable as d → 0,
/ 0
Cov (0) = {[Z (x) − m] [Z (x) − m]} = E [Z (x) − m]2 = Var [Z (x)] = σZ2 .

A second-order stationarity is, however, really justifiable, and a weaker assump-

tion may be adopted instead. The mean value, m, is always unknown, and may not
be constant, so that the variance and covariance cannot be computed directly. It
is, therefore, beneficial to define an alternative statistic, which does not require the
mean value. For instance, successive difference as in Eq. (5.27) may be zero, but
its square is not essentially zero. Considering these characteristics, a consistent set
of assumptions weaker than second-order stationarity have been made and called
by Matheron (1963) as the “intrinsic hypothesis,” which subsumes the following
expressions.
/ 0
Var [Z(x) − Z(x + d)] = E Z (x) − Z (x + d) − E [Z (x) − Z (x + d)]2

Consideration of Eq. (5.27) leads to

228 5 Spatial Modeling

Var [Z(x) − Z(x + d)] = E [Z(x) − Z(x + d)]2 = 2γ (d) , (5.29)

where 2γ (d) is the variogram and γ (d) is the SV or half-variogram. It should be

noted that second-order stationarity implies the intrinsic hypothesis, but the converse
is not true.
In addition, under the hypothesis of second-order-stationarity, the covariance and
SV are two equivalent tools that characterize the autocorrelation between two vari-
ables Z(x+d) and Z(x) separated by d as

γ(d) = Cov(0) − Cov(d). (5.30)

The validity of this expression is true especially if the ReV has a Gaussian (normal)
pdf. Estimation of the SV is preferable to estimation of the covariance, because the
experimental SV does not require a prior estimate of the population mean. Besides,
SV calculation requires a set of measurements with no time variation. Under the
same condition, the relationship between the model autocovariance, ρ(d) and SV is

γ (d) = σZ2 [1 − ρ (d)] . (5.31)

5.5 Geostatistical Estimator (Kriging)

After successful matching of theoretical SV model, one can use Kriging to estimate
what weights should be applied to each surrounding point to estimate the value of
the ReV at an unknown location (e.g., a grid node, see Fig. 5.1). Kriging calculates
the weighted average of neighboring values and assigns it to each grid node. Kriging
can also calculate the error variance for each estimated value, which means that it is
possible to know what the most likely value is at each location and how likely that it
is to be the actual value. In its most basic form, Kriging does not work where there
is a trend in the field (high values on one end and low on another). Universal Kriging
addresses this limitation by first calculating a trend surface, then using Kriging to
estimate the values of the difference between the real surface and the trend surface
(called residuals). The problem of local estimation appears in finding the best esti-
mator of the mean ReV value over a limited domain. The use of Kriging estimations
is advantageous because it is not limited to a simple point estimation of the region
magnitude (Delhomme, 1978; Marsily, 1986).
Kriging is the name given to a local estimation technique. Provided that the basic
assumptions of no trend and a model for the SV (or CSV) are known or determined
in some way from the data, then the Kriging produces always BLUE of the unknown
ReV characteristics. In the case of non-stationarity, several non-bias conditions are
required and this leads to the technique known “universal Kriging” or unbiased
Kriging of order say, n. When groups of variables are correlated, the co-Kriging
technique can be used to estimate any ReV (e.g., response spectra) from the data
available on all the correlated variables (e.g., peak acceleration, velocity, and dis-
placement in addition to the response spectra). Other classes of estimators are also
5.5 Geostatistical Estimator (Kriging) 229

used in practice; for instance, non-linear estimator can be built by prior transforma-
tion of the data. Disjunctive Kriging (Matheron, 1971), Lognornal Kriging (Krige,
1951), and indicator Kriging (Journel, 1983) are examples of non-linear estimators.
All Kriging techniques are based on the simple linear models as

ZE = λ1 x1 + λ2 x2 + ... + λn xn , (5.32)

where ZE is the estimator of the true value at location E, and λi are the weights
allocated to each observation such that


n
λi = 1. (5.33)
i=1

The technique minimizes estimation variables by solving a set of Kriging equa-

tions, which include covariance between the point or volume to be estimated and
the sample points and covariance between each pair of sample points. The weights
calculated through solving the system of equations, which depend upon the size
and shape of the volume to be estimated, the distance and direction of each sam-
ple from the volume to be estimated, the distance between the samples, and the SV
or CSV.
The equations and their full derivation as well as attributes and advantages of
Kriging can be found in numerous publications such as Matheron (1971), David
(1977), Journel and Huijbregts (1978), and Clark (1979). The following points are
some of the characteristics and advantages of Kriging technique.

1) If one has a model for the SV (or CSV), he/she can produce the minimum vari-
ance using the Kriging technique.
2) If the proper models are used for the SV (or CSV) and the system is set up
correctly, then there is always a unique solution to the Kriging system.
3) If one has regular sampling, and hence the same sampling set up at many dif-
ferent positions within the region, it is not necessary to recalculate the Kriging
system each time.
4) Kriging is not limited to a single point estimation of the given magnitude Z, but
can also be used
a) to estimate the mean value, Z, on a given block, e.g., on the mesh of model
or a sub-domain of any shape of watershed,
b) to obtain the estimation variance of magnitude Z, i.e., roughly the confidence
interval of this estimation,
c) to locate the best situation for a new measurement point, e.g., by minimizing
the overall uncertainty in the field under consideration.
5) Kriging is advantageous because it considers the following points explicitly.
a) The number of spatial configuration of observation points within the study
region
230 5 Spatial Modeling

b) The position of the data points within the region of interest

c) The distance between the data points with respect to the area of interest
d) The spectral continuity of the interpolated variable.

For instance, in an earthquake ground motion study, Kriging can be used for
evaluating the earthquake ground motion hazard within the region of interest and
for estimating the tripartite earthquake response spectra at a site of interest. It is
already stated that potential errors in the data collection may lead to nugget effect
(Chapter 4), which becomes evident especially from the sample SV, and its existence
causes smoothing operation through the Kriging and less confidence in individual
data points versus the overall trend of the data. It has the same unit as the SV. There
are two components that give rise to nugget effect, namely variances due to error,
σe2 , and separation, σs2 . The latter variance is a measure of variation that occurs at
separation distances of less than the typical nearest neighbor sample spacing. The
more the random fluctuation of the same data at a given location, the greater is the
error variance and the less is the prediction reliability. Consequently, Kriging tends
to smooth the surface, and therefore it is not a perfect estimator.

5.5.1 Kriging Methodologies and Advantages

Once the degree of spatial dependence has been established, then the SV can be
used to interpolate values for points not measured through the process of Kriging,
which is an interpolation method that uses the sample (empirical, experimental)
SV to weight sample points based on their locations in space relative to the point
value that is to be estimated. Therefore, a first step in Kriging is to fit a theoretical
function to the SV model that describes the theoretical SV. Kriging has many things
in common with traditional point-interpolation methods such as inverse distance
(or square) weighting, triangulation, polygonization, etc. methodologies, which are
explained in Chapter 2. The result of Kriging is an interpolated surface map of ReVs.
There are several types of Kriging methodology.

1) Ordinary Kriging: It is similar to multiple linear regression and interpolates val-

ues based on point estimates.
2) Indicator Kriging: It is used to estimate indicator variables. Indicator Kriging is
simply ordinary Kriging performed on the indicator-transformed data.
3) Punctual Kriging: Both ordinary and indicator Kriging methodologies are forms
of punctual Kriging because they are used to estimate values for exact points
within the sampling unit. Punctual Kriging is the most common method used in
earth sciences. It helps to estimate the value at a point from a set of nearby sam-
ple values using Kriging. The kriged estimate for a point will usually be quite
similar to the kriged estimate for a relatively small block centered on the point,
but the computed Kriging standard deviation will be higher. When a kriged point
happens to coincide with a sample location, the kriged estimate will equal the
sample value.
5.5 Geostatistical Estimator (Kriging) 231

4) Block Kriging: Its interpolations are based on values in a particular finite area.
It is a more accurate and intensive computation that uses point estimates within
a block to derive an average estimate for the block.
5) Co-Kriging is a modification of ordinary Kriging that relies on the fact that many
phenomena are multivariate and that the primary variable of interest is under-
sampled. Co-Kriging estimation is done by minimizing the variance of the esti-
mation error using the cross-correlation (dependence) between several variables.
Estimates are derived for both the primary and the secondary variables. The co-
Kriging technique is a modification of the simpler technique of Kriging. It is
used to merge two variables or more. Estimation of co-Kriging contains a pri-
mary variable of interest and one or more secondary variables. Improvement in
the interpolation of one variable by using other variables is important (David,
1977; Journel and Huijbregts, 1978; Myers et al., 1982). There are two steps in
co-Kriging estimation.
a) Evaluation of the cross-correlation (or co-SVs) between variables to obtain
information about continuity and dependencies.
b) Construction of contour maps for the primary variable.

Seo (1998) used linear co-Kriging to interpret rainfall data from a set of rain
gages and radar information. He concludes that the consistency of the improvement
by gage-radar estimation makes co-Kriging an attractive tool in rainfall estimation.
Martinez (1996) applied co-Kriging to improve the accuracy of evapo-transpiration
estimation over a regular grid by including the effects of topography.
Among other spatial estimation techniques, Kriging methodology has the follow-
ing advantages.

1) It yields BLUE for the ReV modeling in earth sciences.

2) It is best because it minimizes the error variance in the estimate, unbiased
because the weights sum to one, linear because it is a simple weighted average.
3) Also uses a weighted average method to calculate the value at unsampled loca-
tions (Eq. 2.27).
4) Weights are determined by considering the representative sample and then the-
oretical SV (or CSV) and the estimation problem are solved by using matrix
algebra.

In the application of Kriging algorithm, there are four successive essential steps
as follows.

1) When computing the interpolation weights, the algorithm considers the spacing
between the point of estimation and the data sites. The algorithm considers also
the inter-data spacing, which allows for declustering.
2) When computing the interpolation weights, the algorithm considers the inherent
length scale of the data. For example, there are regions where the topography
varies much more slowly than some other regions. If two points’ elevations at
232 5 Spatial Modeling

the same distance are considered in these two different topographies, then in the
slowly changing regional elevation case it would be reasonable to assume a lin-
ear variation between these two observations, while in the other region such an
assumed linear variation would be unrealistic. The algorithm adjusts the inter-
polation weights accordingly.
3) When computing the interpolation weights, the algorithm considers the inherent
trustworthiness of the data. If the data measurements are exceedingly precise and
accurate, the interpolated surface goes through each and every observed value. If
the data measurements are suspect, the interpolated surface may not go through
an observed value, especially if a particular value is in stark disagreement with
neighboring observed values. This is an issue of data repeatability.
4) Natural phenomena are created by physical processes, which have often pre-
ferred orientations. For example, at the mouth of a river the coarse material set-
tles out fastest, while the finer material takes longer to settle. Thus, the closer
one is to the shoreline the coarser the sediments, while the further from the
shoreline the finer the sediments. When computing the interpolation weights,
the algorithm incorporates this natural anisotropy. When interpolating at a point,
an observation 100 m away but in a direction parallel to the shoreline is more
likely to be similar to the value at the interpolation point than is an equidistant
observation in a direction perpendicular to the shoreline.

The last three items incorporate something about the underlying process from
which the observations are taken. The length scale, data repeatability, and anisotropy
are not a function of the data locations. These enter into the Kriging algorithm via
the SV (or CSV). The length scale is given by the SV range (or slope), the data
repeatability is specified by the nugget effect, and the anisotropy is given by the
anisotropy (Chapter 4).

5.6 Simple Kriging

This Kriging method has straightforward mathematical derivations based on the fol-
lowing three basic assumptions.

1) The spatial sampling points are representatives of the ReV at a set of given
locations with measurement values.
2) The ReV is considered as a second-order random field variable with mean, vari-
ance, and SV.
3) The mean of ReV is known, which limits the application of this Kriging model-
ing alternative severely.

In practical applications there are many cases where the areal mean of the ReV
is known and hence direct application of the simple Kriging methodology becomes
5.6 Simple Kriging 233

• • measurement sites
• estimation site
• •
• Zi
•
• • º ZE
•
•
• • Zn
• Z1 • •
•
0 5 10 km

Fig. 5.14 ReV sample sites and estimation site

attractive. Especially, in the case of second-order stationarity (mean and variance

constancy) the whole ReV samples can be standardized according to classical sta-
tistical standardization formulation and hence the standardized ReV has zero mean
and unit variance. Let a set of measurement locations are scattered irregularly over
the study area as in Fig. 5.14, where there are n measurement and one estimation
sites.
The regional variability of the ReV is recognized by a suitable regional depen-
dence (covariance or SV) function between each pair of the measuring sites. Kriging
estimation is considered as the weighted average of the measurement values on the
estimation point with distant-dependent weighting values, λi . The Kriging estima-
tion, ZE , is a linear weighted average of the surrounding measurements Zi (i = 1, 2,
. . ., n) as follows.


n
 
ZE = Z + λi Zi − Z , (5.34)
i=1

where Z indicates the regional constant mean value of the ReV. If there are n neigh-
boring sites for the estimation calculation, then there will be n2 elements in the
covariance (or SV) matrix, with the variances on the main diagonal. Due to the diag-
onal symmetry the number of different elements in the matrix is equal to n(n − 1)/2
as follows.
⎡ ⎤
var(z1 ) cov(z1 ,z2 ) ... cov(z1 ,zn )
⎢cov(z2 ,z1 ) var(z2 ) ... cov(z2 ,zn )⎥
⎢ ⎥
⎢ . . ... . ⎥
C=⎢
⎢
⎥
⎥ (5.35)
⎢ . . ... . ⎥
⎣ . . ... . ⎦
cov(zn ,z1 ) cov(zn ,z2 ) ... var (zn )

In this matrix each element is dependent also on the distance difference (rel-
ative distance) between the two sites. It is obvious that Cov(zi , zj ) = Cov(zj , zi )
and furthermore Cov(zi , zi ) = σi2 , which is the variance at site i. However, if the
234 5 Spatial Modeling

ReV is standardized with constant regional mean, Z and variance, σZ2 then Cov(zi ,
zi ) = 1 and the covariance corresponds to dependence (correlation) coefficient as
ρ(z1 , z2 ) = cov(z1 , z2 ). For standardized ReV, Eq. (5.35) takes the following form.
⎡ ⎤
1 ρ(z1 ,z2 ) ... ρ(z1 ,zn )
⎢ρ(z2 ,z1 ) 1 ... ρ(z2 ,zn )⎥
⎢ ⎥
⎢ . . ... . ⎥
ρ=⎢
⎢ .
⎥ (5.36)
⎢ . ... . ⎥ ⎥
⎣ . . 1 . ⎦
ρ(zn ,z1 ) ρ(zn ,z2 ) ... 1

This is the regional correlation matrix for ReV. Similar to this matrix one can write
also the distance matrix, D, between these n sites, with zero distances along the main
diagonal as follows.
⎡ ⎤
0 dis(z1 ,z2 ) ... dis(z1 ,zn )
⎢dis(z2 ,z1 ) 0 ... dis(z2 ,zn )⎥
⎢ ⎥
⎢ . . ... . ⎥
D=⎢
⎢
⎥
⎥ (5.37)
⎢ . . ... . ⎥
⎣ . . 0 . ⎦
dis(zn ,z1 ) dis(zn ,z2 ) ... 0

This matrix is also symmetrical with respect to the main diagonal. Hence, both the
covariance and the distance matrices provide information in the form of, say, upper
triangular matrices. The plot of distance matrix values on the horizontal axis versus
corresponding regional correlation coefficients from the correlation matrix provides
a general shape as in Fig. 5.15, which may be referred to as the regional dependence
(correlation) function. Logically, as the distance increases the correlation coefficient
between the ReV values decreases, and therefore Fig. 5.15 has a decreasing trend
with distance and theoretically this function should be asymptotic to the horizontal
axis. Consideration of Eq. (5.31) with unit variance yields the corresponding SVs in
the same figure.
The general expression in Eq. (5.34) for the simple Kriging can be rewritten for
a standardized ReV as

γ (d), ρ(d)

1.0

ρ(d)
γ(d)

0.0 Distance

Fig. 5.15 Covariance–distance graph

5.6 Simple Kriging 235


n
zE = λi zi . (5.38)
i=1

Here there are n unknowns and accordingly n equations are necessary for the simul-
taneous solution. For this purpose, both sides of Eq. (5.38) are multiplied by each
measurement ReV variable and then the averages (expectations) are taken. The
resultant set of equations becomes
n
λi ρ (zi ,z1 ) = ρ (zE ,z1 )
i=1
n
λi ρ (zi ,z2 ) = ρ (zE ,z2 )
i=1 (5.39)
....................................
n
λi ρ (zi ,zk ) = ρ (zE ,zk )
i=1

In order to bring this set of simultaneous equations into a matrix form, the following
additional succinct vector definitions are necessary. The unknown column vector is
⎤⎡
λ1
⎢ λ2 ⎥
⎢ ⎥
λ=⎢
⎢ . ⎥
⎥ (5.40)
⎣ . ⎦
λn

Finally, the right-hand side of Eq. (5.39) represents the known part, say, column
vector, B, which is defined as
⎡ ⎤
ρ (zE ,z1 )
⎢ ρ (zE ,z2 ) ⎥
⎢ ⎥
B=⎢ ⎢ . ⎥;
⎥ (5.41)
⎣ . ⎦
ρ (zE ,zn )

with these notations at hand, Eq. (5.39) can be written shortly as

C = B,

or the inversion operation gives the solution implicitly as

 = C−1 B. (5.42)

After the determination of the weighting values, λi , from this last expression their
substitution into Eq. (5.38) leads to the estimation of the standard ReV, which is
then converted to non-standard (original) ReVs as
236 5 Spatial Modeling

ZE = Z + σE2 zE , (5.43)

where zE is an (n×1) matrix of the measured ReV values with zero mean and unit
variance.
The simple Kriging procedure can also be applied by considering the relation-
ship between the SV and the covariance function as given in Eq. (5.31). However,
the application of this transition between the covariance and the corresponding SV
will be reliable only in the case of the normally distributed ReVs. Otherwise, the
results obtained from the use of the covariance will be biased. The reader can see
the difference by applying the simple Kriging once with the covariance and then
with the SV functions of the same set ReV measurements. The simple Kriging is
equivalent to multiple regression procedure where the covariance is used for the
parameter estimation.
All what has been explained in this section is based on the covariance function for
the depiction of spatial dependence. Since there is a relationship between the covari-
ance and the SV functions in the case of standardized ReV according to Eq. (5.31)
as ρ(d) = 1 − γ(d), then the replacement of the covariance terms in all the equations
of this section provides an alternative spatial modeling of ReV based on SV. The
variance of the estimation in case of covariance use is

σE2 = 1 − B . (5.44)

When the SV is used for the spatial modeling, the same estimation variance
becomes

σE2 = B . (5.45)

The critic of the simple Kriging is that it depends on the statistical property of
the covariance (or SV) function preservation in the final estimations. In other words,
the spatial estimation is achieved in such a way that the overall spatial dependence
function (SDF) of the ReV is preserved throughout the procedure. Unfortunately,
neither the cross-validation nor the unbiasedness procedures are applied explicitly
in the simple Kriging procedure.
Example 5.1
The earthquake magnitude measurements at five stations are presented in Table 5.5
with the positions in Fig. 5.16. The spatial estimation, ZE , will be obtained from all
the five measurement sites.
The distance matrix between each pair of data is calculated with the following
results.
⎡ ⎤
0
⎢0.534689 0 ⎥
⎢ ⎥
⎢
D = ⎢0.349005 0.326195 0 ⎥
⎥
⎣0.664208 0.185368 0.370756 0 ⎦
0.483628 1.015764 0.772863 1.129264 0
5.6 Simple Kriging 237

Table 5.5 Spatial data

ReV Easting (km) Northing (km) Magnitude

Z1 24.47950 40.40017 1.87

Z2 24.48400 39.86550 1.92
Z3 24.68783 40.12017 2.2
Z4 24.63183 39.75367 2.15
Z5 24.54383 40.87950 2.07
ZE 24.59567 40.12667 2.65

41
Z5
40.8

40.6
Northing (km)

40.4 Z1

40.2
ZE Z3
40
Z2
39.8
Z4
39.6
24.5 24.55 24.6 24.65 24.7
Easting (km)

Fig. 5.16 Spatial scatter of data locations

On the other hand, the distances between the estimation site, ZE , and measure-
ment sites are as follows.

1 2 3 4 5

E 0.297151 0.284038 0.092392 0.374746 0.754616

Provided that the ReV is standardized, the corresponding half-squared differ-

ences (i.e., SV) matrix can be obtained by substituting Eq. (5.31) with σZ2 = 1 into
Eq. (5.36), which leads to
⎡ ⎤
0 1 − γ(z1 ,z2 ) . . . 1 − γ(z1 ,zn )
⎢1 − γ(z2 ,z1 ) 0 . . . 1 − γ(z2 ,zn )⎥
⎢ ⎥
⎢ . . ... . ⎥
γ=⎢
⎢
⎥
⎥ (5.46)
⎢ . . ... . ⎥
⎣ . . ... . ⎦
1 − γ(zn ,z1 ) 1 − γ(zn ,z2 ) ... 0
238 5 Spatial Modeling

There are two different ways to calculate the SV values in practical works, either
from a given small sample as in Table 5.6 without knowing the basic structure of
the sample SV or after defining the structural form of the sample SV from a large
number of data, which is preferred to be more than 30 data values. The former
approach gives a matrix, which that represents half-squared differences subtraction
from 1 according to Eq. (5.46) between the earthquake magnitudes becomes from
Table 5.5.
⎡ ⎤
0
⎢1 − 0.00125 0 ⎥
⎢ ⎥
γ=⎢ ⎢1 − 0.05445 1 − 0.03920 0 ⎥
⎥
⎣1 − 0.03920 1 − 0.02645 1 − 0.00125 0 ⎦
1 − 0.02000 1 − 0.01125 1 − 0.00845 1 − 0.00320

Likewise, Eq. (4.41) can be written by considering Eq. (5.31) as

⎡ ⎤
1 − γ (zE ,z1 )
⎢1 − γ (zE ,z2 )⎥
⎢ ⎥
B=⎢
⎢ . ⎥
⎥ (5.47)
⎣ . ⎦
1 − γ (zE ,zn )

It is not possible to estimate the SV values in this vector because the earthquake
value at prediction location is not known. Therefore, it is necessary to know the
global SV that would depend on many location records, and it is assumed herein that
a priory structural analysis had produced the sample SV model as a linear model

γ (d) = 0.015 + 0.1d.

Now one can calculate the SV value from the distances between the estimation point
and other surrounding points in Fig. 5.16, which leads to
⎡ ⎤
1 − 0.0447151
⎢ 1 − 0.0434038 ⎥
⎢ ⎥
B=⎢
⎢ 1 − 0.0242392
⎥
⎥
⎣ 1 − 0.0524746 ⎦
1 − 0.0904616

Of course, it is now possible to calculate the matrix in Eq. (5.46) according to the
distance matrix above by using large sample SV equation, which leads to
⎡ ⎤
0
⎢1 − 0.0684689 0 ⎥
⎢ ⎥
⎢
γ = ⎢1 − 0.0499005 1 − 0.0476195 0 ⎥
⎥
⎣1 − 0.0814208 1 − 0.0335368 1 − 0.0520756 0 ⎦
1 − 0.0633628 1 − 0.1165764 1 − 0.0922863 1 − 0.279264
5.7 Ordinary Kriging 239

The final solution can be found by taking inverse of this , which appears as follows.
⎡ ⎤
−0.7747 0.3017 0.2867 0.2763 0.2058
⎢ 0.3017 −0.7773 0.2822 0.2041 0.2866 ⎥
⎢ ⎥
γ−1 =⎢
⎢ 0.2867 0.2822 −0.7600 0.2391 0.2558 ⎥
⎥
⎣ 0.2763 0.2041 0.2391 −0.9102 0.4471 ⎦
0.2058 0.2866 0.2558 0.4471 −0.9829

Hence, application of Eq. (5.42) leads to the final weight values as,
⎡ ⎤
0.2801
⎢ 0.2668 ⎥
⎢ ⎥
λ=⎢ ⎥
⎢ 0.2641 ⎥
⎣ 0.2387 ⎦
0.2526

It is now possible to calculate the prediction value from Eq. (5.32), which gives ZE
= 2.65.

5.7 Ordinary Kriging

This procedure has two assumptions of the simple Kriging, with conflict of the third
one, which is assumed to be constant regionally but unknown. Hence, it is not pos-
sible to apply the standardization procedure to the ReV measurements because the
mean value is unknown. Similar expression to Eq. (5.34) can be written, but with
the consideration of unknown regional mean, m, value as


n
ZE = m + λi (Zi − m) (5.48)
i=1

Comparison of this with Eq. (5.34) indicates that rather than the standardized ReV
variables, non-standardized ReVs are used. If both sides of this last expression is
arithmetically averaged, one can then obtain


n
m=m+m λi − m = 0,
i=1

which yields the restrictive condition, as has already been given by Eq. (5.33).
Hence, the first rule in the ordinary Kriging is that the summation of all the weights
ought to be equal to 1. It indicates that the Kriging weights are independent of the
ReV average. This is referred to as the unbiasedness principle in the Kriging litera-
ture. It is possible to write Eq. (5.48) as
240 5 Spatial Modeling
% &

n 
n
ZE = m 1 − λi + λi Zi , (5.49)
i=1 i=1

where the parenthesis in the first term on the right-hand side is equal to zero by
definition. Equation (5.33) is as a condition to the main estimation expression in Eq.
(5.48), which can be rewritten with the estimation error term, εE as


n
(ZE − m) = (λi Zi − m) + εE .
i=1

In the following this estimation error variance will be minimized. Let us leave the
estimation error as a subject,


n
εE = (ZE − m) − (λi Zi − m).
i=1

It is known by definition that the error estimation overall arithmetic average, εE , is

equal to zero. The square of both sides can be obtained as
 n 2

n 
ε2E = (ZE − m)2 − 2 (λi Zi − m)(ZE − m) + (λi Zi − m) ,
i=1 i=1
or more explicitly


n 
n 
n
 
ε2E = (ZE − m)2 − 2 (λi Zi − m)(ZE − m) + λj Zj − m (λi Zi − m) .
i=1 j=1 i=1

If the estimation is made n times, then the average error estimate square (variance)
will be obtained as

1 ( 2 ) 1 1 
n n n n
εE = (ZE − m)2 − 2 (λi Zi − m)(ZE − m)
n l n n
l=1 l=1 l=1 i=1

1 
n n n

+ λj Zj − m (λi Zi − m)
n
l=1 j=1 i=1

or
 n   n 
1 ( 2 ) 1   1
n n
εE = (ZE − m) − 2
2
(λi Zi − m)l (XE − m)l
n l n n
l=1 l=1 i=1 l=1
 n 
n  n
1  
+ λj Zj − m ((λi Zi − m))
n
j=1 i=1 l=1
5.7 Ordinary Kriging 241

The first average in the big brackets on the right-hand side is equal to the estima-
tion variance, σE2 , the second term average is equivalent to estimation-measurement
covariance, Cov (ZE , Zi ) and the last term is the covariance between two measure-
ments, Cov (Zi , Zj ). In fact, the average terms based on n values are equivalent to
their respective expectations as n goes to infinity theoretically. The last expression
can be written with these new covariance values as

n 
n 
n
ε2El = σE2 −2 Cov (ZE , Zi ) + Cov (Zi , Zj ). (5.50)
i=1 j=1 i=1

This expression must be minimized with the condition in Eq. (5.33) and, therefore,
the minimization equation can be written with the Lagrange multiplier, μ, as


n 
n 
n 
n
ε2El = σE2 − 2 Cov (ZE , Zi ) + Cov (Zi , Zj ) + μ λi . (5.51)
i=1 j=1 i=1 i=1

This can be written in the matrix and vector form after the definition of the following
quantities.
⎡ ⎤
cov(z1 ,z1 ) cov(z1 ,z2 ) . . . cov(z1 ,zn ) 1
⎢cov(z2 ,z1 ) cov(z2 ,z2 ) . . . cov(z2 ,zn ) 1⎥
⎢ ⎥
⎢ . . ... . .⎥
⎢ ⎥
C=⎢
⎢ . . ... . .⎥⎥, (5.52)
⎢ . . ... . .⎥
⎢ ⎥
⎣cov(zn ,z1 ) cov(zn ,z2 ) . . . cov(zn ,zn ) 1⎦
1 1 1 0

which can be written based on the SV formulation alternatively as

⎡ ⎤
λ1
⎢λ2 ⎥
⎢ ⎥
⎢.⎥
λ=⎢ ⎥
⎢ . ⎥, (5.53)
⎢ ⎥
⎣λn ⎦
μ

and finally
⎡ ⎤
cov(zE ,z1 )
⎢cov(zE ,z2 )⎥
⎢ ⎥
⎢ . ⎥
B=⎢
⎢
⎥.
⎥ (5.54)
⎢ . ⎥
⎣cov(zE ,zn )⎦
1
242 5 Spatial Modeling

The measurement vector that includes n nearest values is given as

⎡ ⎤
Z1
⎢Z2 ⎥
⎢ ⎥
⎢.⎥
M=⎢ ⎥
⎢ . ⎥. (5.55)
⎢ ⎥
⎣Zn ⎦
0

With these notations, the first ordinary Kriging weights may be estimated by using
either the covariance or SV values from

 = C−1 B. (5.56)

The ordinary Kriging estimate of the regionalized variable at location E can be

obtained as
ZE = MT  = MT C−1 B. (5.57)

Finally, the ordinary Kriging estimation variance is

σE2 = BT  = BT C−1 B. (5.58)

The estimate and estimation error depend on the weights chosen. Ideally, Krig-
ing tries to choose the optimal weights that produce the minimum estimation error.
In order to derive the necessary equations for Kriging, extensive calculus use is
required, which is not included here; however, information about the derivation
can be found in various textbooks, such as by Clark (1979) and Olea (1975). Opti-
mal weights produce unbiased estimates and have a minimum estimation variance,
which are obtained by solving a set of simultaneous equations. In the case of SV the
corresponding expressions to Eqs. (5.52, 5.53 and 5.54) are
⎡ ⎤
γ(z1 ,z1 ) γ(z1 ,z2 ) . . . γ(z1 ,zn ) 1
⎢γ(z2 ,z1 ) γ(z2 ,z2 ) . . . γ(z2 ,zn ) 1⎥
⎢ ⎥
⎢ . . ... . .⎥
⎢ ⎥
⎢ . . ... . .⎥
⎢ ⎥, (5.59)
⎢ . . . . . . . ⎥
⎢ ⎥
⎣γ(zn ,z1 ) γ(zn ,z2 ) . . . γ(zn ,zn ) 1⎦
1 1 1 0

⎡ ⎤
λ1
⎢ λ2 ⎥
⎢ ⎥
⎢ . ⎥
λ=⎢ ⎥
⎢ . ⎥, (5.60)
⎢ ⎥
⎣ λn ⎦
−μ
5.7 Ordinary Kriging 243
⎡ ⎤
γ (zE ,z1 )
⎢ γ(zE ,z2 ) ⎥
⎢ ⎥
⎢ . ⎥
⎢
B=⎢ ⎥, (5.61)
. ⎥
⎢ ⎥
⎣γ (zE ,zn )⎦
1
respectively.

