Geostatistics: Kriging

8.10.2015
Konetekniikka 1, Otakaari 4, 150
10-12

Rangsima Sunila, D.Sc.

Background
What is Geostatitics
Concepts
Variogram: experimental, theoretical
Anisotropy, Isotropy
Lag, Sill, Range, Nugget
Types of Kriging
Example of kriging interpolation
2

Interpolation
How to estimate unknown values at specific
locations.

Spatial Interpolation
Examples
Trend surfaces
Nearest neighbours: Thiessen(voronoi)
Inverse distance weighting (IDW)
Splines
Kriging

Example:
Site

D to
(5,5)

4.2426

10
2
8

2.8284

4
6
4

5.6569

4
6
2

1.0000
0
0

10

2.0000

We would like to estimate the variable value at (5,5)

Example: IDW
Value of z(x) is estimated from all known
values of z at all n points. (Weighted Moving
Average technique)
n

z ( x) wi zi
i 1

Weights usually add to 1,

w 1
i 1

Example: IDW
In IDW, the weights weights are based on the
distance from each of the known points (i) to
the point we are trying to estimate (k): dik. In
IDW, we consider the inverse distance, 1/dik

1
d ik
wi n
1

11 d ik
7

Example: IDW
Location (x,y)

D to (5,5)

ID

Weights

(2,2)

4.2426

0.2357

0.1040

(3,7)

2.8284

0.3536

0.1560

(9,9)

5.6569

0.1768

0.0780

(6,5)

1.0000

0.4413

(5,3)

2.0000

0.5

0.2207

2.2661

sum

Z(5,5)

= 0.1040(3) + 0.1560(4) + 0.0780(2) + 0.4413(4) + 0.2207(6)

= 4.1814

Historical background
Geostatistics, first developed by Georges Matheron (19302000), the French geomathematician. The major concepts
and theory were discovered during 1954-1963 while he was
working with the French Geological Survey in Algeria and
France.
In 1963, he defined the linear geostatistics and concepts of
variography, varaiances of estimation and kriging (named
after Danie Krige) in the Trait de gostatistique applique.
The principles of geostatistics was published in Economic
Geology Vol. 58, 1246-1266.
Kriging was named in honour of Danie Krige (1919-2013),
the South African mining engineer who developed the
methods of interpolation.
9

What is Geostatistics
Techniques which are used for mapping of surfaces
from limited sample data and the estimation of values
at unsampled locations
Geostatistics is used for:

spatial data modelling

characterizing the spatial variation
spatial interpolation
simulation
optimization of sampling
characterizing the uncertainty

The idea of geostatistics is the points which are close to

each other in the space should be likely close in values.
10

Mining
Geography
Geology
Geophysics
Oceanography
Hydrography
Meterology
Biotechnology
Enviromental studies
Agriculture
11

Geostatistical methods provide

How to deal with the limitations of
deterministic interpolation
The prediction of attribute values at unvisited
points is optimal
BLUE (Best Linear Unbiased Estimate)

12

Geostatistical method for interpolation

Reconigtion that the spatial variation of any
continuous attribute is often too irregular to
be modelled by a simple mathematical
function.
The variation can be described better by a
stochastic surface.
The interpolation with geostatistics is known
as kriging.
13

Variogram cloud

OSullivan & Unwin, 2002)

14

Variogram/Semivariogram
Variogram is the variance of the difference
random variables at two locations
To examine the spatial continuity of a
regionalized variable and how this continuity
changes as a function of distance and direction.
The computation of a variogram involves plotting
the relationship between the semivariance and
the lag distance
Measure the strength of correlation as a function
of distance
Quantify the spatial autocorrelation
15

Variogram

Semivariogram, y(h)

Variability increase

Lag distance (h)

16

Variograms
Half of average squared difference between
the paired data values.
The variogram calculated by

1
2
h
(
x

x
)

i
j
2 N (h) (i , j )|hij h

17

Variogram
Experimental variogram (sample or observed
variogram) :
when variogram is computed from sampled data.
The first step towards a quantitative description of
the regionalized variation.

