Chapter 9

Statistical and Geostatistical Applications

in Geology

Chapter Outline
9.1 Why Geostatistics? 167 9.2.15 The t-Test 177
9.2 Statistical Applications 168 9.2.16 The F-Test 178
9.2.1 Universe 168 9.2.17 The Chi-Square Test 178
9.2.2 Population, Sample, and Sampling Unit 168 9.2.18 Analysis of Variance 179
9.2.3 Probability Distribution 168 9.2.19 Trend Surface Analysis 180
9.2.4 Frequency Distribution 168 9.2.20 Moving Average 183
9.2.5 Normal or Gaussian Distribution 169 9.2.20.1 Moving Weighted Average of Block
9.2.6 Minimum, Maximum, Range, Median, Mode, Mean 183
and Mean 169 9.2.20.2 Moving Arithmetic Average 183
9.2.7 Sample Variance and Standard Deviation 171 9.2.20.3 Other Mathematical Techniques 184
9.2.8 Probability Plot 171 9.3 Misclassified Tonnage 184
9.2.9 Log-Normal or Pareto Distribution 171 9.4 Geostatistical Applications 185
9.2.9.1 Log-Normal Computation 172 9.4.1 Block Variance 185
9.2.9.2 Sichel’s t-Estimate 173 9.4.2 Semivariogram 185
9.2.9.3 Confidence Limits 173 9.4.2.1 Properties of a Semivariogram 187
9.2.9.4 Coefficient of Skewness 174 9.4.2.2 Semivariogram Model 188
9.2.10 Covariance 174 9.4.2.3 Smoothing a Semivariogram 189
9.2.10.1 Computation of Covariance 174 9.4.3 Estimation Variance 190
9.2.11 Correlation Coefficient 175 9.4.4 Kriging 191
9.2.12 Scatter Diagram 175 9.4.5 Kriging Estimation: An Example 192
9.2.13 Regression 175 9.4.6 Benefits of Statistics and Geostatistics 194
9.2.14 The Null Hypothesis 175 References 194

Statistical and geostatistical study enhances the analytical experience factors are applied to compromise as and when
and interpretive skill of the operator to visualize the inside the estimate goes wrong, often resulting in write-off
characteristics of the orebody. payable deposits or overvalued uneconomic deposits that
never deliver the predicted grades.
Author. Workers in the field of mineral resource and ore reserve
estimation (Davis, 1971) corrected this shortcoming by
9.1 WHY GEOSTATISTICS? borrowing techniques from formal statistical theory. These
include analysis of sample population, range, frequency
Geologists and mining engineers advocate that the grade of
distribution, histogram plot, mean, variance, standard de-
a mineral deposit depends on the location of a subblock
viation, correlation coefficients, analysis of variances, t-test,
within the orebody and sample value of a small block. Thus
F-test, trend analysis, and inverse distance for univariate
conventional methods of resource and reserve estimation
and multivariate elements. The statistical techniques
(refer to Chapter 8) are based on spatial position and value
compute the global reserves with an overall confidence
of the surrounding ground. There is no objective way to
limit of estimation (tonnage and grade). The classical
measure the reliability of the estimation. Professional

Mineral Exploration. https://doi.org/10.1016/B978-0-12-814022-2.00009-5 167

Copyright © 2018 Elsevier Inc. All rights reserved.
168 Mineral Exploration

statistical methods assume that sample values are randomly 9.2.2 Population, Sample, and Sampling
distributed and are independent of each other. They do not Unit
include the inherent geological variance within the deposit,
and ignore spatial relationships between samples. Population consists of a well-defined set of elementsd
Krige (1951) and Matheron (1971) developed the finite or infinite. Sample is a subset of elements taken from
theory of regionalized variables to resolve this problem. a population. Geological sample collection, preparation,
Regionalized variances are random quantities that as- and chemical analysis are the most critical issues for the
sume different values depending on their position within correct representation of the extremely variable nature of
a region. In ore reserve estimation, the orebody is the mineral occurrences. The sampling unit is part of the
region and grade, thickness, etc. are the regionalized universe on which the measurement is made. It is 5 kg of
variables. This technique offers the best linear unbiased rock or an ore sample, a truckload of bulk sample, or a 3 m
estimate (BLUE) of the reserves in the context of both piece of drill core. One must specify the sampling unit to
local and global bases. It also provides a direct quanti- recognize different populations, and treat them separately.
tative measure of reliability. The procedure involves the The population of 1 m drill core, channel, or chip samples
creation of a semivariogram, sample cross-validation, is likely to be statistically very different from the popula-
and an estimation technique known as point and block tion of 10
10
10 m3 blocks of ore. The size of the
kriging. sampling unit is significant. One should never make a
statement about the population of a grade without speci-
fying what grade it isdsay 2% Cu.
9.2 STATISTICAL APPLICATIONS
Statistics is a branch of applied mathematics. It deals with
9.2.3 Probability Distribution
the development and applications of procedures to collect,
organize, analyze, interpret, present, and summarize quan- Probability is the determination of the probable from the
titative data for understanding and drawing conclusions. It possible. The possible outcome of a random selection is
uses probability theory to estimate the population parame- described by the probability distribution of its parameter.
ters present, including future prediction. Univariate sta- When a coin is tossed the possible outcomes are either
tistics describes the distribution of individual variables heads or tails, each with an independent equal probability
(Koch and Link, 1986; Davis, 2002; Sahu; 2005). The data of ½. Similarly, when a perfect dice is thrown, the possible
in earth science often demonstrate inherent relationships outcomes are either 1, 2, 3, 4, 5, or 6, each with an equal
and interdependencies between variables, and are called probability of 1/6. This is the case of a discrete variable that
bivariate or multivariate

 statistics. The important param- assumes only integer values; the distribution will associate
eters are the mean X , median, mode, variance (s2 or S2), with each possible value x of probability P(x). The value
standard deviation (s or s), and coefficient of variation will be nonnegative, and summation of all possible P(x)
(CV). Statistical application prefers at least 30 samples for will be equal to one. Probability, in an applied sense, is a
drawing any inference. measure of the likeliness that a random event will occur.
Statistical procedures, tests of hypotheses, and appli- Geological events are typically not discrete, having
cations involve different assumptions, and accordingly are infinite and continuous numbers of possible outcomes. The
assembled in two groups: parametric, and nonparametric range of possible outcomes is finite. The exact result that
(Henley, 1981). Parametric statistics fundamentally relies may appear cannot be predicted within the range. Such
on the assumption that the parent populations are normally events are called continuous random variables. For
distributed, and preserve homogeneity of variance within example, one may wish to know the chance of obtaining a
groups. Nonparametric statistics does not require normal grade in the interval of 2%e4% or 6%e8% in a mineral
distribution or variance assumptions about the populations deposit. In practice, such a distribution will never be
from which the samples are drawn. known. All that can be done is to compute an experimental
probability distribution and then try to infer which theo-
retical distribution may have produced such an experi-
9.2.1 Universe mental sampling distribution, e.g., normal curve.
Universe is the total mass of interest and total source of all
possible data. Universe is characterized by one or more
attributes and can be single or multidimensional. An
9.2.4 Frequency Distribution
example of a geological universe is the area of our The measurements of some attributes, say the analytical
interestda stratigraphic province, a mineral deposit with a value of samples of a mineral deposit, can be classified into
defined boundary. The characteristics of universe change grade groups of 0%e1%, 1%e2%, 2%e3%, and so forth.
with changing boundary limits. The frequency of occurrences (number of assays per grade
 Statistical and Geostatistical Applications in Geology Chapter | 9 169

9.2.5 Normal or Gaussian Distribution

Normal or Gaussian distribution is a continuous proba-
bility distribution that has a bell-shaped probability density
function (Gaussian function), or informally a bell curve.
The frequency distribution plot of Table 9.2 and Fig. 9.1
depicts the bell-shaped curve and stands for the normal or
Gaussian distribution of attributes with a definite function
of spread from the mean (m). In the case of a theoretical
normal distribution curve the rule of thumb is that 68% of
the population or sample assays will be plus or minus one
standard deviation (d or S) from the population or sample
mean (m or X). Similarly, 95% of the population or sample
values will stretch out between plus and minus two stan-
FIGURE 9.1 Relative frequency plot showing the % area covered by 1 dard deviations (precisely 1.96) on either side of the mean.
and 2 standard deviations on either side of a central value or mean grade of Theoretical normal distributions show a single mode, but in
1 m sample population in Table 9.2, which represents a standard normal
natural conditions, mineral distribution can be bimodal and
probability distribution.
multimodal in character. The normal distribution is an
approximation that describes the real-valued random dis-
interval) in the different groups is tabulated. This infor- tribution that clusters around a single mean value.
mation would constitute a grouping or frequency or rela- A cumulative frequency is a process that understands
tive frequency distribution. The grouped information is whether collective information in a data set is less than or
plotted with class intervals on the abscissa (x-axis) and equal to a particular value. The plot will display graphically
frequency is plotted on the ordinate (y-axis). This allows a the observations by number (n) or percentage (%) or pro-
proportionate area for occurrence of each frequency against portion (1). The cumulative frequency curve provides the
the grade interval, and provides a visual picture, known as a continuity of information instead of discrete number of a
histogram (Fig. 9.1). Distribution is a function of sample particular group. A cumulative frequency distribution ob-
numbers. The sampling unit should be mentioned while tained from n number of samples can be transformed into a
defining the histogram. The population may be symmetric, continuous probability distribution (percentage or propor-
asymmetric, normal, log-normal, or bimodal. If there are n tion) by dividing each frequency by n, the total number of
number of samples (Table 9.1), each with a value of xi, observations (Table 9.2, Fig. 9.2). The total area of the
where i ¼ 1, 2, 3, ., n, then the samples can be grouped cumulative frequency distribution curve is transformed into
into a number of classes such as in Table 9.2. continuous probability (Isaak and Srivastava, 1989). This
approach to probability is intuitively appealing to geolo-
gists, as the concept is closely akin to unification.
TABLE 9.1 % Zinc Content of Exploration Drill Hole
Samples at 1 m Uniform Length
9.2.6 Minimum, Maximum, Range,
% Zn % Zn % Zn % Zn % Zn % Zn
Median, Mode, and Mean
5.70 3.90 4.50 3.30 3.60 4.60
1.50 2.60 5.50 3.40 4.50 2.20
Minimum and maximum are the smallest (Xmin) and
largest (Xmax) values in the data set, respectively. The
4.20 4.40 2.40 2.60 7.50 4.70
smallest and largest numbers in Table 9.1 are 0.30% Zn and
4.40 3.50 1.40 3.70 3.60 1.20 7.50% Zn.
2.50 0.50 5.50 4.70 0.70 2.40 Range is the difference between the largest and the
0.50 5.60 2.60 3.30 1.80 4.30
smallest values. It can be expressed that % Zn varies be-
tween 0.30 and 7.50 with a range of 7.20% Zn. The mid-
3.70 2.70 3.40 4.30 1.60 3.60
range is half of the range (xmaxxmin)/2.
3.30 5.10 4.60 3.50 2.70 4.50 Median is the middle value of a group of numbers (say
5.90 3.10 4.80 1.50 2.80 3.50 assay values) that have been arranged by size in ascending
2.30 1.30 3.60 2.40 5.30 5.40
order. The median splits the data set in two equal halves.
The median is the middle entry in the series when the total
2.90 4.90 2.10 4.10 3.70 6.50
number of observations is odd. In an odd series of 1, 2, 3, 4,
0.30 3.40 3.20 3.40 3.80 2.30 5, 6, 7, 8, 9 , 5 is the median. Similarly, the median is
equal to the sum of the two middle numbers divided by two
170 Mineral Exploration