Example 5.2
Ordinary Kriging methodology is demonstrated by considering groundwater levels
at five sites as given in Table 5.6 with easting, northing, and groundwater level
elevation values. The graphical locations are shown in Fig. 5.17.
The distances between each pair are given as in the following matrix.
⎡ ⎤
0
⎢3.512834 0 ⎥
⎢ ⎥
⎢
D=⎢ 2.236068 4.338202 0 ⎥
⎥
⎣1.860108 2.154066 2.19545 0 ⎦
1.529706 5.035871 2.731300 3.310589 0

Table 5.6 Well coordinates and groundwater levels

Well location Easting (km) Northing (km) Elevation (m)

Z1 4.2 3.0 89
Z2 3.7 6.5 75
Z3 1.8 2.6 98
Z4 2.9 4.5 81
Z5 4.3 1.5 105
ZE 3.2 3.0 91.27

7
Z2

5
Z4
Northing (km)

4
ZE Z1
3 Z3

2 Z5

0
Fig. 5.17 Sample 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
geographical locations Easting (km)
244 5 Spatial Modeling

The distances between the estimation point, ZE , and others are given as,

1 2 3 4 5

E 0.8 3.535534 1.456022 1.529706 1.860107524

After calculations the half-squared difference matrix appears as follows.

⎡ ⎤
0
⎢332112.5 0 ⎥
⎢ ⎥
γ (d) = ⎢
⎢313632.0 264.5 0
⎥
⎥
⎣327240.5 18.0 144.5 0 ⎦
308112.5 450.0 24.5 288.0 0

For large sample sizes the structure of the SV is determined as a linear function
as

γ (d) = 0.3 + 3.8d,

and accordingly the SV matrix turns out to be as

⎡ ⎤
0 13.68 8.78 7.37 6.11 1
⎢13.68 0 16.78 8.47 19.93 1⎥
⎢ ⎥
⎢ 8.78 16.78 0 8.86 10.68 1⎥
γ=⎢
⎢ 7.37
⎥.
⎢ 8.47 8.86 0 12.88 1⎥⎥
⎣ 6.11 19.93 10.68 12.88 0 1⎦
1 1 1 1 1 0

The inverse of this matrix becomes

⎡ ⎤
−0.1489 0.0122 0.0159 0.0465 0.0742 −0.1032
⎢ 0.0122 −0.0603 −0.0016 0.0524 −0.0027 0.4696 ⎥
⎢ ⎥
⎢ 0.0159 −0.0016 −0.0788 0.0393 0.0252 0.2699 ⎥
γ =⎢
−1
⎢ 0.0465
⎥.
⎥
⎢ 0.0524 0.0393 −0.1317 −0.0065 −0.0512 ⎥
⎣ 0.0742 −0.0027 0.0252 −0.0065 −0.0902 0.4149 ⎦
−0.1032 0.4696 0.2699 −0.0512 0.4149 −10.9519

Likewise, from Eq. (5.61) realizing the distances between the estimation site and
other sites as given above, it is possible o obtain
⎡⎤
3.34
⎢13.74⎥
⎢ ⎥
⎢ 5.83 ⎥
B=⎢ ⎥
⎢ 6.11 ⎥ .
⎢ ⎥
⎣ 7.37 ⎦
1
5.8 Universal Kriging 245

It is possible to find from Eq. (5.56) the lambda values as

⎡ ⎤
0.4916
⎢ −0.0275 ⎥
⎢ ⎥
⎢ 0.2676 ⎥
⎢
λ=⎢ ⎥.
⎥
⎢ 0.2004 ⎥
⎣ 0.0679 ⎦
−0.5255

One can evaluate the estimation value from Eq. (5.57), which leads to ZE = 91.27 m.
The Kriging estimation variance is the weighted sum of the SV for the distance
from the points to the estimation location, which can be calculated from Eq. (5.58)
leading to σE2 = 4.02 m2 . This implies that the standard error of estimation is σE =
2.00 m.
It is possible to make such calculations with estimation and its estimation vari-
ance at every point. For this purpose it is useful to make cross-validation by con-
sidering each measurement as if it does not exist and Kriging methodology gives
the estimation and error variance. Hence one can construct two maps for the Krig-
ing estimates as a best guess of the mapped variable configuration and an error
map showing the confidence envelope that surrounds this estimation. All these
are based on the measurement and estimation sites configuration and distances
between measurements used in the estimation process, and on the degree of spa-
tial continuity of the ReV as expressed by the spatial covariance and preferably SV
models.

5.8 Universal Kriging

Unfortunately, the ordinary Kriging cannot be used unless the ReV has a con-
stant mean although it is not known. Because the successive differences conceal
the regional mean in the SV calculations. This means that in the case of a sys-
tematic variation in the regional average such as abrupt changes (Fig. 4.9), trends
(Fig. 4.11), oscillations, the ordinary Kriging assumption of constant mean is vio-
lated. It is, therefore, necessary to develop another ReV estimation procedure, which
should take into consideration these systematic mean variations. The universal Krig-
ing considers first-order stationary ReV as consisting of two components, namely
the regional drift and the residuals. The drift is the long-term average (expectation)
value of the ReV within the estimation site neighborhood. It slowly varies and it is
the non-stationary component of the local region. On the other hand, the residuals
are the differences between the measurements and the corresponding drift values. If
the drift is removed from the ReV, then the residuals can be modeled by the ordinary
Kriging principles. Hence, in the procedural structure of the universal Kriging there
are three steps.
246 5 Spatial Modeling

1) Removal of the drift values from the measurements of the ReV.

2) Application of the ordinary Kriging to establish the residuals at non-
measurement points
3) Addition of these residual values to the original drift values.

The difference between the trend surface as defined in Chapter 3, Section 3.10,
and the drift covers partial local area whereas the trend surface extends over all
the ReV variability domain. The underlying drift component can be removed from
the original measurements of the concerned ReV through different methodologies.
Among these are linear and non-linear trend surface fitting, double-dimensional
Fourier analysis, etc. These methodologies depict the drift component, and its sub-
traction from the original measurements leaves the residuals with almost zero arith-
metic average. Under the statistical theory that includes universal Kriging, a single-
valued, continuous, mapable property is called an ReV and is considered to consist
of two parts, a drift, or expected value, and a residual, or deviation from the drift.
The drift may be modeled by a local polynomial function within a neighborhood,
which is analogous to a local trend surface. If the drift is removed, the residual sur-
face can be regarded as first-order stationary in a statistical sense. Hence, again the
simple or ordinary Kriging models can be used for the residual data set. After the
estimation of residuals again, the summation with the convenient drift values, the
original ReV can be estimated.
Apart from this rather complicated and piecewise application, it is also possible
to develop universal Kriging equations including another condition of drift in the
derivations with another Lagrange multiplier. Hence, the original measurements can
be used directly in the calculations. For example, similar set of equations can be
derived for the universal Kriging to Eqs. (5.57, 5.58, and 5.59) and Eq. (5.53) as
⎡ ⎤
γ(z1 ,z1 ) λ(z1 ,z2 ) ... γ(z1 ,zk ) 1 L1,1 L2,1
⎢ γ(xz2 ,z1 ) γ(z2 ,z2 ) ... γ(z2 ,zk ) 1 L1,2 L2,2 ⎥
⎢ ⎥
⎢ . . ... . . . . ⎥
⎢ ⎥
⎢ . . ... . . . . ⎥
⎢ ⎥
C=⎢
⎢ . . ... . . . . ⎥ ⎥ (5.62)
⎢ γ(zk ,z1 ) γ(zk ,z2 ) ... λ(zk ,zk ) 1 L1,1 L2,1 ⎥
⎢ ⎥
⎢ 1 1 ... 1 0 0 0 ⎥
⎢ ⎥
⎣ L1,1 L1,2 ... L1,k 0 0 0 ⎦
L2,1 L2,2 ... L21,k 0 0 0

⎡
⎤
λ1
⎢ λ2 ⎥
⎢ ⎥
⎢ . ⎥
⎢ ⎥
⎢ . ⎥
λ=⎢ ⎥
⎢ λk ⎥ , (5.63)
⎢ ⎥
⎢ μ0 ⎥
⎢ ⎥
⎣ μ1 ⎦
μ3
5.8 Universal Kriging 247

and finally
⎡ ⎤
γ (zE ,z1 )
⎢γ (zE ,z2 )⎥
⎢ ⎥
⎢ . ⎥
⎢ ⎥
⎢ . ⎥
⎢
B=⎢ ⎥, (5.64)
⎥
⎢γ (zE ,zk )⎥
⎢ 1 ⎥
⎢ ⎥
⎣ L1,k ⎦
L2,k
respectively. Herein, zi represents a vector of the coordinates of point I, while L1,i
(L2,i ) is the scalar value representing the location of this point along coordinate axis
horizontal (vertical) or 1 (2), for instance, east–west (north–south) direction. The
vector of universal Kriging weight  is found by Eq. (5.56), except that W and B
are given by Eqs. (5.62) and (5.64). The measurement vector that includes k nearest
values is given as
⎡ ⎤
Z1
⎢Z2 ⎥
⎢ ⎥
⎢.⎥
⎢ ⎥
⎢.⎥
M=⎢ ⎥
⎢Zk ⎥ . (5.65)
⎢ ⎥
⎢0⎥
⎢ ⎥
⎣0⎦
0
In this vector there are (k + d + 1) elements, where the last d+1 elements that corre-
spond to the Lagrange multipliers are zero.
Universal Kriging is a procedure that can be used to estimate values of a surface
at the nodes of a regular grid from irregularly spaced data points. If the surface
is second-order stationary, or can be made stationary by some transformation, the
spatial autocorrelation will express the degree of dependence between all locations
on the surface, and most particularly between observations and grid nodes.

Example 5.3
The problem given in Example 5.2 can be extended for the application of Universal
Kriging. By making use of Table 5.6, the numerical form of the coefficients matrix
can be obtained similar to the previous example but with additional location val-
ues as
⎡ ⎤
0 13.68 8.78 7.37 6.11 1 4.2 3.0
⎢13.68 0 16.78 8.47 19.93 1 3.7 6.5⎥
⎢ ⎥
⎢ 8.78 16.78 0 8.86 10.68 1 1.8 2.6⎥
⎢ ⎥
⎢ 7.37 8.47 8.86 0 12.88 1 2.9 4.5⎥
γ=⎢ ⎢ ⎥.
⎥
⎢ 6.11 19.93 10.68 12.88 0 1 4.3 1.5⎥
⎢ 1 1 1 1 1 0 0 0⎥
⎢ ⎥
⎣ 4.2 3.7 1.8 2.9 4.3 0 0 0 ⎦
3.0 6.5 2.6 4.5 1.5 0 0 0
248 5 Spatial Modeling

The inverse of this 8×8 matrix can be obtained as

⎡ ⎤
−0.1248 0.0212 −0.0192 0.0342 0.0887 −0.9626 0.2474 0.0031
⎢ 0.0212 −0.0328 −0.0222 0.0552 −0.0184 −0.5517 0.0815 0.1934 ⎥
⎢ ⎥
⎢ −0.0192 −0.0222 −0.0252 0.0559 0.0108 1.7447 −0.3575 −0.0652 ⎥
⎢ ⎥
⎢ 0.0342 0.0522 0.0559 −0.1246 −0.0177 0.2611 −0.1285 0.0333 ⎥
 −1 =⎢
⎢ 0.0887
⎥.
⎢ −0.0184 0.0108 −0.0177 −0.0634 0.5085 0.1571 −0.1646 ⎥
⎥
⎢ −0.9626 −0.5517 1.7447 0.2611 0.5085 40.1311 −8.5379 −5.6915 ⎥
⎢ ⎥
⎣ 0.2474 0.0815 −0.3575 −0.1285 0.1571 −8.5379 2.5455 −0.0474 ⎦
0.0031 0.1934 −0.0652 0.0333 −0.1646 −6.6915 −0.0474 1.5258

On the other hand, the elements of the right-hand side vector in Eq. (5.39) becomes
as follows.
⎡ ⎤
3.34
⎢13.74⎥
⎢ ⎥
⎢ 5.83 ⎥
⎢ ⎥
⎢ 6.11 ⎥
⎢
B=⎢ ⎥
⎥
⎢ 7.37 ⎥
⎢ 1 ⎥
⎢ ⎥
⎣ 3.2 ⎦
3.0

The universal Kriging weight vector can be found after the necessary algebraic cal-
culations according to Eq. (5.56) as
⎡ ⎤
0.4625
⎢ −0.0362 ⎥
⎢ ⎥
⎢ 0.3094 ⎥
⎢ ⎥
⎢ 0.2157 ⎥
λ=⎢ ⎥
⎢ 0.0486 ⎥ .
⎢ ⎥
⎢ 0.4546 ⎥
⎢ ⎥
⎣ −0.2998 ⎦
0.0123

Finally, after all these calculations the estimation value can be obtained from
Eq. (5.57) as ZE = 91.34 m.
One can notice that for the given example there is no major difference between
the ordinary and universal Kriging methodologies. Ordinary Kriging, however, in
common with other weighted-averaging methods, does not extrapolate well beyond
the convex hull of the control points. That is, most estimated values will lie on the
slopes of the surface and the highest and lowest points on the surface usually will
be defined by estimation (control) points.
5.9 Block Kriging 249

5.9 Block Kriging

All above mentioned Kriging methods (simple, ordinary, and universal) can be con-
sidered as punctual Kriging methodologies because their estimations are obtained at
the support of individual sites. In the block Kriging, however, line, areal, or volume
supports are considered and a single estimation value is obtained, which is repre-
sentative for the whole support. This Kriging procedure also starts to work on a set
of ReV measurement values, but its final product is valid for a certain domain in
1, 2, or 3 dimensions according to the type of problem. For instance, in ore grade
reserve, pollution concentration, etc. estimations 3D shapes are considered. Rainfall
maps and the hydrographs are of 2D.
The early versions of Kriging techniques were all concerned with the block Krig-
ing since the ore reserves and grade degrees were the major concentrations for
research in the domain of geology. In fact, the most commonly used block Krig-
ing is similar to the ordinary Kriging with the interpretation of known vector, B,
elements, which are taken as the average of the spatial covariance or SV values
within the interest of block domain, A. Actually, the elements of vector Bn represent
the covariances or SVs between the observations and A, integrated over the domain
of 2D or 3D.
The major problem prior to the block Kriging application is the determination of
block sizes. Each smaller subarea or sub-volume is represented by the geographical
coordinates of its center point, and spatial covariance or SV is determined between
an observation, Zi , and each of the center points within a block. Then all of these
covariance or SV is averaged in order to determine the point-to-block spatial covari-
ance, cov(A, Zi ) or SV, γ (A, Xi ) between block A and observation. If possible, the
subdivisions of blocks should be regular and should be the same. It has been sug-
gested by Isaaks and Srivastava (1989) that 4×4 = 16 and 4×4×4 = 64 subdivisions
are adequate for the area and volume blocks. The fundamentals of block Kriging
are presented in a simple manner by Olea (1999). It is commonly the counterpart of
ordinary Kriging and a set of normal equations must be solved similar to Eq. (5.59).
In this new formulation the spatial dependence function (SDF) either in the form of
covariance or SV (or CSV) does not represent the relationship between two points,
but they are the averages of the spatial covariance or SV (or CSV) between the point
observations, Zi , and all possible points within an area, A.
Figure 5.18 shows the line and areal blocks and measurement point i. In Fig.
5.18a the line is drawn into four pieces and the distances between the center of each
piece and the measurement point I as d1,i , d2,i , d3,i and d4,i . Likewise, in Fig. 5.18b
there are 16 sub-areas in a given block with distances d1,i , d2,i , . . ., d16,i .
Since the distances are known, their corresponding SV values can be calculated
easily from a given theoretical SV function. The arithmetic averages of the relevant
SV values give the representative SV value for line or areal block. Furthermore,
now the block is considered as a point with this average SV value, and then accord-
ingly one of the convenient Kriging techniques as mentioned above can be applied
suitably.
250 5 Spatial Modeling

d1,i
i
i

d16,i
d1,i d2,i d3,i d4,i

1 2 3 4
b
a

Fig. 5.18 Block pieces (a) line, (b) area

5.10 Triple Diagram Model

Since gradual (trend) or abrupt (shifts) climatic change questions have gained par-
ticular attention in recent years most of the researches on lake level changes are con-
cerned with meteorological factors of temperature and precipitation data. Along this
line of research Hubert et al. (1989), Vannitseni and Demaree (1991), and Sneyers
(1992) used statistical methods to show that temperature, pressure, and flow series
in Africa and Europe have altered several times during the present century. On the
other hand, as stated by Slivitzky and Mathier (1993), most of the modelling of
levels and flow series on the Great Lakes has assumed stationarity of time series
using either Markov or ARIMA (AutoRegressive Moving Average) processes pre-
sented by Box and Jenkins (1976). These models may work on lags of one, two,
three, or more, but they consider the linear structure in the lake level fluctuations.
Since, lake level fluctuations do not have stationarity property, classical models such
as Markov and ARIMA processes cannot stimulate lake levels reliably. Multivari-
ate models using monthly lake level variable failed to adequately reproduce the
statistical properties and persistence of basin supplies (Loucks, 1989; Iruine and
Eberthardt, 1992). On the other hand, spectral analysis of water levels pointed to
the possibility of significant trends in lake level hydrological variables (Privalsky,
1990). Almost all these scientific studies relied significantly on the presence of an
autocorrelation coefficient as an indicator of long-term persistence in lake level time
series. However, many researchers have shown that shifts in average lake level might
introduce unrealistic and spurious autocorrelations. This is the main reason why the
classical stochastic and statistical models often fail to reproduce the statistical prop-
erties. However, Mathier et al. (1992) were able to reproduce adequately the statis-
tical properties of a shifting-mean model. In the following sequel a version of the
Kriging methodology is adopted and used for the lake level estimations. For this
purpose, the world’s largest soda lake, Lake Van on the Anatolian high plateau in
eastern Turkey (38.5◦ N and 43◦ E) is adopted for application (Fig. 5.19). Lake Van
area has very severe winters with frequent temperatures below 0◦ C. Most of the
5.10 Triple Diagram Model 251

Fig. 5.19 Location map

precipitation falls during winter season in the form of snow and toward the end of
spring heavy rainfalls occur. High runoff rates occur in spring during snowmelt, and
more than 80% of annual discharge reaches the lake during this period. The sum-
mer period (July to September) is warm and dry with average temperatures of 20◦ C.
Diurnal temperature variations are about 20◦ C.
Human beings can visualize at the maximum three-dimensional variations. The
best configuration and interpretation of such variations can be achieved in three-
dimensional Cartesian coordinate systems through contour maps. Generally, maps
are regarded as the variation of a variable by location variables that are either longi-
tudes and latitudes or eastings and northings (Isaaks and Srivastava, 1989; Cressie,
1993; Kitanidis, 1997). Hence, it is possible to estimate the concerned (mapped)
variable value for a given pair of location variables. Since one wants to predict the
current lake level from previous records, it is suggested that two previous records
replace the two location variables. In this manner, it is possible to map the current
values of a variable based on two previous values of the same variable. The first
step in any Kriging methodology prior to mapping is to determine the sample SV,
which guides for the theoretical model that will be employed in the classical Kriging
modelling. For this purpose, the scatter of SV values versus distance is obtained for
lag-one, -two, and -three. In order to depict the general trend of the scatter diagram,
the distance range is divided into nine intervals, and the average of the SV values
that fall within each interval is considered as the representative SV value within
the mid-point distance of that interval as suggested by Myers et al. (1982). Dif-
ferent theoretical SV models such as linear, power, spherical, and Gaussian types
are tried for the best fit, and at the end the Gaussian SV is seen to have the best
match with the sample SV trend (see Fig. 5.20). The Gaussian model is the most
suitable in all lags and the properties of fitted Gaussian SV model are presented in
Table 5.7.
252 5 Spatial Modeling

Fig. 5.20 Empirical and 6007.401

Semivariance
theoretical SV for three lags 4505.551
3003.701
1501.850
0.000
0.00 48.51 97.02 145.53 194.04
Distance

Lag-2, Gaussian
Semivariance 4998.219
3748.664
2499.110
1249.555
0.000
0.00 48.51 97.02 145.53 194.04
Distance
Lag-3, Gaussian
Semivariance

4248.569
3186.427
2124.284
1062.142
0.000
0.00 48.51 97.02 145.53 194.04
Distance

Such a mapping technique is referred to as triple diagram method (TDM) (Şen,

2008). Such maps are based on three consecutive lake levels. TDMs help to make
interpretations in spite of extremely scattered points. Although Davis (1986) has
suggested for mapping the application of various simple regional techniques such as
inverse distance, inverse distance square, which consider the geometrical configura-
tion of the scatter points only without the use of a third variable; herein, preparation
of the TDM is based on classical Kriging technique.
The construction of a TDM requires three variables, two of which are referred to
as independent variables (predictors) and they constitute the basic scatter diagram.
The third is the dependent variable, which has its measured values attached to each
scatter point. The equal value lines are constructed by the Kriging methodology
concepts explained in earlier sections of this chapter.

Table 5.7 Theoretical Gaussian SV parameters

Lag Nugget (cm2 ) Sill (cm2 ) Range (cm) Correlation coefficient

1 70.0 6,250 213.20 0.977

2 270.0 5,555.0 586.80 0.990
3 250.0 4,327.0 136.0 0.977
5.10 Triple Diagram Model 253

Lake Van water level records are used for the implementation of the Kriging
methodology so as to obtain triple diagrams that give the common behavior of three
variables, which are taken consequently from the historical time series data. The first
two variables represent the two past lake levels and third one indicates the present
lake levels. Hence, the model has three parts, namely observations (recorded time
series) as input, triple diagram as response, and the output as prediction. It is pos-
sible to consider lags between the successive data at one, two, three, etc. intervals.
Such an approach is very similar to a second-order Markov process, which can be
expressed as

Hi = αHi−1 + βHi−2 + εi , (5.66)

where Hi , Hi−1 , and Hi−2 are the three consecutive lake levels; α and β are model
parameters; and finally, εi is the error term. The application of such a model requires,
prior to any prediction procedure, the parameter estimations from the available data.
Furthermore, its application is possible under the light of a set of assumptions,
which includes linearity, normality (Gaussian distribution of the residuals, i.e., εi ’s),
variance constancy, ergodicity, and independence of residuals. The triple diagram
replaces Eq. (5.66) without any restriction in the form of map. Such a map presents
the appearance of natural relationship between three consecutive time values of the
same variable.
In order to apply the triple diagram approach, it is necessary to divide the data
into training and testing parts. Herein, the past 24 months (two years) are left for
the test (prediction) whereas all other data values are employed for training, which
is the TDM as in Fig. 5.21.
Prior to any prediction, it is possible to draw the following interpretations from
these figures.

1) In the case of lag-one there is a strong relationship between Hi−1 and Hi−2 with
increasing contour values of Hi along almost 45◦ line (see Fig. 5.21a). The small
Hi values are concentrated at small Hi−1 and Hi−2 values; this implies the clus-
tering of small values of the three consecutive lake levels. Similarly, high lake
level values of the three consecutive levels also constitute high values cluster.
This means that small values follow small values and high values follow high
values, which indicates positive correlations. Local variations in the contour
lines appear at either low (high) Hi−1 or high (low) Hi−2 values. Consequently,
better predictions can be expected within a certain band around the 45◦ line
(Fig. 5.22). It is possible to deduce the following set of logical rules from Fig.
5.21b.
IF Hi−1 is low and Hi−2 is low, THEN Hi is low,
IF Hi−1 is medium low and Hi−2 is medium, THEN Hi is medium,
IF Hi−1 is high and Hi−2 is high, THEN Hi is high.
These rules can be used for fuzzy logic inference system as suggested by Zadeh
(1968).
254 5 Spatial Modeling

Hi Hi

400 400

350 350

300 300
Hi – 1

Hi – 2
250 250

200 200

150 150

100 100
100 150 200 250 300 350 400 100 150 200 250 300 350 400
Hi – 2 Hi –3
a b
Hi

400

350

300
Hi – 3

250

200

150

100
100 150 200 250 300 350 400
Hi – 4
c

Fig. 5.21 Lake level TDMs (a) lag-one, (b) lag-two, (c) lag-three

450
400
350
300
Predicted

250
200
150
100
50
0
0 100 200 300 400
Observed

Fig. 5.22 Lag-one model verification

5.10 Triple Diagram Model 255

2) In Fig. 5.21b the variations in the contour lines become very distinctive and
rather haphazard compared to Fig. 5.21a. This implies that with the increment
in the lag value, present time lake level prediction will have more relative error.
There is also a distinctive 45◦ line, but with comparatively narrower band of
certainty around it.
3) Finally, at lag-three case (Fig. 5.21c) the contour pattern takes even more hap-
hazard variation. This implies increase in the relative error of predictions.

Şen et al. (1999, 2000) identified suitable models and estimates for lake level
fluctuations and their parameters for trend, periodic, and stochastic parts. A second-
order Markov model is found suitable for the stochastic part. As explained before,
TDM of lake levels can replace the second-order Markov process. In this manner, it
is not necessary to use first- and second-order autocorrelation coefficients, in order
to take into account more persistence. In order to make predictions for the past
24 months that are not used in the triple diagram constructions in Fig. 5.21, it is
necessary to enter Hi−1 and Hi−1 for each month on vertical and horizontal axes,
respectively. The prediction value of Hi can be either read from this map approxi-
mately, or calculated by using Kriging prediction equations. The prediction results
are shown in Table 5.8 with corresponding relative error amounts. Individual errors

Table 5.8 Lag-one lake level prediction (cm)

Hi−2 Hi−1 Hi Prediction Relative error (%)

119 112 110 109.32 0.62

107 111 114 118.40 3.72
125 130 125 134.87 7.32
130 125 118 128.43 8.12
125 118 105 118.60 11.47
120 138 142 145.50 2.41
138 142 138 139.53 1.10
142 138 131 134.16 2.35
138 131 117 131.64 11.12
127 141 151 144.08 4.58
141 151 151 146.85 2.75
151 151 144 144.46 0.32
137 140 144 138.76 3.64
140 144 159 140.22 11.81
144 159 182 157.03 13.72
199 202 195 203.62 4.23
202 195 185 191.33 3.31
195 185 177 182.29 2.90
189 193 202 202.01 0.00
193 202 221 209.94 5.00
202 221 245 229.48 6.34
262 254 244 252.90 3.52
254 244 239 243.16 1.71
244 239 241 231.89 3.78
Average 4.83
256 5 Spatial Modeling

400 400
Lake level (cm)

Lake level (cm)

300 300

200 200
Observation Observation
100 Prediction Prediction
100

0 0
0 10 20 30 0 10 30 40
(Month) (Month)
a b

400
Lake level (cm)

300

200
Observation
100 Prediction

0
0 10 30 40
(Month)
c

Fig. 5.23 Observed and predicted lake levels (a) lag-one, (b) lag-two, (c) lag-three

are slightly greater than 10%, but the overall prediction relative error percentage is
about 4.83%.
Figure 5.23 indicates the observed and predicted Hi values. It is obvious that they
follow each other very closely, and on the average observed and predicted lake level
series have almost the same statistical parameters.
The triple diagram model depicts even the increasing trend, which is not possi-
ble directly with the second-order Markov process. During the prediction procedure
there is no special treatment of trend, but even so it is modeled successfully. How-
ever, in any stochastic or statistical modeling, it is first necessary to make trend
analysis and separate it from the original data. In order to further show the verifi-
cation of the triple diagram approach for lake level predictions, in Fig. 5.22 the test
data are plotted versus the predictions. It is obvious that almost all the points are
around 45◦ line and hence the model is not biased. Predictions are successful at low
or high values.

5.11 Regional Rainfall Pattern Description

The mean annual and seasonal rainfall records in the southwest of Saudi Arabia are
adapted from the reports published by the Hydrology Division, Ministry of Agricul-
ture and Water in Saudi Arabia, and Al-Jerash (1985). Rainfall records at 63 stations
from 1971 to 1990 are selected for the Kriging application (Fig. 5.24). These stations
are chosen based on four criteria (Subyani, 2004, 2005).
5.11 Regional Rainfall Pattern Description 257

38 42 46 50 54 58

Taif
26°
Turabah Riyadh

o '
22°
21 00
STUDY
AREA
18°

14°
0 500 Km

o ' L ith Alaqiq

20 00 Bishah

Tathlith
900
T

o '
I

19 00
H

Kiyat
R

A
E

Abha
D

Qa
M

o ' ma
S

18 00 h
A
E
A

Najran
H

o ' Raingage Location

17 00 Jizan
500 Elevation (m.a.s.l)
YEMEN
0 1 2 3 4
0.0 100 Km
o ' o ' o ' o '
41 00 42 00 43 00 44 00

Fig. 5.24 Topographic map in southwest Saudi Arabia

1) They represent the best spatial coverage of the region.

2) They maximize the same monthly rainfall records.
3) They have continuous monthly rainfall.
4) They reflect the wide variety of environments within the study area – literally
from coast (Tihamah), mountains, and to leeside of the mountains (desert).

Descriptive statistics for the 63 stations are listed in Table 5.9 for these three
regions. All together 25 stations are located within the coast; 24 within mountains;
and 14 stations within the leeside.
Problems in the data, such as non-normality, trend, and outliers, should be fixed
before developing any kind of model. Normality of the sample data distribution
is known to improve the results from Kriging. Transformation is very important
to make the data more symmetric, linear, and constant in variance. Since annual
 258

Table 5.9 Dataset grouped into three regions

Region Number of stations Winter (mm) Spring (mm) Summer (mm) Fall (mm) Annual (mm)

Coast (Tihamah) 25 141a 199 238 152 684

13b 6 5 3 55
53.5c 51.4 58.6 53.7 217.2
33.4d 51.3 58.2 40.5 160.3
Mountains 24 169 264 148 69 527
20 80 16 10 150
78.5 155.3 57.1 30.8 321.8
47 53.5 35.4 18.8 120.4
Lee Side (Desert) 14 28 99 30 37 150
7 29 3 3 44
17.4 64.4 11.9 9.6 103.3
6.6 17.8 8.22 8.3 27.7

a Sample maximum; b Sample minimum; c Sample mean; d Sample standard deviation

5 Spatial Modeling
5.11 Regional Rainfall Pattern Description 259

and seasonal rainfall data are considered, it is pragmatic to find one transformation,
which works reasonably well for all. The Box-Cox transformation is widely used
and can be easily managed so that the skewness of transformed data Z(x,t) becomes
close to zero (Salas, 1993).
However, rainfall histogram in arid regions, as stated by Hevesi et al. (1992),
behaves as lognormal distribution. Hence, the transformation as Y = ln Z(x)
is applied for determining approximately normal annual and seasonal data. It is
accepted as normally distributed, if the computed Kolmogorov-Smirnov statistic
(Dmax ) is less than the corresponding critical value. The critical value for the 5%
level of significance is D0.05 = 0.171, which is greater than Dmax of transformed
data. Thus the null hypothesis of the transformed data normality cannot be rejected
at 0.05 level of significant. Further investigation can be done by visually observing
the normal probability plots, and most of the data lie on a straight line for the trans-
formed rainfall values. In addition, the skewness coefficients are reduced close to
zero (Subyani, 2005). Table 5.10 shows the statistics and normality test for original
and transformed annual and seasonal data.
The back-transformed value, i.e., exp (Y(x)) is biased predictor. However, the
unbiased expression for the Kriging estimates Z(x) is given by Aboufirassi and
Mariño (1984), Gilbert (1985), and Deutsch and Journel (1992). As
/ 0
Z∗ (x) = exp Y∗ (x) + σy2 /2 , (5.67)

where Z∗ (x) is the original data in millimeters; and Y∗ (x) is the natural logarithm,
and σy2 is the lognormal Kriging variance. The estimation variance is given as

 2 " #
σZ2 ∗ = Z∗ exp (σy2 ) − 1 . (5.68)

These two last expressions are used for constructing the rainfall isohyetals and
their variances. Figure 5.25 shows the sample and fitted SV for the natural log of
average annual rainfall (LnAAR). An isotropic spherical model with no nugget, but
with a sill equal to the sample variance of 0.48, and a range of 110 km, is selected
as the best representation of the spatial structure.