Theoretical variogram or variogram model:

when it is modelled to fit the experimental
variogram.
18

19

Omnidirectional variogram
Omnidirectional variogram is a test for erratic
directional variograms
The omnidirectional variogram contains more
sample pairs than any directional variogram so it
is more likely to show a clearly interpretable
structure.
If the omnidirectional variogram is messy, then
try to discover the reasons for the erraticness,
e.g. Examine the h-scatterplots may reveal that a
single sample value shows large influence on the
calculations.
20

Isotropy
The spatial correlation structure has no
directional effects, the resulting variogram
averages the variogram over all directions.
The covariance function, correlogram, and
semivariogram depend only on the magnitude of
the lag vector h and not the direction
The empirical semivariogram can be computed by
pooling data pairs separated by the appropriate
distances, regardless of direction.
The semevariogram describes omnidirectional.
21

22

23

24

Anisotropy
Spatial variation is not the same in all
directions
The variogram is computed for specific
directions
If the process is anisotropy, then so is the
variogram

25

26

27

Fitting variogram models

Why do we need a variogram model?
We need a variogram value for some distance or
direction for which we do not have a sample
variogram value.

28

Fitting variogram models

Fitting variogram models can be difficult
The accuracy of the observed semivariance is not
constant
The variation may be anisotropic
The experimental variogram may contain much
point-to-point fluctuation
Most models are non-linear in one or more
parameters
Both visual inspection and statistical fitting are
recommended
29

Fitting variogram models

Fitting variogram models may be poor
Chose unsuitable model in the first place
Give poor estimates of the parameters at the start
of the iteration
A lot of scatter in the expertimental variogram
The computer program was faulty

30

Description
Lag The distance between sampling pairs.
Range The point where the semivariogram
reaches the sill on the lag h axis. Sample points
that are farther apart than range are not spatially
autocorrelated.
Nugget The point where semivariance
intercepts the ordinate.
Sill The value where the semivariogram first
flattens off, the maximum level of semivariance.
The points above the sill indicate negative spatial
correlation and vice versa.
31

Fitting variogram models

h

Spherical

range

Exponenial

range

sill

nugget

nugget
Lag (h)

sill

Lag (h)

Linear

Gaussian
range

nugget

sill

nugget
Lag (h)

Lag (h)

32

Fitting variogram models

Spherical

Linear

Exponential

Gaussian

Source: Longley, P.A., Goodchild, M.F., Maguire, D.J. And Rhind, D.W., 2001, Geographic Information Systems and Science

34

Spherical model

35

Exponential model

Gaussian model

Linear model

36

Nested structure
One variogram model can be created by
several variogram models
n

(h) i i (h)
i 1

and
t (h) 1 (h) 2 (h)

37

Nested Structure
Example: the nested spherical or double
spherical

38

Ordinary kriging
In ordinary kriging, a probability model is used
in which the bias and error variance can be
computed and select weights for the
neighbour sample locations that the everage
error for the model is 0 and the error variance
is minimized.
The procedure of ordinary kriging is similar to
weighted moving average except the weights
are derived from geostatistical analysis.
39

Ordinary kriging
The estimationi by ordinary kringing can be
expressedn by:
n
i 1
z ( x0 ) i z ( xi ) where
i 1
i 1
The minimum variance of z(x0) is
n

2 i ( xi , x0 )
i 1

And it is obtained when

x , x ( x , x )
n

i 1

for all j
40

Example:
Site

D to
(5,5)

4.2426

10
2
8

2.8284

4
6
4

5.6569

4
6
2

1.0000
0
0

10

2.0000

We would like to estimate the variable value at (5,5)

41

Example:Ordinary kriging
Computing kriging weights for the unsampled point x = 5,
y = 5. Let the spatial variation of the attribute sampled at
the five points be modelled by a spherical variogram with
parameters c0=2.5, c1=7.5 and range a = 10. The data at
the five sampled points are:
We need to solve

A b

1

Where A is the matrix of semivarainces between of pairs of data

points, b is the vector of semivariances between each data point
and the point to be predicted, is the vector of weights and is
a lagrangian for solving the equations.

Example from Principles of Geographical Information Systems by Burrough and McDonnell, 1998, Oxford University Press

42

Example:Ordinary kriging
Value at (5,5) = weights * z
= 4.3985
With estimation variance = (weights*b)+
= 4.2177 + (-0.1544)
= 4.0628
Note: The estimation error variance is also known as kriging
variance.

43

Comparison of the results

Method

Estimate value z(5,5)

IDW

4.1814

Ordinary Kriging
(Kriging variance)

4.3985
(4.0628)

44

Block kriging
The modification of kriging equations to
estimate an average value z(B) of the variable
z over a block of area B.
10
z3
8
z4
6
z2
4
z1
2

z5

0
0

10

45

Block kriging

Example showing a regular 2x3 grid of point locations within a

block. Each discretizing point accounts for the same area.