TABLE 9.2 Frequency Table of 1 m Borehole Assay Data of Table 9.1

Sample Class of % Zn

Lower Upper Number of Cumulative Cumulative

Limit Limit Midpoint Samples Frequency Frequency (%)
0 1 0.5 4 4 5.6
1 2 1.5 7 11 15.3
2 3 2.5 15 26 36.1
3 4 3.5 21 47 65.3
4 5 4.5 15 62 86.1
5 6 5.5 8 70 97.2
6 7 6.5 1 71 98.6
7 8 7.5 1 72 100.0

the fifth decile and is equal to the median. Percentiles split

the data set into 100th divisions and the 50th percentile is
equal to the median. The 25th and 75th percentiles are equal
as first and third quartiles, respectively. The order data set
is plotted in the x-axis in normal scale and cumulative
frequency percent is plotted in the y-axis in normal (Fig. 9.2)
or logarithmic scale. Quantiles are generalizations of the
data-splitting concept. They represent the points taken at
regular intervals from the cumulative distribution function of
a random variable. A quantile-quantile or q-q plot is a
graphical presentation on which the quantiles from two
distributions are plotted versus one another. A q-q plot
compares two distributions, and is particularly useful when
FIGURE 9.2 Cumulative plots by number, percent, proportion, quartile, there are sufficient reasons to expect that the two distributions
deciles, and percentiles of 1 m sample grades in Table 9.2. are similar.

Sample mean, arithmetic mean, or average value X
conceptualizes the central tendency of distribution pa-
when the numbers of the series are even. In an even number rameters around which it disperses on either side. It is
series of 1, 2, 3, 4, 6, 7, 8, 9 , (4 þ 6)/2 ¼ 5 is the me- computed by the sum of the values of all observations
dian. The odd and even rule is important. In an order data within the population divided by the number of samples:
string, half of the population is below the median value. ( )
Mode is the value that occurs most frequently in a set of
 Xn

numbers or the highest frequency in grouped data or in a Sample mean grade X ¼ ð1=nÞ
Xi
i¼1
sample population. In a series of 1, 2, 3, 4, 4, 4, 5, 5, 6 , 4
is the most frequently occurring number. There may be The % zinc contents of an exploration drill hole
more than one mode when two or more numbers have an sampled at 1 m uniform length have been tabulated in
equal or relatively high number of instances (bimode or Table 9.1 for the computation of mean grade:
multimode). A mode does not exist if no number has more P
n
than one instance. xi ¼ 248:80
i¼1
Quartiles split ascending or descending order data sets
into quarters. A quarter (25%) of the data falls below the n ¼ 72
lower or first quartile (Q1). Similarly, a quarter (75%) of

Mean X ¼ 248:80=72 ¼ 3:45 % Zn
the data lies above the upper or third quartile (Q3). Deciles
split the order data set into 10 divisions. One-tenth of the The computation of arithmetic mean assumes that
data falls below the lower or the first decile, and 1/50th is all observed values are composed of the same size of
 Statistical and Geostatistical Applications in Geology Chapter | 9 171

FIGURE 9.3 Schematic diagram showing (A) asymmetric and (B) symmetric probability distribution.

pffiffiffi
sample length. Therefore unequal sampling units cannot be Confidence limit at 95% probability level ¼ S= n

utilized together to calculate valid statistical parameters. 1:96 ¼ 0:34 % Zn
Transformation of raw data to equal length of the sample
CV or relative standard deviation is the ratio between
unit is a prerequisite in statistical applications.
standard deviation and mean. CV represents the alternative
The three measures of central tendency, i.e., mode,
to skewness to describe shape of the distribution. The CV
median, and mean, will be different in an asymmetric dis-
will be 1 for normal distribution. A CV of 1 indicates
tribution curve (Fig. 9.3A). The normal distribution curve is
the presence of some erratic high sample values that have a
symmetric, and the three statistical measures of central
significant impact on estimated mean grade. The erratic
tendency will coincide (Fig. 9.3B).
high values may be due to sampling error. These samples
are sorted, located, and rejected if they do not fit in the
9.2.7 Sample Variance and Standard geological system:
Deviation 

CV ¼ S X ¼ 1:476=3:455 ¼ 0:427
A reasonable way to measure sample variation about the
The computation of mean from grouped data
mean value is to subtract the mean from the observed value
(Table 9.2) is:
and sum as follows: .X

 X
X Mean X ¼ ðF
mpÞ F ¼ ð249=72Þ
n


Xi e X
i¼1
¼ 3:458 % Zn

As the algebraic sum of the deviation of a set of where F ¼ frequency and mp ¼ midpoint:
numbers from its mean is always zero, the foregoing for-
 n X
. o
2
mula becomes meaningless as a measure of variance. The Sample variance S2 ¼ F
mp2 N  X
average variance is thus expressed as the sum of the
¼ ð1012=72Þeð3:458Þ
2
squared deviation of the mean from each observed value
(Table 9.1) divided by the number of observations 1,
¼ 14:05e11:95 ¼ 2:10
representing the degree of freedom. The population sam-
ple variance estimator is expressed as:
Pn 
2 
Sample varianceðS2 Þ ¼ Xi e X ðn  1Þ 9.2.8 Probability Plot
i¼1
( ! 2 ) Mean and standard deviation can be obtained from the plot
Pn X n
or; ¼ Xi2 e Xi ðn  1Þ of the upper class value limit (grade) versus the cumulative
i¼1 i¼1 frequency percentages on linear probability paper
r S2 for Zn distribution ¼ 154:638=71 ¼ 2:178 (Fig. 9.4). A normal distribution will create a straight line
on normal probability paper. The grade corresponding to
The square root of sample variance is the sample the 50% probability line gives the mean. The difference in
standard deviation estimator: grade between 84% and 50% probability gives an estima-
pffiffiffi2 pffiffiffiffiffiffiffiffiffiffiffi
¼
pffiffiffi S ¼ 2:178
pffiffiffiffiffi ¼ 1:476
Sample standard deviation 
(S) tion of standard deviation.
Variance of the mean ¼ S2 n ¼p2:178ffiffiffi 72 ¼ 0:256
pffiffiffiffiffi
Standard error of the mean ¼ S= n ¼ 1:476 72 ¼ 9.2.9 Log-Normal or Pareto Distribution
0:174 pffiffiffi
pffiffiffiffiffi limit at 68% probability level ¼ S= n ¼
Or confidence The log-normal or Pareto distribution plot exhibits
1:476 72 ¼ 0:174 exponentially and is skewed to the right or left. The
172 Mineral Exploration

FIGURE 9.4 Probability plot of Table 9.2 for computation of sample mean and standard deviation of Zn grade.

skewed log-normal distribution corresponds to spotty or 9.2.9.1 Log-Normal Computation

low-grade mineral deposits. High values are occasionally
A total of 2229 samples from a surface borehole at 0.50 m
encountered, while the majority of samples are of lower
uniform length from Baroi lead-zinc deposit, Rajasthan,
grade. The log-normal distribution generally corresponds
India, have been tabulated in Table 9.3.
to the normal distribution by logarithmic transformation
The frequency diagram and three-cycle logarithmic plot
of these values. The plot of group values often depicts a
have been depicted in Figs. 9.5 and 9.6, respectively.
straight line on log probability paper. Alternatively,
a three-parameter log-normal paper is used to obtain a N ¼ 2229
straight line. A constant can be added to each class for Ln at 50% ¼ Ln (5.00) ¼ 1.6094
better results. The mean value for a logarithmic distribu-
Ln at 84% ¼ Ln (11.05) ¼ 2.4024
tion is estimated by the following relationship from
Table 9.3, Figs. 9.5 and 9.6: b ¼ Ln 84% e Ln 50% ¼ 0.79299
 b2 ¼ 0.6288
Mean ¼ ea þ b2 2 ¼ egeometric mean
eb =2
2

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi b2/2 ¼ 0.3144


b2 ffi
Standard deviationðSÞ ¼ ea þ b2 2
e 1 Mean ¼ e(1.6094þ0.3144)
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

b2 ffi ¼ e(1.9238)
¼ Mean
e 1
.pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ 6.848 e 3.0 (constant added)
Confidence LimitðCLÞ ¼ ðS
1:96Þ n at 95% probability. ¼ 3.85% Pb
Standard
Geometric mean (a) is the logarithm (Ln) of mean deviation ¼ 3.85
0.932
grade at 50% point frequency. Geometric standard devia-
¼ 3.58%
tion of the log value (b) is the difference between the
natural logarithm of the grade at the 50% and 84% cumu- Confidence pffiffiffiffiffiffiffiffiffiffi
lative frequency ordinates. It corresponds to one standard limit ¼ ð3:58
1:96Þ 2229
deviation. ¼  0.149% Pb
 Statistical and Geostatistical Applications in Geology Chapter | 9 173

TABLE 9.3 Grouped Borehole Sample Data of Baroi Lead-Zinc Mine, India

Class Interval Cumulative Cumulative Class Interval Cumulative Cumulative

(% Pb) f Frequency Frequency (%) (% Pb) f Frequency Frequency (%)
0e1 811 811 36.38 15e16 54 2183 97.94
1e2 296 1107 49.66 16e17 2 2185 98.03
2e3 229 1336 59.94 17e18 5 2190 98.25
3e4 202 1538 69.00 18e19 3 2193 98.38
4e5 135 1673 75.06 19e20 4 2197 98.56
5e6 102 1775 79.63 20e21 4 2201 98.74
6e7 82 1857 83.31 21e22 e 22.01 98.74
7e8 59 1916 85.96 22e23 5 2206 98.97
8e9 49 1965 88.16 23e24 e 2206 98.97
9e10 44 2009 90.13 24e25 1 2207 99.01
10e11 38 2047 91.83 25e26 2 2209 99.10
11e12 31 2078 93.23 26e27 e 2209 99.10
12e13 20 2098 94.12 27e28 2 2211 99.19
13e14 21 2119 95.07 28e29 2 2213 99.28
14e15 10 2129 95.51 29e30 2 2215 99.37
>30 14 2229 100