Table 5.10 Descriptive statistics and normality test for annual and seasonal data

Season Mean (mm) Median (mm) St. Dev. (mm) Skewness CV(%) K-S (Dm )

Annual 232 178 151 0.8 0.65 0.18

Winter 55.0 41.0 43.1 1.15 0.78 0.20
Spring 94.0 81.0 68.0 0.72 0.72 0.14
Summer 47.6 32.0 47.3 1.84 0.99 0.17
Fall 35.2 21.0 33.2 1.44 0.95 0.19
ln-Annual 5.22 5.18 0.7 −0.7 0.13 0.10
ln-Winter 3.71 3.71 0.79 −0.051 0.21 0.07
ln-Spring 4.21 4.39 0.92 −0.65 0.22 0.11
ln-Summer 3.38 3.46 1.06 −0.24 0.31 0.08
ln-Fall 3.1 3.04 1.02 −0.15 0.33 0.08
260 5 Spatial Modeling

0.6

0.5

0.4
γ (h)Ln P.

Range = 110 km
0.3 Sill = 0.48

0.2
Experimental Variogram
Fitted Spherical Model
0.1

0.0
0 100 200 300 400
Distance (km)

Fig. 5.25 Experimental and fitted SV model for Ln-annual rainfall

Cross-validation is applied to check whether the spherical model is adequate

to describe the spatial correlation of the annual rainfall. For diagnostic check,
the mean estimation error (MEE) and the root-mean-square errors (RMSE) for
LnAAR are 0.067 and 0.94, respectively, which suggest the validity of the spherical
model. Similar tests are performed for seasonal values and the results are shown in
Table 5.11.
Figure 5.26 presents the sample and theoretical SVs and their parameters for
winter, spring, summer, and fall seasons. In winter, the sample SV model is fitted
with the small nugget as 0.05. The sill is 0.67 and the practical range of dependency,
i.e., radius of influence, is 70 km (Fig. 5.26a).
In spring sample SV behaved as an exponential model with nugget of 0.1. The
sill is 0.95 with a practical range of dependency equal to 90 km (Fig. 5.26b). In the
summer and fall, the sample SV behaves as a spherical model. The sill for these two
seasons is around 1.00, and the ranges of correlation are within 140 km and 110 m,
respectively (see Fig. 5.26c and d).
Kriging estimates are computed for the average annual and seasonal rainfall
amounts. The sample SV represents the continuity structure quite well. The cross-
validation supports selection of models and their parameters. For Kriging estimation
and variance, the back-transformed values are applied.

Table 5.11 SV
Cross-validation Season Model N MEE RMSE

Ln-Annual Spherical 60 0.067 0.94

Ln-Winter Exponential 60 0.027 0.84
Ln-Spring Exponential 61 0.08 0.71
Ln-Summer Spherical 62 0.07 0.91
Ln-Fall Spherical 62 0.04 1.08
5.11 Regional Rainfall Pattern Description 261

1.4 0.8

1.2 0.7
0.6
1.0
0.5

γ (h) Ln P.
γ (h) Ln P.

0.8 Range = 70 km
0.4 Sill = 0.67
Range = 90 km
0.6 Nugget = 0.05
Sill = 0.95
0.3
Nugget = 0.1
0.4 Expremintal Variogram
Experimental Variogram
0.2
Fitted Exponential Model
0.2 Fitted Exponential Model 0.1

0.0 0.0
0 100 200 300 400 0 100 200 300 400
Distance (km) Distance (km)

(a) Winter (b) Spring

1.6 1.6

1.2 1.2

γ (h) Ln P.
γ (h) Ln P.

0.8 Range = 140 km 0.8 Range = 110 km

Sill = 1.1 Sill = 1.03

Experimental Variogram Experimental Variogram

0.4 0.4 Fitted Spherical Model
Fitted Spherical Model

0.0 0.0
0 100 200 300 400 0 100 200 300 400
Distance (km) Distance (km)

(c) Summer (d) Fall

Fig. 5.26 Experimental and fitted SV models for Ln-seasonal rainfall

The kriged isohyets for annual rainfall show a rapid increase in average from the
Red Sea shore line up to the mountains and a gradual decrease to the north and east
parts of the study area (Fig. 5.27a).
Orographic effects are produced toward the mountain area with the maximum
kriged estimate exceeding 350 mm/year. In the east and northeast parts of the
study area, Kriging estimates are around 100 mm/year. In the northern part with
a moderate elevation reaching more than 1,000 m, Kriging estimates also exceed
100 mm/year. This figure reflects the topographic variation similar to Fig. 5.24 with
annual rainfall that generally increases with elevation.
Kriging variances indicate similar behavior to the average annual rainfall esti-
mates. Small values near the clusters of stations in the mountain area (Fig. 5.27b)
indicate high estimation accuracy, whereas large values in the north, east, and north-
east indicate low estimation accuracy areas owing to the scarcity of sample loca-
tions. Generally, high estimation variances appear at areas of lacking data.
In winter (December–February), rainfall is associated most of the time with moist
and cold air of northerly Mediterranean origin, which is coupled with the local
effects of the Red Sea convergence zone the Scarp mountains as well as orographic
rainfall occurrences. Figure 5.28a shows that the Kriging estimates exceed 120 mm
in the middle and northern section of the mountainous areas.
However, they do not exceed 30 mm in the south of Tihamah, because there is no
Mediterranean or Red Sea effect, and the elevation is not high enough for orographic
rainfall occurrence. On the other hand, the southern part of the study area receives
less than 100 mm of rainfall due to the absence of monsoons. The Plateau area (east
262 5 Spatial Modeling

600000 600000
Taif N Taif N

500000 500000

Lith Bishah Lith Bishah

400000 100 400000 100

300000 300000

R
R

E
E

D
D

200000 Abha 200000 Abha

S
S

150 150

E
E

Najran Najran

A
A

100000 100000
Jizan Jizan
0

0.0
1 2 3 4

100 Km YEMEN 0.0

0 1 2 3 4

100 Km YEMEN
0 0
0 100000 200000 300000 400000 0 100000 200000 300000 400000

a b
Fig. 5.27 Isohyetal map of Kriging for annual rainfall (mm)

and northeast parts of the study area) receives less than 20 mm, because it is located
in the shadow (lee-side) area. Kriging estimation variance map has a similar trend
throughout the study area as shown in Fig. 5.28b.

600000 600000
Taif N Taif N
500

500000 500000

Lith Bishah Bishah

Lith
400000 400000

300000 300000
R

R
E

Abha
D

Abha
200000 200000
S

S
E

Najran Najran
A

100000 100000
Jizan Jizan
0

0.0
1 2 3 4

100 Km YEMEN 0

0.0
1 2 3 4

100 Km YEMEN
0 0
0 100000 200000 300000 400000 0 100000 200000 300000 400000
a b

Fig. 5.28 Isohyetal map of Kriging for winter rainfall (mm)

5.11 Regional Rainfall Pattern Description 263

600000 600000
Taif N Taif N
80

500000 500000

Bishah Bishah
Lith Lith
400000 400000

300000 300000

R
R

E
E

D
Abha
D

Abha
200000

S
200000
S

E
E

Najran
5000

A
Najran
A

100000
100000
Jizan
Jizan
YEMEN 0.0 100 Km YEMEN
0.0 100 Km
0
0 0 100000 200000 300000 400000
0 100000 200000 300000 400000
a b

Fig. 5.29 Isohyetal map of Kriging for spring rainfall (mm)

In spring (March–May), the whole region comes under the influence of south-
east monsoon air stream flow, the Red Sea convergence zone, and Mediterranean
depression, which distribute the rainfall in all regions. The Kriging estimates give
more detailed information about the rainfall distribution as shown in Fig. 5.29a.
Rainfall in this figure increases gradually from the Red Sea coast (40 mm) to
the mountain where the highest amount of rainfall falls (more than 160 mm) and
decreases to the plateau area, which receives about 100 mm. Generally, the south-
west region of the Arabian Peninsula receives the highest amount of rainfall during
the spring season compared to other seasons. This high amount of rainfall is a result
of increasing African–Mediterranean interaction effect, where rainfall occurs oro-
graphically in the mountains and southeast monsoon effect where the Plateau and
eastern slope receive more rainfall than the Red Sea coast. Kriging variances show
an increase in estimation accuracy as shown in Fig. 5.29b.
In summer (June–August), the southwest monsoon flow from the Indian Ocean
and the Arabian Sea is the predominant factor, which increases the rainfall along
the Scarp mountains and low elevation areas in the south of the study area. Kriging
estimates for summer season exceed 120 mm in the mountains and 160 mm in the
foothills near the Yemen border at the southwestern corner of the Arabian peninsula
(Fig. 5.30a).
Rainfall decreases toward the northern part of the study area due to its distance
from the monsoon effect, even though this area has a high elevation. Moreover,
the Kriging variances show no change in estimation accuracy in the foothills and
mountains, but they cannot be calculated with reliability in the plateau areas due to
the paucity of data (see Fig. 5.30b).
264 5 Spatial Modeling

600000 600000
Taif N Taif N
20
500000 500000 1000
10

Lith Bishah Lith Bishah

400000 400000

1000

300000 300000 500

R

R
E

E
D

D
Abha Abha
200000
S

200000

S
E

E
Najran
A

Najran

A
20

100000 100000
Jizan Jizan
0.0 100 Km YEMEN YEMEN
0.0 100 Km
0 0
0 100000 200000 300000 400000 0 100000 200000 300000 400000
a b

Fig. 5.30 Isohyetal map of Kriging for Summer rainfall (mm)

In fall (September–November), the local diurnal circulation and the southern air
stream weakens. In other words, it is a transition period from summer to winter and,
in general, the area receives little amount of rainfall. The Kriging estimation in the
foothills and the mountains in the southern part of the study area shows that they
receive higher amount of rainfall than the northern areas, similarly to the fall mon-
soon flow effects as shown in Fig. 5.31a. The Kriging variances show an increase in

600000 600000
Taif N Taif N
30

500000 500000

Lith Bishah Lith Bishah

400000 400000

300000 300000
R
R

E
E

D
D

Abha Abha
200000 200000
S
S

E
E

Najran Najran
A
A

100000 100000
Jizan Jizan
0.0 100 Km YEMEN 0 1 2 3 4 YEMEN
0.0 100 Km
0 0
0 100000 200000 300000 400000 0 100000 200000 300000 400000
a b

Fig. 5.31 Isohyetal map of Kriging for fall rainfall (mm)

5.11 Regional Rainfall Pattern Description 265

N
TER
WIN

RED SEA

G
RIN
SP

RED SEA

ER
MM
SU

RED SEA

L
FAL

RED SEA

Fig. 5.32 Spatiotemporal Kriging maps for rainfall in southwest Saudi Arabia
266 5 Spatial Modeling

estimation accuracy in the northern part of the study area, whereas there is no clear
change in the southern part as shown in Fig. 5.31b.
Generally, rainfall is predominant in the northern mountain areas during winter as
a result of the Mediterranean effect, and it is widespread in all regions during spring
because of the local diurnal circulation effects. Orographic conditions are clear in
winter and spring seasons. This orographic factor is also clear for the appearance
of the nugget effect in the exponential models in both winter and spring seasons.
During summer, rainfall moves toward the south due to the monsoon flow effect,
with its southwesterly wind. However, during fall, as a transition season, the area
comes under the influence of monsoon as well as the local diurnal circulations.
Figure 5.32 illustrates these spatio-seasonal variations of rainfall in the southwest
of Saudi Arabia.
The Kriging estimation variances are also investigated concerning the spatial and
temporal variation of rainfall in the study area for these four seasons. In space varia-
tion, the small value or high estimation accuracy of Kriging variances occurs in the
mountainous areas in all seasons. Towards the east, north, and northeast of the study
area, there is a consistent increase in variances implying low estimation accuracy.
In time variation, Kriging variance increases from winter to fall. These variations in
space and time are due to several factors such as

1) Clusters of stations in the mountain areas reflect higher estimation accuracy.

2) Scarcity of stations in north, east, and northeast areas reflects lower estimation
accuracy.
3) High variability in summer and fall rainfall gives rise to low accuracy estima-
tion compared with the somewhat lower variability in rainfall during winter and
spring seasons.

The description of the rainfall variability in space and/or in time is among the fun-
damental requirements for a wide variety of human activities and water resources
project design, management, and operation. Geostatistical methods are applied to
develop new maps for the prediction of rainfall over different seasons. The assigned
objectives of this study are to predict the magnitude and variation of the rainfall
in space as well as during different time periods. These techniques are applied to
rainfall data gathered from meteorological station network covering the southwest
region of the Arabian Peninsula. Rainfall in this area is characterized by high varia-
tion in spatial and temporal distributions.

References
Aboufirassi, M., and Mariño, M. A., 1984. A geostatistically based approach to the identification
of aquifer transmissivities in Yolo Basin, California. Math. Geol. 16(26), 125–137.
Al-Jerash, M., 1985. Climatic subdivisions in Saudi Arabia: an application of principle component
analysis, J. Climate 5, 307–323.
Barnes, S. L., 1964. A technique for maximizing details in numerical weather map analysis. J.
Appl. Meteor. 3, 396–409.
References 267

Box, G. E. P., and Jenkins, G. M., 1976. Time Series Analysis Forecasting and Control. Holden
Day, San Francisco, 560 pp.
Buell, C. E., 1958. The correlation between wind and height on a isobaric surface II: summer data.
J. Meteorol. 15(12), 502–512.
Carr, J., and Glass, C., 1984. Estimation of earthquake response spectra using Kriging. In: G. Verly
et al. (Eds.), Geostatistics for Natural Resources Characterization. Proceedings of the 2nd
NATO Advanced Study Institute of Geostatistics for Natural Resources Characterization: Part
2, D. Reidel, Dordrecht, Holland, pp. 745–752.
Carr, J., Bailey, R., and Deng, E., 1985. Use of indicator variogram for an enhanced spatial analysis.
J. Math. Geol. 18, 197–213.
Carr, J., and Glass, C., 1985. Treatment of earthquake ground-motion using regionalized variables.
J. Math. Geol. 17, 221–241.
Clark, I., 1979. Practical Geostatistics. Applied Science Publishers, London, 281 pp.
Cressie, N. A. C., 1993. Statistics for Spatial Data (revised edition). Wiley, New York, 900 pp.
Cressman, G. P., 1959. An operational objective analysis system. Monthly Weather Rev. 87(10),
367–374.
David, M., 1977. Geostatistical Ore Reserve Estimation. Elsevier Scientific Publishing Company,
364 pp.
Davis, J. C., 1986. Statistics and Data Analysis in Geology. John Wiley & Sons, Inc., New York,
646 pp.
Delhomme, J. P., 1978. Kriging in the hydrosciences. Adv. Resour. 1, 251.
Deutsch, C. V., and Journel, A. G., 1992. GSLIB Geostatistical Software Library and User’s Guide.
Oxford University Press, New York, TIC: 224174.
Dixon, R., 1969. Orthogonal polynomial as a basis for objective analysis. Meteorological Office,
Scientific Paper No. 30, Her Majesty’s Stationery Office, London, 20 pp.
Dixon, R., 1976. An objective analysis using orthogonal polynomials. Proceedings of the JOC
Study Group Conference on four-dimensional Data Assimilation, Paris, GARP Program on
Numerical Experimentation, Report No. 11, pp. 73–85.
Eddy, A., 1964. The objective analysis of horizontal wind divergence fields. Quart. J. Roy. Meteo-
rol. Soc. 90, 424–440.
Eliassen, A., 1954. Provisional report on spatial covariance and autocorrelation of the pressure
field. Inst. Weather and Climate Research Academy of Science Oslo Report No. 5.
Fisher, R. A., 1935. The Design of Experiments. Oliver and Boyd, Edinburg.
Flattery, T. W., 1970. Spectral models for global analysis and forecasting, Proceedings of the Sixth
AWS Technical Exchange Conference, U. S. Naval Academy, 21–21 September, Air Weather
Service Technical Report 242, pp. 42–54.
Fritsch, J. M., 1971. Objective analysis of a two-dimensional data field by the cubic spline tech-
nique. Monthly Weather Rev. 99(5), 1122–1143.
Gandin, L. S., 1963. Objective Analysis of Meteorological Fields. Leningrad, Gidromet; Translated
from Russian, Jerusalem, Israel Program for Scientific Translations, 1965, 242 pp.
Gandin, L. S., 1965. Objective Analysis, Lectures on Numerical Short-range Weather Prediction,
WMO, Regional Training Seminar, pp. 633–677.
Gilbert, S., 1985. Introduction to Applied Mathematics. Wellesley-Cambridge Press, Wellesley,
MA 02181.
Glass, C. E., 1978. Application of regionalized variables to micro-zonation. Proc. 2nd Int. Conf.
Micro-zonation for Safer Construction 1, 509–512.
Habib, Z., 1999. Optimum interpolation modeling supported by the cumulative semivariogram
for spatial-temporal meteorological variables. Unpublished Ph.D. Thesis, Istanbul Technical
University, 156 pp.
Hevesi, J. A., Istok, J. D., and Flint, A. L., 1992. Precipitation estimation in mountainous terrain
using multivariate geostatistics, Part I. Structural analysis. J. App. Met. 31, 661–676.
Hubert, P., Carbomnel, J. D., and Chaouche, A., 1989. Segmentation des series hydrometeoralo-
ques: application a des series de precipitation et de debits de L’Afrique de L’ouest. J. Hydrol.
110, 349–367.
268 5 Spatial Modeling

Iruine, K. N., and Eberthardt, A. K., 1992. Multiplicative seasonal ARIMA models for lake Erie
and Lake Ontario water levels. Water Res. Bull. 28(3), 385–396.
Isaaks, E. H., and Srivastava, R. M., 1989. An Introduction to Applied Geostatistics. Oxford Uni-
versity Press, Oxford, 561 pp.
Journel, A. G., and Huijbregts, C. I., 1978. Mining Geostatistics. Academic Press, London, 710 pp.
Journel, A. G., 1983. Non-parametric estimation of spatial distribution. Math. Geol. 15(13),
445–468.
Kalman, R. E., 1960. A new Approach to linear filtering and prediction problem. J. Basic Eng.
Series D 82(1), 35–45.
Kitanidis, P. K. 1997. Introduction to Geostatistics: Applications in Hydrogeology. Cambridge
University Press, Cambridge, 249 pp.
Krige, D. G., 1951. A statistical approach to some basic mine evaluation problems on the Witwa-
teround. J. Chimic. Min. Soc. South-Africa 52, 119–139.
Ledimet, F., and Talagrand, O., 1986. Variational algorithms for analysis and assimilation of mete-
orological observation: theoretical aspects. Tellus 38A(2), 97–110.
Lorenc, A. C., 1981. A global three-dimensional multivariate statistical interpolation scheme.
Monthly Weather Rev. 109, 701–721.
Loucks, F. D., 1989. Modeling the Great Lakes hydrologic-hydraulic system. Ph D Thesis, Univer-
sity of Wisconsin, Madison.
Marsily, G. D., 1986. Quantitative Hydrogeology, Groundwater Hydrology for Engineers. Aca-
demic Press, New York.
Martinez, C. A., 1996. Multivariate geostatistical analysis of evapo-transpiration and precipitation
in mountainous terrain. J. Hydrol. 174, 19–35.
Matheron, G., 1963. Principles of geostatistics. Econ. Geol. 58, 1246–1266.
Matheron, G., 1971. The theory of regionalized variables and its applications. Ecole de Mines.
Fontainbleau, France.
Mathier, L., Fagherazzi, L., Rasson, J. C., and Bobee, B., 1992. Great Lakes net basin supply
simulation by a stochastic approach. INRS-Eau Rapp Scient (fique 362, INRS-Fau, Sainte-
Foy), 95 pp.
Meleschko, V. P., and Prigodich, A. E., 1964. Objective analysis of humidity and temperature.
Trudy Sympoziuma Pochislenyum Melodam Prognoza Pagody, Gidrometeoizdat, Moscow,
U.S.S.R.
Myers, D. E., Begovich, C. L., Butz, T. R., and Kane, V. E., 1982. Variogram models for regional
groundwater chemical data. Math. Geol. 14, 629–644.
Olea, R. A., 1975. Optimum mapping techniques using regionalized variable theory: series on
spatial analysis no. 3. Lawrence, Kansas, Kansas Geological Survey, 137 pp.
Olea, R. A., 1999. Geostatistics for Engineers and Earth Scientists. Kluwer Academic Publishers,
Boston, MA, 303 pp.
Panofsky, H. A., 1949. Objective weather map analysis. J. Meteorol. 6, 386–392.
Privalsky, V., 1990. Statistical analysis and predictability of Lake Erie water level variations. In:
H. C. Hurtmann and M. J. Donalhue (Eds.), Proc Great Lakes Water Level Forecasting and
Statistics Symposium. Great Lake Comission, Ann Arbor, Michigan, pp. 255–264.
Salas, M., 1993. Analysis and modeling of hydrologic time series. In: D. A. Maindment (Ed.),
Handbook of Hydrology. McGraw-Hill, New York, pp. 19.1–19.72.
Sasaki, Y., 1958. An objective analysis based on variational method. J. Meteorol. Soc. Japan 36(3),
77–88.
Schlatter, T. W., and Branstator, G. W., 1987. Experiments with a three-dimensional statistical
objective analysis scheme using FGGE data. Monthly Weather Rev. 115, 272–296.
Schlatter, T. W., 1988. Past and present trends in the objective analysis of meteorological data for
now-casting and numerical forecasting. Eight Conference on Numerical Weather Prediction,
American Meteorological Society, pp. j9–j25.
Seo, D. J., 1998. Real-time estimation of rainfall fields using radar rainfall and rain gage data. J.
Hydrol. 208(1–2), 37–52.
Şen, Z., 1980. Adaptive Fourier analysis of periodic-stochastic hydrologic series. J. Hydrol. 46,
239–249.
References 269

Şen, Z., 1983. Predictive hydrologic models in hydrology. Nordic Hydrol. 14, 19–32.
Şen, Z., 1997. Objective analysis of cumulative semivariogram technique and its application in
Turkey. J. Appl. Meteorol. 36, 1712–1720.
Şen, Z., Kadıoğlu, M., and Batur, E., 1999. Cluster regression model and level fluctuation features
of Van Lake, Turkey. Ann. Geophysicae, 17, 273–279.
Şen, Z., and Habib, Z., 2000. Spatial precipitation assessment with elevation by using point cumu-
lative semivariogram technique. Water Resour. Management 14, 311–325.
Şen, Z., Kadıoğlu, M., and Batur, E., 2000. Stochastic modeling of the Van Lake monthly fluctua-
tions in Turkey. Theor. Appl. Climatol. 65, 99–110.
Şen, Z., 2008. Wadi Hydrology. CRC Lewis Publishers, Boca Raton, 386 pp.
Slivitzky, M., and Mathier, L., 1993. Climatic changes during the 20th century on the Lauren-
tian Great Lakes and their impacts on hydrologic regime. NATO Advanced Study Institute,
Deauaille, France.
Sneyers, R., 1992. On the use of statistical analysis for the objective determination of climate
change. Meteorol Z. 1, 247–256.
Student, 1907. On the error of counting with a hemacytometer. Biometrika 5, 351–360.
Subyani, A. M., 2004. Geostatistical study of annual and seasonal mean rainfall patterns in south-
west Saudi Arabia. Hydrol. Sci. J. 49(5), 803–817.
Subyani, A. M., 2005. Hydrochemical identification and salinity problem of groundwater in Wadi
Yalamlam basin, Western Saudi Arabia. J. Arid Environ. 60(1), 53–66.
Vannitseni, S., and Demaree, G., 1991. Detection et modelisation des secheresses an Sahel-
proposition d’une nouvelle methodologie. Hydra! Continent 6(2), 155–171.
Yates, F., 1938. The comparative advantages of systematic and randomized arrangements in the
design of agricultural and biological experiments. Biometrika 30, 444–466.
Zadeh, L. A., 1968. Probability measures of fuzzy events. Oikos. 23, 421–427.
Chapter 6
Spatial Simulation

Abstract Earth sciences phenomena are heterogeneous, anisotropic, non-uniform,

random, especially in their spatial behaviors, and their exhaustive sampling is not
possible and the only scientific way for the regional and global assessment of these
phenomena is through suitable mathematical modeling techniques. The main pur-
pose is to deduce from limited and small samples the general characteristics and
finally the generating mechanism of the phenomenon. Development in computer sci-
ences enhanced time-consuming simulation studies, which developed into a feasi-
ble and practical tool to reveal non-linear and non-equilibrium evolution of an earth
sciences phenomenon based on the fundamental laws governing the nature. Non-
linearity and causality of a natural phenomenon are demonstrated to be consistently
explained by simulations based on fundamental laws. Simulation methodology is
timely born with practically usable starts of digital computers during 1950s. Sim-
ulation studies are essential because the existing analytical methods are not suffi-
ciently powerful to prove the validity of the fundamental laws for naturally occurring
phenomena, but simulation could do this effectively. Spatial pattern simulations in
one-, two-, and three-dimensional space are explained through autoregressive mod-
els, which are based on the spatial correlation function. Rock masses are fractured
into different sizes and they are intact from each other due to fracturing. Different
simulation techniques are presented for dependent (persistent) and independent frac-
ture patterns, which are exemplified in the field by scanline measurements. Along
this line, Rock Quality Designation (RQD) quantification is furnished through var-
ious simulation models and final products are presented in terms of various charts
that can be used in practical applications. Porous medium is very significant for
water, oil, and gas storages, and therefore its spatial behavior is significant for effec-
tive assessments, planning, operation, and management. Autorun spatial modeling
technique is developed and applied for the simulation of porous medium. Finally,
the application of the cumulative semivariogram technique is presented for intact
length simulation in rock masses.

Keywords Autoregressive models · Autorun · Simulation · Fracture

network · Rock quality designation · Intact length · Persistence · Markov
process · ARIMA process · Multi-directional simulation

Z. Şen, Spatial Modeling Principles in Earth Sciences, 271

DOI 10.1007/978-1-4020-9672-3_6,  C Springer Science+Business Media B.V. 2009
272 6 Spatial Simulation

6.1 General
Spatial simulation models of geological phenomenon such as ore grades,
groundwater level elevations, porosity, chemical compositions of different litholog-
ical units, fracture spacing are bound to be increasingly important, due to their abil-
ity to model the underlying generating mechanism of these phenomena. The first
approximations in quantifying the geological phenomena were rather deterministic
and they did not take into account any factor of chance in their descriptions. In fact,
the variability in these phenomena was known to geologists for many decades, but
due to the deterministic training the solutions sought were also deterministic. For
instance, at most the arithmetical average of the data concerned is calculated and
then this value is treated as the best estimator of the phenomena. This way of calcu-
lation gave rise to over-estimations, which consequently made the geologists aware
of the fact that the variability within the data should definitely be taken into con-
sideration (Krige, 1951). This awareness led geologists to consider the frequency
distribution functions of any spatial variable. Hence, a new trend of statistical meth-
ods in analyzing the geological data was started (Krumbein, 1970; Agterberg, 1975).
It is a mere application of the statistical methods to process geologic data.
Later, it was recognized that, in addition to variability inherent in the data, the
interdependence either within the data themselves or with other variables is very
important due to the continuity of spatial variables. In order to account for the depen-
dence either the correlation techniques (Matern, 1960; Switzer, 1965) or the SVs are
employed (Matheron, 1965).
As a consequence of the aforementioned developments, the geologists started to
think about a convenient model to simulate the spatial variables so as to be able to
control them in the case of any change as well as in assessing the risks associated
with the data. Of course, developments of digital computers had a great impact on
this trend since without computers any simulation study is very tedious, needs great
patience, and is rather time-consuming, i.e., not practical.
Although simulation studies in other branches of science like engineering and
economics have started long time ago, the simulation of geological phenomena is
postponed. This is due to the fact that in other disciplines simulation is needed only
along one axis, for instance, time axis. Therefore, the simulation of one dimensional
variable is extensively available in the literature (Box and Jenkins, 1970).
General simulation models are possible for the generation of anisotropic as well
as isotropic synthetic patterns in 1D, 2D, or 3D, which have significance for the
purpose of modeling geologic properties such as ore grades, reservoir porosity,
mineral distribution, fracture spacing, aperture, orientation. General procedures for
such simulations by the autoregressive processes (including Markov processes) are
given for the model parameters estimation and synthetic pattern generation. The
model works on the square net basis and generates sequential pattern, first along
any desired direction for 1D simulation and then 2D patterns are constructed with
reference to two orthogonal 1D sequences. Applications to synthetic 2D pattern are
shown for isotropic cases with different model parameters. The extension of model
to 3D space is readily available.
6.2 3D Autoregressive Model 273

Because any geologic phenomenon occurs in space, in general, it has to be simu-

lated in 3D space. The first of spatial simulation models was proposed by Cliff and
Ord (1973), who took into consideration an overall spatial correlation coefficient, a
set of physical constants, and a random component all within a linear model concept.
Later, Journel (1974) proposed a turning-band method whereby 2D or 3D geologi-
cal variables are projected onto equally spaced lines and then 1D simulation along
these lines is transformed into 3D variables. On the other hand, Sharp and Aroian
(1985) have presented herringbone method of simulation for 2D and 3D geological
variables. However, as explained by Şen (1989a) there are some practical problems
with their model.