Example from An introduction to applied Geostatistics by Edward H. Isaaks and R.Mohan Srivastava

46

Block kriging
The average value of z(B) over the block B is
given by

is estimated by

z ( x)dx
z ( B)
B area B
n

with

i 1

z( B) i z ( xi )
i 1

i 1

47

Block kriging
The minimum variance is
n

2 ( B) i ( xi , B) ( B, B)
i 1

and is obtained with

n

( x , x ) ( x , B)
i 1

for all j

48

1: (+531)

2: (+75)

4: (+333)

5: (+280)

3: (+326)

Example from An introduction to applied Geostatistics by Edward H. Isaaks and R.Mohan Srivastava

49

1: (+531)

2: (+75)

4: (+333)

5: (+280)

3: (+326)

50

Example: Block kriging

Point

Estimate

Kriging weights for samples

1

336

0.17

0.11

0.09

0.60

0.03

361

0.22

0.03

0.05

0.56

0.14

313

0.07

0.12

0.17

0.61

0.03

339

0.11

0.03

0.12

0.62

0.12

Average

337

0.14

0.07

0.11

0.60

0.08

51

Simple kriging
It is similar to ordinary kriging except that the
weights sum equation (=1) is not added.
The mean is a known constant.
It uses the average of the entire data set.
(ordinary kriging uses local average : the
average of the points in the subset for a
particular interpolation point)

52

Cokriging
It is an extension of ordinary kriging where two or
more variables are interdependent.
How:
U and V are spatial correlated
Variable U can be used to predict variable V that is information
about spatial variation of U can help to map V.

Why:
V data may be expensive to measure or collect or have some
limitations in data collection process so the data may be
infrequent.
U data, on the other hand, may be cheap to measure and
possible to collect more observations.
53

Indicator kriging
Binary value
From a continuous variable z(x), an indicator
can be created by indicating it 1 for z(x) is less
than or equal to a cut-off value, zc, and 0
otherwise
if z ( x) zc
1
( x)
0

otherwise

54

Kriging: Step by step

Studying the gathered data: data analysis
Fitting variogram models: experimental
variogram and theoretical variogram models
Estimating values at those locations which have
not been sampled (kriging) e.g. ordinary kriging,
simple kriging, indicator kriging and so on
Examining standard error which may be used to
quantify confidence levels
Kriging interpolation
55

Example
Elevation data set in Rastila
2000 laser scanning points
Minimum

Maximum

18

Mean

4.7247

Median

Skewness

0.64651

Kurtosis

2.0344

Standard deviation

5.543

Variance

30.725
56

Example

57

Example
Variogram model

range

nugget

sill

length

Exponential

0.5

0.10847

1.1232

0.18064

Linear

0.5

0.50754

1.806

Gaussian

0.5

0.4546

1.0447

0.22811

Spherical

0.75

0.35265

1.0666

0.48576

Exponential

0.75

0.13161

1.1388

0.19144

Linear

0.75

0.58671

1.5711

Gaussian

0.75

0.48089

1.0728

0.24717

Spherical

0.95

0.54285

1.9597

1.8513

Exponential

0.95

0.36872

1.4082

0.40794

Linear

0.95

0.55676

1.6424

Gaussian

0.95

0.57831

1.1922

0.34484
58

Example
Variogram model

RMSE

Exponential (0.5)

3.281

Linear (0.5)

3.367

Gaussian (0.5)

3.431

Spherical (0.75)

3.307

Exponential (0.75)

3.272

Linear (0.75)

3.397

Gaussian (0.75)

3.439

Spherical (0.95)

3.381

Exponential (0.95)

3.302

Linear (0.95)

3.386

Gaussian (0.95)

3.466
59

Example

60

Example

61

References:
Geographic information Analysis by Sullivan, D. And
Unwin, D.
Geostatistics for Environmental Scientists by Richard
Webster and Margaret Oliver
Principle of Geographical Information Systems, Chapter 5
and 6 by Peter Burrough and Rachael McDonnell
An Introduction to Applied Geostatistics by Edward Isaaks
and Mohan Srivastava
Quality Aspects in Spatial Data Mining by Alfred Stein,
Wenzhong Shi and Wietske Bijker
62