9.2.9.2 Sichel’s t-Estimate

Sichel formulated tables to simplify the estimation pro-
cedure, especially in the case when only a few samples are
available such as:
N ¼ 2229
a ¼ Ln at 50%¼Ln (5.00)¼1.6094
Ln at 84%¼Ln (11.05)¼2.4024
b ¼ Ln 84% e Ln 50% ¼ 2.4024 e 1.6094 ¼ 0.7929
Log
variance ¼b2¼0.6288
Read the variance (b2) for corresponding t-value from FIGURE 9.5 Frequency plot of % lead grade at 0.50 m uniform sample
mean of lognormal population available at standard length showing log-normal probability with positively skewed distribution.
Sichel’s Table-B for stated number of samples (n). If the
variance is in between the table value interpolated value can
be computed between higher and lower table values: 9.2.9.3 Confidence Limits
Mean ¼ Log value at 50%
teconstantðif anyÞ
Read the lower and upper limits of t-value of log variance
¼ 5
1:3638e3 corresponding to stated number of samples (n) from stan-
dard Sichel’s t-Table
¼ 6:82  3
Lower limit ¼ Mean
t lower value ¼ 3.82
0.554 ¼ 3.649
¼ 3:82% Pb
Upper limit ¼ Mean
t upper value ¼ 3.82
1.048 ¼ 4.00
174 Mineral Exploration

FIGURE 9.6 Plot and computation of X, S2, and confidence limit of lead sample values by double logarithmic paper.

9.2.9.4 Coefficient of Skewness parameters. Covariance (COV) is the joint variation of two
variables about their common mean:
The quantity of symmetry of the sample distribution is the ( )
coefficient of skewness or skewness. The statistics are X n


plotted as a histogram and defined as: COVðX; YÞ ¼ ðxi  xÞ
yi  y ðn  1Þ
" # i¼1
X n
3 3 ( ! )
Coefficient of skewness ¼ 1=n ðxi  xÞ d X
n X
n X
n
i¼1 ¼ x i yi e xi
yi n ðn  1Þ
i¼1 i¼1 i¼1

Annual mill feed data by month of a process plant is

9.2.10 Covariance tabulated in Table 9.4 for computation of covariance be-
In typical geological situations, more than one variable is tween lead and silver content.
measured on each observational point, e.g., % Pb, % Zn, %
Fe, g/t Ag, and g/t Cd content of mill feed. The data can be 9.2.10.1 Computation of Covariance
arranged in an n
m array, where n is the number of
observations (months) and m is the number of variables
Pb Ag
(% or g/t metals). The statistical parameters of each variable P
such as X, S2, S, CL, and CV can be computed. However, X 16.32 470
P
different variables measured on the same object usually tend X2 23.749 19,000
to change together in some manner. The variables, which X 1.36 39
have no relation to each other, are said to be mutually in- S 0.3758 7.334
dependent, i.e., an increase or decrease in one variable is not
affected by a predictable change in another variable. Some X
n

measure of their mutual interface or relationship is expected Pbi
Agi ¼ 663:92

i¼1
in the case of variables not being totally or always inde-
pendent of each other. The procedure employed to compute COVðPb; AgÞ ¼ f663:92eð16:32
470Þ=12g=11
the variance of a single parameter can be extended to
calculate the measure of mutual variability for a pair of ¼ 2:247
 Statistical and Geostatistical Applications in Geology Chapter | 9 175

The plot is known as a scatter diagram, which provides an

TABLE 9.4 Monthly Average Mill Feed Grade of a instant sense of the expected relationship (Fig. 9.7).
Zinc-Lead-Silver Beneficiation Plant During 2010e11

Month % Zn % Pb % Fe g/t Ag 9.2.13 Regression

April 3.67 1.67 5.50 49
The linear mathematical relationship between two variables
May 3.34 1.12 4.60 38 based on the interdependability function is formulated as: Y
June 3.30 0.86 4.30 33 is a function of X, i.e., Y ¼ f (X):
July 3.55 1.28 4.40 41 ei ¼ b0 þ b1 Xi
Y
August 3.76 1.40 3.60 43
where Y ei is an estimated value of Yi at a specified value of
September 3.50 2.25 4.30 52
Xi, Yi is a dependent or regressed variable, and Xi is an in-
October 3.80 1.40 4.30 45 dependent or repressor variable
November 4.12 1.50 4.50 41 The fitted line will intercept the y-axis at a point b0 and
December 3.97 1.60 3.70 36
a slope of b1. The equation must satisfy the condition that
the deviation of Yi from the line is minimal such as:
January 3.32 1.12 3.20 33
X
n

2
February 2.92 0.93 3.60 30 e i  Yi
Y ¼ Minimum.
March 3.37 1.19 3.80 29 i¼1
( ! ),
X
n X
n X
n

b1 ¼ Xi Yi  Xi
Yi n
i¼1 i¼1 i¼1
9.2.11 Correlation Coefficient ( !2 )
X
n X
n

The correlation coefficient (r) is the measure of degree of Xi2 e Xi n
interrelationship between variables. The computation is not i¼1 i¼1
influenced by the unit of measurement of variables. Cor-
relation is the ratio between the covariance of two variables ¼ SPxy =SSx
and the product of their standard deviation: .
¼ ðSum of products for covarianceÞ ðSum of
rðj; kÞ ¼ COVJK =SJ SK
squares for varianceÞ
The correlation coefficient is a ratio and is expressed as
a unitless number. Covariance can be equal but cannot b0 ¼ Yeb1 X
exceed the product of the standard deviations of its vari- The silver (dependent) and lead (independent) contents
ables. The correlation coefficient will range between þ1 from Table 9.4 are plotted in a scatter diagram (Fig. 9.8).
(perfect direct relationship) and 1 (perfect inverse rela- The regression equation is computed as:
tionship). The spectrum of the r exists between these two  2 
extremities of perfect positive or negative relationships b1 ¼ ð663:92eð16:32
470Þ=nÞ 23:749  ð16:32Þ n
passing through zero (no relation) in a linear function. The
correlation between Pb and Ag from the data of Table 9.4 ¼ 24:720=1:554 ¼ 15:907
will be:
b0 ¼ 39:167eð15:907
1:36Þ ¼ 17:533
rðPb; AgÞ ¼ fCOVPb; Ag =ðSPb
SAg Þg
r The equation is: Agðg=tÞ ¼ 17:533 þ ð15:907
% PbÞ
¼ 2:247=ð0:3758
7:334Þ ¼ 0:815

Goodness of fit ¼ r2 ¼ 0:66 or 66% 9.2.14 The Null Hypothesis

The null hypothesis is a characteristic arithmetic theory
9.2.12 Scatter Diagram suggesting that no statistical relationship and significance
The interrelation between two variables can be visualized exists in a set of given, single, observed variables between
by plotting an independent variable along the x-axis and two sets of observed data and measured phenomena. The
its corresponding dependent variable along the y-axis. hypotheses play an important role in testing the significance
176 Mineral Exploration

FIGURE 9.7 Plot of scatter diagrams of two variables showing different types of correlations between them depending on mutual interdependability.

of differences in experiments and between observations.

H0 symbolizes the null hypothesis of no difference. It
presumes to be true until evidence indicates otherwise. Let
us take two sets of mill feed silver samples from Table 9.5
and compare the mean grade between set and population
and between two sets. The null hypothesis presumes and
states that:
H 0 : m1 ¼ m0
where H0 ¼ null hypothesis of no difference, m1 ¼ mean
of population 1, and m0 ¼ mean of population 2.
The null hypothesis states that the mean m1 of the parent
population from which the samples are drawn is equal to or
FIGURE 9.8 Plot of scatter diagram showing the regression line of Pb not different from the mean of the other population m0. The
versus Ag from Table 9.4. samples are drawn from the same population such that the
 Statistical and Geostatistical Applications in Geology Chapter | 9 177

TABLE 9.5 Average Monthly Mill Feed Silver Grade

(g/t) of a Zinc-Lead-Silver Mine

Difference (d)
Sample Set I Set II Set (IeII) (Difference)2
April 2010 49 51 2 4
May 38 31 7 49
June 33 33 0 0
FIGURE 9.9 Schematic diagram showing analysis of the t-test.
July 41 41 0 0
August 43 40 3 9
where X ¼ mean of the sample, m0 ¼ hypothetical mean
September 52 43 9 81
of the population, n ¼ number of observations, and
October 45 51 6 36
S ¼ standard deviation of observations.
November 41 32 9 81 If the computed value exceeds the table value of t at a
December 36 35 1 1 specific degree of freedom, level of significance, or lies in
January 33 30 3 9
the critical region (Fig. 9.9) or region of rejection, the null
2011 hypothesis (m1  m0) is rejected, leaving the alternative that
m1 > m0. If the computed value of t is less than the table
N 10 10 10 10
value, it can only be stated that there is nothing in the
SUM (Ʃ) 411 387 24 270 sample to suggest that m1 > m0, but it cannot specify that


AVG X 41.1 38.7 2.4 27.0 m1 < m0. This indecisiveness is a consequence of the
2 manner in which statistical tests are formulated. They can
VAR (S ) 40.77 61.57 23.60 1078.67
demonstrate, with specified probabilities, what things are
STD (S) 6.38 7.85 4.86 32.84 not. They cannot stipulate what they are.
The average monthly mill feed silver grade (g/t) of a
zinc-lead mine over 10 months is given in Table 9.5. There
are two sets of samples collected at different time intervals.
variance and shape of the distributions are also equal. The hypothetical mean grade (m0) of the deposit is 40 g/t
Alternative statistical applications such as t, F, and chi- Ag. The t-value of samples at Set I is computed to
test  the
square can only reject a null hypothesis or fail to reject it. hypothesis of likelihood of the means of samples X and
The evidence can state that the mean of the population from population (m0):
which the samples are drawn does not equal the specified

pffiffiffi
population mean and is expressed as: t ¼ X  m0 S n
. .pffiffiffiffiffi
H0 : m1 sm0 ¼ ð41:10  40:00Þ 6:38 10 ¼ ð1:10=2:02Þ
¼ 0:54