6.2 3D Autoregressive Model

The general form of the simulation model for 3D is similar to autoregressive models
(Markov processes), which can be given as

xi,j,k = αxi−1,j,k + βxi,j−1,k + γxi,j,k−1 + εi,j,k, (6.1)

where xi,j,k is the spatial variable at a point with coordinates i, j, and k with respect
to a reference point as in Fig. 6.1.
Herein, α, β, and γ are the model parameters and εi,j,k, is a random component
with zero mean and variance, σε 2 , and it is independent from xi–1,j,k , xi,j−1,k , and
xi,j,k−1 . This model has four parameters, namely α, β, γ, and σε 2 , to be determined
from an available set of data. In Eq. (6.1) the geologic variable, xi,j,k is a standard
variable with zero mean and unit variance. This does not cause any loss of generality
in simulation studies, since after the simulation of standard variables the final form
of them can be found as

Xi,j,k = μi,j,k + σi,j,k xi,j,k , (6.2)

in which Xi,j,k is the spatial variable at position (i, j, k) with mean, μi,j,k , and stan-
dard deviation, σi,j,k . Standardization procedure renders the original variable into

k k
k

ρ'k xi,j–
ρj(1) xi–
ρi(1) xi,j,k
ρk(1) ρ'j
j ρ'i j
i

i i xi,j,k
Autocorrelation Cross-correlation j

Fig. 6.1 3D space

274 6 Spatial Simulation

a second-order stationary variable (Şen, 1979b). Any simulation model has three
major stages in arriving at its final goal. These are

1) Given a set of observations xi,j,k (i, j, k = 1, 2, . . ., n) how to express model

parameters as functions of data? This is referred to as parameters estimation
stage.
2) How to check the validity of model for given data? This is the identification
stage, which is achieved by calculating the residuals as

εi,j,k = xi,j,k − αxi−1,j,k − βxi,j−1,k − γxi,j,k−1 . (6.3)

If the model is valid, then the sequence of εi,j,k should have an independent
(completely independent) structure. This stage is referred to as the diagnostic
checking for model suitability.
3) Usage of the suitable model to generate equally likely synthetic sets of data that
are statistically indistinguishable from the original data, i.e., they should have
the same distribution functions and average statistical parameters. This corre-
sponds to the simulation of the underlying geological phenomenon. In real-time
estimations, this is equivalent to prediction stage.

6.2.1 Parameters Estimation

The model parameters must be estimated from available data such that the simulated
variables yield statistically indistinguishable parameters in the long run. One of the
basic requirements in any simulation is to derive meaningful relationships between
the model and the statistical parameters of the available data. The number of model
parameters should be equal to the number of statistical parameters and accordingly
Eq. (6.1) allows the preservation of four statistical parameters. The expectations
(arithmetic averages) provide obsolete information because of the standardization
procedure as E(xi,j,k ) = E(xi−1,j,k ) = E(xi,j−1,k ) = E(xi,j,k−1 ) = 0, and E(εi,j,k ) = 0.
Further low-order statistical parameters are the spatial variance and three lag-one
autocorrelation coefficient each along i, j, and k axes. For the variance, the multi-
plication of both sides in Eq. (6.1) by xi,j,k and then taking expectations leads to
( )        
E x2i,j,k = αE xi−1,j,k xi,j,k + βE xi,j−1,k xi,j,k + γE xi,j,k−1 xi,j,k + E εi,j,k xi,j,k ,
(6.4)
( )
where E x2i,j,k = 1, E(xi−1,j,k xi,j,k ) = ρi (1), E(xi,j−1,k xi,j,k ) = ρj (1) and
E(xi,j,k−1 xi,j,k ) = ρk (1). Because εi,j,k is dependent on xi,j,k only, one can write
E(εi,j,k xi,j,k ) = E(εi,j,k )2 = σε 2 , the substitution of these last expressions into Eq.
(6.4) yields

αρi (1) + βρj (1) + γρk (1) + σε2 = 1, (6.5)

6.2 3D Autoregressive Model 275

in which ρi (1), ρj (1), and ρk (1) are the lag-one autocorrelation coefficients along the
i, j, and k axes, respectively. It is obvious that Eq. (6.4) reduces to Sharp and Aroian
(1985) simulation expression if α = β = γ = φ and with their notations ρi (1) =
ρ100 , ρj (1) = ρ010 , ρk (1) = ρ001 and σε 2 = σa 2 . Multiplication of both sides in Eq.
(6.1) by xi−1,j,k , xi,j−1,k , and xi,j,k−1 , respectively, and taking expectations leads to
three additional equations as

ρi (1) = α + βρk + γρj , (6.6)

ρj (1) = αρk + β + γρi , (6.7)

ρk (1) = αρj + βρi + γ, (6.8)

in which E(xi−1,j,k xi,j−1,k ) = ρk  , E(xi−1,j,k xi,j,k−1 ) = ρj  , and E(xi,j,k−1 xi,j−1,k ) =

ρi  , where ρk  , ρj  , and ρi  indicate the lag-one cross-correlation coefficients on the
k = j = i = constant planes, respectively. Furthermore, because εi,j,k is independent
of xi−i,j,k , xi,j−1,k , and xi,j,k−1 . Hence, E(εi,j,k xi−1,j,k ) = 0, E(εi,j,k xi,j−1,k ) = 0, and
E(εi,j,k xi,j,k−1 ) = 0. A sketch has been presented in Fig. 6.2 to show the auto- and
cross-correlations.
The four unknown model parameters can be obtained as a result of simultaneous
solution of Eqs. (6.5, 6.7, and 6.8). It is possible to arrange these equations in matrix
form explicitly as
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ρi (1) ρj (1) ρk (1) 1 α 1
⎢ 1
⎢  ρk ρj 0⎥
⎥
⎢ β ⎥ ⎢ ρi (1) ⎥
⎢ ⎥=⎢ ⎥
⎣ ρ 1 ρi 0⎦ ⎣ λ ⎦ ⎣ ρj (1) ⎦ , (6.9)
k
ρj ρi ρk (1)
1 0 o2ε

or implicitly as

AX = C, (6.10)

k k

ρ
ρ
ρ
j j

i i
ρi(1) = ρj(1) = ρk(1) ρi = ρj = ρk = ρd
Autocorrelation Cross-correlation

Fig. 6.2 Auto- and cross-correlation representations in anisotropy model

276 6 Spatial Simulation

where A is a (4×4) matrix of coefficients with its elements estimated from a given
set of data, hence it is known and dependent on the available data only; X is the
unknowns vector and includes the model parameters only; and, finally C is the vector
of lag-one autocorrelations. The solution of Eq. (6.8) requires matrix inversion of
A, after which the solution can be written as

X = A−1 C. (6.11)

It is necessary that A be a positive semi-definite matrix. Alternatively, a point worth

to notice is that Eqs. (6.6, 6.7, and 6.8) include three unknowns (α, β, and γ) and
after their simultaneous solution the substitution of these model parameters into Eq.
(6.5) yields σε 2 . The matrix notation of these three equations is
⎡ ⎤⎡ ⎤ ⎡ ⎤
1 ρk ρk α ρi (1)
⎣ρk 1 ρi ⎦ ⎣β⎦ = ⎣ ρj (1) ⎦ , (6.12)
ρj ρi 1 γ ρk (1)

in which the coefficient matrix is symmetric and positive semi-definite. Application

of Cramer’s rule leads to the following results:

ρi (1) + (ρk ρi − ρj )ρk (1) + (ρj ρi − ρk )ρj (1) − ρ2
j ρi (1)
α= , (6.13)
1 − (ρ2 2 2   
i + ρj + ρk ) + 2ρi ρj ρk

ρj (1) + (ρi ρj − ρk )ρi (1) + (ρj ρk − ρi )ρk (1) − ρ2
j ρj (1)
β= , (6.14)
1 − (ρ2 2 2  
i + ρj + ρk ) + 2ρi ρj ρk

and

ρk (1) + (ρj ρk − ρi )ρj (1) + (ρk ρi − ρj )ρi (1) − ρ2
k ρk (1)
γ= , (6.15)
1 − (ρ2 2 2   
i + ρj + ρk ) + 2ρi ρj ρk

It is clear from Eqs. (6.13, 6.14, and 6.15) that the model parameters (α, β, and γ)
are functions of the correlation structure of the geological phenomenon concerned.
There are three autocorrelation and three cross-correlation coefficients to be esti-
mated from the data. Hence, the total number of statistical parameters including
the average value and the variance to be extracted from the data is equal to eight,
provided that the geological phenomenon considered is homogeneous.

6.2.2 2D Uniform Model Parameters

In a homogeneous geologic phenomenon the mean and standard deviation values of

the samples are independent of spatial variations. Such a phenomenon is said to be
second-order stationary. In addition, if it is also isotropic, then the autocorrelations
6.2 3D Autoregressive Model 277

j j
ρi ρ

ρ'd
ρj ρ 1 2
( 1 + 8ρ2 −1)
ρ'd
i i
Anisotropic Isotropic

Fig. 6.3 Auto- and cross-correlation representations in isotropic model

and cross-correlations are equal to each other among themselves. This is tantamount
to saying that ρi (1) = ρj (1) = ρk (1) = ρ and ρi  = ρj  = ρk  = ρd where ρ and
ρd are isotropic serial and diagonal correlation coefficients. In other words, these
correlations do not depend on axial directions as shown in Fig. 6.3. Furthermore, in
an isotropic medium the autocorrelations along any direction is dependent only on
the variability along this direction. Hence, the dimensions of cubic blocks in spatial
simulation become significant. For the sake of simplicity with no loss of generality,
herein, the dimensions are assumed as units. Hence, the isotropic autocorrelation,
ρ, in Fig. 6.3 should be interpreted as the correlation between the adjacent corner
values of the phenomenon on a square mesh.
In 2D simulation this value will reduce into a square unit as shown in Fig. 6.4.

ρ = 0.10; α = 0.10; β = 0.10

Fig. 6.4 2D isotropic model

parameter
278 6 Spatial Simulation

The model parameter estimations are found from Eqs. (6.13) and (6.14) by sub-
stituting the isotropic medium auto- and cross correlations, i.e., ρk  = ρk (1) = 0,
ρi 1 = ρj (1) = ρ and ρi  = ρj  = ρd ,

ρ
α= , (6.16)
1 + ρd

and β has value identical to α. With this the 2D model mathematical expression
becomes

xi, j = α(xi−1, j + xi,j−1 ) + εi,j . (6.17)

The main problem in a 2D isotropic simulation is to satisfy the equality of cross-

diagonal correlations. This has been treated by Şen (1989b) and with the current
notations the diagonal correlation coefficient should have the following form

.1 + 8ρ2 − 1
ρd = . (6.18)
2

The variance of the independent random variable, εi,j , can be found from Eq.
(6.9) with necessary substitutions as

2ρ2
σε2 = 1.0 − . (6.19)
(1 + ρd )

Different interpretations of Eqs. (6.16) and (6.18)−(6.19) indicate already known

facts in the simulation literature as follows.

1) If the diagonal correlation is unity (ρd = 1), then Eqs. (6.16) and (6.19) reduce to
the 1D case for which α = ρ and σε 2 = 1 – ρ2 . These correspond to the properties
of lag-one Markov process.
2) For zero autocorrelation (ρ = 0), it is possible to see that α = β = 0; ρd = 0 and
σε 2 = 1. These properties imply independent random variable with zero mean
and unit variance.

For the square net with dimensions (n×n), the generation procedure requires
first the generation of n normally distributed independent random variables ε1,j (j =
1, 2, . . ., n) with zero mean and variance equal to (1 – ρ2 ). Subsequently, these
variables are converted into an auto-correlated row sequence by means of the first-
order Markov process as

x1,j = αx1,j−1 + ε1,j (j = 1, 2, . . . , n). (6.20)

The first column sequence is also generated in the same way except that the initial
(n–1) normally distributed independent variables with zero mean and variance equal
6.2 3D Autoregressive Model 279

to (1–ρ2 ) are generated. It is important that the first row value, ε1,1 , is used as the
initial value for column Markov model, which is expressed as

xi,1 = αxi−1,1 + εi,1 (i = 2, 3, . . . , n), (6.21)

in which x1,1 is taken equal to ε1,1 . Then the remainder rows (or columns) are gen-
erated through the use of 2D model as

xi,j = α(xi−1,j +xi,j−1 ) + εi,j (i, j = 2, 3, . . . , n). (6.22)

Prior to the generation of related 2D isotropic variables, it is necessary to generate

an array of (n–1)×(n–1) independent normal variables, εi,k , (i, j = 2, 3, . . ., n) with
zero mean but variance σε 2 given in Eq. (6.19). These independent variables are con-
verted into autoregressive variables through Eq. (6.22) by taking into consideration
the two previously generated 1D sequences. The simulation in 2D can be completed
either row by row- or column-wise.
The generation procedure is applied with different autocorrelation coefficients
and the sample plots are presented in Fig. 6.5. The points in these figures indicate
generated variables that have values greater than the mean value, which is taken as
zero. Correspondingly, blanks are the locations of variables less than the mean value.
The visual inspection of these figures indicates that they are isotropic, which means
to say that there is no preferential direction or stratification along any direction.

6.2.3 Extension to 3D

The simulation procedure extension into 3D space is achieved by the execution of

the following three stages.

1) 1D variables are generated along orthogonal i, j, and k directions according to

lag-one Markov process. The most important point at this stage is that these three
orthogonal direction sequences have a common original random variable which
is ε1,1,1 .
2) Three layers are generated from the three unidirectional sequences in the previ-
ous stage according to the 2D generation scheme through similar equations to
Eq. (6.22).
3) 3D isotropic simulation model parameters can be found from Eqs. (6.13) to
(6.15), by considering that ρi (1) = ρj (1) = ρk (1) = ρ and ρi  = ρj  = ρk  = ρd ,
which leads to the same value of model parameters α=β=γ as

ρ(ρd − 1)
α= . (6.23)
2ρd − ρd − 1

Hence, the 3D isotropic simulation model takes the form

280 6 Spatial Simulation

a) ρ = 0.20; α = 0.19; β = 0.19 b) ρ = 0.30; α = 0.26; β = 0.26

c) ρ = 0.40; α = 0.32; β = 0.32 d) ρ = 0.50; α = 0.37; β = 0.37

e) ρ = 0.60; α = 0.53; β = 0.53

Fig. 6.5 Two-dimensional isotropic pattern for (a) ρ = 0.1; (b) ρ = 0.2; (c) ρ = 0.3; (d) ρ = 0.4;
(e) ρ = 0.5
6.3 Rock Quality Designation Simulation 281

xi,j,k = α(xi−1,j,k + xi,j−1,k + xi,j,k−1 ) + εi,j,k . (6.24)

The substitution of the parameters into Eq. (6.5) leads after some algebra to the
variance of 3D independent variables as

3ρ2 (ρd − 1)
σε2 = 1 − . (6.25)
2ρ2d − ρd − 1

Hence, to complete 3D simulation first normal independent array (n–1)×(n–1) is

generated with zero mean and variance σε 2 as given in Eq. (6.25). These variables
are combined together with the layer auto-correlated variables in the previous stage
through Eq. (6.24) parallel to any one of the (i, j), (i, k) and (j, k) layer, and the
procedure is repeated likewise until the desired domain is covered.

6.3 Rock Quality Designation Simulation

A number of workers have devised various approaches that attempt to standardize

and quantify descriptions of rock masses, which are dissected by sets of disconti-
nuities. The very first work was due to Terzaghi (1946), who was mainly concerned
with the rock defects and loads on tunnel supports. Later, Deere (1964), Coates
(1964), Barton et al. (1974), Bieniawski (1974), Goodman (1976), Şen and Kazi
(1984), Şen (1984, 1990a), Ege (1987), and many others proposed either statistical
or empirical or analytical methodologies or modifications of existing ones in order to
describe the discontinuities quantitatively. All of the aforementioned studies sought
average values for discontinuities and their probabilistic descriptions were not men-
tioned at all. It is the main purpose of this section to seek the pdf for some of the
basic rock quality descriptions.

6.3.1 Independent Intact Lengths

Three important geometric properties of discontinuities that are of common interest
to engineering geologists are their density (spacing frequency), size (trace length
area), and orientation (strike dip direction cosines). The measure that is adopted
herein is the spacing between two successive discontinuities, which is referred to as
intact length. The intact lengths along any scanline may be measured by the separa-
tion of the intersections of discontinuities with sample lines in different directions
on the same rock outcrop as shown schematically in Fig. 6.6.
Although there exist several descriptions of rock masses proposed by Interna-
tional Society for Rock Mechanics (1978), the discontinuity spacing, i.e., intact
lengths, deserves special attention and detailed study for effective assessment of
rock quality. A simple mean for indicating the rock quality is proposed by Deere
(1964) as rock quality designation, (RQD). It is, in fact, the ratio of total intact
282 6 Spatial Simulation

Fig. 6.6 Scanline survey of Discontinuities Scanline

discontinuities

Intact lengths

lengths, each of which is greater than a pre-designated threshold value to the total
length of scanline. Due to its relative simplicity, RQD has been used extensively in
the rock classification for engineering purposes. Directly RQD or its slight modifi-
cations have been employed in a variety of engineering applications. For instance,
Piteau (1970) has used RQD for rock slope stability, Barton et al. (1974) in the
design of tunnel support, Louis and Pernot (1972) in dam foundations analysis of
permeability, and Cording and Mahar (1978) in underground chamber design in
rocks, Bieniawski (1974) and Kulhawy (1978) in estimating the strength of rock
materials.
Most of the researchers so far in the literature have concentrated in the scanline
measurement evaluations analytically (Hudson and Priest, 1979; Priest and Hudson,
1976, 1981; Şen and Kazi, 1984; Şen, 1984; Kazi and Şen, 1985) or empirically
(Cruden, 1977; Wallis and King, 1980; or simulation on digital computers by Good-
man and Smith, 1980) and Şen (1990b). In order to alleviate some drawbacks in
RQD, Şen (1990a) has proposed the concept of rock quality percentage (RQP) and
rock quality risk (RQR).
Although the quality of rock based on these measurements can be done by visual
inspection, any quantitative method is always better in unifying different opinions
about the same rock mass. Therefore, the method of RQD was adopted by Deere
(1964, 1968) and was expressed as

100 
n
RQD = Li , (6.26)
L
i=1

where the summation term includes all intact lengths Li (i = 1, 2, . . ., n) of sound

rock between discontinuities that are 0.1 m (0.3 ft) or more apart and L is the total
length of the scanline considered. It is obvious from the definition of RQD that its
values range between 0 and 100. Furthermore, Deere classified any rock based on
the RQD value as in Table 6.1.
6.3 Rock Quality Designation Simulation 283

Table 6.1 RQD classification

RQD value Classification

RQD < 25 Very poor

25 < RQD < 50 Poor
50 < RQD < 75 Fair
75 < RQD < 90 Good
90 < RQD < 100 Excellent

Initial analytical expressions for RQD based on negative exponentially dis-

tributed intact lengths are presented by Priest and Hudson (1976) for very long
scanline. The effects of finite length scanline on RQD value are described by Şen
and Kazi (1984). Further analytical studies are presented by Şen (1984) for differ-
ent intact probability distributions, such as uniform normal lognormal and gamma
pdfs. However to the authors’ knowledge no work has appeared on RQD simulation
based on the pdfs.
Although a relatively uncomplicated and fast procedure, RQD has limited appli-
cations in quantifying rock mass quality. This drawback is due to the following
disadvantages.
1) It does not reveal anything about the intact length properties, which are shorter
than 0.1 m.
2) Two or more scanline measurements with intact lengths (i.e., equally spaced
discontinuities) longer than 0.1 m would have all RQD of 100%.
3) It provides directional rock qualities, but fails to give a global quality description
for the whole rock mass. For instance, RQDs of 67% and 92% along the north–
south and west–east directions, respectively, describe the same rock mass as fair
and excellent quality. Hence a dilemma arises as to the global RQD.
4) As already noticed by Şen (1990a), RQD does not provide answers to questions
such as in what percentages does each rock quality (excellent good fair poor and
very poor) occur within the same rock mass? Or what is the probability (risk)
that the RQD value will be less than any given RQD value?
The only way to avoid these problems is to obtain the pdf of RQD itself on the
basis of given intact length pdf. Such a task is very difficult by analytical means, and
therefore Monte Carlo simulation technique is adopted in this chapter for reaching
to the desired goal.
The value of RQD along any scanline is calculated from Eq. (6.26) for the intact
lengths, which are longer than a certain threshold value t. In any rock mass the intact
lengths x are random variables, which have a certain pdf. In addition, the number
n of fractures is also an integer-valued random variable. Consequently, in a statis-
tical context RQD can be regarded as the random summation of random variables
that are truncated at the threshold value. The pdf of intact lengths can be obtained
empirically in the field from intact length measurements along very long scanline.
However, determining the pdf of the RQD requires many scanlines with different
284 6 Spatial Simulation

directions and at different places, which is quite tedious and not practical. Therefore,
the question arises how to determine the RQD distribution function from the basic
pdf of the intact lengths. It has been already shown by Şen (1990a) that an analytical
derivation of the RQD distribution function is almost impossible, and therefore the
only way to obtain it is by numerical methods using Monte Carlo techniques.
Simulation of stochastic variables is rather similar to the numerical solution
methods in mathematics. That is to say provided that the underlying properties of a
phenomenon are known, then simulation gives a way of reaching the desired goal
numerically. The desired goal here is the RQD distribution. For such a simulation
study, the following steps must be considered.
1) Determine the underlying pdfs of the intact lengths within a rock mass. In previ-
ous studies, this pdf has been taken to be either negative exponential or lognor-
mal pdf, which have the mathematical forms

f (x) = λ e−λx (0 ≤ x ≤ ∞), (6.27)

and
  
1 1 1 x
f (x) = √ exp − Ln (0 <x<∞), (6.28)
x 2π σLnx 2 σLnx mx

respectively, where λ is the average number of discontinuities, mx is the mean

of the intact lengths, and σLnx is the standard deviation of logarithmic intact
lengths. Equation (6.27) has been used in most evaluations of RQD. The loga-
rithmic distribution has not been used as often. The latter has more flexibility in
representing the intact lengths, because it takes into account the standard devia-
tion of intact lengths independently of their mean value.
In the simulation study herein, first uniformly distributed random variables
u with a range from 0 to 1 are generated and then they are transformed into
negative exponentially distributed intact lengths with parameter λ by

2.3 1
s=− log u = − Ln u. (6.29)
λ λ
However, for the logarithmic distributions first the uniformly distributed pairs of
variables (u1 and u2 ) are transformed into a normally distributed random vari-
able pairs (s1 and s2 ) with a procedure already presented by Hammersely and
Handscomb (1964) as

s1 = −2 log u1 cos (2πu2 ) . (6.30)

s21 = −2 log u1 s11 (2πu2 ) (6.31)

Herein, u1 and u2 are uniformly distributed random variables within the range
of 0 to 1. The values of s are then transformed into logarithmically distributed
intact lengths x using
6.3 Rock Quality Designation Simulation 285

x = e−x , (6.32)

in which s represents s1 or s2 . In this study 1,000,000 identically distributed

intact lengths were generated separately with negative exponential or lognormal
pdfs.
2) These intact lengths are classified into two groups depending on whether they
are greater or smaller than a threshold value t. In the RQD definition as given
by Eq. (6.26), the summation of xs over the scanline length is calculated. The
scanline length is in fact equal to the summation of the whole intact lengths,
provided that it starts and stops at discontinuity.
3) The RQDs for a set of pre-determined numbers are identified.
4) The relative and cumulative frequency distribution functions for RQD values
from (3) are calculated.

The cumulative pdf of RQD values resulting from negative exponentially dis-
tributed intact lengths with a set of average fracture number were presented in
Fig. 6.7 for five different threshold values.
On the same figure, classification of RQD values is also shown. There is a sim-
ilarity between the grain size distribution of granular rocks and these graphs which
show the rock quality distribution of the fractured hard rocks. Field experiences
show that any fractured rock is heterogeneous and accordingly more than one type
of rock quality exists within the same rock mass. The percentages of these different
qualities can be found quantitatively from Fig. 6.7. Inspection of this figure leads to
the following significant conclusions.
1) The smaller the average number of discontinuities, the smaller is the range,
which indicates the uniformity of the rock quality. For instance, invariably what-
ever the threshold value the rock has an excellent quality when the average num-
ber of discontinuities is equal to unity.
2) As the average number of discontinuities increases, the rock becomes heteroge-
neous. For instance, in Fig. 6.7 when the threshold value is 0.05, the curve that
represents 20 average intact lengths has three different qualities, namely excel-
lent, good, and fair portions. It is clear from the same curve that the majority
of values are in the “good” quality zone whereas other qualities are less likely
to occur.
3) An increase in the average number of discontinuities leads to deteriorating rock
qualities as shown in Fig. 6.8.
4) As the threshold value decreases, the rock quality increases (see Fig. 6.8). On the
other hand, for a given threshold value the deterioration rate in the rock quality
is higher at small discontinuity numbers. For instance, at 0.20 truncation level
the reduction in the rock quality is almost 12% between discontinuity numbers
2 and 4 whereas 3% from 18 to 20.
In addition, the pdfs of RQD for a set of discontinuity numbers and different
threshold values are presented in Fig. 6.9 as negative exponentially distributed intact
lengths.
286 6 Spatial Simulation

c
Fig. 6.7 RQD chart for negative exponential pdf with (a) t = 0.05; (b) t = 0.15; (c) t = 0.20
6.3 Rock Quality Designation Simulation 287

10–1 2 4 6 8 100 2 4
0

0.5

1.0

10%

10%
1.5
Depth ,d, (km)

2.0

2.5

3.0

Observed data
Regression
3.5 Gay
Haimson
Worothicki
and Deham
4.0
K(d)

Fig. 6.8 RQD threshold value discontinuity number chart (negative exponential pdf)

One of the most striking properties of these pdfs is that irrespective of the discon-
tinuity number and threshold value, they are invariably symmetric. The positions of
the maximum points on any one of these curves along the RQD axis shows the most
likely rock quality within the rock mass.
Due to the aforementioned symmetry, the average RQD value coincides with the
most likely rock quality value. This point indicates that the average RQD value is
equal to the maximum likelihood estimation of the averages resulting from these
pdfs. This value corresponds with the classical RQD value as defined by Deere.
Besides the mean mode and the median values of the RQD are equal to each other.
This last statement suggests that the pdf of RQD within a rock mass can be approx-
imated by a normal distribution as
  2 
1 1 r−μ
f (r) = √ exp − , (6.33)
2π σ 2 σ
288 6 Spatial Simulation

c
Fig. 6.9 RQD negative exponential pdfs with (a) t = 0.05; (b) t = 0.15; (c) t = 0.20

in which r is a dummy variable representing RQD; μ and σ are the population mean
and standard deviation of RQD, respectively. For the negative exponential distribu-
tion, Şen and Kazi (1984) have shown already that the RQD has the same mean and
standard deviation value, which are expressible as
6.3 Rock Quality Designation Simulation 289

c
Fig. 6.10 RQD description cart for Log-normal PDF (a) (t = 0.05, m = 0.0, s = 1.0); (b) (t =
0.15, m= 0.0, s = 1.0); (c) (t = 0.25, m = 0.0, s = 1.0)
290 6 Spatial Simulation

μ = σ = 100 (1 + λt) e−λt . (6.34)

Finally, the cumulative pdf of RQD for the underlying intact length distribution as
the lognormal pdf are given in Fig. 6.10.
All of the conclusions for the negative exponential pdfs are equally valid for these
curves. In addition, comparisons of various graphs in Fig. 6.10 indicate different
standard deviations, and an increase in the intact length standard deviation improves
the rock quality designation. In other words, the less uniform the fracture spacing
the stronger the rock mass. The following significant conclusions can be drawn from
this study.

1) Any rock mass might have different rock qualities at the same time in different
directions.
2) Rock quality deteriorates with increase of the average number of discontinu-
ities for any intact length distribution, but an increase in the standard deviation
of log-normal pdf of intact lengths leads to improvements in the rock quality
designations.
3) The pdfs of RQD are unimodal in any case, but symmetrical for the negative
exponentially distributed intact lengths.
4) Any dominant type of rock quality has almost the same maximum frequency,
confined within 0.20–0.25 for threshold levels more than or equal to 0.10 m.

6.3.2 Dependent Intact Lengths

All of the aforementioned studies have a common point in that they give an RQD
estimation without consideration of the intact length correlation. However, it is a fact
that even on the same outcrop of the rock, there might be correlated intact lengths
along scanline taken at various orientations (Eissa and Şen, 1990). Although RQD
calculations, according to Deere’s (1964) definition directly from the scanline mea-
surements, implicitly account for the intact length correlation, unfortunately, the
analytical formulations do account for this correlation neither implicitly nor explic-
itly. Accordingly, for correlated intact lengths the existing analytical results are in
error. It is obvious then that the comparison of analytical and empirical RQD esti-
mations is possible accurately only for the cases of independent intact lengths. Oth-
erwise, such a comparison is meaningless.
An important factor in the analysis of rock quality assessments from discon-
tinuity measurements along a scanline is the correlation of the intact lengths. The
autorun model and first-order autorun coefficient are proposed as a method of objec-
tively quantifying the intact length correlation structure and discontinuity occur-
rences within a rock mass (Şen, 1978, 1984).
Any straight line through the rock mass encounters random number of discon-
tinuities. An intact length is defined as the length of scanline or drill-core between
two successive discontinuities. In general, if there are n+1 discontinuity, the number
6.3 Rock Quality Designation Simulation 291

of intact lengths is n provided that the start and end of the scanline are at discontinu-
ities. A first step in rock mass classification is to consider two types of intact lengths,
namely those whose lengths are greater than a pre-designate threshold value or oth-
erwise (Eq. 6.26).
For the sake of convenience, alternative intact lengths will be grouped into two
sets as elements ai , (i = 1, 2, . . ., k) in set A and bj (j = 1, 2, . . ., l) in set B, where
k and l are the number of intact lengths in each set. It is obvious that k+l = n, which
is the total number of intact lengths. Furthermore, in an alternate sequence either
k = l–1 or l = k–1; however, practically one may assume with no loss of generality
that k = l = n/2. In short, the intact lengths along a scanline will be an alternative
combination of elements from two sets, namely A = {a1 , a2 , . . ., ak } and B = {b1 ,
b2 , . . ., bl } as shown in Fig. 6.11a. In such a combination, the correlation structure of
sequence al , bl , a2 , b2 , . . ., an/2 , bn/2 (see Fig. 6.11b) is utmost important in addition
to various statistical descriptions of intact lengths.
Besides, it may well be that the intact lengths in set A have different PDF than B.
However, this point lies outside the scope of this chapter. Of course, the assumption
of uncorrelated intact lengths simplifies the analytical derivation of RQD but at a
sacrifice of precision. Due to such an assumption, there is no term representing the
correlation of intact lengths in any RQD formulations that are available so far in
the literature. The major elements that effect the RQD calculations are in detail as
follows:

1) The number of discontinuities: As mentioned earlier, practically, half of this

number will be attached to intact lengths of set A and the second half to set B.
2) Intact length PDF: It has been assumed in the majority of RQD studies as nega-
tive exponential type (Priest and Hudson, 1976, 1981). However, the log-normal
(Roulean and Gale, 1985) or uniform and gamma pdfs (Şen, 1984) are also
employed to a certain extend.

A b1
a1 b2
a7 b
b15
a4 a10
b5
ak b1

A B
B
a1 a2 a3 a4 a5
b1 b2 b4 b5
b3
C
Fig. 6.11 Alternative intact
length concepts HHHTTTTTTHH T T T T THHHHHHTTT.......................
292 6 Spatial Simulation

3) Correlation structure: Any significant correlation effects not only the discon-
tinuity occurrences but also the intact lengths. This element is ignored invari-
ably in any of the previous RQD studies in the literature. However, only some
indirect procedures have been proposed for accounting the intact length correla-
tion (Higgs, 1984; Şen, 1991).
4) Threshold value: It is a fixed value as 0.1 m or 4 inches, below which the intact
lengths are not considered in the RQD calculations.

The first two elements are stochastic variables and in nature they are both serially
and crossly correlated. Hence, the probabilistic laws of these stochastic parts lead
to meaningful analytical expressions for RQD only after the consideration of cor-
relation structure. For instance, logically any increase in the correlation will imply
the occurrence of longer intact lengths along a scanline than the case where the
intact lengths are independent. It also implies that the number of discontinuities
decreases with increasing correlation. Hence, in general, the existence of relatively
longer intact lengths, (or lesser number of discontinuities) along a scanline implies
improvement in its quality. Consequently, the key in the analytical RQD formulation
for correlated intact lengths is the expression of correlation by an objective measure,
which is adopted herein conveniently as the autorun coefficient.

6.3.2.1 Correlation Measurement

The main question is whether the intact lengths along any direction are serially
correlated or not? As was mentioned above, different authors have shown that
the intact lengths are distributed according to various distributions among which
the normal pdf has the least significant role. On the contrary, the classical corre-
lation techniques, (auto- and cross-correlation) are valid for normally distributed
variables only. Although by suitable transformations intact lengths may be ren-
dered into normally distributed variables, such a transformation distorts the original
genuine correlation function to a greater extent (Şen, 1978, 1979a). However, the
autorun analysis is robust to any pdf or transformation and yields unbiased as well
as consistent estimates of correlation. In general, lag-k autorun coefficient is defined
by Şen (1978) as

2nk
rk = , (6.35)
n−k

in which nk is the number of overlapping successive the same type of events lag
k-apart; n is the number of unit intact lengths. From the definition, it is obvious that
0 < rk < 1. In the case of purely independent observations, whatever the underlying
pdf Eq. (6.41) becomes equal to 0.5. Therefore, 0.5 shows the fact that the two
observations separated by lag-k are independent from each other. On the other hand,
if the observations are perfectly correlated, then rk = 1.0.
The autorun coefficient application is very suitable for binary type of data; there-
fore, prior to its application the variable concerned such as the intact length must
6.3 Rock Quality Designation Simulation 293

be rendered into a binary form. For such a purpose the analogy suggested by Priest
and Hudson (1976) as an unbiased coin tossing sequence of heads and tails will be
adopted herein for alternating intact lengths, where a head represents a unit length
of intact rock of type A and a tail represents a unit length of type B. With such an
analogy, the scanline in Fig. 6.11b can be considered as a sequence of heads and
tails (see Fig. 6.11c). The following significant points emerge in such an analogy:
1) The succession of uninterrupted sequence of heads(tails) represent intact lengths
of type A(B).
2) Each appearance of alternate successive events, i.e., head–tail or tail–head suc-
cession corresponds to a discontinuity. It is obvious that two successive head–
head or tail–tail events represent two units from overall intact lengths. These
explanations indicate the suitability of lag-one autorun coefficient, r, in quanti-
fying the intact length correlation structure.
3) The percentages of heads(tails) along a scanline is equal to the probability of
type A(B) intact length. Let these probabilities be denoted by p and q, respec-
tively, then obviously p + q = 1. In terms of total length, LA (LB ) for set A(B)
intact lengths, the probability can be expressed as p = LA /L(q = LB /L).
Assuming uncorrelated intact lengths, Priest and Hudson (1976) have presented
the analytical formulation of expected RQD as

E (RQD) = 100 (1 + λt) e−λt , (6.36)

in which λ is the average number of discontinuities. The discontinuities are also

assumed to have Poisson PDF with λ as its sole parameter. Equation (6.36) should
be used only after the confirmation that the intact lengths are independent.
For instance, a scanline composed of 18, 4, 5, 3 cm intact lengths can be repre-
sented on the basis of 1 cm as an intact unit length by analogy to coin tossing as
HHHHHHHHHHHHHHHHHHTTTTHHHHHTTT
In such a sequence the probability of an intact length unit of type A (or H) is p =
23/30 = 0.77 and the probability of type B (or T) is q = 7/30 = 0.23, with p + q =1.
These probability values do not tell us whether the intact lengths are independent or
not. However, the first autorun coefficient becomes r = 19/29 = 0.66; and since it is
bigger than 0.5, there is a positive correlation between intact lengths. This implies,
in general, that long intact lengths follow long intact lengths and short ones follow
short intact lengths.

6.3.2.2 RQD Formulation and Discussion

Mathematical modeling of a scanline can be achieved by considering either the
occurrence of discontinuity numbers or intact lengths, both of which are interdepen-
dent random variables. Modeling only one of them is sufficient because it implies
the properties of the other. For instance, whatever the PDF, there is only one unique
relationship between the expected intact length, E(x), and the expected number of
discontinuities, E(n) along any scanline as
294 6 Spatial Simulation

L
E (n) = , (6.37)
E (x)

in which L is the scanline length. On the other hand, RQD as appears in Eq. (6.26)
is equivalent to the summation of random number of random variables; and first by
taking the expectations of both sides and then by considering Eq. (6.37), one can
write

100 E(x∗ )
E(RQD) = E(n)E(x∗ ) = 100 , (6.38)
L E(x)

in which E(x∗ ) is the expectation of intact lengths greater than a threshold value, t.
Due to the fact that E(x∗ ) < E(x), the ratio of expectations in the expression always
assumes a value between 0 and 100. The expectations on the right-hand side of
Eq. (6.38) can be found provided that the PDF of random variables concerned are
known.
It can be shown similar to autorun modeling (Şen, 1985) that the pdfs of k suc-
cessive heads and tails are of geometric distribution types as

P(nh = k) = (1 − r1 )rk−1
1 , (6.39)

and
 k−1
p p
P(nt = k) = (1 − r1 ) 1 − (1 − r1 ) , (6.40)
q q

respectively, where nh is the number of uninterrupted successive heads and nt suc-

cessive tails. It is worthy to notice that for r = p these equations give the independent
intact length case as available in the literature (Priest and Hudson, 1976). Further-
more, expectations of intact lengths in sets A and B can be obtained from Eqs. (6.39)
and (6.40), respectively, as

1
E(nh ) = , (6.41)
1 − r1

and
q
E(nt ) = . (6.42)
p(1 − r1 )

As mentioned above the number of set A intact lengths is one less or more than
set B intact lengths. In other words, practically they may be assumed as equals,
and therefore each type of intact length has its probability of occurrence equal to
0.5. With this information, the overall expectation of intact lengths, E(x), without
distinction between sets A and B can be seen to be E(x) = 0.5E(nh ) + 0.5E(nt ),
which yields by consideration of Eqs. (6.41) and (6.42) to
6.3 Rock Quality Designation Simulation 295

1
E(x) = (6.43)
2p(1 − r1 )

or from Eq. (6.37) one can find the expected number of discontinuities as

E(n) = 2Lp(1 − r1 ). (6.44)

In order to verify the validity of this expression, extensive computer simulations

have been carried out through the Monte Carlo techniques and the results are shown
in Fig. 6.12. It is obvious that a very good agreement has been observed between
the simulation and the analytical formulation.
Finally, the average number of discontinuities, λr1 = E(x)/L, per unit length
becomes

λr1 = 2p(1 − r1 ), (6.45)

where the Subscript signifies the correlatedness of the intact lengths. The probability
P(k) of k discontinuity occurrences along a scanline of length x at an average rate
of λr1 becomes, according to the Poisson process,

k−1
−2xp(1−r1 ) 2xp(1 − r1 )
P(k) = e . (6.46)
ký

1.0
SIMULATION
ANALYTICAL
0.7

0.8
r=
6
0.

0.6
r1

5
0.
4
0.

0.4
0.3

0.2 0.2

Fig. 6.12 Average number of

discontinuity versus scanline 0.0
0.0 0.2 0.4 0.6 0.8 1.0
length for different autorun
coefficients P
296 6 Spatial Simulation

Since the interest lies in the discontinuity spacing pdf, by considering the distance,
d, from one discontinuity to the following, one can write that P(d < x) = 1 – P
(k = 0), and hence substitution of Eq. (6.45) leads to cumulative PDF as

P(d ≤ x) = 1 − e−2xp(1−r1 ) .

By taking its derivative with respect to x the pdf, f(x), of intact lengths can be derived
as

f(x) = 2p(1 − r1 )e−2xp(1−r1 ) .

Subsequently, the expectation of intact lengths that are more than t can be found
according to

!∞
∗
E(x ) = xf (x)dx,
t
which leads after the substitution of Eq. (6.46) to

1 + 2p(1 − r1 )t −2p(1−r1 )t
E(x∗ ) = e . (6.47)
2p(1 − r1 )

Finally, the substitution of Eqs. (6.43) and (6.47) into Eq. (6.38) leads to


E(RQD) = 100 1 + 2p(1 − r1 )t e−2p(1−r1 )t . (6.48)

For independent intact lengths, r1 = 0.5 and the occurrences of unit intact lengths
comply with the Binomial distribution, which leads to geometric intact length distri-
bution with E(x) = 1/p or λ = p = 0.5, and therefore Eq. (6.48) becomes identical
to what was suggested by Priest and Hudson (1976) as in Eq. (6.41). Under the light
of aforementioned discussions, one can rewrite Eq. (6.48) as

E(RQD) = 100 [1 + λ(1 − r1 )t] e−2λ(1−r1 )t . (6.49)

The validity of this formula is checked with extensive Monte Carlo simulation
technique by using autorun model for generating correlated intact lengths as pro-
posed by Şen (1985). First of all, estimates of average intact lengths of sets A and
B are calculated as

1 
k
nA = (nA )i , (6.50)
mA
i=1
6.3 Rock Quality Designation Simulation 297

and

1 
l
nB = (nB )i , (6.51)
mB
i=1

respectively. Herein, mA and mB are the number of intact lengths; (nA )i and (nB )i
are i-th intact length in sets A and B, respectively. The geometric pdf parameter,
which is the first-order autorun coefficient, can be estimated from Eq. (5.41) as

rA = (nA − 1) / nA . (6.52)

Similarly, the geometric pdf parameter, rB , for set B intact lengths turns out to be

rB = (nB − 1) / nB . (6.53)

Hence, it is possible to generate geometrically distributed and integer-valued alter-

nate lengths, y, with parameters rA and rB , respectively, on a digital computer
through

y = 1 + log (ε) / log (r1 ), (6.54)

in which ε is the uniformly distributed random variable between 0 and 1 and r1

is the geometric distribution parameter, which assumes either the value of rA or
rB as required. The simulation results are presented in Fig. 6.13 together with the
analytical solutions and they show a very good agreement.
The following significant points can be drawn from this figure:

1) The formulation provided by Priest and Hudson (1976) for E(RQD) yields
underestimated results if the intact lengths are positively correlated, which is the
case in most of the natural rocks as will be presented in the application section
of this section.
2) Increase in the correlation structure gives rise to increase in the E(RQD) values.
3) Relatively better RQD values are obtained for the same number of discontinu-
ities but correlated intact lengths.
4) The difference between the dependent and the independent intact length RQD
values is relatively less significant at small λ values than big λs. In fact, at 5%
relative error level the correlated intact lengths do not lead to significantly dif-
ferent RQD values, provided that λ < 3 and 10% error level when λ < 10. In Fig.
5.3 upper and lower confidence limits at 5 and 10% significance levels are shown
around the Priest and Hudson independent intact length solution. It is obvious
that for small average number of discontinuities their solution gives confident
RQD estimates even though the intact lengths are dependent. However, for big
average number of discontinuities the significance of intact length correlation
becomes very pronounced in the RQD estimations.
 298
6 Spatial Simulation

Fig. 6.13 RQD – average number of discontinuity – autorun coefficient chart

6.3 Rock Quality Designation Simulation 299

6.3.2.3 Applications
The applications of the methodology developed herein are carried out for the field
data from various parts of the world. The first field data are recorded along the
exposed outcrop surfaces of granitic rocks in the western part of the Kingdom of
Saudi Arabia. Extensive geological field survey by Otaibi (1990) showed that the
area consists of one rock unit, which is granite of light pink color on fresh surface
and dark brown on weathered surface, medium to coarse grained and equi-granular.
This area was selected since it has a good combination of well-exposed bedrock
and relatively simple fracture pattern. Three sets of fracture orientations can be seen
distinctively in this area. Each one of the fracture set is measured by the scanline
perpendicular to the fracture traces. The fracture measurements are carried out at
three sites. These sites are selected such that they give rather random characteriza-
tion of the fracture geometry, i.e., they are quite independent from each other. In
order to be able to apply the methodology developed herein, the relevant values are
calculated and presented in Table 6.2.
It is obvious that the use of independent intact length RQD formulation does not
yield significant deviations from the dependent intact length case. This is due to two
major reasons. First of all, since the average number of discontinuities is all less
than 3, and therefore, as already explained in the previous section, even if the intact
lengths are strongly correlated there will not be practically significant difference,
i.e., the relative error will be less than 5%. In addition to this major reason, the
autorun coefficients are rather close to 0.5, which also confirms the approximation
in the results of E(RQD) calculations either by the use of Eqs. (6.36) or (6.47).
The second set of data for the implementation of the methodology are extracted
from a previous study by Ryckes (1984) on modes of failure and stability of rock
slopes in Tytherington quarry, which is located halfway between Thurnbury and
Tytherington for about 16 km north of Bristol in the county of Avon, England. The
Paleozoic rocks of the lower coal series in the area are affected by many movements
in the past, which led to faults, folds, and unconformities, each of which led to dif-
ferent pattern of intact lengths. Due to these different mechanisms, it is not possible

Table 6.2 Intact length characteristics (Saudi Arabia)

Site Scanline λ (1/m) p q r1 E(ROD) Error (%)

Equation (5.3) Equation (5.16)

x 1.56 0.52 0.48 0.51 98 98 0.0
1 y 2.56 0.54 0.46 0.51 97 97 0.0
z 1.04 0.62 0.38 0.63 99 99 0.0
x 2.75 0.53 0.47 0.51 97 97 0.0
2 y 1.83 0.59 0.41 0.57 98 99 1.0
z 1.98 0.63 0.37 0.61 98 99 1.0
x 1.78 0.48 0.52 0.47 98 98 0.0
3 y 1.80 0.54 0.46 0.52 98 99 1.0
z 1.56 0.57 0.63 0.55 99 99 0.0
300 6 Spatial Simulation

Table 6.3 Intact length characteristics (England)

Scanline λ (1/m) p q r1 E(RQD) Error (%)

Equation (5.3) Equation (5.16)

SL1 6.46 0.46 0.54 0.40 86 81 6
SL2 10.01 0.57 0.43 0.48 74 72 3
SL3 6.52 0.50 0.50 0.43 86 83 4

to expect that these intact lengths have independent correlation structure. In order to
depict the regional discontinuity pattern in this area, three scanlines were set up at
different directions. The bedding plane orientation is almost horizontal for the first
scanline, which will be referred to SL1. The second scanline SL2 has a direction of
20◦ toward southwest whereas the third, SL3, has an intermediate inclination to the
former. The necessary parameters as well as the E(RQD) calculations are presented
in Table 6.3.
The major difference between Tables 6.2 and 6.3 is that the average discontinuity
numbers of scanlines in Table 6.2 are far bigger than Saudi Arabian measurements;
however, the autorun coefficients in Table 6.3 are invariably less than 0.5, indicating
that there are negatively correlated intact lengths. Consequently, Priest and Hudson
formulation (Eq. 5.3) gives overestimation.

6.4 RQD and Correlated Intact Length Simulation

Generally, in open rock mechanics literature persistence implies the areal extent or
size of a discontinuity within a plane (Brown, 1981). It can be quantified crudely by
observing the discontinuity trace lengths on the surface of exposures. However, the
type of persistance herein is related to the sequential occurrences of intact lengths,
which may constitute obvious clusters. Within the contest of this chapter, persis-
tence can be defined as the tendency of short intact lengths to follow short intact
lengths and long intact lengths to follow long intact lengths. As Priest and Hudson
(1976) say, a high frequency of low spacing values occurs within clusters and a low
frequency of high spacing values occur between clusters.
The simple ways of expressing persistance in a sequence of observations is
through either the classical serial correlation coefficient or the autorun function
(Şen, 1978). Although the former requires the measurements to be normally dis-
tributed, the latter is robust and applicable in any distribution case. However, for
normally distributed intact lengths both give exactly the same result. In fact, in such
a situation the autorun coefficient r is convertible to the correlation coefficient ρ by
ρ = sinπ(r – 0.5). As a consequence, only for the normally distributed intact lengths
one can use interchangeably the autorun and autocorrelation terminologies.
In the following sequel, the intact length persistence is quantified with the help
of autorun coefficient rk for lag-k defined as in Eq. (6.35). In order to apply this
formulation to a scanline, the following steps must be executed on any scanline as
presented in Fig. 6.14:
6.4 RQD and Correlated Intact Length Simulation 301

Fig. 6.14 Scanline and e d e d e d e

models (a) effective; (b)
defective; (c) model interval (a)

1
(b)
0
1 (c)
0
Δs Δs Δs Δs Δs

1) By considering the threshold value equal to 0.1 m (or 4 in) as proposed by Deere
(1964), the effective intact lengths are assigned uniform value of +1 and defec-
tive lengths as 0 (see Fig. 6.14b). In fact, the resulting modified scanline has
zone of square waves, which are separated by zero valued intervals.
2) Divide the scanline into fixed length of finite intervals s as shown in Fig. 6.14c
where the intervals of fixed lengths are adopted as +1 cm, which is convenient
for any practical purpose. The number of such intervals along the scanline is
denoted by n.
The ratio of numbers n within the effective intervals along the scanline to the
total interval number n is equivalent to the estimate of RQD, i.e., RQD = 100
(n1 /n).
3) Find the number n of overlapping pairs of 1 and estimate r from Eq. (3.36) for
different lags from a sequence of r1 , r2 , r3 , r4 , . . .

The plot of autorun coefficient versus lag value gives rise to graphs that represent
the persistance existing in the original intact sequence. A very significant result at
this point is that if practically all of the autorun are equal to zero, then the intact
lengths originate from an independent random process and hence the calculation
of RQD as defined by Deere is reliable, otherwise persistance appears implying
clusters in intact lengths and unreliability in RQD calculations.
The applications of the abovementioned persistance procedure to actual data
are achieved with considering field measurements from South England near Bris-
tol (Ryckes, 1984) and measurements from the western province of Saudi Arabia
collected by authors.
The Paloezoic rocks of the lower coal series in the area are affected by vari-
ous earth movements in the past, which led to faults folds and unconformaties.
The location of intact length measurement sites are on the axis of a syncline,
which inclines toward the south. The bedding planes orientation is almost hori-
zontal for the first scanline (SL1). The second scanline (SL2) has a direction of
20o toward southwest and the third scanline (SL3) has an intermediate inclination
to the former. Further detailed information on these scanline measurements can
be found in Ryckes (1984). The fourth scanline (SL4) was an illustrative exam-
ple adopted from Brady and Brown (1985). However, the fifth scanline (SL5) is
a representative scanline measurement in crystalline dioritic rocks in southwest of
Saudi Arabia.
302 6 Spatial Simulation

Fig. 6.15 Autorun function 1.0

for (a) SL1; (b) SL2; (c) SL3;
(d) SL4; (e) SL5
0.5

0
0 5 10 15 20 25
lag
1.0 a

0.5

0
0 5 10 15 20 25
lag
b
1.0

0.5

0
0 5 10 15 20 25
1.0 lag
c

0.5

0
0 5 10 15 20 25
1.0 lag
d

0.5

0
0 5 10 15 20 25
lag
e
6.4 RQD and Correlated Intact Length Simulation 303

Three scanline measurements from England lead to the autorun functions as

shown in Fig. 6.15, and for the illustrative example and the Saudi Arabia scanline
measurements autorun functions are given in Fig. 6.15d and e.
A common property in all these figures is that the first autorun coefficient has
the greatest value among other lags, and there appears an exponential or power type
of decrease with increasing lag value. It is clear that the classical RQD calculations
without considering persistance, i.e., assuming that rk = 0 for all lag-k are erroneous
and should be rectified accordingly. Logically, the same RQD value may correspond
to different persistance levels or vice versa.
The RQD and persistance values for different scanlines considered herein are
shown in Table 6.4

Table 6.4 Scanline rock qualities

Scanline RQD (%) Quality Persistence Quality

SL1 88 Very good 0.68 Fair

SL2 74 Fair 0.82 Very good
SL3 86 Very good 0.73 Fair
SL4 89 Very good 0.78 Very good
SL5 86 Very good 0.70 Fair

A striking conflict appears from this table; so far, the rock qualities on RQD and
persistance basis are concerned. For the fourth scanline although the same quality
conclusion is arrived, other scanlines are in conflict to some extent. Hence a dilenma
arises as to which one is to be chosen for decision-making. The view taken in this
research is that persistance-based qualities should be preferred over the RQD-based
descriptions since the former takes into consideration not only the mean and vari-
ance of intact lengths but also their correlation structure, i.e., clustering effects.

6.4.1 Proposed Models of Persistance

Provided that the autorun functions are available, the question arises as to what type
of dependent process represents best the intact length occurrences within a rock
mass. A straight forward answer to this question can be given, provided that the
autocorrelation structures of different theoretical stochastic processes are known
beforehand. For this purpose three stochastic processes will be described herein,
namely the independent random process, the lag-one Markov process, and the
autoregressive integrated moving average process ARIMA (1, 1).

6.4.1.1 The Independent Process

It does not have any autorun function, i.e., all of the autocorrelation coefficients
at each lag except lagzero are equal to zero as shown in Fig. 6.16a. It is the easi-
est and simplest stochastic process employed in representing natural phenomenon.
304 6 Spatial Simulation

1.0 1.0

0.8 0.8

0.6 ρ = 0.9
0.6
ρk

ρk
0.4 0.4

0.2 0.2

0.0 0.0
0 2 4 6 0 2 4 6 8 10 12 14 16
Lag, k Lag, k
a b
1.0 θ = 0.1
φ = 0.1
0.8

0.6
ρk

0.4

0.2

0.0
0 2 4 6 8 10 12 14 16
Lag, k
c

Fig. 6.16 Autocorrelation structure of different processes (a) independent process; (b) Markov
process; and (c) ARIMA(1, 1) process

Two parameters, namely the mean and standard deviation, are enough to describe
the phenomenon completely. In fact, all of the RQD, RQR, RQP, etc., have been
based on the understanding that the phenomenon has an independent structure (Şen,
1990a). Since there has been so many studies in the past concerning this process, its
repetition is avoided herein.

6.4.1.2 First-Order Markov Process

In general, geologic variables possess dependent structure, especially when they are
at close intervals from each other. The simplest way of representing any dependence
is through autoregressive processes such as the lagone Markov process (Şen, 1974).
The mathematical form of this process can be written as

lı = μ + (li−1 − μ) + σ 1 − ρ2 εi , (6.55)

in which μ, σ, and ρ are the mean, standard deviation, and first-order serial correla-
tion coefficient, respectively, of intact lengths li and, finally, εi is a normal random
shock with zero mean and unit variance. The autocorrelation structure of this process
is given as
6.4 RQD and Correlated Intact Length Simulation 305

ρ0 = 1
(6.56)
ρi = ρi ,

which has been presented for different ρ values in Fig. 6.16b.

6.4.1.3 ARIMA (1, 1) Process

It is used in order to represent more persistently dependent processes, which are
noticed during intact length analysis in this study. These processs were prepared
first by Box and Jenkins (1970) with the following expression,

lı = μ + φ (li−1 − μ) + εi − θεi−1 , (6.57)

in which φ and θ are model parameters and εi is again normal random process with
zero mean and unit variance. The autocorrelation structure of this model is given as

ρ0 = 1
(1 − φθ) (φ − θ)
ρ1 =   (6.58)
1 + θ2 − 1φθ
ρk = φρk−1 (for k ≥ 2) ,

which is drawn for same set of parameters in Fig. 6.16c. Comparisons of graphs
in Fig. 6.16 indicate that only ARIMA (1, 1) processes lead to significantly more
persistant correlations at large lags than other processes.

6.4.2 Simulation of Intact Lengths

The purpose of this section is to derive the relationship between the RQD and
the first-order autorun correlation coefficient as a representative of persistance. As
already mentioned above, since the autorun functions in Fig. 6.15 indicate persis-
tance structure in the intact length occurrences, they cannot be represented by clas-
sical processes such as the negative exponential distribution, which is based on the
assumption that the intact lengths are generated from an independent process. This
implies autocorrelation structure as in Fig. 6.16a. It is obvious from Fig. 6.15 that
all of the scanlines considered in this study have autorun functions with signifi-
icant persistances. In order to be able to simulate these persistant patterns, the intact
lengths are considered to be normally distributed so that the autorun functions can
be considered as equivalents to the autocorrelation functions, which are necessary
ingradients for simulation by use of the Markov and ARIMA(1, 1) processes.
Hence the first step in the simulation is to find the best stochastic model that
fits the experimental autocorrelation functions. Therefore, the experimental auto-
correlation funct ions that are already given in Fig. 6.15 are compared with a set
of theoretical autocorrelation structures as obtained from Eqs. (6.50) and (6.57) for
Markov and ARIMA(1, 1) models, respectively, by considering different parameter
sets. In fact, the identification of model parameters such as ρ, φ, and θ are made
306 6 Spatial Simulation

Fig. 6.17 Experimental and 1.0

theoretical autocorrelation
function for (a) SL1; (b) SL2; Markov
(c) SL3; (d) SL4; (e) SL5 θ = 0.85
0.5 ARIMA
θ = 0.80

0
0 5 10 15 20 25
lag

1.0
a

Markov
θ = 0.85 ARIMA
0.5

0
0 5 10 15 20 25
lag
b
1.0

Markov
θ = 0.85
0.5 ARIMA
θ = 0.80

0
0 5 10 15 20 25
lag
1.0 c

Markov
θ = 0.86
0.5 ARIMA
θ = 0.82

0
0 5 10 15 20 25
lag
1.0 d

Markov
θ = 0.80
0.5
ARIMA
θ = 0.75

0
0 5 10 15 20 25
lag
e
6.4 RQD and Correlated Intact Length Simulation 307

by visual comparison of experimental and theoritical autocorrelation functions as

presented in Fig. 6.17.
It is obvious that for the Markov process there appears to be only one theoretical
autocorrelation function based on the first-order autocorrelation coefficient whereas
ARIMA(1, 1) process has different functions for the same correlation coefficient
depending on the values of φ and θ. In deciding on the best fit only the first onethird
of the experimental autocorrelat ion must be considered as reliable. On this basis,
it appeared that only SL2 had a very good match with the Markov process and the
others accord by the ARIMA(1, 1) process but with different set of parameters.
Table 6.5 exhibits the model types and parameter values for each scanline.
It is obvious from this table that the least persistant scanline on the basis of auto-
correlation function is SL2. However, positive φ–1 values for the other scanlines
indicate that the rock quality is comparatively better than SL1.
After the availability of the model parameters from Table 6.5 it is possible to
generate synthetic scanlines by using the appropriate model on digital computers.
The benefit from such a simulation is to gain detailed insight into the possibilities of
the fracture distribution within the rock mass. In fact, simulation results in as many
scanlines as required, which are statistically indistinguishable from each other.
This is tantamount to saying that along each synthetic scanline the occurrence of
intact lengths will be definitely different, but in the long run they will have the same
statistical parameters such as the mean and variance as well as the same autocorre-
lation structure. In order to explore the relationship between the persistance and the
RQD, the following set of different simulations are undertaken.
1. Irrespective of the scanlines considered in this chapter, first of all the relation-
ship between the RQD and the persistance parameters (ρ, φ, and θ) are sought for
standard intact lengths, i.e., the mean and variance of intact lengths are assumed
to be zero and one, respectively. Hence, the generated synthetic scanlines are not
affected by any other parameter except the persistance. However, these relationships
are obtained for different treshold values as 0.05, 0.10, and 0.15 m. The resulting
standard curves for the Markov process are presented in Fig. 6.18.
2. It is obvious that in the Markov process case the RQD is proportionally related
to the persistance (in this case ρ), but the relationship is non-linear. RQD is not
sensitive to changes in ρ for small ρ-values, for instance, in the case of threshold

Table 6.5 Models and parameters

Parameters
Scanline No. Model ρ φ θ μ σ

SL1 ARIMA(1,1) 0.68 0.85 0.333 13.77 14.78

SL2 MARKOV 0.82 – – 9.98 10.13
SL3 ARIMA(1,1) 0.73 0.80 0.152 15.33 14.05
SL4 ARIMA(1,1) 0.78 0.82 0.103 19.78 17.34
SL5 ARIMA(1,1) 0.70 0.80 0.200 14.38 9.49
308 6 Spatial Simulation

Fig. 6.18 RQD–persistence 100

relation (Markov process)

80

60 1 = 5cm

RQD
40
10

20
15

0 0.2 0.4 0.6 0.8 1.0

ρ

0.10 m when ρ < 0.5. However, for ρ >0.5 the sensitivity increases enormously
and therefore the classical RQD calculations may not be reliable, especially when
0.5 < ρ < 1.0, provided that the underlying generating mechanism of intact lengths
is of Markov type. On the other hand, Fig. 5.9 also indicates that the rock quality
improves if a low threshold value such as 0.05 m is adopted for the basic RQD
definition. On the other hand, the comparison of three curves in the same figure
shows that the rock quality improvement from threshold value 0.15 to 0.10 m is less
than the transition from 0.10 to 0.05 m.
The same type of simulation, but for the ARIMA(1, 1) model, leads to standard
RQD–persistance relationship as in Figs. 6.19 and 6.20, which present only two
samples from an infinite number of such relationships.