9.2.15 The t-Test The calculated value of 0.54 is much less than the table
value of t (1.83) for 9 degrees of freedom and 5% (a ¼ 0.05
The t-test follows parametric statistics assuming certain for a one-tail test) level of significance (Box 9.1). The
conditions of normal probability distribution and equality t-computed value lies away from the region of rejection.
of variance between groups and parent populations. The The test is indecisive to suggest that the population mean is
t-test or Student’s t-test is based on t-probability distribu- greater than 40 g/t Ag.
tion, similar to normal distribution with widespread, and The two sets of samples (I and II) collected at different
dependent upon the size of the samples. This is powerful time intervals are unpaired and independent. The pooled
for testing the hypothesis and useful for establishing the
likelihood that a given sample could be a member of a
population with specified characteristics (mean), or for
testing a hypothesis about the equivalency of two samples. BOX 9.1 CRITICAL VALUE OF T
The t-statistic is computed by: Critical Values of t for Various Degrees of Freedom and


pffiffiffi Selected Levels of Significance from Table 21,
t ¼ X  m0 S n Penguin-Honeywell Book of Tables (Kellaway, 1968).
178 Mineral Exploration

t-estimate tests the null hypothesis and the equality of the The estimated t-value of 1.56 is less than the table value
mean of two normally distributed populations. The test of 1.83 for 9 degrees of freedom and the 5% (a ¼ 0.05 for
statistical form is stated as: the one-tail test) level of significance. It does not fall into

.n pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi o the critical region and it will be concluded that there is no
t ¼ X1  X2 Sp
ð1=n1 þ 1=n2 Þ evidence to suggest that the two samples came from pop-
ulations having different means. Three assumptions are
where Sp is the pooled estimate of the population standard necessary to perform this test:
deviation based on both sample sets:
  1. Both samples are selected at random.
Sp2 ¼ ðn1  1Þ
S1 2 þ ðn2  1Þ
S2 2 ðn1 þ n2  2Þ 2. Populations from which the samples were drawn are
. normally distributed.
Pooled t ¼ ð41:10e38:70Þ 3. Variances of the two populations are equal.
hnpffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi o pffiffiffiffiffiffiffi i
ð9
40:77Þ þ ð9
61:57Þ=18
0:2 9.2.16 The F-Test
.pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffi The F-test pertains to parametric statistics and follows the
¼ 2:40 fð366:93 þ 544:13Þ=18g
0:20 assumptions and conditions of normal probability distribu-
. pffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffi tion of populations from which the samples are drawn.
¼ 2:40 51:17
0:20 The equality of variances of two data sets can be found and
tested based on a probability distribution known as the
¼ 2:40=ð7:15
0:44Þ ¼ 2:40=3:146 ¼ 0:76 F-distribution under the null hypothesis. The F-test is sen-
sitive to nonnormality. F is the ratio between the variances of
The table value of t for a two-tailed test with 18 or
two populations. The F-distribution is dependent upon two
9 degrees of freedom and the 10% (5% in each tail) level of
values: each associated with each variance in the ratio. The
significance are 1.73 and þ1.73. The calculated t-value of
samples are expected to be randomly collected from a
0.76 does not fall into either critical region. The null hy-
normal distribution to satisfy the condition of F-distribution.
pothesis cannot be rejected. There is no evidence to suggest
The F-hypothesis test is defined as:
that the two samples came from populations having
different means. H 0 : d1 2 ¼ d2 2
The paired t-test provides a hypothesis test of the
differences between population means for a pair of random d1 2 < d2 2 for a lower one  tailed test
samples whose differences are approximately normally
distributed. Let us assume that the two sets of samples in d1 2 > d2 2 for a lower one  tailed test
Table 9.5 are a repeat analysis of silver at two different
laboratories. The paired t-value can be computed for testing d1 2 sd1 2 for a two  tailed test
equivalency of the means: F ¼ Explained variance=Unexplained variance
.
Paired t ¼ d Sd
 
F ¼ S1 2 S2 2
where di ¼ difference of each paired sample and d ¼ mean where S1 2 and S2 2 are the sample variances of the population.
difference: It is concluded that the greater this ratio deviates from 1,
( !2 ) the stronger the evidence for unequal population variances.
2 X n Xn
 In other words, the variance in the populations is not same
Sd ¼ n
2
di e di nðn  1Þ
i¼1 i¼1
in the two groups if the computed F-value exceeds the table
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi value. If the calculated value is less than the table value we
 would have no evidence that the variances are different.
Sd ¼ ðSd2 nÞ
The two-sample data sets will represent the same popula-
From Table 9.5: tion in the case of S1 2 ¼ S2 2 .
d ¼ 2:4 Sd ¼ 24 Sd2 ¼ 270
9.2.17 The Chi-Square Test
Sd2 ¼ fð10
270Þe576g=ð10
9Þ ¼ 23:60
The chi-square (c2) test relates to nonparametric statistics,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffi
Sd ¼ ð23:60=10Þ ¼ 2:36 ¼ 1:536 is typically easy to compute, and is capable of processing
data measured on a nominal scale (categorical variables).
Paired t ¼ 2:40=1:536 ¼ 1:56 c2 is a statistical test frequently used to compare observed
 Statistical and Geostatistical Applications in Geology Chapter | 9 179

data with data expected to be obtained according to a 9.2.18 Analysis of Variance

specific hypothesis. If a sample of size n is taken from a
population having a normal distribution with mean m and The analysis of statistical variance (ANOVA) compares
standard deviation a, it may allow a test to be made of groups of samples. The observed variance in a particular
whether the variance of the population has a predetermined variable is partitioned into components attributable to
value. Each observation can be standardized to the standard different sources of variation. There are several models and
normal form with 0 mean and unit variance following procedures to perform analysis of variances due to the

 effects of categorical factors. The models are one-way
Zi ¼ Xi  X S) for samples or Z ¼ (Xi  m)/a for
ANOVA, multifactor ANOVA, variance component anal-
population parameters. The values of Z are squared to
ysis, and the general linear model. One-way ANOVA deals
eliminate the sign and summed to formed a new statistic:
P 2 P with single categorical factors equivalent to comparing
Z ¼ ni¼ 1 fðXi  mÞ=ag2 . multiple groups of data sets. Multifactor ANOVA uses
Data processing, interpretation, and testing of the hy- more than one categorical factor arranged in a crossed
pothesis are similar to parametric t- and F-tests. The chi- pattern. The variance component analysis model utilizes
square test computes a value from the data using the c2 multiple factors arranged in a hierarchical manner. The
procedure. The value is compared to a critical value from a linear model is a complex type and uses categorical and
c2 table with a degree of freedom equivalent to that of the quantitative, fixed and random factors arranged in crossed
data (Box 9.2). If the calculated value is greater than or and nested patterns. The general format of the one-way
equal to the table value the null hypothesis is rejected. If the model is presented in Table 9.7.
value is less than the critical value the null hypothesis is Five ore samples are collected from a zinc-lead deposit.
accepted. c2 is an important test in nonparametric statistics. The samples are analyzed for % Pb content at six standard
It is used to test the difference between an actual sample laboratories (Table 9.8). Each assay value of one sample
and another hypothetical or previously established popu- analyzed at different laboratories is called a replicate. In
lation, which is expected due to chance or probability: one-way analysis, total variance of the data set is broken
X
n
2
into two parts: variance within each set of replicates and
c2 ¼ ðOi  Ei Þ Ei variance among the samples.
i¼1 The total variance of all observations (all replicates of
where Oi ¼ observed frequency of the ith class and all samples) can be computed as:
Ei ¼ expected frequency of the ith class. !2 
X m X n Xm X n
Example: A balanced coin is tossed 20 times and yields SST ¼ Xij 
2
Xij N
12 times heads and 8 times tails against expected proba- j¼1 i¼1 j¼1 i¼1
bility values of 10 and 10 (Table 9.6).
The calculated value of c2 is 0.80 (without any 
¼ 96:66  ð52:77Þ2 30 ¼ 3:84
correction factor) and is much less than the table value of
3.84 at 1 degree of freedom, which is not statistically sig- The variance among the samples and within replicates is:
nificant at the 0.05 level. It can be concluded that the de- ( ) !2 
Xm X n
 Xm X n
viation between observed and expected frequency is due to SSA ¼ 2
Xij n e Xij N
sampling error or to chance alone. j¼1 i¼1 j¼1 i¼1

2
¼ 95:20eð52:77Þ 30 ¼ 2:37
BOX 9.2 CRITICAL VALUES OF c2
Critical Values of c2 for Various Degrees of Freedom and SSW ¼ SST  SSA ¼ 3:84 e 2:37 ¼ 1:47
Selected Levels of Significance from Table 24,
Penguin-Honeywell Book of Tables (Kellaway, 1968).

TABLE 9.7 Model of One-Way Analysis of Variance

of Grouped Data
TABLE 9.6 Results of Coin Toss and c2 Calculation
Source of Sum of Degree of
Observed Excepted Variation Squares Freedom F-Test
(Oi) (Ei) (Oi L Ei)2 (Oi L Ei)2/Ei Among samples SSA me1 MSA/MSW
Head 12 10 2 0.40 Within replicates SSW Nem
Tail 8 10 2 0.40
Total variation SST Ne1
Total 20 20 0.80
m, number of samples; n, number of replicates and N ¼ m
n.
180 Mineral Exploration

TABLE 9.8 % Pb Assay Value of Five Samples Analyzed at Six Standard Laboratories

m[5

n[6 Replicate (Lab) Sample 1 Sample 2 Sample 3 Sample 4 Sample 5 N [ 5 3 6 [ 30

1 1.92 1.87 1.25 2.03 1.99
2 1.87 1.43 1.43 2.25 2.43
3 2.13 2.02 0.87 1.76 1.76
4 1.65 1.76 1.14 1.84 2.02
5 1.73 1.93 0.95 1.59 1.84
6 2.24 1.61 1.65 1.90 1.91
P
x 11.54 10.62 7.29 11.37 11.95 ¼ 52.77
P 2
x 22.45 19.03 9.29 21.80 24.07 ¼ 96.66

9.2.19 Trend Surface Analysis

TABLE 9.9 Results of Analysis of Variance
Trend surface mapping is a mathematical technique for
Source of Sum of Degree of Mean computing an empirical 2D plane or 3D curves, contours,
Variation Squares Freedom Squares F-Test and wire meshes derived by the regression method. It
Among 2.37 4 0.59 10.17 provides the overall structure of spatial variation present in
samples the data set. It is a linear function of geographic coordinates
Within 1.47 25 0.058 of total scattered observations; it constructs global func-
replicates tional relations and estimates new values on regular grid
Total 3.84 29 points. The estimated values must minimize the deviations
variation from the trend functioning as a wide range operator. It is a
global pattern recognition technique in mineral exploration.
The mathematical equation expressing the functional rela-
tionship is stated as:
Zn ¼ a0 þ a1 x þ a2 y þ e ðfirst  degree equationÞ