100

80

60 1 = 5cm
RQD

40
10

15
20

Fig. 6.19 RQD–persistence

relation (ARIMA process 0 0.2 0.4 0.6 0.8 1.0
φ = 0.9) ρ
6.4 RQD and Correlated Intact Length Simulation 309

Fig. 6.20 RQD–persistence 100

relation (ARIMA process
φ = 0.7)
80

1 = 5cm
60

RQD
40

10
20
15

0 0.2 0.4 0.6 0.8 1.0

ρ

First comparison of Figs. 6.19 and 6.20 with Fig. 6.18 shows that the ARIMA
(1, 1) process implies better rock quality for the same threshold value. This is indeed
the logically expected result from the previous discussions. On the other hand, the
sensitivity of RQD to persistence parameter changes decreased significantly for
ρ > 0.5, but increased for ρ < 0.5. In other words, the curvature of the curves is
very small compared with the Markov process to the extent that even in same cases
with such threshold values of 0.05 m the RQD–persistance relationship appears as
a straight line. Furthermore, comparison of Fig. 6.19 with Fig. 6.20 indicates that
increase in the φ value means improvement in in the rock quality, which is due to
the fact that large φ-values imply longer intact lengths.
3. In order to see the effects of genuine parameters, i.e., mean and variance
of the scanlines, a second set of simulation with the same models are performed
on digital computers. Some of the representative results for SL1, SL2, and SL3
are presented in Fig. 6.21. A general conclusion from these figures is that the
mean and standard deviation of the intact lengths are not sufficient in calculat-
ing RQD values. For instance, consideration of Fig. 6.21 leads to the conclusion
that the least RQD value as almost 20 will appear, provided that the intact lengths
along SL1 occur independently (ρ = 0). However, in addition to these two basic
parameters, the value of φ plays a dominant role by causing increment in the
RQD value. Similar statements are valid for other scanlines. Besides, compari-
son of Fig. 6.21 with Figs. 6.19 and 6.20 shows that as the mean and standard
deviation values increase, the sensitivity of persistance parameter decreases. It is
therefore plausible to conclude that the longer and less variable become the intact
lengths the more reliable will be the classical RQD calculations by Deere’s original
definition.
The main theme of this study was to emphasize that in practical studies the RQD
calculations are not reliable without the consideration of intact lengths correlation
structure. It is a well-known fact that the classical RQD evaluations are based on the
310 6 Spatial Simulation

100 100

F = 0.9 F = 0.9
80 = 0.7 80 = 0.7
= 0.5 = 0.5
= 0.3 = 0.3
= 0.1 = 0.1
60 60

RQD
RQD

40 40
Markov
Markov
20 20

0 0
0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
a b

100
F = 0.9
= 0.7
= 0.5
80 = 0.3
= 0.1
60
RQD

40 Markov

20

0
0 0.2 0.4 0.6 0.8 1.0
c
Fig. 6.21 RQD–persistence relation for (a) SL1; (b) SL2; (c) SL3

mean value percentage of the intact lengths that are greater than a threshold value
only. However, under the light of the simulation study performed in this chapter, the
following important points are worth to notice.
1) The persistance structure of intact lengths gives rise to additional rock quality
increments.
2) There exists a proportional relationship between the RQD and the persistance
parameter, which is adopted as the first-order autocorrelation coefficient in
this study.
3) İt is possible to identify the underlying generating mechanism of the intact
lengths by comparing the experimental and theoritical autocorrelation functions
of stochastic processes, such as the Markov or ARIMA(1, 1) processes.

6.5 Autorun Simulation of Porous Material

Any porous medium will have either a solid or a void at each point in space. Quanti-
tatively, solid point is represented by –1 and a void point by +1. Figure 6.22 indicates
voids as white squares and solids as black patches, and such a spatial distribution of
digitized numbers is a bivariate random field or ReV.
6.5 Autorun Simulation of Porous Material 311

Fig. 6.22 Schematic representation of a line characteristic function

Hence, ReV function, ξ(x), is a two-valued random variable in space as (Şen,

1990b)

+1 if x ∈ A
ξ (x) = , (6.59)
−1 if x ∈ Ac

where ∈ means “belongs to” A and Ac are the sets of voids and solids, respectively,
within a porous medium bulk volume, V. These two sets are mutually exclusive and
complementary, so it follows that

A ∪ Ac = V (6.60)

and

A ∩ Ac = , (6.61)

where ∪ and ∩ are union and intersection of sets, respectively, and is the empty
set. The expression “random function of coordinate” must be understood in the sense
that at each point of the 3D space, the value ξ(x) is a random variable, and conse-
quently it cannot be predicted exactly. The values of ξ(x) are subject to a certain pdf.
If the pdf is invariant with respect to a shift of the system of points, then the ReV
and corresponding porous medium are homogeneous. The same ReV is statistically
homogeneous and isotropic when the pdfs are invariant with respect to an arbitrary
rotation of the system points (such as a solid body) and to a mirror reflection of the
system with respect to an arbitrary plane passing through the origin of the coordi-
nate system. In other words, the statistical moments depend upon the configuration
of the grain–void system for which they are formed, but not upon the position of the
system in space. In practical terms, the porous medium is isotropic if the properties
of any point are the same in all directions from that point. The medium is of hetero-
geneous composition if its nature or properties of isotropy or anisotropy vary from
one point to another in the medium. If the porous medium is statistically homoge-
neous and isotropic, then the moments do not depend on any preferred direction
within the medium.
312 6 Spatial Simulation

6.5.1 Line Characteristic Function of Porous Medium

Any internal property of porous medium can be quantified by an arbitrary sampling
line, which passes through a rock in any direction. Each point on this line corre-
sponds to either a void or a solid and the sample line characteristic function, f(s),
can be defined similarly Eq. (6.59) as

+1 if s ∈ A
f (s) = , (6.62)
−1 if s ∈ Ac

where s is the distance along the sampling line from any arbitrary origin. The graph-
ical representation of f(s) forms a square wave as it passes alternately from void to
grain (Fig. 6.22). This function represents one of the possible realizations of the
ensemble of the porous medium. It reflects the size distribution of solids and voids,
their orientation, and packing within the sandstone. In practice, f(s) may be deter-
mined by preparing a high-contrast photographic image of a thin section whose
pores have been filled with epoxy.
In order to treat f(s) with classical time-series techniques, such as auto-
covariance, spectral, autorun analysis, and so forth, it must be defined as random
variable set at n points equally spaced s apart along the sampling line. Hence, ns
is equal to the sampling line length. The ordered set of f(s) values at ‘ = 1, 2, . . ., ∞,
is called a stochastic process. If the pdf is the same for all I, the stochastic process
is said to be weakly stationary. In addition, strict stationarity implies that all of the
possible joint pdfs are functions of the distance between pairs. However, in practice,
strictly stationary processes are not usually encountered. Non-stationarities in sed-
imentary rock units may arise from the presence of distinctive layering or gradual
grading. In general, sedimentary units are considered as stochastically stationary if
the pattern of variation in a property is similar in each sampled area of a bed.

6.5.2 Autorun Analysis of Sandstone

The autorun function is especially capable of investigating the sequential properties

of random series when only two distinct values exist. Due to its distribution-free
behavior, computational simplicity, and robustness to various transformations, it is
more flexible than the classical autocorrelation function (Şen, 1978). In general, the
autorun function is defined for discrete processes as a conditional probability


r (ks) = P f (ks) > m / f (s) > m , (6.63)

where k is referred to as the lag and m as the truncation level, taken in most cases
as the median value. From the definition, it is obvious that values of the autorun
function vary between 0 and +1 for any lag-k. Let p be the probability of +1 in a
characteristic function (see Eq. 6.62) of infinite length. For a line characteristic func-
tion the truncation level can be taken as any value between +1 and –1, exclusively.
6.5 Autorun Simulation of Porous Material 313

For independent processes all of the autorun values at different lags are equal to p.
However, for any process the following autorun function properties are valid:
⎫
r (0) = +1 ⎬
0 < r (ks) < +1 . (6.64)
⎭
r (∞) = p

The small sample estimate of the lag-k autorun coefficient from finite length
characteristic functions can be obtained similar to Eq. (6.35) as

2nk
r (ks) = , (6.65)
n−k

where nk shows the number of overlapping successive +1 pairs at a distance of

ks apart and n is the number of equally spaced points in the porous medium line
characteristic function. In the case of independent process r(ks) = p for any lag
(see Fig. 6.23). If l ≥ r (ks) > p, then on the average the points lag-k apart fall
on void space; otherwise, when p > r (ks) ≥ p, they fall on grain space.
The former situation indicates positive dependence and sometimes is referred to
as persistence. It implies physically that grain points and void points follow void
points. Şen (1978) has shown the analytical relationship between the autorun and
the autocorrelation coefficients for p = 0.5 as

1 1
r (ks) = + arcsin ρks , (6.66)
2 π
where ρks is the lag-k autocorrelation coefficient. Assuming a first-order Markov
process where ρks = ρks , this last expression gives autorun functions as shown
in Fig. 6.24. It is obvious that for this case 0 ≤ ρks ≤ +1 corresponds to
0.5 ≤ ρks ≤ +1. Figure 6.24a–d represents persistence, and the autorun function
converges 0.5 asymptotically. These continuous decreases indicate the stationary
nature of the underlying sample characteristic function. On the other hand, negative
dependence is characterized by autorun functions. In Fig. 6.24, rk implies r(ks)
and r(s).
The autorun coefficient is robust and is not dependent on any particular distribu-
tion function. Furthermore, it is very convenient for binary random variables. The
physical interpretation of the auto-covariance function may be difficult or impossi-
ble to make (Jenkins and Watt, 1968). Contrarily, the autorun function has phys-
ical meaning as conditional probabilities and is usually simple to interpret. The

rk

Fig. 6.23 Autorun function 0.5

of an independent process
[r(ks) = 0.5, p = 0.5] k
314 6 Spatial Simulation

(a) (b) (c)

1.0 1.0 1.0
.8 p = 0.5 .8 p = 0.5 .8 p = 0.5
r = 0.6 r = 0.7 r = 0.8
.6 .6 .6
rk

rk

rk
.4 .4 .4
.2 .2 .2
0 0 0
0 4 8 0 4 8 0 4 8 12 16 20 24
k k k

(d)
1.0 p = 0.5
.8 r = 0.9
.6
rk

.4
.2
0
0 4 8 12 16 20 24 28 32 36
k

Fig. 6.24 Autorun function of the dependent process for different r(ks) values and p = 0.5

asymptotic value of the autorun function corresponds to the porosity of the porous
medium. However, in practice autorun coefficients for relatively large lags estimated
from Eq. (6.65) yield the approximate porosity.
Another very important physical parameter that is directly related to the autorun
function is the specific surface, σ. The specific surface is defined as the ratio of
the total surface, S , of solids (or voids) to the bulk volume of the porous medium
concerned. Hence, generally

S
σ = (6.67)
V.

Knowledge of the specific surface is important in cases of adsorption of materi-

als from the fluid flowing through the porous medium. It is also important in the
design of sand and gravel filter columns. In hydrogeology it is related to the specific
retention of water in sandstone. The finer the grains, the greater will be the specific
retention and specific surface.
The random variable, η (x), identifying this surface, can be obtained from the
basic ReV as

+1 for x ∈ Sc
η (x) = lim ξ (x) ξ (x + s) = , (6.68)
s → 0 −1 for x ∈ S

where S is the set of points in the porous medium resting on the grain surface and Sc
is the complementary set including the remaining points. Hence, –1 is important in
6.5 Autorun Simulation of Porous Material 315

the definition of grain surface. In fact, points on S physically correspond to “cross-

ings” from a grain to a void or vice versa, that is, grain boundaries. Along any sam-
pling line, these crossings occur as random points after each square wave. In the sta-
tistical literature these points are referred to as the zero crossing (Mood, 1940). The
total number of such crossings in a continuous sequence has been given originally
by Rice (1945). An approximate but practical calculation to total cross numbers,
T, in any discrete process has been given by Şen (1980) in terms of the first-order
autorun coefficient (k = 1), as

T = 2n [1 − r (s)] p. (6.69)

Here T represents the points of solid surface along any sampling line, provided that
the porous medium is isotropic and homogeneous. Hence, the estimate of specific
∧
surface σ along such a line can be determined after dividing Eq. (6.69) by the length
of this line ns
∧ 1 − r (s)
σ =2 p,
s
or more conveniently

∧ r (0) − r (s)
σ =2 p. (6.70)
s
It is important to notice at this stage that the ration in Eq. (6.60) is the slope of
the autorun function at the origin. A similar relationship has been obtained with
the autocorrelation function slope at the origin by Watson (1955). The population
(asymptotic) value of the specific surface can be found from Eq. (6.70) as s → 0,
which leads to

σ = −2r (0) p, (6.71)

where r (0) is the derivative (slope) of the autorun function at the origin. For inde-
pendent processes, r (0) = −∞ and for completely dependent processes r (0) = 0.
It is therefore expected theoretically that the specific surface is isotropic and homo-
geneous materials may take any positive value. Since processes in earth sciences fall
between the two aforementioned extremes, their specific surface values fall between
zero and infinity. It has been shown by Cramer (1938) that the specific surface of
sandstones varies in the range of 150–320 cm−1 . The finer the sandstone, the greater
the number of crossings on any sampling line, and therefore the slope of the autorun
function will be greater, leading to greater specific surfaces. It thus becomes very
obvious that fine materials will exhibit much greater specific surface than will coarse
materials. Some fine materials contain an enormous grain surface area per unit vol-
ume. The generalization of Eq. (6.71) for any preferred direction, say α, within the
anisotropic porous medium is possible after dividing it by the surface area of the
sphere with unit radius. Hence, Eq. (6.71) becomes
316 6 Spatial Simulation
!
1
σ =− rα (0) pα dα, (6.72)
2π

where rα (0) and pα are the derivatives of the autorun function at the origin and the
porosity along this preferred sampling line, respectively. The numerical calculation
of Eq. (6.72) can be achieved by taking various thin sections on different directions.

6.5.3 Autorun Modeling of Porous Media

Since the porous medium has been regarded as a realization of a stochastic process,
it is necessary to develop a model for generating pore structure that is more rep-
resentative than those based deterministically on assemblages of spheres or tubes.
However, the spherical beads and capillary tube models are assumed for analytical
purposes in solving for fluid flow on the scale of a few pores. Therefore, they are
not convenient for generating the stochastic characteristics of the porous medium.
A simulation model of the grain-void size distribution can be achieved through the
autorun technique. Although such a model does not give, on the average, new infor-
mation about the medium, it helps to generate all of the possible line characteristics
function realizations of the medium. The first simulation model in this direction has
been proposed by Roach (1968) for independent processes.
However, in general, the porous medium composition of voids and solids has
dependent structure. This is tantamount to having clusters of voids and/or solids;
that is, as a general tendency voids follow voids and solids follow solids. The first
autorun coefficient, r(s), provides a criterion to decide whether the porous medium
composition has dependent composition; that is, occurrence of any void or solid
does not affect others. However, r(s) is significantly different from the porosity,
then dependence and, therefore, clustering exists. When r(s) > p a positive depen-
dence exists, which means physically that the clustering of voids is predominant.
On the contrary, for r(s) < p clustering of solids is effective.
The probabilities of having at least two successive voids is r(s), of having three
successive voids r2 (s), and in general having at least j successive voids is

P (nν ≥ j) = [r (s)]j−1 . (6.73)

It has been shown by Feller (1967) that

P (nν = j) = P (nν ≥ j) − P (nν ≥ j + 1) . (6.74)

Hence, substitution of Eqs. (6.73) and (6.74) yields the probability of having j unin-
terrupted successive voids on an infinite sampling line as

P (nν = j) = [1 − r (s)] [r (s)]j−1 . (6.75)

6.5 Autorun Simulation of Porous Material 317

It indicates that void lengths are geometrically distributed with parameter r(s). The
expectation of void length in an infinite characteristic function can be found from
Eq. (6.75) as

1
E (nν ) = . (6.76)
1 − r (s)

For independent processes, p = r(s); this equation gives E(nυ ) = 1/(1 − p),
which has been presented by Feller (1967). The purpose of a simulation model
is to generate statistically indistinguishable synthetic characteristic functions from
the observed line characteristic function. In other words, on the average, statistical
parameters such as the mean and variance must be preserved in the synthetic char-
acteristic functions. It is possible to generate geometrically distributed void lengths,
υ, with parameter r(s) as

log ξ
ν=1+ , (6.77)
log r (s)

where υ is an integer-valued random variable and ξ is a uniformly distributed ran-

dom variable between 0 and +1. The solid lengths, g, can be generated similarly to
Eq. (6. 77) as

log ξ
g=1+ , (6.78)
log rg (s)

where rg (s) is the first autorun coefficient calculated from the line characteristics
after its multiplication by −1.
Since the void and solid lengths occur alternatively on a sampling line, the initial
length can either be selected randomly or according to the final length type on the
observed sampling line, and then void and solid lengths are arranged in sequence.
However, whichever way is adopted does not make any difference in the final
product.
The aforementioned modeling technique for the porous media based on the
autorun model has been implemented on a digital computer for parameters p = 0.5;
r(s) = rg (s) = 0.5, 0.7, and 0.9. Very long sequences of synthetic line character-
istic functions (10,000 points) have been generated, but three samples with length
of 125 points each are presented in Figs. 6.25a, 6.26a, and 6.27a, together with
their sample autorun functions in Figs. 6.25b, 6.26b, and 6.27b, respectively. Fig-
ure 6.25a represents the independent structure of void-solid occurrences resulting
from the autorun model, whereas Figs. 6.26a and 6.27a are samples of dependent
structure.
It is clear that on the average the lengths of voids or grains are shorter in Fig.
6.25a than in others; in addition Fig. 6.25a is richer in the number of crossings
than others. As the autorun coefficient increases, these lengths as well as the num-
ber of crossings increase. Line characteristic functions in Figs. 6.25a, 6.26a, and
318 6 Spatial Simulation

Fig. 6.25 Synthetic line characteristic function [r(ks) = 0.5, p = 0.5]

Fig. 6.26 Synthetic line characteristic function [r(ks) = 0.7, p = 0.5]

6.27a are the possible realizations of fine, medium, and coarse-grained sandstones,
respectively.
On the other hand, generalization of the stochastic model proposed by Roach
(1968) for finite sampling lines can be achieved by the autorun coefficient so as
to cover dependent porous medium configuration. For this purpose, the following
conditional probabilities of the combinations of two successive events at equally
spaced points along a sampling line can be written as

P (ν/ν) = r (s) ,
P (g/ν) = 1 − r (s) ,
p
P (g/g) = 1 − 1−p [1 − r (s)] , (6.79)
p
P (ν/g) = 1−p [1 − r (s)] .
6.5 Autorun Simulation of Porous Material 319

0.10 0.20
p = 0.1 p = 0.2
0.08 0.16

0.06 0.12
Nv /N

Nv /N

n=
n=

0.04 0.08

1
1

0.02 2 0.04
2
3 3
0 0
0 2 4 6 8 10 0 2 4 6 8 10
r(Δs) r(Δs)
a b

0.30 0.40

p = 0.3 p = 0.4
0.24 0.32

0.18 0.24
Nv /N
Nv /N

n=

n=

0.12 0.16
1

0.06 2 0.08 2
3 3
0 0
0 2 4 6 8 10 0 2 4 6 8 10
r(Δs) r(Δs)
c d
Fig. 6.27 (a–d) Number of voids on a line characteristic function

On the average, the probability P(υ) that a point lies in a void space is equal to
p, whereas the probability P(g) that a point is within a solid space is q = 1 − p.
The void-solid sequence probability is given in Eq. (6.79) as P(υ/g), which defines
a crossing point in the line characteristic function. The probability that a solid is
followed by a void, P(g, υ), is equal to P(g)P(υ/g) and explicitly

P (g,ν) = p [1 − r (s)] .

However, the probability P(g, υ, g) of having void length equal to 1s can be
obtained as P(g, υ)P(g/υ), or as

P (g,ν, g) = p [1 − r (s)]2 . (6.80)

320 6 Spatial Simulation

Similarly, in general, the probability of having the void length equal to ns becomes

P (g,ν, ν, ν, ν . . . , ν, g) = p [1 − r (s)]n−1 [1 − r (s)]2 . (6.81)

The number of voids, Nυ , of length ns in a set of N equally spaced points along
the sampling line can be obtained as

Nν = Np [1 − r (s)]n−1 [1 − r (s)]2 . (6.82)

For the independent process, r(s) = p, and then Eq. (6.82) yields the same result
as originally proposed by Roach (1968). Furthermore, a similar expression for the
solid can be found as

n−1
2
Ng = N(1 − p) 1 − rg (s) 1 − rg (s) . (6.83)

Figure 6.28 shows the change in the number of voids per unit length, Nυ /N with the
autorun coefficient for given void length of 1s, 2s, and 3s at different porosi-
ties, p = 0.1, 0.2, 0.3, and 0.4. It is obvious that the number of voids with length 1s
decreases continuously with the increase of the autorun coefficient. However, such
a relationship is not valid for other void lengths, because the number of voids with
lengths other than 1s has maximum values for autorun coefficients other than zero.
For instance, this maximum occurs at about r(s) = 0.3 for p = 0.3. Furthermore,
whatever the porous medium parameters are in the long run, the number of voids
of any length will be more than the voids with smaller lengths. It can be concluded
from Fig. 6.28 that the difference between the numbers of voids per unit length for

14

12

10

8
E(γck)

0 2 4 6 8 10 12 14 16 18 20
Log-K

Fig. 6.28 Theoretical standard CSV model of independent intact lengths

6.6 CSV Technique for Identification of Intact Length Correlation Structure 321

voids of length 1s and others is relatively larger at low autorun coefficients. For
a line characteristic function with N = 10,000 points, p = 0.3, and r(s) = 0.2,
the number of voids for void lengths of 1s, 2s, and 3s are 1920, 384, and 77,
whereas with r(s) = 0.8 they are 120, 96, and 77, respectively.

6.6 CSV Technique for Identification of Intact Length

Correlation Structure

Different rock quality description indices that are presented into the rock mechanics
literature, which are based on intact lengths, assume that the occurrences of discon-
tinuities and, accordingly, intact lengths are independent from each other. Unfortu-
nately, the validity of such an assumption is questionable, and therefore the final
equations for rock quality assessments need to be validated prior to their use.
Otherwise, the results can be regarded only as initial approximations, which
might be rough estimations of the rock quality. It is rather obvious that due to dif-
ferent geological phenomena the rock discontinuities appear in local cluster forms,
which shows that the intact lengths cannot be considered as independent from each
other. In order to check whether the intact lengths along a scanline are indepen-
dent or dependent from each other, herein, the CSV method has been presented in
detail. The basic concept of standard CSV is presented and so simplified that in
the case of independent intact length occurrences its variation with the total intact
length appears as a straight line, which passes from the origin. Any systematic devi-
ation from straight line implies that the intact length are dependent, and therefore
classical rock quality description index must not be used and instead measures that
take into account the correlation of intact lengths must be preferred in rock qual-
ity studies. Besides, theoretical CSV models are developed for the independent,
Markov, and ARIMA(1, 1) processes. These models provide a general guide in the
identification of intact length correlation structure as well as the numerical value of
such a correlation.
The application of the methodology developed, herein, is presented for scanline
measurements at Tytherington quarry in England. The Markovian type of correlation
structure is identified for all of the scanlines from this quarry.
Mechanical and/or hydraulic behavior of jointed rock masses require an accu-
rate representation of joint geometry including intact and trace lengths, orientation,
intensity, etc. The spacing between two successive joints along an arbitrary sam-
pling line, namely scanline, is referred to as the intact length. Most of the mechani-
cal behavior as well as the quality classification of jointed rock masses are directly
related to the intact length occurrences along different directions. According to
Terzaghi (1965), the intact lengths along the mean pole direction can be regarded
as the true spacing, whereas along any other directions they can be expressed as
a function of the true spacing and the angle between the sampling and the mean
pole directions. In case of several joint sets, within the same rock mass, it is rather
322 6 Spatial Simulation

difficult to measure true spacing for an individual set. In practice, however, scanline
measurements include intact lengths between joints belonging to different sets.
Many investigators have examined empirical joint spacing distributions based on
measured intact lengths, and generally two different types of theoretical distribution
functions were proposed, namely lognormal (Steffen, 1975; Bridges, 1975; Barton,
1977; Şen and Kazi, 1984) and negative exponential distribution (Call et al., 1976;
Priest and Hudson, 1976; Einestein and Baecher, 1983; Wallis and King, 1980).
Besides, Gamma distribution of intact lengths has been proposed by Şen (1984).
However, it seems that the negative exponential distribution appears to be the most
frequently used one due to its simplicity. In order to decide whether the intact lengths
fit the negative exponential or lognormal distribution, most often the conclusions are
based on visual comparison, but Baecher et al. (1977) have used goodness-of-fit tests
in the decision-making about the best distribution. For such a test the primary requi-
site is the histogram of intact lengths. However, any test of goodness-of-fit ignores
the internal structure in the intact lengths occurrence along a scanline, i.e., the intact
length occurrences are assumed to have a complete random behavior without any
dependence on each other. We cannot really assume that nature is nice enough to
present us with independent identically distributed intact lengths. In nature the cre-
ation of fractures depend on many interactive phenomenanon such as pressure, tem-
perature, volcanic activity, earthquakes. As a consequence, prior to any assessment
of intact lengths with the classical techniques, one should confirm their random
occurrences. All of the aforementioned studies imply that the intact lengths occur
randomly, which might not be the case in actual situations and subsequently overes-
timations result in the evaluations. Hence, a key question in the intact length evalu-
ation is whether they have an independent structure or not? Provided that the inde-
pendence is verified, then the use of classical techniques leads to reliable answers.
Otherwise, new techniques should be devised so as to take into consideration the
dependence structure of intact lengths. This point has been observed first by Higgs
(1984), who suggested an empirical technique called “the profile area method”
by which the dependence structure within the intact length measurements series
is taken into consideration indirectly but in an effective way. Later, Şen (1990b)
provided analytical formulation for the profile area method, which showed explicitly
that the rock quality measures are not only functions of simple statistical parameters
such as the mean and variance values but additionally the serial correlation coeffi-
cient. The classical techniques appear as a special case of the analytical formula-
tions when the serial correlation coefficient is practically equal to zero. On the other
hand, Eissa and Şen (1990) made an extensive computer simulation by using Monte
Carlo techniques so as to generate serially dependent intact lengths according to a
very simple Markov model. The calculations of RQD for different serial correlation
coefficients indicated that increase in correlation coefficient results in increase in
rock quality.
However, neither the empirical method (Higgs, 1984) nor the simulation tech-
nique (Eissa and Şen, 1990) is suitable for practical purposes. It is, therefore, sug-
gested herein the use of CSV concept as suggested by Şen (1989a) in evaluating
the intact length occurrences and accordingly the rock quality classification. The
6.6 CSV Technique for Identification of Intact Length Correlation Structure 323

CSV technique provides graphs that give additional interpretations in intact length
occurrences, and especially in their regional behaviors.

6.6.1 Intact Length CSV

Intact length measurements in the field are dependent on the relative positions of
the scanline within the study area. The variability of intact lengths along a scan-
line leads to the concept of regional variability of intact length with the change in
scanline orientation. This variability determines the regional behavior as well as the
predictability of the intact lengths on the basis of which the rock quality appreci-
ations can be assessed. Large variability implies that the degree of dependence of
intact length on each other might be rather small even for scanlines close to each
other. Such a variability may be a product of either one of the active geological
phenomenon such as tectonic, volcanic, depositional, erosion, recharge activities.
In order to quantify the degree of variability of regional changes, variance tech-
niques have been used already (Higgs, 1984). On the other hand, Eissa and Şen
(1990) and Şen (1991) have employed autocorrelation methods in the intact length
assessment. However, these methods cannot account for the regional dependence
due to either non-normal distribution and/or irregularity of sampling positions (Şen,
1978).
The SV method has been proposed by Matheron (1963) to get rid of the afore-
mentioned drawbacks. Its elegancy is that the regionalized variable PDF is not
important in obtaining the SV, and furthermore it is effective for irregular data
positions. It is to be recalled herein that the classical variogram, autocorrelation,
and autorun techniques all require equally spaced data values. However, the dis-
continuities along a scanline are irregularly spaced, and therefore use of classical
techniques is highly questionable, except that these techniques provide approximate
results only. Although the SV technique is suitable for irregularly spaced data, it
has practical difficulties as summarized by Şen (1989a). Among such difficulties is
the grouping of distance data into classes of equal or variable lengths for SV con-
struction, but resulting SV results in an inconsistent pattern and does not have a non
decreasing form.
However, adaptation of CSV gives with the same data always a non-decreasing
pattern without grouping of distances, but rather their arrangement in ascending
order. By this arrangement each one of the distances is considered individually in
the regional variability of the intact lengths. In general, CSV, γs (dk ) is defined as a
successive summation of squared-differences as


k ( ) 1 k
γc (dk ) = d di = (Zi − Zi−1 )2 , (6.84)
2
i=1 i=1
324 6 Spatial Simulation

in which d(di ) indicates the half-squared difference at i-th order in an ordered intact
length arrangements where superscript i indicates the rank; Zi is the intact length
corresponding to rank i.

6.6.2 Theoretical CSV Model

Equation (5.1) in its present form is useful to obtain experimental CSV from a given
set of intact lengths, which will be performed in the application. However, the inter-
pretation of these CSVs require a set of relevant theoretical models for their com-
parison. It is possible to derive theoretical models, provided that the underlying
generating mechanism of the intact lengths is known or assumed. To start with, let
us take the expectation on both sides in Eq. (6.84), leading to (after the expansion
of parenthesis term on the right-hand side)

 1 " ( 2) ( )#
E γc (hk ) = E Zi − 2E (Zi Zi−1 ) + E Z2i−1 . (6.85)
2

Let us assume that the intact lengths are second -order stationary, which implies
that E(Zi 2 ) = E(Z2 i−1 ) = σ2 and E(Zi Zi−1 ) = cov(Zi Zi−1 ) + σ2 where cov(Zi ,
Zi−1 ) indicates the covariance between Zi and Zi−1 and finally σ2 is the variance
of intact lengths. It is to be noted that second-order stationary implies that the sta-
tistical parameters are independent from i, i.e., from rank. Consideration of these
conditions and the substitution of the relevant values into Eq. (6.2)lead to
 

 
k
cov (Zi ,Zi−1 )
E γc (hk ) = σ 2 k − , (6.86)
σ2
i=1

or since from statistical time series analysis (Box and Jenkins, 1970) the ratio term
in the parenthesis is defined as the lag-k autocorrelation function, ρi , then Eq. (6.86)
can be rewritten succinctly as
% &

 
k
E γc (kk ) = σ 2
k− ρi , (6.87)
i=1

in which ρi represents the correlation within the intact length sequence and this
expression is a general formulation of the theoretical CSV. The following specific
models of CSV can be derived from the available stochastic processes in the litera-
ture as employed by Eissa and Şen (1990).
1. Independent model CSV: This is the most widely used assumption in the rock
quality assessments that have appeared in the literature (Priest and Hudson, 1976;
Hudson and Priest, 1979; Şen 1984, 1990b). The uncorrelatedness of intact lengths
implies ρ = 0, and consequently the simplest model of all emerges as
6.6 CSV Technique for Identification of Intact Length Correlation Structure 325


E γc (kk ) = kσ 2 , (6.88)

which means that in the CSV model of such intact length occurrences, the vari-
ance plays the dominant role only. The graphical representation of independent CSV
model appears as a straight line that passes through the origin as shown in Fig. 6.29.
The slope of this straight line gives the variance of intact lengths.
As the variance value becomes smaller and smaller, the independent process CSV
becomes closer to the horizontal axis, which represents the distances. For instance,
in the case of equal intact lengths along a scanline, the variance becomes zero and
hence the horizontal axis represents the CSV. Otherwise, for large variances the
CSV model becomes closer to the vertical axis, indicating significant and random
differences between successive intact lengths.
2. Markov model CSV: When the intact lengths are serially dependent and the
correlation coefficient decreases according to a power law with the increase of lag,
then the generating mechanism of them is Markov process, which has the autocor-
relation structure as ρ0 = 1, but ρi = ρ1 i (i =1, 2, . . .), where ρi is the lag-one
correlation coefficient, which may assume any value between −1 and +1. For this
model the CVS model becomes
% &

 1 − ρk1
E γc (kk ) = k 1 − ρ1 σ 2, (6.89)
1 − ρ1

which reduces to Eq. (6.87) for ρ1 = 0. Positive ρ1 value means that, the average,
long intact lengths follow long intact lengths and short intact lengths follow short
intact lengths. On the other hand, negative ρ1 implies that long intact lengths follow

14

12 5
0. 7
0.
10
9
d 0.
8 ρ=
E(γck)

0 2 4 6 8 10 12 14 16 18 20
Log-K

Fig. 6.29 Theoretical standard CSV model for Markovian intact lengths
326 6 Spatial Simulation

short intact lengths or vice versa. Equation (6.84) has power form for small k values
but becomes straight line for large k-values as
 

 ρ1
E γc (kk ) = k 1 − σ 2. (6.90)
1 − ρ1

Division

of both sides by σ2 gives rise to the definition of standard CVS,
E γc (kk ) /σ as
2

 
ρ1
 (k) = k 1 − . (6.91)
1 − ρ1

The graphical representation of (k) is given in Fig. 6.30.