Zn ¼ a0 þ a1 x þ a2 y þ a3 x2 þ a4 xy þ a5 y2 þ e
ðsecond  degree equationÞ
and by similar types of equations, where:
Zn ¼ dependent variable ðsample valueÞ
X; Y and Z ðelevationÞ ¼ independent variables ðlocation
of sample coordinatesÞ
e ¼ random error component;
FIGURE 9.10 Analysis of the F-test from the table value at a 5% level of a0 .an ¼ unknown coefficients.
confidence.
The number of coefficients is dependent on the degree
(NDEG) of the polynomial used, and is computed as
fðNDEG þ 1Þ
ðNDEG þ 2Þg=2. The higher degree of
The results of the analysis of variances as derived from polynomial attempts to reduce the noise level in data
data set in Table 9.8 are given in Table 9.9. minimizes the sum of squares of residuals, and allows
The computed F-value of 10.17 is greater than the table definition of the trend in a unique manner. However, it is
value of 2.76 at a 5% level of confidence (Fig. 9.10). observed that after reaching a certain degree, the linear fit
Therefore it can be concluded that the variation is not the becomes static due to error propagation and the noisy na-
same in the two groups. ture of the large matrices.
 Statistical and Geostatistical Applications in Geology Chapter | 9 181

In a first-degree equation we need three normal equa- The dependent variables, e.g., Pb, Zn, Fe, and width, are
tions to find the three coefficients: assumed as a numerical functional relation of criteria var-
P X X iables of x (strike) and y (level) coordinates. The trend
Zn ¼ a0 n þ a1 x þ a2 y
surface maps of lead and zinc at the fourth degree are
P X X X
generated on a longitudinal vertical section (Figs. 9.11 and
xZn ¼ a0 x þ a1 x2 þ a2 xy
P 9.12, respectively) considering local coordinates of x and y.
X X X
yZn ¼ a0 y þ a1 xy þ a2 y2 The zinc value has been computed at every observed
drill intersection point of the orebody using the fourth-
The equation in matrix form is:
degree trend surface equation. The comparison between
2 P P 3 2 3 2 P 3
n x y a0 Zn actual sample and computed values is given in Table 9.10.
6P P 2 P 7 6 7 6P 7 The analysis and observations from actual samples and
4 x x xy 5
4 a1 5 ¼ 4 xZn 5
P P P 2 P the trend surface equation (Table 9.10) are:
y xy y a2 yZn
The coefficients for the best-fitting linear trend surface Total variance (TV) ¼ 71.80
are obtained by solving this set of simultaneous equations. Unexplained variance (UV) ¼ 12.10
The trend surface technique is illustrated by a 2D case
study. A zinc-lead orebody extends along N42 WeS42 E Explained variance (EV) ¼ (TV  UV) ¼ 59.70
over a strike length of 90 m. It is lens shaped and widest at
the center. The width varies between 2 and 16 m with an Goodness of fit (GF) ¼ (EV/TV)
100 ¼ 83.15%
average of 7 m. It pinches at both ends. Mineralization dips GF is a measure of efficiency of the trend equation
to the west and plunges 55 NW. The orebody has been applied, and depends on input data and inherent structure.
drilled from underground at four levels at 15 m intervals. The extreme values will lower the GF even with a large
The mineralized intersections are predominantly zinc rich database. GF is a useful measure for a homogeneous
with moderate iron and subordinate amounts of lead. The population, but is not appropriate for skewed or mixed types.
grades vary between 3.96 and 11.18 with an average of Hence lack of significance or high significance needs to be
6.87% Zn, 0.00 and 5.98 with an average of 0.40% Pb, and supported by geological interpretation. It determines the
2.44 and 15.25 with an average of 6.05% Fe. The drill functional relationship between predictor and criteria variates.
intersection values are plotted along a longitudinal section.

FIGURE 9.11 Fourth-degree trend surface map of lead content of orebody 7W, Balaria deposit, Rajasthan, India.
182 Mineral Exploration

FIGURE 9.12 Fourth-degree trend surface map of zinc content of orebody 7W, Balaria deposit, Rajasthan, India.

TABLE 9.10 Computed Value of Zinc Content of Orebody 7W, Balaria Deposit, Rajasthan, India

% Zn % Zn % Zn % Zn
Level Latitude (Sampled) (Computed) Level Latitude (Sampled) (Computed)
343 68 6.11 6.05 288 102 4.85 4.27
343 86 7.17 7.46 263 15 6.44 6.67
343 102 11.18 10.24 263 38 8.28 8.26
343 105 10.01 10.62 263 54 6.81 7.92
343 134 4.41 4.49 263 67 8.54 7.04
314 46 6.63 6.75 263 84 4.51 4.81
314 60 6.37 6.21 250 5 7.30 7.15
314 82 7.70 7.55 250 35 8.22 8.23
314 91 8.73 7.88 250 64 4.13 4.27
314 101 7.44 7.17 317 59 6.34 6.17
288 28 6.20 6.19 286 34 6.29 6.41
288 49 7.58 6.98 330 81 6.88 7.18
288 74 8.74 7.92 309 80 6.11 7.56
288 90 5.34 6.96

Unpublished Ph.D. Thesis, Haldar, 1982.

The accuracy of map generation is centered on the midpoint The trend surface map can indicate (1) possible conti-
of the 2D grid. The computed trend value and actual values nuity of mineralization in depth and strike directions, (2)
may show a strong divergence away from the midpoint. inherent pattern of mineralization like metal zoning and
The margins are the most inaccurate zones. Extrapolation of association, and (3) the grain of longer and shorter axes of
the trend equation beyond the control point will result in variations. It may often generate false anomalies since the
erroneous values. separation of noise into positive and negative sets is closely
 Statistical and Geostatistical Applications in Geology Chapter | 9 183

linked and together they sum to zero. Smoothing of higher to extreme grade fluctuation due to the volume/variance
values and enhancement of lower values are an inbuilt part relations.
of the procedure. Subset trend maps show no relation to The weight factor (W) of a block is to be determined to


composite trend maps. In spite of those constraints, these establish a moving average based on block mean value Y .
are utilized for a generalized pattern of global variation. A block moving average design is shown in Fig. 9.13. Y e is
the estimated grade of block 1 from the averages of the
9.2.20 Moving Average exploration sample in blocks 1e9. Ye is the estimated grade
Moving average is a technique of smoothing data that occur of ore to be produced from the center block and Y 1 is the
infrequently at high and low values. It is an estimator mean assay value of the drill core in the same block. The
operating within a pattern of adjacent values by designing sample design will move to a new region once the equation
appropriate weighting criteria. The pattern may be linear is obtained:
combinations, circular, rectangular, or square window. X
n
fij ¼
Y Wk Yk
k¼1
9.2.20.1 Moving Weighted Average of
X
n
Block Mean Y ¼ bo þ bi Y i
i¼1
A sizable block of ore is removed during a mining opera-
tion. The minimum size depends on the orebody configu- Y ¼ Actual from past record
ration and extraction method. It varies from deposit to Y ¼ Mean of exploration data
deposit and mine to mine. Large blocks are less subjective
b ¼ Coefficients
The estimated Y e is based on the weighted sum of
adjacent observation Y:
X
n
e ij ¼
Y Wk Y k
k¼1

e ¼ estimated value
Y
W ¼ weight
Y ¼ adjacent block value
The sum of total weight must be equal to 1:
X
n
Wk ¼ 1
k¼1

9.2.20.2 Moving Arithmetic Average

The arithmetic average of adjacent samples on either side
controlled by a window can give an estimation of near-
FIGURE 9.13 Diagram showing concept of “block” estimation by the homogeneous ore distribution as in the case of under-
moving average method. ground working or grid samples (Fig. 9.14A and B).

FIGURE 9.14 Concept of moving arithmetic average grade computation of (A) underground workings and (B) geochemical grid samples on the surface.
184 Mineral Exploration

9.2.20.3 Other Mathematical Techniques sample grade. In reality, a number of blocks would have a
similar estimated grade with wide variation from true
The occurrence of erratic high values in the spatial distri-
grade. This indicates that any estimated value has a bias
bution, as in the case of gold, silver, platinum, and similar
(error) when compared with the true block values except
precious metals, can be smoothed by log transformation
where the line of best fit intersects the 45 degree line. The
(Fig. 9.15A and B).
average of all the blocks’ estimates would not necessarily
be much different than the true average grade. If we apply
9.3 MISCLASSIFIED TONNAGE a desired cutoff grade, four regions, namely, I, II, III
and IV, will be formed on the estimated true grade
Krige (1951) observed that blocks of ore estimated as high
distribution:
grade using long-standing traditional methods repetitively
indicate lower true grade after mill feed-grade reconcilia- 1. Region I is comprised of all those blocks that have both
tion. Similarly, estimated low-grade blocks signify a con- an estimated and true grade above the cutoff grade,
stant higher true grade. The true versus estimated grades of which would be and should be mined.
all blocks is plotted in Fig. 9.16. An elliptical envelope is 2. Region II includes all those blocks that have an esti-
observed with a major axis (least square fit) not parallel to mated and true grade below the cutoff grade, which
the 45 degree line from the origin. would not and should not be mined.
All the estimated versus actual points would fall on or 3. Region III includes blocks estimated to be above the
close to the 45 degree line if the blocks had been estimated cutoff grade but have a true grade below the cutoff
correctly. The estimation technique is expected to provide grade. These blocks would be mined when they should
close proximity between estimated and actual production not and would thus reduce the mine head grade.

FIGURE 9.15 Transformation of (A) fluctuating raw samples to (B) smooth values.

FIGURE 9.16 Misclassified tonnages as conceived by Krige (1951) portray that the “estimated” and “true production” grades vary considerably due to
inadequate sampling, estimation flaw, and complexity of mineralization.
 Statistical and Geostatistical Applications in Geology Chapter | 9 185

4. Region IV is comprised of blocks estimated to be below

the cutoff grade, but have a true grade above the cutoff
grade. These blocks would not be mined when they
should have been and thus will reduce the mine head
grade.
The misclassified-tonnage curve represents an envelope
of dispersion or variance of block grades. The difference
between the expected or planned mine-head grade and
true grade can be identified after the application of cutoff
grade criteria. It is therefore necessary to employ that
technique, which decreases the variance in block esti-
mates, i.e., decreases the spread of block estimates so that
the regression line moves to 45 degrees. In addition, it is
necessary to employ an estimating technique that will FIGURE 9.17 Concept of mutual relationship between sample volume
reduce variance of the error of estimation or estimation size and variances.
variance so that the envelope narrows and ideally becomes
the 45 degree line. statistical information. Some geological concepts that need
to be considered are:
9.4 GEOSTATISTICAL APPLICATIONS 1. The better the orebody continuity, the smaller the error
associated with extrapolating a given grade over a larger
These problems of grade/tonnage mismatch and wider area.
grade variances of estimated blocks have been resolved by 2. Influence of a sample over an area may be related to a
developing regionalized stationarity and variability of geological zone of influence that may vary in different
metal distribution within the deposit (Matheron, 1971, directions (anisotropy).
Hans, 2013). 3. Some ore deposits, such as gold, silver, and platinum,
may exhibit a nugget effect (large variation in grade
9.4.1 Block Variance over a small distance) and therefore larger samples
may be required since a sample is a volume (support)
Let us assume that we have a large number of evenly and not a point.
spaced sample points distributed within the orebody. The 4. Demarcate geologically distinct areas within a deposit
body is divided into a number of blocks of equal size. An so that different estimation procedures may be required
average grade can be computed for each block by taking the (nonstationarity).
arithmetic or weighted average of all the sample points
falling in that block. The simplest case would be an equal 9.4.2 Semivariogram
number of blocks with one sample in each block. The other
extreme case would be only one block representing the The most natural way to compare two values is to consider
whole deposit that contains all sample points. If the vari- their differences. Let us assume that two samples Z(x) and
ances of the block grades are plotted against size of the Z(x þ h) are located at two points, x and x þ h, with two
block, the block variance relationship is obtained (Fig. 9.17). different % metal grades. The second sample is h meters
Obviously, the variance of block grades decreases with away from the first. This value expressing the dissimilarity
increase in block size. As each block value is an average of of grade between two particular points is insignificant. The
all sample points in the block and the mean of the ore average difference for all possible pairs of samples at h
deposit is the mean of all block values, the mean of the meters apart throughout the deposit will be geologically
deposit remains constant. The variance of the errors be- significant. The difference between any pair of samples will
tween the true grade of a block and that estimated from a be either positive or negative, and the sum of all pairs will
center sample point generally decreases with increase in be misleading. It is therefore logical to square the differ-
block size. ences, sum them, and divide by the number of pairs to
The estimation of block size will depend on mine understand the average variance of paired samples at
planning and it is necessary to select an estimation tech- certain distances apart. This dissimilarity function is
nique that will take into account all surrounding samples to expressed by the model equation:
reduce the variance of the estimation error. An orebody
model has to be developed based on geological and 2g (h) ¼ Average S{Z(x)  Z(x þ h)}2
186 Mineral Exploration

FIGURE 9.18 Concept of semivariogram by sliding sample data string at

certain space intervals.

FIGURE 9.19 Drawing of standard semivariogram along the drill hole

samples.
where 2g (h) is the semivariogram or simply variogram
function. It is the function of a vector; in other words, a
This can be represented in a semivariogram plot of lag
distance and the orientation of that distance. It articulates
against g(h) function along the x- and y-axes, respectively
how the grades differ on average according to the distance
(Fig. 9.19). The freehand or fitted curve is extended
in that direction (Fig. 9.18).
downward to intersect the variance axis. If it touches above
The semivariogram is thus stated as:
the origin, this part is known as the nugget effect (C0). The
X
N
20
curve then rises up to the maximum variance (population
gðhÞ ¼ 1=2N fZðxÞeZðx þ hÞg variance d2) at a particular lag equivalent distance and
i1
levels out or flattens. This distance is known as range (a).
where g(h) ¼ semivariogram function value, N ¼ number Each sample value has an influence up to about 2/3 of the
of sample pairs, Z(x) ¼ value at point (x), may be grade, range. The extension variance C1 is the difference between
width, accumulation, etc., and Z (x þ h) ¼ value at point population variance and nugget effect.
(x þ h), i.e., h distance away from point (x). In the case of 2D grid sample data the variogram is
The simplest semivariogram computation can be by computed in different directions. Sample points on either
sliding mineralization in one direction along a borehole side of the variogram line, within acceptable limits, are
(Fig. 9.18). Let there be a set of five samples of 1 m length projected on it. The variance between points is computed
each along a borehole as 2%, 6%, 3%, 9%, and 12% Zn. and grouped under a similar lag range. This process is
The sample string slides by one step (1 lag), i.e., at 1, 2, 3 m repeated on all samples and arranged in serial order as per
lag, to compute the average variance and continue. lag. The number of participating pairs contributing to the

Sample Lag Difference Difference2

2 1m
6 2 4 16 Ʃ difference ¼ 702
3 6 3 9 No. of pairs ¼ 4
9 3 6 36 Average variance ¼ 70/(2
4)
12 9 3 9 Semivariance ¼ 8.75
2 2m
6
3 2 1 1 Ʃ difference2 ¼ 91
9 6 3 9 No. of pairs ¼ 3
12 3 9 81 Average variance ¼ 15.16
2 3m
6 Ʃ difference2 ¼ 85
3 No. of pairs ¼ 2
9 2 7 49 Average variance ¼ 21.25
12 6 6 36
 Statistical and Geostatistical Applications in Geology Chapter | 9 187

pffiffiffi
sum of differences square is known, and a variogram is The next shortest sample distance is 2 unit distance
computed. Similarly, variograms in other directions within apart with total sample pairs of 12:
the study area are computed and compared. pffiffiffi pffiffiffiffiffiffiffi
In practice, a theoretical semivariogram is never real- r g 2 ¼ 26=24 or g 1:4 ¼ 1:08
ized. The gamma function g(h) is estimated from limited
points and is called the experimental variogram (Fig. 9.20). The gamma functions for all the possible lags are
Fig. 9.20 displays a number of sample points at equal computed and plotted as g(h) against lag. The semivario-
distances apart with their sample values, say thickness next gram is shown in Fig. 9.21.
to each point. We wish to compute g(1), i.e., the semi-
variogram value of all samples points at 1 unit distance 9.4.2.1 Properties of a Semivariogram
(lag) apart, irrespective of direction. There are 17 sample
9.4.2.1.1 Continuity
pairs at 1 unit apart:
Continuity is reflected by the growth rate of g(h) for a
gð1Þ ¼ 1=2Nð1  2Þ2 þ ð2  2Þ2 þ ð2  3Þ2 þ ð2  1Þ2 þ small value of h. The growth curve exhibits the region-
alized element of samples. The steady and smooth in-
ð1  4Þ2 þ ð4  3Þ2 þ ð2  1Þ2 þ ð1  1Þ2 þ crease is indicative of the high degree of continuity of
mineralization until it plateaus off at some distance. This
ð1  2Þ þ ð1  2Þ þ ð2  2Þ þ ð2  1Þ þ is known as the structured variance or explained variance
2 2 2 2

(C or C1) read on a g(h) scale. A typical semivariogram

ð1  1Þ þ ð2  4Þ þ ð4  1Þ þ
2 2 2 of a coal deposit showing good continuity is given in
Fig. 9.22.
ð3  3Þ þ ð3  2Þ
2 2

9.4.2.1.2 Nugget Effect (C0)

¼ 31=34 ¼ 0:91 The semivariogram value at zero separation distance
(lag ¼ 0) is theoretically 0. However, a semivariogram
often exhibits a nugget effect (>0) at an infinitely small
distance apart. The complex mineral deposits, like base and
noble metals, may occur as nuggets, blobs, and are often
concentrated as alternate veinlets resulting rapid changes
over short distances. The gamma functions are extrapolated
back to intersect the y-axis (Fig. 9.23). The nugget effect
(C0) is the positive measure of g(0) representing random or
unexplained elements of samples. Noble and precious
metals like gold, silver, and platinum-group elements often
exhibit very high nugget effects even up to total unex-
plained variance. This creates a lot of uncertainty in the
FIGURE 9.20 Computation of experimental semivariogram of grid continuity of mineralization and grade estimation leading to
sample data along various directions such as 0 and 90 degrees. extensive sampling.

FIGURE 9.21 Typical semivariogram plot with curve fitting following the population variance after reaching the zone of influence.
188 Mineral Exploration

9.4.2.2.1 Features of a Semivariogram

The sill ðC1 þ C0 Þ.
The growing semivariogram curve normally reaches a
plateau after a certain lag, and equals the population vari-
ance. This is known as the sill. In practice the gamma
function may fluctuate on the sill line because of metal
zoning, layering, and the hole effect:
SillðCÞ ¼ C1 þ C0

9.4.2.2.2 The Zone of Influence (a)

The zone of influence is that neighborhood beyond which
FIGURE 9.22 Typical spherical semivariogram from coal deposits the influence of a sample disappears and samples become
without any nugget effect.
independent of each other. The zone of influence in a given
direction is characterized by the distance at which the
There are three possible reasons for the nugget effect of semivariogram eventually reaches a plateau and levels out.
various magnitudes: The first few points of g(h) are joined by freehand or
polynomial regression, and extended to the sill. The inter-
1. Sampling and assaying errors.
section with the sill is w2/3 of the zone of influence or
2. Smaller microstructures or nested variograms at shorter
range (a).
distances where no borehole data exist.
3. Combination of both.
9.4.2.2.3 The Isotropic Anisotropies
9.4.2.2 Semivariogram Model The isotropic anisotropies are defined by computing semi-
variograms in different directions. The semivariogram that
The computed experimental points of a semivariogram can points in the eastewest orientation of Fig. 9.20 will be:
be tailored to a standard type to fit a model. There are many
semivariogram models that fulfill certain mathematical gð1Þ ¼ 15=18
constraints (Rendu, 1981; Clark, 1982). The models can be gð2Þ ¼ 12=12
broadly divided into two groups: (1) those with a sill and
(2) those without sill. A sill implies that, once a certain gð3Þ ¼ 5=6
distance h is reached, the values of g(h) do not increase,
although they may oscillate around the population variance. Semivariograms in the northesouth or any other di-
Models without a sill have growing values of g(h) with rection can be computed and compared. The semivariogram
increasing values of h. is isotropic if the underlying structure exhibits the same
features irrespective of the directions. The semivariograms
is anisotropic if the structure displays different features in
various directions. An average semivariogram can be
considered in such a condition.
The types of semivariogram with and without sills are
given in Figs. 9.24 and 9.25 and as follows:
1. Spherical model


 
gðhÞ ¼ C0 þ C1 1:5h=ae0:5 h3 a3 ; for h < a; and

gðhÞ ¼ C0 þ C1 ; for h  a
where C0 ¼ nugget value, C1 ¼ explained variance,
C0 þ C1 ¼ sill, h ¼ distance between points, and
a ¼ zone of influence or range.
2. Random model
FIGURE 9.23 Typical semivariogram from massive base metal deposits gðhÞ ¼ C0 þ C1 ; for all h
with moderate unexplained variance (C0).
 Statistical and Geostatistical Applications in Geology Chapter | 9 189

FIGURE 9.25 Semivariogram models without sill: (A) parabolic,

(B) logarithmic, and (C) linear.