3. ARIMA model CSV: These models have been proposed by Box and Jenkins
(1970) and applied to the intact length simulation by Eissa and Şen (1990). The
autocorrelation structure of this model is given as

ρ0 = 1,
(1 − φθ) (φ − θ)
ρ1 = ,
1 + θ2 − 2φθ
i−1
ρi = φρ1 ,
for i > 2. ARIMA model represents intact lengths, which are more persistent than
the Markov model case. The substitution of autocorrelation structure into Eq. (6.87)
yields, after necessary algebraic manipulation,

14

ρ = 0.9 1
12 0.
5
0.
10 d
9
8 0.
E(γck)

5
6 0.9

4 9
φ=9

0 2 4 6 8 10 12 14 16 18 20
Log-K

Fig. 6.30 Theoretical standard CSV model for ARIMA intact lengths
6.6 CSV Technique for Identification of Intact Length Correlation Structure 327
 

 1 − φk
E γc (kk ) = k − ρ1 σ 2, (6.92)
1−φ

where φ is the model parameter representing extra persistence on the Markov model.
The standard CSV model is

1 − φk
 (k) = k − ρ1 , (6.93)
1−φ

which provides power type of curves for small k values; but as k becomes big-
ger and bigger, it asymptotically approaches to a straight line portion that can be
expressed as

kρ1
 (k) = k − . (6.94)
1−φ

The graphical representation of ARIMA (1, 1) model standard CSVs are presented
in Fig. 6.31.
The comparison of this figure with the Markovian standard CSVs in Fig. 6.30
indicates that in the case of ARIMA (1, 1) model the attendance of CSV to a straight
line appears at bigger distances. It is obvious that ARIMA model CVS reduces to
Markov and independent model cases for φ = ρ1 and ρ1 = 0, respectively. Equation
(6.94) helps to identify the correlation structure of intact lengths. In order to find the
correlation coefficient, it is sufficient to take the derivative of Eq. (6.94) with respect
to k, which leads to

d (k) ρ1
γc = =1− , (6.95)
dk 1−φ

in which γc indicates the slope of standard CSV at large distance. For independent
process case ρ1 = 0 and hence γc =1.
This is the case where the RQD formulations in terms of the average discontinu-
ity number as given by Priest and Hudson (1976) or Şen (1984) for different distri-
butions is valid. Otherwise, deviations from the straight line that passes through the
origin on a standard CSV plot indicates invalidity of all these formulations since the
intact lengths are correlated.
If the correlation structure is of Markovian type, i.e., φ = ρ1 , then Eq. (6.95) can
be rearranged for the first-order (lag-one) autocorrelation coefficient estimation as

1 − γc
ρ= . (6.96)
2 − γc

Example 6.1
The field data for the implementation of the methodology developed herein are
extracted from a previous study by Ryckes (1984) on modes of failure and the sta-
bility of rock slopes. The scanline measurements are carried out at the Tytherington
328 6 Spatial Simulation

40 15

30 12

Frequency
Frequency

9
20
6
10
3

0 0
0 30 60 90 120 150 0 20 40 60 80 100
Intact length (cm) Intact length (cm)
a b

30 16

25
12
20
Frequency

Frequency

15 8

10
4
5

0 0
0 20 40 60 80 0 40 60 80 100 120
Intact length (cm) Intact length (cm)
c d
20

16
Frequency

12

0
0 20 40 60 80 100
Intact length (cm)
e
Fig. 6.31 Sample and theoretical pdfs for intact lengths along (a) SL1; (b) SL11; (c) SL12; (d)
SL2; (e) SL3

quarry, which is located halfway between Thurnbury and Tytherington about 16 km

north of Bristol in the county of Avon. The Paleozoic rocks of the lower coal series
in the area are affected by many movements in the past, which led to faults, folds,
and unconformities, each of which led to different pattern of discontinuity spacing,
6.6 CSV Technique for Identification of Intact Length Correlation Structure 329

Table 6.6 Scanline data characteristics

Parameter SL1 SL11 SL12 SL2 SL3

Length (m) 13.30 4.82 6.19 7.80 11.19

Intact length number 75 35 40 79 73
Average 15.48 13.78 15.48 10.00 15.33
Median 8.00 10.00 7.50 7.00 11.00
SD 219.5 14.87 22.18 10.14 14.05
RQD (%) 84.5 88 79.5 74 86.7
Distribution Log-normal Exponential Log-normal Exponential Log-normal
(Fig. 6.4) (Fig. 6.5) (Fig. 6.6) (Fig. 6.7) (Fig. 6.8)
χ2 -square 0.260 0.350 7.880 9.046 2.34

i.e., intact length distributions. Due to these different mechanisms it is not possible
to expect that the intact length sequence along any scanline in this area has inde-
pendent correlation structure. In order to depict the regional discontinuity pattern
in this area, three scanlines were set up at different directions. The bedding plane
orientation is almost horizontal for the first scanline, which will be referred herein
as SL1. The second scanline, SL2, has a direction of 20◦ toward southwest whereas
the third scanline, SL3, has an intermediate inclination to the former. All of the
scanlines were set up horizontally as two joint sets were recognized to be vertical.
The geometrical and statistical summaries of these three scanlines are presented in
Table 6.6. Due to its heterogeneous structure the first scanline will be examined in
two subsets, which will be referred to as SL11 and SL12.
The frequency distribution functions for scanlines are shown in Fig. 6.32.
The theoretical pdfs have been fitted to the experimental counterparts by using
χ2 tests. For five different scanlines, two types of intact length distributions, namely
negative exponential and logarithmic normal, emerge as representative. This dif-
ference in the intact length distribution indicates very clearly that their occurrence
within the rock mass is heterogeneous, i.e., direction dependent, and consequently
even from this observation one can qualitatively conclude that intact lengths have
dependent structure. However, classical statistical tests, such as the one used for dis-
tribution function identification, do not give any clue about the correlation structure
of intact lengths.
Eissa and Şen (1990) suggested the use of correlation function technique in
intact length structure exploration. However, correlation techniques are valid only
for the normally distributed intact lengths (Şen, 1978). As shown above, none
of the intact lengths at Tytherington quarry are distributed normally and conse-
quently the use of correlation techniques is not meaningful. However, the non-
normal intact lengths, especially log-normally distributed ones, can be transformed
into normal distribution easily, but the transformation of negative exponential dis-
tribution exposes great difficulties. Nevertheless, even if the transformation is pos-
sible, the transformed intact lengths will not reflect genuine properties of original
lengths.
330 6 Spatial Simulation

Cumulative semivariogram 4000 600

Cumulative semivariogram
500
3000
d 400

2000 300

200
1000
100
d

0 0
50 100 150 200 250 300 350 400 450 500 50 100 150 200 250
R R
Distance (cm) Distance (cm)
a b

1000 4000

Cumulative semivariogram
Cumulative semivariogram

800
3000

600
2000
400
d d
1000
200

0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
R Distance (cm) R Distance (cm)
c d

4000
Cumulative semivariogram

3000

d
1000

0
0 50 100 150 200 250 300 350 400
R

Fig. 6.32 Sample standard CSV of (a) SL1; (b) SL11; (c) SL12; (d) SL2; (e) SL3

As mentioned in the previous section, the CSVs are robust and valid for any
pdf. In fact, the central limit theorem of classical statistics states that whatever the
underlying probability distribution of a random variable, its successive summations
or averages will have normal distribution. The experimental CSVs for each one of
the scanlines considered in this study are presented in Fig. 6.33. At a first glance
one can observe the following significant points.
1) None of the CSVs appear as a straight line passing through the origin. This is
tantamount to saying that the intact length occurrences along any scanline cannot
be considered as independent processes, but they all emerge from dependent
6.6 CSV Technique for Identification of Intact Length Correlation Structure 331

1.0 1.0

0.8 0.8

0.6 0.6
Y Y
0.4 0.4

0.2 0.2

0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
X X

1.0

0.8

0.6
Y
0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0
X

Fig. 6.33 (a) A fracture set with 30o orientation from the north; (b) Two fractures sets with 0 and
30o orientation from the north; (c) Three fractures sets with 0, 30, and 130o orientation from the
north

processes. This further implies that in the creature of discontinuities within the
rock mass, uniform conditions did not prevail, but rather a complex combination
of multitude geological events such as tectonics, cooling, volcanic activities took
place jointly.
2) The initial portions of each experimental CSVs shifts toward the distance axis.
Such a shift implies the existence of positive correlation between successive
intact lengths. It further implies that, in general, big intact lengths follow big
intact lengths and small intact lengths follow small intact lengths. Otherwise, if
negative correlation should prevail in the occurrence of intact lengths, then the
initial CSV portion would shift toward the vertical axis.
3) Each experimental CSV becomes to fluctuate about a straight line for large dis-
tance values. As stated before, the existence of such a straight line as a portion
in the CSV implies that over the distance range of this straight line portion the
intact lengths are independent from each other. This is the only range where the
classical rock quality designation formulations keep their validity. Local devi-
ations from the straight line indicate the hidden or minor dependencies in the
intact length evaluation.
4) Over an initial distance range, R, the experimental CSV appears as a curve. This
range is defined quantitatively as the distance between the original point and the
332 6 Spatial Simulation

abscissa of the initial point of the late straight line portion, as mentioned in step
(3). It is worth mentioning, herein, that as long as the threshold value necessary
in rock quality designation (RQD) calculation is greater than this range value,
then the calculation is theoretically sound and valid. Otherwise, threshold val-
ues less than this range value will give RQD values, which are unreliable. Hence,
CSV provides a criterion for checking the validity of classical RQD values based
on 0.1 m (4 inches). Such a valuable criterion cannot be obtained with the corre-
lation techniques. These techniques are valid only for the distance domain over
which the CSV appears as a straight line.
5) Existence of straight line at large distances identifies the underlying generating
mechanism of the intact lengths as a Markov process. It is already mentioned
in the previous section that if the curvature continues even at reduced rates at
large distances, then the ARIMA(1, 1) process becomes effective. It is observed
that whole the scanlines considered in this chapter have the Markovian nature of
dependence, i.e., correlation.
6) The slope of the large distance straight line portion is related to the standard
deviation of the underlying intact lengths. It is possible to consider this slope as
the population standard deviation of the intact lengths.
7) The vertical distance between the late distance straight line and the one drawn
parallel to it passing through the origin reflects the magnitude of correlation coef-
ficient of the intact lengths. Obviously, the smaller this distance the smaller will
be the intact length correlation. Under the light of the above mentioned points,
the relevant numerical values concerning the initial range, R, vertical distance, d,
slope, γc of final straight line, and the correlation coefficient estimation from
Eq. (5.13) are summarized in Table 6.7.

It is obvious from this table that the correlation coefficient along each scanline
has more or less the same magnitude, and therefore the intact lengths dependence
may be regarded as regionally isotropic. It implies that only one type of model,
which is already identified as Markovian type, can be used in intact length descrip-
tion. The use of this model in intact length generation is outside of this chapter’s
scope. Last but not the least, the following particular points for each scanline can be
inferred from the CSV graphs comparison from Fig. 6.33.

1) The CSVs of SL1, SL2, and SL3 have the similar pattern in that they are com-
posed of initial curvature portion and large distance straight line. However, local
deviations around this straight line are rather more persistent in the case of SL1

Table 6.7 CSV parameters

Parameter SL1 SL11 SL12 SL2 SL3

Range, R(m) 0.74 0.37 1.00 0.68 0.52

Vertical distance, d(m2 ) 475.0 54.0 465.0 650.0 325.0
Slope, yc (m) 8.26 2.00 5.70 15.38 8.06
Correlation coefficient, ρ 0.495 0.497 0.493 0.460 0.489
6.7 Multidirectional RQD Simulation 333

than others. This indicates that the intact lengths along SL1 have same secondary
internal structure. In order to discover this structure SL1 is divided into two
mutually exclusive portions, namely SL11 and Sl12.
2) The CSVs for SL11 and SL12 in Fig. 6.33b and c are distinctively different than
others. They exhibit not a single straight line but two successive straight lines.
This implies that along the SL1 direction there is another secondary geological
event, which plays a role in the occurrence of discontinuities along this scanline.
A reliable explanation on this point can be arrived only after a detailed geological
study of the quarry considered.
One of the most recent and significant question about the intact lengths is whether
they have independent or dependent occurrences along a given scanline? In the
past, without consideration of such a question all of the rock quality classifica-
tions derived from the intact length properties were based on the assumption that
the intact lengths are identically and independently distributed. The difficulty was
due to the lack of a reliable technique in quantifying the correlation structure of
intact lengths. However, in this chapter standard CSV are proposed as a practical
tool in measuring the intact length correlations. The CSV calculations are straight-
forward, without any difficulty or ambiguity. The CSV graphs show the change
of half squared-differences between intact lengths with ordered distance. The intact
lengths are independent only when the standard CSV variation with distance appears
as a straight line that passes through the origin. Otherwise, they are dependent and
according to the dependence structure standard CSV graphs take different shapes.
However, they have some common property such as that at small distances they
appear as curves whereas at big distance again straight lines dominate, but their
extensions do not pass through the origin. The slope of the straight line portions
on standard CSV plot gives opportunity in calculating the intact length correlation
coefficient. For instance, if this slope is close to unity, then and only then the intact
lengths can be assumed as independent and consequently theoretical RQD relation-
ships with the average number of discontinuities are reliably used in any rock eval-
uation project. The application of methodology developed in this chapter has been
performed for the field data obtained at Tytherington quarry in England.

6.7 Multidirectional RQD Simulation

A simple fracture network model is proposed based on the negative exponential
pdf of the discontinuities in all directions. In the fracture-generating procedure, the
midpoint coordinates of each fracture is generated by a uniform pdf; the fracture
directions are simulated by normal pdfs and the fracture lengths are derived from a
log-normal pdf. Three different sets of fractures are presented on a planar surface
with different random seeds. The main purpose of this chapter is to asses the rock
quality designation, (RQD), on the basis of one, two, and three directional fracture
sets. All previous studies up to date theoretically or practically were confined to
one direction only. A graphical procedure is presented for depicting the population
334 6 Spatial Simulation

estimate of the average number of discontinuities along a scanline. The substitution

of this estimate into the relevant RQD equation provides the rock quality estimation
in a better representative manner than the classical estimates.
Fracture network models are useful tools for understanding the geotechnical and
geohydrological behaviors of rock masses. These networks can be constructed by
adopting a model for the fracture geometry (position, direction, trace length, aper-
ture, roughness, etc.), estimating the statistical distribution of the appropriate geo-
metric parameters through field measurements and then generating realizations of
statistically indistinguishable networks. Once the geometry of a particular realiza-
tion is specified, then either flow through the network or the geotechnical properties
can be studied. For example, one might benefit from such a procedure in the study of
average fracture lengths in various directions, extreme values of apertures or traces,
average permeability of fracture network under various boundary conditions, etc.
Last but not least, fracture network models might be a part of a larger study of
hydrologic response to a perturbation such as the construction of an underground
opening containing nuclear waste.
Simple simulation studies were initiated by Goodman and Smith (1980) for uni-
directional discontinuity spacing, with the purpose of obtaining an experimental
relationship between the RQD and the average number of discontinuities. On the
other hand, Baczynski (1980) has proposed a zonal concept for spatial distribu-
tion of fractures in rock masses. However, many fracture network studies have been
undertaken especially for the water flow through interconnected fractures. Among
such studies are the equivalent porous medium permeability in network of discontin-
uous fractures by Long (1983), the groundwater flow through fractures by Rouleau
(1984), mass transport in fractured media by Smith and Schwartz (1984). Recently,
extensive computer simulation studies were performed by Şen (1984, 1990b) and
Eissa and Şen (1990) in assessing the rock quality classifications. To the best of
author’s knowledge so far all of the fracture simulation studies in the geomechanics
domain are performed unidirectionally for RQD.

6.7.1 Fracture Network Model

In general, the dimensions, orientation, trace lengths, apertures, and other geome-
chanical properties of each fracture are randomly distributed in nature. Therefore,
their replications through Monte Carlo simulations lead to realizations of a possible
portion of the fractured medium with the same statistical properties as the observed
ones. Once a representative simulation model is developed, it is then possible to
obtain the geomechanical properties of the fracture network in detail. Herein, the
main geomechanical property is the directional RQD. In order to achieve such a
goal in this chapter, the discontinuity network properties are assumed and defined
by the following characteristics.
1) The discontinuities within a fracture network are assumed to occur randomly
with their midpoint coordinates, obeying the uniform pdf within a unit frame
of 1×1. It is possible to enlarge such a basic frame by suitable horizontal and
6.7 Multidirectional RQD Simulation 335

vertical scales in order to obtain desired frame dimensions. The distribution of

discontinuity midpoint location in the frame obeys a 2D Poisson process.
2) Throughout this study the discontinuities are assumed to be planar and of finite
trace lengths so that intersection by any arbitrary plane gives rise to a group of
linear traces. These traces can be represented by an average length and a measure
of variability in terms of standard deviation. In this study, the trace lengths are
assumed to originate from a lognormal pdf. Besides, this trace length distribution
is independent of the intersection plane orientation.
3) The discontinuity orientations for each set are assumed to have a normal pdf with
an average orientation from the north direction, in addition to a reasonable varia-
tion expressed by the standard deviation. The occurrences of different directions
within each set are independent from each other.
The plausible values of each one of the pdfs mentioned above can be estimated
from field studies by examining either borehole cores or exposed rock faces. How-
ever, reliable measurements of trace lengths can be obtained from extensive rock
faces only. Parameter estimations for the pdfs in describing discontinuities should be
determined from both linear and areal measurements. Priest and Hudson (1981) and
Hudson and Priest (1983) have presented brief explanations of the scanline measure-
ments, i.e., linear measurements, which provide information about the frequency of
discontinuity occurrences along a scanline as well as variability of intact lengths
and aperture sizes to a certain extent. However, areal measurements provide signif-
icant supplementary information about the discontinuity directions, trace lengths,
and midpoints.

6.7.2 RQD Analysis

In its simplest form the RQD is defined as the percentage of intact length summa-
tion, S, that is greater than a threshold value such as 0.1 m (4 in) along a scanline
length, L, as

S
RQD = 100 . (6.97)
L
The rock mass is classification is already given in Table 6.1. The practical diffi-
culties in using Eq. (6.97) are as follows

1) It gives a single sample value of RQD along one direction. The practical dif-
ficulty may arise in trying to get a different scanline direction so as to assess
the possible heterogeneities within the rock mass. For instance, along a highway
cut or tunnel excavation, only longitudinal scanlines can be measured for RQD
calculations. Lateral scanlines are possible only in large diameter tunnels and
large highway cuts. Therefore, it is necessary to set up a fracture network model
from the available scanline measurements; and then by using this model many
scanlines in a multitude of desired directions can be taken synthetically.
336 6 Spatial Simulation

2) Most often in nature, either due to narrow rock surface exposure, or weathering
or geomorphologic features, it may not be possible to make measurements along
rather long scanlines. This gives rise to another source of error or bias in the
RQD calculations according to Eq. (5.1) (Sen and Kazi, 1984). However, a well-
defined representative fracture network model provides unlimited opportunities
in making scanline measurements as long as desired.
3) The RQD calculations from Eq. (6.97) do not give any insight into the intact
length distribution. As mentioned by Şen (1984) different distributions provide
different formulations of the RQD calculation. However, in a fracture network
model, as developed in this chapter, the intact length distributions are known;
and therefore, one can make more accurate RQD estimations.
4) In a scanline measurement, different sets of fractures are measured irrespective
of their individual probability distribution functions. Therefore, it is not possible
to assess the effect of each set separately from others. However, in a fracture
network model such assessments can be made rather easily and the contribution
of each fracture set on the RQD value can be quantified individually.
5) Along the scanline adopted in the field, there may occur only intact lengths that
are all greater (or smaller) than the threshold value. In such a situation, Eq. (6.97)
leads to the rock mass as either excellent (or very poor). In order to account
completely for all of the possible intact lengths, the population distribution of
fractures at every direction should be known. Such a task can be achieved easily
once the fracture network model is established for the rock mass.

Because of the aforementioned points it is worth to investigate the RQD value

behavior within a rock mass based on a specific intact length distribution, which is
adopted herein as the negative exponential distribution.
The negative exponential distribution function, f(x) of intact lengths can be
expressed through Eq. (6.27), with λ as the average number of discontinuity, i.e.,
fractures, say along one direction. Priest and Hudson (1981) derived the expected
RQD value based on this distribution as in Eq. (6.34) with the threshold value, t.
Difference between Eqs. (6.97) and (6.27) is such that the former gives only sample
estimate of RQD whereas the latter provides population estimation. Of course, in
any study the latter type of estimation must be preferred, but its calculation is rather
difficult. Theoretically, very long scanline measurements might yield approximately
the population RQD value with the use of Eq. (6.97). However, practical difficulties
such as small rock surface exposures make such an approximation almost impossi-
ble. Besides, the short scanline estimates lead to biased estimations, i.e., underes-
timation of the RQD value even if many scanlines were available (Şen and Kazi,
1984). Therefore, in the RQD analysis, as proposed herein a new technique of its
estimation by using Eq. (6.27) is suggested. The field measurements will not be
processed as usually done by Eq. (6.97), but rather by Eq. (6.27).
To this end, let us consider the cumulative probability distribution of the intact
lengths by taking integration of Eq. (6.27) from 0 to any x value. After necessary
calculations this leads to
6.7 Multidirectional RQD Simulation 337

F (x) = 1 − e−λx , (6.98)

where F(x) is the cumulative probability, which is a non-decreasing function of the

intact length, x. It assumes any value in the range 0–1. However, in order to bring
this equation into a more practical form, first of all the probability, P(x), of an intact
length to be greater than a given length x is considered as P(x) = 1 F(x) and then
natural logarithms of both sides in Eq. (6.98) is leading to

ln P (x) = ln [1 − F (x)] = −λx. (6.99)

By taking the differentiation of both sides in this expression, it is possible to

obtain dln P(x) = −λdx hence, finally,

dx 1
=− . (6.100)
d ln P (x) λ

This equation shows a linear relationship between ln P(x) and x and the slope of
straight line is equal to 1/λ. Consequently, the following points can be advised for a
practicing geologist in order to find the λ value.

1) Find the probability, P(x), of each intact length measurement.

2) Plot on a semi logarithmic chapter intact lengths versus the probabilities on the
logarithmic axis.
3) Fit a straight line through the scatter of field points.
4) Calculate the slope value of the straight line. This slope value is, in fact, equal to
1/λ.
5) Substitute the λ value into Eq. (6.34) and find the population RQD value, μ.

The only practical difficulty in this procedure is encountered in step (1) for cal-
culating the cumulative probability. However, there appears to be two distinctive
approaches. One is the usual way of constructing a histogram for the intact lengths,
and then transforming it into a cumulative PDF. However, this application raises
practical difficulties, especially when the number of intact lengths is small, which is
almost the case in any field survey. Another way of obtaining the cumulative prob-
ability is an empirical approach, which is based on the rank, m, of any intact length
within a set of available data with size n. Hence, the cumulative probability can be
expressed as

m
P (x) = . (6.101)
n+1

This empirical approach is used very often in the practical studies, and it does
not give any difficulty in the applications. The RQD analysis based on this empirical
law, together with the generated intact lengths, is given in the following section.
338 6 Spatial Simulation

6.7.3 RQD Simulation Results

The simulation studies in this chapter are based upon a computer program that has
been developed to generate in two dimensions a random network of discontinuity
pattern. Three different orientations are adopted for the fracture orientation, each of
which is assumed to be normally distributed about a given value. Table 5.1 shows the
relevant parameters taken for the simulation of fractures in this chapter. However, in
any future study, they may be changed according to the field conditions.
This program can typically handle up to any number of sets, each containing
any desired number of discontinuities. The program finds the beginning and ending
points’ coordinates of any trace. The results of this program appear as a numerical
representation of a 2D discontinuity network. Such a network is one of the example
samples of random discontinuity geometry generated according to the input data
shown in Table 6.8.
The sample area size in these generations is considered as 1×1. However, its
transformation into any desired size is possible by the multiplication with any
desired scale. Some sample realization of generated discontinuity sets are shown
in Fig. 6.34 in detail as individual and combined fracture sets.
It is important to notice in the individual sets, i.e., Fig. 6.34a, that some fracture
traces intersect with the general frame of the area. These fractures are referred to as
censored trace lengths. The generation scheme gives their total trace lengths, but in
practice they are never known. On the other hand, some short length traces appear
entirely within the sample area. Consideration of the whole set on one sample area,
as in Fig. 6.34c, makes the picture rather complex as it is the case in nature.
In order to assess the effect of different fracture sets individually or collectively
on the RQD calculations, the sample area is overlain with scanlines in the east–
west direction at different levels between 0 and 1. Of course, it is possible to take a
scanline at any level, but the discussion here will concentrate on five different levels,
which are at 0.1, 0.3, 0.5, 0.7, and 0.9. The computer program lays these scanlines,
at these levels, on the sample area of a generated fracture network; subsequently, it
finds individually and collectively the intact lengths along them, ordering them in an
ascending order, and attaches to each intact length an order to be used in Eq. (6.101).
The results are presented on semi-logarithmic chapter as shown in Fig. 6.34.
A visual inspection of these figures indicate that there appear wide ranges of
data scatter; and with the consideration of more than one set of fracture, the

Table 6.8 Input data for fracture network generation

Set 1 Set 2 Set 3

Property Distribution Mean Variance Mean Variance Mean Variance

Orientation Normal 30∗ 9 130∗ 9 0∗ 4

Trace length (m) Log-Normal 0.2 0.05 0.18 0.04 0.16 0.04
Intact lengths (m) Negative Exponential .013 2 × 10−4 .013 2 × 10−4 .013 2 × 10−4
Location Poisson 0.5 0.5 0.5 0.5 0.5 0.5
6.7 Multidirectional RQD Simulation 339

1.0
Cumulative probability 1.0

Cumulative probability
0.1 0.1

(A) (B)
0.01 0.01
0 5 10 15 25 0 5 10 15 25
Intact length (×10–2 m) Intact length (×10–2 m)

1.0
Cumulative probability

0.1

1/λ = 0.106
λ = 9.434
(C)
0.01
0 5 10 15 25
Intact length (×10–2 m)

Fig. 6.34 Cumulative probability plots of intact lengths at the 0.1 level. (A) single set at 30o ; (B)
two sets at 0 and 30o ; (C) three sets at 0, 30 and 130o

scatter becomes less. The best straight lines are fitted through the scatter of points;
their slopes are found by considering a full cycle on the logarithmic axis. The
corresponding length on the horizontal axis is equal to the slope, i.e., 1/λ parameter.
This slope calculation procedure is presented in Fig. 64C. The results are summa-
rized in Table 6.9.

Table 6.9 Slope and RQD values for t = 0.1 m

Level One Set Two Sets Three Sets

0.1 Slope 0.267 0.120 0.106

RQD % 94.5 79.8 75.6
0.3 Slope 0.214 0.135 0.128
RQD % 92.0 83.0 81.0
0.5 Slope 0.227 0.117 0.079
RQD % 92.7 79.0 63.9
0.7 Slope 0.212 0.122 0.050
RQD % 91.8 80.0 40.6
0.9 Slope 0.188 0.092 0.076
RQD % 90.0 70.4 62.0
Average RQD % 92.2 78.4 64.6
340 6 Spatial Simulation

As is obvious from Table 6.4, the best rock quality is obtained when only one
fracture set is considered. As expected, an increase in the number of fracture sets can
cause a decrease in the RQD value. There appears a practically significant difference
between the RQD values based on the single set of fracture and those on two or three
fracture sets. The rock quality is of an excellent type when one set of fractures is
considered; whereas the quality deteriorates and becomes very good and fair for
two and three fracture sets, respectively. The last row in Table 6.4 gives average
RQD values for all the scanlines considered in this study. It is obvious from the
comparison of these values that when two sets are considered, the RQD deterioration
relative to one set is almost 15%, whereas for three sets consideration is almost 30%.