The following points are to be observed when con-

structing an experimental semivariogram:
1. Most semivariograms (>95%) in mineral applications
FIGURE 9.24 Semivariogram models with sill: (A) spherical, are spherical models. The sill is considered equal to
(B) random, and (C) exponential. d2 to fit the variogram. A straight line through the initial
few points should intersect the line of the sill at w2/3 of
the slope.
3. Exponential model 2. A minimum of 50 pairs of samples is used for the

 computation of experimental points. Approximately
gðhÞ ¼ ðC0 þ C1 Þ 1eeh=a ; for h < a 100 sample values are required to achieve this, although
satisfactory models can be fitted with fewer data.
The types of variogram without sills are (Fig. 9.25): 3. As distance h increases, the number of sample pairs de-
creases, indicating less reliability of g(h) for larger dis-
1. Linear model tances. The initial 5e10 experimental points are enough


gðhÞ ¼ a2 h=2 to fit a model.

where a2 is a constant. 9.4.2.3 Smoothing a Semivariogram

2. Logarithmic model
The sample data in Fig. 9.20 are equally spaced. However,
gðhÞ ¼ 3a logh the coordinates of borehole collars are irregular to some
3. Parabolic model odd meters away from the gridlines, even in a regular
drilling pattern. If a semi-variogram is to compute at a
gðhÞ ¼ 1=2a2 h2 specific distance h, it may not find a sample pair at exactly
190 Mineral Exploration

h distance apart. The g(h) computation for calculated

values of h will be incomplete due to inadequate sample
pairs. This will result in a random pattern of experimental
semivariograms and no model can be fitted to a statistically
valid experimental prototype.
The effect of data at irregular intervals can be preserved
as a noise, and the data need to be smoothed to find the
underlying semivariogram structures. Optimal smoothing
can be done by (1) step interval and (2) angular tolerance.

9.4.2.3.1 Step Interval

The simple procedure is to select experimental points for a
range of distances Dh. If the samples are irregularly spaced
at an average distance of w100 m, a step interval of 110 m
can be opted to select all sample pairs from 0 to 100 m, 100
to 200 m, 200 to 300 m, etc. Semivariograms are computed
at different step intervals of þ5, 10, and 20 m with most
variables occurring at short intervals. As the step value
increases, the filtering effect reduces the noise, and a FIGURE 9.26 Angular regularization-sample falling in sketch area are
smoother and more statistically viable model is obtained. accepted by tolerance in distance (David, 1988).
Oversmoothing will mislay the underlying structure, and a
straight line of experimental points at population variance
will appear incorrectly, suggesting a random model with no
The greater the distance for each step and the wider the
special correlation.
angle around a specific direction, the greater is the
Referring to sample data from Fig. 9.20 the g(h) in an
smoothing. Insufficient smoothing could be difficult to fit a
eastewest direction is:
model. Oversmoothing will destroy the actual character.
gð0  2Þ ¼ 27=30

gð2  4Þ ¼ 5=6 9.4.3 Estimation Variance

The average of the distance is used for the single value Estimation variance is the variance of error made in esti-
of g(h): mating the grade of a panel or block of ore by assigning the
gð1:4Þ ¼ 27=30 sample values lying in and around it within the zone of
influence. It is apparent from the semivariogram that the
gð3Þ ¼ 5=6 relative location of samples to block will have an influence
upon the weighting coefficients that can be assigned to each
sample value.
9.4.2.3.2 Angular Regularization Z(Vi) are true unknown grades of blocks V i, and Z*(Vi)
P
Angular regularization is based on the fact that irregular are the linear combinations, where Z*(Vi) ¼ nj¼ 1 aj ZðXj Þ
samples may not fall along specific directions when of the sample grades at locations Xj (j ¼ 1.n). Then
anisotropic behavior is apprehended. No pair of samples Z*(Vi)  Z(Vi) will be the error made in assuming that the
may lie along the eastewest axis (0e180 degrees). This value Zi extends over V when the true value is Vi, and
method of smoothing is not to calculate the semivariogram VAR(Zi eZi ) is the variance of this error.
for a single discrete direction, but for a direction and range

VARðErrorÞ ¼ VAR Zi eZi
Da to either side of it (Fig. 9.26). Therefore with a direction
of 0 degrees, angular regularization of 45 degrees, and a If a block value is estimated by a sample or a number of
step interval of 1 unit distance in Fig. 9.20, the values of samples each given an equal weight, then the variance of
g(h) are for the exclusive/inclusive range: the error d2e associated with the estimated mean is:
gð0  1Þ ¼ gð1Þ ¼ 15=18 d2e ðWs to WÞ ¼ þ2yðW2 ; WÞ  yðWs ; Ws Þ  ðW; WÞ
gð1  2Þ ¼ gð1:7Þ ¼ 28=36
where Ws ¼ sample support and W ¼ block.
gð2  3Þ ¼ gð2:3Þ ¼ 33=30 Or, in the presence of the nugget effect, C0/n is to be
gð3  4Þ ¼ gð3:3Þ ¼ 6=12 added, where n ¼ number of samples.
 Statistical and Geostatistical Applications in Geology Chapter | 9 191

the estimator kriging. The principles of the estimation

procedure are:
1. It should be a linear function of the sample value xi
(Armstrong, 1998). The block value of:
X
n
W ¼ b1 x1 þ b2 x2 þ /bn xn
i¼1

where bi is the weight given to sample wi.

2. It should be unbiased. The expected value (mk) should
be equal to the true block value (mW):
EfðmkemWÞg ¼ 0
3. The mean squared error of estimation of mW should
be a minimum:
 2
E ðmkemWÞ ¼ a minimum

FIGURE 9.27 Block (W) estimation representing nine subblocks with

The kriging estimator (mk) satisfies these conditions of
one sample in the center (w1) and one sample at the corner (w2). linear function, unbiased estimation, and minimum vari-
ance. The corresponding error of estimation

 from sample
(ws) to block (W) is the kriging error d2e . The kriging
In the general case where the weights are different for estimator is also known as BLUE.
each sample that is used to estimate the block value, the A block W of 3
3 is estimated using two samples: w1
weights still add up to 1.0 so that the estimation is not in the center and w2 in the corner (Fig. 9.27). A set of
biased, and the formula for the estimation variance is: simultaneous equations is used to resolve for the optimum
set of weights assuming the population mean is unknown:
X
n n X
X n


d2e ¼ 2 bi yðwi ; WÞ  bi bj y w i ; w j  n X
X n

 X
n

i¼1 i¼1 j¼1

bi bj y wi ; wj þ l ¼ bi yðwi ; WÞ
i¼1 j¼1 i¼1
yðW; WÞ
where l ¼ the lag range multiplier.
where bi or bj ¼ the weights for each sample and n ¼ 1, I,
The error of variance d2e can be estimated from the
or j.
kriged weights by:
Therefore P
Block value ¼ ni¼ 1 bi Zðwi Þ X
2 X
2 X
n
P ¼ yðw; WÞ þ bi yðwi ; WÞ þ l
where Z(wi) ¼ sample value and ni¼ 1 bi ¼ 1.
e k i¼1
It is possible to vary the two weights applied to w1 and
w2 in an infinite number of combinations (Fig. 9.27). A For example, and ignoring l for the moment:
new d2e can be estimated for each combination, but only one
b1 y11 þ b2 y12 ¼ y1 W
combination will minimize d2e . Kriging is the method of
determining b1 and b2 to minimize d2e . The optimal weights b1 y21 þ b2 y22 ¼ y2 W
are derived from Krige’s equation dependent on a semi-
The simultaneous equations when written in matrix
variogram or spatial correlation of distance function and
form are:
direction.
yðw1 w1 Þ yðw1 w2 Þ b1 yðw1 ; WÞ

¼
9.4.4 Kriging yðw2 w1 Þ yðw2 w2 Þ b2 yðw2 ; WÞ

Prof. D.G. Krige initiated and developed the application of or:

geostatistics in the valuation and optimization of ore in 0 1 b1 0:536
South African gold mines in the early 1970s (Krige, 1951,
¼
1962, 1978; David, 1977, 1988; Journel and Huijbregts, 1 0 b2 0:882
1978; Sinclair and Blackwell, 2002). Prof. Matheron called Therefore b2 ¼ 0.536 and b1 ¼ 0.882.
192 Mineral Exploration

However, b1 þ b2 must be equal to 1, but in this case 9.4.5 Kriging Estimation: An Example
(0.536 þ 0.882) ¼ 1.1418. Therefore the lag range multi-
plier is to be introduced to ensure b1 þ b2 ¼ 1. The new Point and block kriging estimates the value at a point or
matrix is: block, exclusively supported by a semivariogram. The ac-
2 3 2 3 2 3 curacy of kriging estimation, to a large extent, is propor-
yðw1 w1 Þ yðw1 w2 Þ 1 b1 yðw1 ; WÞ tional to the precision factor of the variogram. Point kriging
6 7 6 7 6 7
4 yðw2 w1 Þ yðw2 w2 Þ 11 5
4 b2 5 ¼ 4 yðw2 ; WÞ 5 has been suggested as a test for a semivariogram model.
1 1 0 1 Each sample point is kriged by a selected semivariogram
l
model without actual sample values being used in the
or: matrix. The kriged point values are compared with actual
2 3 2 3 2 3 sample values and the model is adjusted until the errors are
0 1 1 b1 0:536
6 7 6 7 6 7 minimized. This is known as the jackknife test.
41 0 1 5
4 b2 5 ¼ 4 0:882 5 The stepwise kriging computation for part of an ore-
1 1 0 l 1 body (sublevel stope) is demonstrated by seven sample
points. The sample details, semivariogram for kriging at a
r b2 þ l ¼ 0:536 central point “Q” of block ABCD (25 m
25 m), are given
in Table 9.11, Figs. 9.28 and 9.29.
b1 þ l ¼ 0:882
The semivariogram of the total population provides
b1 þ b2 ¼ 1 C0 ¼ 0.8, C1 ¼ 2.5, and a ¼ 50 m. Five samples, namely,

Solving b1, b2, and l:

b1 ¼ 0:673 TABLE 9.11 Sample Coordinates and Grade From a
b2 ¼ 0:327 Sublevel Stope
l ¼ 0:209 Coordinate (m)
b1 þ b2 ¼ 1 as required Sample No. Easting Level % Zn
The mean estimate for the block is: 1 314 83 2.40

0:673Zðw1 Þ þ 0:327Zðw2 Þ 2 314 93 2.76

3 314 107 4.84
And the error of estimation is:
4 288 74 5.98
d2e ¼ 0:683 þ 0:673
0:536 þ 0:327
0:882 þ 0:209 5 288 90 6.38
¼ 0:175 6 288 105 7.04
7 298 76 3.63
where de ¼ 0.418%, the confidence limit, which is indepen-
dent of grade.
In summary, we observe that if we estimated the block
value by different sample configurations the following er-
rors were made:
In the block estimated from the corner sample,
d2e ¼ 1:081.
In the block estimated from the center sample,
d2e ¼ 0:389.
In the block estimated from the corner and center
samples with equal weights, d2e ¼ 0:235.
In the block estimated from the corner and center
samples using kriged weights, d2e ¼ 0:175.
The kriging method is best with minimum error vari-
ance, unbiased as weights sum to 1.0 denoting expected
block value equal to true block value, and a linear estimator
in its application. FIGURE 9.28 Semivariogram parameters of samples from Table 9.9.
 Statistical and Geostatistical Applications in Geology Chapter | 9 193

FIGURE 9.29 Conceptual block kriging from surrounding samples.

1, 2, 3, 5, and 6, are selected that fall within the radius of This relates interassay variance to variance contribution
influence of 30 m from computation point Q. The distances to point Q. The solution to the foregoing equations will
d1 (1 to Q), d2 (2 to Q), d3 (3 to Q), d5 (5 to Q), and d6 (6 to provide five weighting coefficients for five sample assays,
Q) are necessary for interpolation. The kriging technique and another coefficient l for the kriging estimator. The
computes distances and semivariances between 1 and 1, 1 coefficients are computed as:
and 2, 1 and 3, 1 and 5, 1 and 6, 2 and 1, 2 and 2, 2 and 3, 2
and 5, 2 and 6, 3 and 1, 3 and 2, 3 and 3, 3 and 5, 3 and a1 ¼ 0:10; a2 ¼ 0:31; a3 ¼ 0:07; a4 ¼ 0:31; a5 ¼ 0:21
6, 5 and 1, 5 and 2, 5 and 3, 5 and 5, 5 and 6, 6 and 1, 6 and and l ¼ 0:27
2, 6 and 3, 6 and 5, and 6 and 6, i.e., with five samples an
Therefore the estimated value at Q will be:
additional 5
5 intersample distances are computed.
Intersample variances eliminate the shadowing effect pre- a1
2:4 þ a2
2:76 þ a3
4:84 þ a4
6:38 þ a5
7:04
sent in sample distribution within the zone of influence.
The equations are formulated in matrix form to compute ¼ 0:1
2:4 þ 0:31
2:76 þ 0:07
4:84 þ
a semivariogram for expressing the relation between dis- 0:31
6:38 þ 0:21
7:04
tances and variances:
2 3 2 3 2 3 ¼ 4:82 % Zn
V11 V12 V13 V15 V16 a1 V1 Q
6V 7 6 7 6V Q7 The estimated value of 4.82% Zn has an associated
6 21 V22 V23 V25 V26 7 6 a2 7 6 2 7
6 7 6 7 6 7 variance. The process of computation provides extension
6 V31 V32 V33 V35 V36 7 6 a3 7 6V Q7
6 7
6 7 ¼ 6 3 7 variance of 2.2 from Q to 1 (blockepoint covariance),
6V 7 6 7 6V Q7
6 51 V52 V53 V55 V56 7 6 a4 7 6 5 7 while we utilized only 0.10 (a1) of this variance to get an
6 7 6 7 6 7
4 V61 V62 V63 V65 V66 5 4 a5 5 4 V6 Q 5 estimate of Q. Similarly, 0.31 has been used for the second
a1 a2 a3 a4 a5 l 1 sample out of 1.9 and so on. Summation of the variances
utilized for estimation at Q will be:
where a1 þ a2 þ a3 þ a4 þ a5 ¼ 1 (weights for coeffi-
cients) and V11 . Vnn represent the variances instead of a1
2:2 þ a2
1:9 þ a3
2:3 þ a4
1:8 þ a5
1:8
the distances, which were measured earlier. In the present ¼ 1:94
data set the following equation is obtained:
The total covariance utilized for the estimation, i.e.,
a1 0:0 þ a2 1:8 þ a3 2:5 þ a4 2:7 þ a5 1:9 ¼ 2:2
blockepoint covariance (1.94) and pointepoint kriging
a1 1:8 þ a2 0:0 þ a3 2:8 þ a4 2:5 þ a5 2:1 ¼ 1:9 covariance l (0.27), is 2.21. The semivariogram of the
orebody can explain 1.72 in terms of variance for a
a1 2:5 þ a2 2:8 þ a3 0:0 þ a4 1:9 þ a5 1:6 ¼ 2:3 block size of 25 m
25 m (an average distance of 12.5 m
from any point to another point within the block). The
a1 2:7 þ a2 2:5 þ a3 1:9 þ a4 0:0 þ a5 1:8 ¼ 1:8 residual variance of the estimate is 2.21  1.72 ¼ 0.49,
which is not explained by the estimate at Q and will be
a1 1:9 þ a2 2:1 þ a3 1:6 þ a4 1:8 þ a5 0:0 ¼ 1:8
equivalent to 0.7 in terms of standard deviation. The mean
a1 þ a2 þ a3 þ a4 þ a5 ¼ 1:0 grade (4.82% Zn) has a possible variability of 0.7 over the
194 Mineral Exploration

mean.pffiffiThe
ffi confidence interval at 95% probability will be directions have to be blended with the mathematical for-
(0.7/ 5)
1.96 or  0.61. The error of the mean is 4.82% mula supported by statistical and geostatistical tests and
Zn 0.61. Keeping other conditions constant, if we change case studies (Haldar et al., 1990).
the block size to 15 m
15 m the explained variance
equivalent to 7.5 m is 1.5 and the error of the mean by
computation will be 0.74. It can be concluded that as the
REFERENCES
block size decreases, the confidence limit value (range) will Armstrong, M., 1998. Basic Linear Geostatistics. Springer, p. 155.
increase. Clark, I., 1982. Practical Geostatistics. Applied Science Publishers, Lon-
Kriging methodology relies largely on the precision of don, p. 129.
the variogram, which in reality is not smooth and indicative David, M., 1977. Geostatistical ore reserve estimation. Developments in
of disturbances inherent within the samples. Since the Geo-mathematics 2. Elsevier Scientific Publishing Company,
Amsterdam, p. 364.
semivariogram used is a generalized average, the estimate
David, M., 1988. Handbook of applied advanced geostatistical ore reserve
may differ from the true value. Local trends existing in the
estimation. In: Developments in Geo-mathematics, vol. 6. Elsevier
data set should be removed before kriging. Scientific Publishing Company, New York, p. 232.
Davis, J.C., 1971. Statistics and Data Analysis in Geology, third ed. John
9.4.6 Benefits of Statistics and Wiley & Sons, New York, p. 550.
Geostatistics Davis, J.C., 2002. Statistics and Data Analysis in Geology, third ed. John
Wiley & Sons, New York, p. 656.
1. Statistical techniques (mean, variance, standard devia- Haldar, S.K., 1982. A study of the Balaria base metal deposit, Zawar
tion, confidence limit, correlation coefficient, and Group of Mines, Rajasthan, using numerical procedures with special
various tests) will provide sequential analysis of an reference to derivation of parameters for mine planning unpublished
exploration sample database with broad parameters at Ph. D thesis. Indian Institute of Technology, Kharagpur, India, p. 641.
any stage of exploration. It will focus on the adequacy Haldar, S.K., Bhatnagar, S.N., Paliwal, H.V., 1990. Application of geo-
statistics to exploration data for sample optimization, mine sub-block
of the exploration input and the power of decision-
estimation, and grade forecast system at Rampura-Agucha lead-zinc
making toward investing in mining, continuing explora-
mine, India. In: XXII APCOM-90, Berlin, West Germany, p. 11.
tion, or keeping the project in abeyance and waiting for Hans, W., 2013. Multivariate Geostatistics: An Introduction with Appli-
the next opportune time. cations. Springer Science & Business Media, p. 257.
2. The semivariogram parameters (g(h)) in geostatistics Henley, S., 1981. Nonparametric Geostatistics. Applied Science Pub-
quantify the range of influence of the samples in all di- lishers, London, p. 145.
rections, indicating the isotropic/anisotropic nature of Isaak, H.E., Srivastava, R.M., 1989. An Introduction to Applied Geo-
the deposit and selection of sample for estimation. statistics. Oxford University Press, Inc., p. 559
3. The semivariogram also extracts the explained (C1) and Journel, A.G., Huijbregts, C.J., 1978. Mining Geostatistics. Academic
unexplained nugget (C0) variances of the orebody that Press, London, p. 600.
control the estimation of block grades. Kellaway, F.W., 1968. Penguin-Honeywell Book of Tables. Penguin
Books Limited, p. 76.
4. Point, area, and block estimation by kriging techniques
Koch, G.S., Link, R.F., 1986. Statistical Analysis of Geological Data,
represent BLUE, which minimizes error variance.
second ed. Wiley, New York, p. 375.
5. Statistical and geostatistical applications compute asso- Krige, D.G., 1951. A statistical approach to some basic mine valuation
ciated estimation errors that are risk factors for invest- problems on the Witwatersrand. J. Chem. Metall. Min. Soc. S. Afr. 52,
ment decision. 119e139.
6. The advance estimation methods estimate reserves and Krige, D.G., 1962. Statistical application in mine valuation. J. Inst. Mine
resources (tonnage), and average grade of individual Survey. S. Afr. 12 (2 and 3), 82.
blocks for planning of selective mining and grade Krige, D.G., 1978. Lognormal-de Wijssian Geostatistics for Ore Evalua-
control. tion. South African Institute of Mining and Metallurgy, Johannesburg,
7. The techniques can optimize sampling design and den- p. 91.
sity (refer to Chapter 15, Section 15.10). Matheron, G., 1971. The Theory of Regionalized Variables and its Ap-
plications. Les Cahiers du Centre de Morphologies Mathematique de
8. The techniques are capable of analyzing the estimated
Fontainebleau, Ecole Nationale Superieure des Mines de Paris, p. 211.
and actual performances.
Rendu, J.-M., 1981. An introduction to geostatistical methods in mineral
9. Finally, a good knowledge of statistics, geostatistics, and evaluation, South African institute of mining and metallurgy mono-
the application of geology and mining software will not graph series. Geostatistics 2, 84.
be enough to estimate the reserves and resources of a de- Sahu, B.K., 2005. Statistical Models in Earth Sciences. B, S. Publication,
posit with full confidence. An inherent and in-depth un- Hyderabad, India, p. 211.
derstanding of mineralization by extensive fieldwork, an Sinclair, A.J., Blackwell, G.H., 2002. Applied Mineral Inventory Esti-
unbiased and unique sample database, and visualization mation. Cambridge University Press, UK, p. 381.
of physical and chemical variations of grade in various