References
Agterberg, F., 1975. New problems at the interface between geostatistics and geology. Geostat. 75,
403–421.
Baczynski, N. R. P., 1980. Zonal concept for spatial distribution of fractures in rock. In Proceed-
ings of the Thrid Australian and New Zealand Conference on Geomechanics. Wellington, New
Zealand, pp. 29–33.
Baecher, G. B., Lanney, N. A., and Einstein, H. H., 1977. Statistical description of rock properties
and sampling. 18th U.S. Symposium on Rock Mechanics, Keystone, Colorado, pp. 5C1-1–
5C1-8.
Barton, C. M., 1977. Geotechnical analysis of rock structure and fabric in C.S.A. Mine, Cobar,
New South Wales. Applied Geomechanics, Technical Paper 24, CSIRO, Australia.
Barton, N., Lien, F., and Lunde, J., 1974. Engineering classification of rock masses for the design
of tunnel support. Rock Mech. 6(4), 189–236.
Bieniawski, Z. T., 1974. Geomechanic classification of rock masses and its application in tun-
neling. Proceedings of the 3rd International Congress on Rock Mechanics, Denver, Colorado,
September 1974, Vol. 2A, pp. 27–32.
Box, G. E. P., and Jenkins, G. M., 1970. Time Series Analysis, Forecasting and Control. Holden
Day, San Francisco, 475 pp.
Brady, B. H. G., and Brown, E. T., 1985. Rock Mechanics for Underground Mining. Allen-Unwin
Press, London.
Bridges, M. C., 1975. Presentation of fracture data for rock mechanics. Second Australia New
Zealand Conference on Geomechanics, Brisborne, Australia, pp. 144–148.
Brown, E. T., 1981. Rock Characterization Testing and Monitoring. Pergamon Press, London,
221 pp.
Call, R. B., Savely, J., and Nicholas, D. E., 1976. Estimation of joint set characteristics from surface
mapping data 17th U.S. Symposium on Rock Mechanics, pp. 2B21–2B29.
Cliff, A. D., and Ord, J. K., 1973. Spatial Autocorrelation. Pion, London.
Coates, D. F., 1964. Classification of rocks for rock mechanics. Int. J. Rock Mech. Min. Sci.
Geomech. Abst. 1(3), 421–429.
Cording, E. J., and Mahar, J. W., 1978. Index properties and observations for design of chambers
in rock. Eng. Geol. 12, 113–142.
Cramer, H., 1938. Random Variables and Probability Distribution Mixtures of alcohols, acids and
amines. J. Chem. Phys. 6, 847.
Cruden, D. M., 1977. Describing the size of discontinuities. Int. J. Rock Mech. Min. Sci. Geomech.
Abstr. 14, 133–137.
Deere, D. U., 1964. Technical description of rock cores for engineering purposes. Rock Mech.
Eng. Geol. 7(7), 16–22.
Deere, D. U., 1968. Geologic considerations. In: K. G. Stagg and O. C. Zienkiewicz (Eds.), Rock
Mechanics in Engineering Practice. John Wiley and Sons, New York, pp. 4–20.
References 341

Ege, J. R., 1987. Core index. A numerical corelogging procedure for estimating rock quality. US
Geol. Surv. 954, 1–15.
Einestein, H. H., and Baecher, G. B., 1983. Probabilistic and statistical methods in engineer-
ing geology, specific methods and examples. Part I: Exploration. Rock Mech. Rock Eng. 16,
39–72.
Eissa, E. A., and Şen, Z., 1990. Intact length persistance in relation to rock quality designation. Int.
J. Rock Mech. Min. Sci. Geomech. Abstr. 28, 411–419.
Feller, W., 1967. An Introduction to Probability Theory and its Application. John Wiley and Sons,
New York, 509 pp.
Goodman, R. E., 1976. Method of Geological Engineering in Discontinuous Rocks. West, San
Francisco, CA.
Goodman, R. E., and Smith, H. R., 1980. RQD and fracture spacing. J. Geotech. Eng. Division,
ASCE 106(GT2), 191–193.
Hammersely, J. M., and Handscomb, D. C., 1964. Monte Carlo Methods. Methuen and Co. Ltd.,
London, 178 pp.
Higgs, N. B., 1984. The profile-area and fracture frequency methods: two quantitative procedures
for evaluating fracture spacing in drill core. Bull. Assoc. Eng. Geol. 21(3), 377–386.
Hudson, J. A., and Priest, S. D., 1979. Discontinuities and rock mass geometry. Int. J. Rock Mech.
Min. Sci. Geomech. Abstr. 16, 339–362.
Hudson, J. A., and Priest, S. D., 1983. Discontinuity frequency in rock mass. Int J Rock Mech,
Mining Sci Geomech Abstracts. 20(2), 73–89.
International Society for Rock Mechanics, 1978. Standardization of laboratory and field tests. Sug-
gested methods for the quantitative description of discontinuities in rock masses. Int. J. Rock
Mech. Min. Sci. Geomech. Abstr. 15, 319–368.
Jenkins, G. M., and Watt, D. G., 1968. Spectral Analysis and Its Applications. HoldenDay, 523 pp.
Journel, A. G., 1974. Simulation conditionelle de gisements miniers-theorie et pratique: These de
Doctor Ingenieur, Universite de Noncy.
Kazi, A., and Şen, Z., 1985. Volumetric RQD; An index of rock quality. Proceedings of the Inter-
national Symposium on Fundamentals of Rock Joints, Bjorkliden, pp. 99–102.
Krige, D. G., 1951. A statistical approach to some basic mine valuation problems on the Witwater-
strand. J. Chem. Metall. Min. Soc. South Africa 52, 119–139.
Krumbein, W. C., 1970. Geological models in transition to geostatistics. Geostatistics. Plenum
Press, New York, pp. 143–161.
Kulhawy, F. H., 1978. Geomechanical model for rock foundation settlement. J. Geotech. Eng.
Division, ASCE 104(GT2), Proc Paper 13547, 211–228.
Long, J. C. S., 1983. Investigation of equivalent porous medium permeability in networks of dis-
continuity fractures. Unpublished Ph. D. Thesis, University of California, Berkeley, CA.
Louis, C., and Pernot, M., 1972. Three-dimensional investigation of flow conditions of Grand
Maison Dam Site. Proceedings of Symposium International Society of Rock Mechanics, Per-
colation through Fissured Rocks.
Matern, B., 1960. Spatial variation. Medd Statons Skogsforsknings institut, 144 pp.
Matheron, G., 1963. Principles of geostatistics. Econ. Geol. 58, 1246–1266.
Matheron, G., 1965. Les variables regionalises et leur estimation. Masson et Cie., Paris, 306 pp.
Mood, A. M., 1940. The distribution theory of runs. Ann. Math. Stat. 11, 427–432.
Otaibi, A., 1990. Geotechnical investigation and engineering geological maps of Al-Nagabah area.
Unpublished Thesis, Faculty of Earth Sciences, King Abdulaziz University, Jeddah.
Piteau, D. R., 1970. Analysis of the genesis and characteristics of jointing in the Nchanga Open
pit for the purpose of ultimately assessing the slope stability. Report, Nchanga Consolidated
Copper Mines Ltd.
Priest, S. D., and Hudson, J., 1976. Discontinuity spacing in rock. Int. J. Rock Mech. Min. Sci.
Geomech. Abstr. 13, 135–148.
Priest, S. D., and Hudson, J. A., 1981. Estimation of discontinuity spacing and trace length using
scanline surveys. Int. Rock Mech. Min. Sci. Geomech. Abstr. 18(3), 183–197.
342 6 Spatial Simulation

Rice, S. O., 1945. Mathematical analysis of random noise. Bell. Sys. Tech. J. 24, 46–156.
Roach, S. A., 1968. The Theory of Random Clumping. Methuen, London.
Rouleau, A., 1984. Statistical characterization and numerical simulation of a fracture system –
Application to groundwater in the Stsipe Granite. Unpublished Ph. D. Thesis, University of
Waterloo, Ontario, Canada.
Roulean, A., and Gale, J. E., 1985. Statistical characterization of the fracture system in the Stripa
Granite, Sweden. Int. J. Rock Mech. Min. Sci. Geomech. Abstr. 22(6), 353–367.
Ryckes, K. D., 1984. A routine method to evaluate the modes of failure and the stability of rock
slopes. Unpublished Thesis, Imperial College of Science, Technology and Medicine, 246 pp.
Şen, Z., 1974. Small sample properties of stationary stochastic models and the Hurst phenomenon
in Hydrology. Unpublished Ph.D. Thesis, Imperial College of Science and Technology, Uni-
versity of London, 286 pp.
Şen, Z., 1978. Autorun analysis of hydrologic time series. J. Hydrol. 36, 75–85.
Şen, Z., 1979a. Application of autorun test to hydrologic data. J. Hydrol. 42, 1–7.
Şen, Z., 1979b. Effect of periodic parameters on the autocorrelation structure of hydrologic series.
Water Resour. Res. 15(6), 1639–1642.
Şen, Z., 1980. Statistical analysis of hydrologic critical droughts. J. Hydraul. Div. ASCE, Proc.
Pap. 14134, 106(HY1), 99–115.
Şen, Z., 1984. RQD models and fracture spacing. J. Geotech. Eng. Amer. Soc. Civ. Eng. 110(2),
203–216.
Şen, Z., and Kazi, A., 1984. Discontinuity spacing and RQD estimates from finite length scanlines.
Int. J. Rock Mech. Min. Sci. Geomech. Abstr. 21(4), 203–212.
Şen, Z., 1985. Autorun model for synthetic flow generation. J. Hydrol. 81, 157–170.
Şen, Z., 1989a. Cumulative semivariogram models of regionalized variables. Int. J. Math. Geol.
21(3), 891–903.
Şen, Z., 1989b. Comment on “The generation of multi-dimensional autoregressive series by her-
ringbone method”. Math. Geol. 21(2), 267–268.
Şen, Z., 1990a. RQP RQR and fracture spacing: Technical note. Int. J. Rock Mech. Min. Sci.
Geomech. Abstr. 22, 135–137.
Şen, Z., 1990b. Spatial simulation of geologic variables. J. Math. Geol. 22(2), 175–188.
Şen, Z., 1991. The profile-area method for statistical evaluation of rock mass quality. Bull. Assoc.
Eng. Geol. XXVIII, 351–357.
Sharp, W. E., and Aroian, L. A., 1985. The generation of multidimensional autoregressive series
by herringbone method. Math. Geol. 17, 67–79.
Smith, L., and Schwartz, F. W., 1984. An application of the influence of fracture geometry on mass
transport in fracture media. Water Resources Res. 20(9), 1241–1252.
Steffen, O., 1975. Recent developments in the interpretation of data from joint surveys in rock
masses. Sixth Regional Conference for Africa on Soil Mechanics and Foundations, II, pp.
17–26.
Switzer, P., 1965. A random set process with a Markovian property. Ann. Math. Stat. 36,
1859–1863.
Terzaghi, K. C., 1946. Introduction to tunnel geology: Rock defects and loads on tunnel supports.
In: Protor, R. V., and White, T. L., Rock Tunnelling with Steel Supports: The Commercial
Shearing and Sampling Co., Youngs Town, Ohio, pp. 15–99.
Terzaghi, R. D., 1965. Sources of error in joint surveys. Geotechnique15, 237–304.
Wallis, P. F., and King, M. S., 1980. Discontinuity spacing in a crystalline rock. Int. J. Rock Mech.
Min. Sci. Geomech. Abstr. 17, 63–66.
Watson, 1955. Serial correlation in regression analysis I. Biometrika, 42, 327–341.
Index

A Barchet, W. R., 139–140

Aboufirassi, M., 135, 259 Barnes, R. J., 158
Agterberg, F., 272 Barnes, S. L., 103–104, 133, 134, 206
Akin, J. E., 48 Barros, V. R., 139
Al-Jerash, M., 256 Barton, C. M., 322
Anisotropy model, 130 Barton, N., 281, 282
auto- and cross-correlation in, 275 Bayer, A. E., 133
ratio and angle, 131 Bayraktar, H., 63
Areal average estimation (AAE), 47 Benjamin, J. R., 94, 96, 136
Areal coverage probability (ACP), 67 Bergman, K., 129
Area of influence methods, 128 Bergthorsson, P., 103, 133
ARIMA model, 36, 304, 305 Bernoulli distribution theory, 73
RQD–persistence relation, 308, 309 Bieniawski, Z. T., 281, 282
Aroian, L. A., 273, 275 Block Kriging, 231
Atmosphere, 2 Box, G. E. P., 36, 40, 135, 250, 259, 272, 305,
Atwater, M. A., 8–9, 55 324, 326
Autocorrelation structure, 304 Brady, B. H. G., 301
Auto-Regressive Integrated Moving Average Branstator, G. W., 218
(ARIMA) models, see ARIMA Bras, R. L., 47
model Bratseth, A. M., 134
Autoregressive models, 273 Bridges, M. C., 322
Autorun coefficient, 292–293, 313 Brown, E. T., 300, 301
discontinuity vs. scanline length, 295 Brown, R. G., 111
vs. autorun function, 314 Bruce, J. P., 61, 62
Autorun function, 302, 312 Bucy, R., 111
of dependent process, 314 Buell, C. E., 206
for discrete processes, 312 Buzzi, A., 134
of independent process, 313
vs. autorun coefficient, 313 C
Autorun simulation of porous material, Call, R. B., 322
310–311 Carr, J. R., 135, 225
analysis of sandstone, 312–316 Chow, V. T., 48
line characteristic function of porous Chu, P. S., 36
medium, 312 Clark, I., 135, 145, 148, 153, 188, 193, 206,
modeling of porous media, 316–321 225, 229, 242
Clark, R. H., 61, 62
B Classical spatial variation models
Baczynski, N. R. P., 334 cluster sampling, 98–99
Baecher, G. B., 322 multisite Kalman filter methodology

343
344 Index

KF application, 115–126 intact length CSV, 323–324

KF iterative procedure, 111 theoretical CSV model, 324–333
one-dimensional KF, 112–114 Markov model, 325
ongoing discrete KF cycle, 114 model for ARIMA intact lengths, 326
nearest neighbor analysis, 100–102 obtaining from ReV data, 164–165
ratio statistics, 101 parameters, 332
random field, 95–98 practicalities and limitations of, 149–150
search algorithms vs. empirical method/simulation technique,
geometric weighting functions, 322–323
103–106 Cumulative semivariogram (CSV), see CSV
quadrant and octant, 103
weighting functions, 105 D
spatial data analysis needs, 87–93 Daley, R., 129, 134
directional variations, 93 Data collection, 85
Easting, northing, and ReV, 88 Data types and logical processing methods
areal coverage probability
relative frequency diagram, 89
dry durations and areas, 68
representative ReV values, 91
extreme value probabilities, 72–73
ReV sampling locations, 91
theoretical treatment, 69–72
spatial pattern, 89
time and spatial units, 68
univariate statistical
number of data
parameters, 90
dependent/independent models, sample,
spatial pattern search, 85–86
33–34, 35–38
data variation, 86
sample points scatter diagrams, 32
systematic and unsystematic
standard Gaussian pdf, 34
components, 86
numerical data types, 25–27
spatio-temporal characteristics, 84 pixel sampling, 25, 26
trend surface analysis (global interpolator) point sampling, 25
calculation table, 109 observations, 22–25
model parameter estimations, polygonizations
108–110 Delaney, Varoni, and Thiessen
uniformity test, 93–95 polygons, 52–53
Cliff, A. D., 273 percentage-weighted polygon (PWP)
Cluster sampling, 98–99 method, 55–57
Coates, D. F., 281 triangular coordinate paper, 56
Co-Kriging, 231 regional representation
Cooley, R. L., 135 inverse distance models, 45–46, 46
Cording, E. J., 282 variability range, 42–44
Cornell, C. A., 94, 96, 136 sampling, 27–31
Covariogram analysis, 93 areal patterns, 28
Cramer, H., 33, 276, 315 calcium, sodium, and chloride
Cressie, N. A. C., 159, 160, 162, 251 measurements, 31
Cressman, G. P., 19–20, 103, 104, 105, elevation data, 30
133, 206 irregular sample points scatter, 30
Cross-validation techniques, 188, 222 quadrangle, 29
Cruden, D. M., 282 spatio-temporal drought theory and
CSV, 164–175, 323, 330 analysis
ARIMA model, 326–327 areal drought variance percentage, 79
attributes and advantages, 166–167 average areal drought coverage, 80
experimental, 330 average variance, 80
graphs comparison, 332–333 drought area probability, 76
independent model, 324–325 drought parameters, 76–81
and intact length correlation structure, drought percentage areal
321–323 coverage, 77
Index 345

drought time change, 78 E

probability of drought area, 75 Earth, environment, and atmospheric
sub-areal partition researches, 16–17
monthly averages, 51 Earth sciences event, 12
seasonal averages, 52 Earth sciences phenomena, 2–7
triangularization, 47–52 data-processing technique, 3
triangular mesh, 50 geological phenomenon, characteristics
values of i, j, k, 51 of, 4–7
3D autoregressive model, 273–274 rock subunit map, 5
2D uniform model parameter, stratigraphic sequence, 6
276–279 structural non-determinacy, 7
extension to 3D, thin-section spatial variability, 6
279–281 three rock type map, 4
parameters estimation, 274–276 Eberthardt, A. K., 250
David, M., 18, 135, 225, Eddy, A., 129, 207
229, 231 Ege, J. R., 281
Davis, A., 38, 41, 98, 100 Einstein, H. H., 322
Davis, J. C., 148, 225, 252 Eissa, E. A., 290, 322, 323, 324, 326, 329, 334
Davis, W. E., 139–140 Eliassen, A., 132, 207
2D discontinuity network, 338 Empirical linear interpolations, 206
Dee, D. P., 110, 134 Empirical technique, 322
Deere, D. U., 281, 282, 287, 290, see also The profile area method
301, 309 Erdik, M., 176
Delhomme, J. P., 18, 228 Error theory, 128
Demaree, G., 250 Estevan, E. A., 139
Determinism versus uncertainty Expected analysis error
zero intercept value, 219
cause, environment, and effect
triplicate, 15
F
diurnal temperature variation, 13
Feller, W., 70, 73, 316, 317
estimation, identification, and
Field quantity, 17
prediction, 16
First-order Markov model, 36
Deutsch, C. V., 259 Fisher, R. A., 22, 205
DeWijs, H. J., 174 Flattery, T.W., 206
Dirichlet, G. L., 52 Fracture network, 334–335
Discontinuities, 281 Franke, R., 134
average number of, 295 Fritsch, J. M., 206
vs. scanline length for autorun
coefficients, 295 G
average number per unit length, 295 Gale, J. E., 291
occurrences along scanline, Gandin, L. S., 18–19, 104, 129, 132, 206, 207
probability, 295 Gauss, K. F., 22
scanline survey of, 282 General Circulation Models (GCM), 27, 36
Discontinuity network properties, Geometric weighting functions, 103–106
334–335 drawbacks, 104–105
2D isotropic model Geometric weights, 103–106
parameter, 277 Geometry, 5
3D isotropic simulation Geostatistical approaches, 206
model, 279 Geostatistics
Dixon, R., 206 defined, 225
Donnelly, K. P., 100 see also ReV
Döös, B. R., 103, 133 Gilbert, S., 259
Drought occurrences, 81 Gilchrist, B., 103, 133
Droughts, 21 Glass, C. E., 225
346 Index

Global interpolator, 106 K

Goodin, W. R., 104, 133 Kalman, R. E., 110–111, 206
Goodman, R. E., 281, 282, 334 Kalman filter, 110–126
Gustafsson, N., 129 application, 115–126
annual precipitation, 121–122
H annual rainfall values, 117,
Habib, Z. Z., 136, 175, 207, 211, 212, 213 118–119, 120
Hammersely, J. M., 284 areal rainfall values, 125
Handscomb, D. C., 284 geographic locations of precipitation
Harrison, P. J., 111 stations, 116
Hevesi, J. A., 47, 259 percentage error of annual rainfall,
Higgs, N. B., 292, 322–323 123–124
Hoeksema, R. J., 135 rainfall distribution, 115
Hohn, M. E., 153, 154 difficulties, 111
Homogeneity, 6, 129–132 estimation execution steps, 114
Hubert, P., 250 iterative procedure, 111
Hudson, J. A., 282, 283, 291, 293, 294, 296, measurement update equations, 113
297, 300, 322, 324, 327, 335, 336 time update equations, 113
Huff, F. A., 55 Kalman filtering techniques, 134
Huijbregts, C. I., 135, 149, 154, 167, 171, 206, Kalman gain matrix, 120
225, 229, 231 Karz, R.W., 36
Hwang, Y. C., 111 Kazi, A., 281, 282, 283, 288,
322, 336
I Kedem, B., 48
Independent process, 303–304 Kendall, M. G., 36
Indicator Kriging, 230 Kinematics similarity, 11
Infinite sampling line, 316 Kinematics variability, 10
Inman, R. L., 133 King, M. S., 322
Intact length, 281, 290 Kitanidis, P. K., 135, 164, 251
characteristics, 299 Koch, S. E., 56, 90, 104, 133
correlation structure, CSV technique, Kolmogorov, A. N., 18–19, 22, 259
321–323 Krige, D. G., 135, 229, 272
CSV model, 320 Kriging, 27, 228
of independent, 320 estimation variances, 266
dependent, 290–292 methodology, 205, 225, 230–231, 250,
applications, 299–300 251, 253
correlation measurement, 292–293 advantages, 231
expectations of, 294 application of algorithm, 231–232
simulation of, 305 types, 230–231
International Society for Rock Mechanics, 281 techniques
Interpolation point value, 212–213 based on simple linear models, 229
Intrinsic hypothesis, 227 characteristics and advantages, 229–230
Intrinsic property, 226–228 intrinsic property, 226–228
Iruine, K. N., 250 Kruger, H. B., 129
Isaaks, E. H., 18, 222, 249, 251 Krumbein, W. C., 272
Isotropic model, 277 Kulhawy, F. H., 282
Isotropy, 129–131
L
J Lag-k autorun coefficient, 292
Jackson, I. J., 55 Lagone Markov process, 304–305
Jenkins, G. M., 36, 40, 135, 250, 272, 305, Latif, A. M., 111
313, 324, 326 Ledimet, F., 206
Journel, A. G., 135, 149, 154, 167, 171, 206, Liebhold, A. M., 150
225, 229, 231, 259, 273 Line characteristic function, 318
Index 347

number of voids on, 319 Optimum interpolation, 211

schematic representation of, 311 data and application, 211–224
Link, R. F., 56, 90 cross-validation of model, 222–224
Long, J. C. S., 334 data search and selection procedure,
Lorenc, A. C., 129, 133–134, 218 219–222
Loucks, F. D., 250 equal correlation lines, 220
Louis, C., 282 estimation of zero intercept value, 219
execution steps, 214
M expected error, 218–219, 220
Mahar, J. W., 282 grid points and measurement sites, 212
Maps, 128 number of influencing stations, 221
Mariño, M. A., 135, 259 observed and predicted monthly values,
Markov processes 223, 224
first-order, 304–305 observed versus estimated values, 223
see also Autoregressive models spatial correlation function,
Marsily, G. D., 228 215–218
Martinez, C. A., 231 spatial distributions, 212
Matern, B., 272 univariate statistics, 222
Matheron, G., 18–19, 83, 135, 139, 145, location of grid point and measurement
164, 179, 206, 225, 227, 229, site, 209
272, 323 Ord, J. K., 273
Mathier, L., 250 Ordinary Kriging, 230
Meleschko, V. P., 221 Otaibi, A., 299
Mercalli scale, 24
Modeling, 204–205
P
Mood, A. M., 315
Pannatier, Y., 159, 161, 162
Multidirectional RQD simulation,
Panofsky, H. A., 132, 206
333–334
Pedder, M. A., 129, 133, 134, 136
fracture network model, 334–335
Percentage-Weighted Polygon (PWP) method
RQD analysis, 335–337
basic tasks, 55
RQD simulation results, 338–340
step-by-step algorithm, 56–57
Myers, D. E., 135, 150, 154, 231, 251
Pernot, M., 282
N Perrie, W., 165
Nearest neighbor, 100–102 Persistence, 313
Negative exponential distribution, 98 Pham, T. D., 188, 190, 193–194
Neill, J. C., 55 Piteau, D. R., 282
North, G. R., 8 Pixel sampling, 25
Nugget effect, 142, 143 Point cumulative semivariogram (PCSV), 175
Nugget (zero distance jump), 150, 156 Point sampling, 25
Numerical data, 21, 25 Polygonalization, 53
Popper, K., 13
O Powers, R. W., 169
Objective modeling, 206–207 Priest, S. D., 282, 283, 291, 293, 294,
Observation, 22 296, 297, 300, 322, 324, 327,
Octant search, 102 335, 336
Olea, R. A., 242, 249 Prigodich, A. E., 221
Operational objective analysis, 103 Privalsky, V., 250
Optimal interpolation model (OIM), 207 The profile area method, 322
assumptions, 208 Punctual Kriging, 230
execution steps, 214
points to remember, 208 Q
steps necessary for practical applications, Quadrant search, 102
213, 215 Quenouille, M. H., 36
348 Index

R Scanlines, 281, 293, 301

Random Field (RF), 17–18, 95–96 frequency distribution functions, 329
statistically homogeneous and isotropic, 18 mathematical modeling of, 293
Randomness, 12, 14 rock qualities, 303
Random variables (RV), 11, 314 Schlatter, T. W., 129, 206, 218
Range (radius of influence), 146, 150 Schubert, S. D., 36
Regional dependence (correlation) Schwartz, F.W., 334
function, 234 Seaman, R. S., 134
Regionalized Variable (ReV), 11, 18–19 Second-order stationarity, 227
theory of, 18 Semivariogram (SV), 135
Regional rainfall pattern information required, 147
ReV, 83, 129, 225 models
data scatter, 87–90 development of, 156
function, 311 patterns, 140–143
Rice, S. O., 315 reasoning, 145–146
Roach, S. A., 316, 318, 320
technique
Rock quality designation (RQD),
practical difficulties, 145
281–282, 335
en, Z., 32, 34, 39, 55, 67, 74, 90, 106, 136, 139,
average number of discontinuity, 298
140, 145, 149, 164, 166, 167, 170,
classification of, 285
171, 175, 176, 179, 183, 206, 207,
and correlated intact length simulation,
252, 255, 273, 274, 278, 281, 282,
300–303
283, 284, 288, 290, 291, 292, 294,
scanline and models, 300
296, 300, 304, 311, 312, 313, 315,
disadvantages, 283
322, 323, 324, 326, 327, 329, 334,
formulation and discussion, 293–298
336
method adopted by Deere, 282
Seo, D. J., 231
negative exponential pdfs, 288
Sharp, W. E., 273, 275
persistence relation, 310
Shenfield, L., 133
threshold value discontinuity number
Sill (regional variance), 142, 146
chart, 287
Rock quality designation simulation, 281 Similarity, 11
dependent intact lengths, 290–300 Simple iterative method, 134
independent intact lengths, 281–290 Simple Kriging, 232–236
Rock quality percentage (RQP), 282 Simulation model
Rock quality risk (RQR), 282 extension into 3D space, stages, 279–281
Rodriguez-Iturbe, I., 47 stages, 274
Rouleau, A., 334 Simulation of stochastic variables, 284
RQD and correlated intact length simulation, Skibin, D., 139
300–303 Slivitzky, M., 250
proposed models of persistance, 303–305 Smith, H. R., 282, 334
simulation of intact lengths, 305–310 Smith, L., 334
RQD classification, 283 Sneyers, R., 250
Rutherford, I. D., 129 Solid length, 317
Ryckes, K. D., 299, 301, 327–328 Spatial correlation function, 135–140
average and theoretical, 138
S empirical, 137
Salas, J. D., 47 model parameters, 215
Salas, M., 259 negative-exponential model, 215, 217
Sample SV pertinent statistics, 218
interpretation of, 150 Sivas city, 216
Sampling, 27–31 Sivas correlation contour lines, 217
Sasaki, Y., 105, 206 Spatial dependence function (SDF),
Sashegyi, K. D., 134 132–135, 183
Index 349

obtaining weighting function from sample exponential, 159

CSV, 183 Gaussian, 159–160
Spatial dependence measures linear, 157–158
cumulative semivariogram logarithmic, 163–164
representative, 165 power, 161–162
sample, 167–169, 171, 172, 173 quadratic, 160
theoretical models, 169–175 rational quadratic, 160–161
function, 183 simple nugget, 156–157
isotropy, anisotropy, and homogeneity, spherical, 162–163
129–132 total dissolved solids, 155
average annual temperature, 131 triple surface of chloride, 154
spatio-temporal depth variations, 130 triple surface of TDS, 155
point cumulative semivariogram, 175 wave (hole effect), 162
seismic categories, 178 Spatial estimation of ReV
seismic events, 177 elements, 205
similarity map, 182 objective analysis, 206
similarity measures at fixed level, spatial data modeling techniques,
180–181 206–207
sample SV, 151–153 Spatial modeling
chloride, 153 block Kriging, 249–250
forms, 151 block pieces, 250
triple surface of chloride, 152 estimation of ReV, 205–207
semivariogram regional dependence geostatistical analysis
measure Kriging technique, 225–228
characterizing spatial correlation, ReV theory and concepts, 225
148–149 geostatistical estimator (Kriging)
continuous linear trend, 142 Kriging methodologies and advantages,
definition, 144–149 230–232
directional SV, 147 optimum interpolation model
discontinuous trend surface, 141, 143 data and application, 211–224
global SV and elements, 146 ordinary Kriging, 239–243
homogeneous, isotropic, and uniform estimation variance, 242
ReV, 141 unbiasedness principle, 239
homogeneous and isotropic ReV SV, regional rainfall pattern description,
141 256–266
independent spatial data, 143 dataset, 258
limitations, 149–150 descriptive statistics and normality
linear trend surface SV, 142 test, 259
philosophy, 140–144 experimental and fitted SV models,
random ReV SV, 142, 144 260, 261
spatial correlation function isohyetal map, 262, 263, 264
correlation coefficient drawback, Kriging variance, 266
136–140 spatiotemporal Kriging
evaluating wind power, 139 maps, 265
spatial dependence function, 132–135 SV cross-validation, 260
objective methods and drawbacks, 133 topographic map, 257
semivariogram (SV) technique, 135 simple Kriging, 232–236
simple iterative method, 134 assumptions, 232
successive correction method, 133, 134 covariance–distance graph, 234
theoretical SV ReV sample sites and estimation
application of Kriging modeling to site, 233
ReV, 156 standardized ReV, 234–235
bicarbonate sample, 154 triple diagram model, 250–256
350 Index

empirical and theoretical SV, 252 Subyani, A. M., 167, 256, 259
interpretations, 253, 255 Successive correction method,
lag-one lake level prediction, 255 133, 134
lag-one model verification, 254 Summer, G., 9, 47, 55
lake level TDMs, 254 Surface fitting methods, 206
location map, 251 SV, see Semivariogram (SV)
observed and predicted lake Switzer, P., 272
levels, 256
theoretical Gaussian SV T
parameters, 252 Tabios, G. O., 47
universal Kriging, 245–247 Talagrand, O., 206
Spatial pattern, 85–86 Tase, N., 74, 77
Spatial prediction study, steps, 129 Taylor, G. I., 144, 179
Spatial simulation, 272–273 Temporal variations, 84
autorun simulation of porous material, Terzaghi, K. C., 281
310–311 Terzaghi, R. D., 321
autorun analysis of sandstone, Theoretical CSV models
312–316 exponential, 172–173
autorun modeling of porous media, Gaussian, 174–175
316–321 linear, 169–171
line characteristic function of porous logarithmic, 173–174
medium, 312 power, 171–172
CSV technique, 321–323 Thiebaux, H. J., 129, 133,
intact length CSV, 323–324 134, 136
theoretical CSV model, 324–333 Thiessen, A. H., 52, 54
3D autoregressive model, 273–274 Thiessen polygons, 52, 53
2D uniform model parameters, 276–279 obtaining sub-polygons, 53
extension to 3D, 279–281 Time series, 84
parameters estimation, 274–276 Toulany, B., 165
multidirectional RQD simulation, 333–334 Trend surface, 106–107
fracture network model, 334–335 calculation table, 109
RQD analysis, 335–337 variables, 107
RQD simulation results, 338–340 Triangularization, 47–52
rock quality designation simulation, 281 Triple diagram method (TDM), 252
dependent intact lengths, 290–300
independent intact lengths, 281–290 U
RQD and intact length simulation, 300–303 Uncertainty techniques, 14
proposed models of persistance, Uniformity test, 93–94
303–305 Universal Kriging, 228, 245–247
simulation of intact lengths, 305–310 procedural structure, 246
Spatial variability, 8, 9, 203
Specific surface, 314 V
Srivastava, R. M., 18, 222, Vannitseni, S., 250
249, 251 Variability, 1
Statistical interpolation, 211 deterministic and stochastic
see also Optimum interpolation variations, 11
Statistical objective analysis, 206 geometric similarities, 10
Steffen, O., 322 kinematics similarity and, 10
Stevens, C. F., 111 quantitative consideration, 11
Stochastic process, 312 spatial, 8, 9
Stout, G. E., 55 Variational techniques, 206
Stratigraphic variation, 4 Void length, 317
Student, 205 Voronoi, G., 52
Index 351

W Y
Wallis, P. F., 282, 322 Yates, F., 205
Watt, D. G., 313 Yevjevich, Y., 77
Weighted-averaging analysis
scheme, 104
Wiener, N., 22
Wiesner, C. J., 61, 62 Z
Wilson, J. W., 8, 55 Zadeh, L. A., 23, 90, 253