Objectives:

At the end of this unit, you must be able to:

• Explain how sample values correlate and vary spatially.
• Explain how the differences between sample values at specific lag distance apart and
orientations assist in overcoming the limitation of quantifying the estimation error
using the conventional method.
• Calculate the semi-variogram values at different lag separation distances.
• Explain the purpose of a semi-variogram.

4. Experimental Variogram - Introduction

In the previous inverse distance (IDW) study unit, we explored how we can integrate distances
(from the point of interest to the samples) in the estimation of the value at the unknown point.
We did this by calculating distance-related weights such that the further the sample is from
the unknown value, the lower its weighing will become when estimating the unknown point.
This is a very important step because it is in line with geological processes, i.e. the geological
properties at two points which are closer are usually similar (low variability) than properties
of points which are further apart (high variability). Hence, we want values that are closer to
the point of estimation to have more weighting in the estimation than those that are further
away. Another way of thinking about this is by looking at vegetation. At shorter distances,
vegetation in nature will tend to be almost similar (low variation). However, the further apart
two areas are, you would most likely find a large variation in the type of vegetation. In other
words, the closer the samples, the better the correlation between them.

4.1. Estimation Error

However, although we could integrate impact distance in IDW, there is no way of quantifying

the error ( . In other words, how reliable is our estimate? Alternatively, what confidence can
we attach to the estimate? This is an important question which we will attempt to answer for
the most part of the following chapters/units.

16
The error is the difference between the true value (Z) and the estimated value (Z*). That is:

(1)

(2)

Substitute equation (2) into (1):

(3)

(4)

Because

(5)

Meaning that for three samples:

Total Error = .

are error contributions towards the total error .

e previous equation for estimating error poses a problem because we do not know the true
value (Z); we can only know the true value after mining. So what do we do in order to estimate
the error? We need look at what we know, i.e.: we can investigate the differences between
sample values at two different locations for a particular direction and specific distance apart.
For example, we can investigate the difference of two sample values at 50m (East-West),
100m (South-North), 45 degrees, omni-directional (all directions), etc.

17
Given 25 sample points with measured calorific values (CV) on a regular grid, we can
investigate the differences between two samples values at the following directions and
distances as follows:

What can we learn from investigating the differences between two sample points (pairs) at
each direction and specific distance apart?

18
 • If we drew a histogram of the differences in sample values, we can investigate the
distribution from which they are drawn.

• We can therefore make an intrinsic assumption that, the difference in value between
two samples at different positions across a deposit depends only on distance between
them and their relative orientation.

• If the data is drawn from a normal distribution, it is likely that the differences will also
be normally distributed and are therefore defined by a mean and standard deviation.

• If all of the orebody was drilled, some of the differences (zi – zj) will be negative and
others will be positive. The sum of the negative and positive pair difference would be

zero, ( . Therefore, average of the differences ( ) over the whole

deposit should be zero ( if the is no underlying trend in the deposit (i.e.

geologically homogeneous zone).

• If the distance between samples and the relative orientation is denoted by h, the
difference in grade between two samples depends on h, and the distribution of the
differences depend only on h. Hence, the mean and the variance of the differences
also only depends on h.

4.2. Calculation of Semi-variogram Values

We now have an understanding of the differences between pairs and their relationship

to h and orientation. This is still not as useful because we expect . However,

by square the differences we can ensure that we do not get zero when we sum the squares

of differences. Furthermore, can be used to calculate the variance of the

differences between the samples at each. Now we are getting somewhere, from trying to

estimate the error as and subsequently

failing to Z being unknown. We are at-least now at a place where we can calculate the
variance of differences between samples.
Therefore, we can now estimate the variance of the differences between sample values
across the entire mineral deposit, given by:

19
 Variance of the differences =

Where, are sample values at two locations; is the number of pairs for

distance h and direction ϴ. If no trend in data .

Therefore, the variance of the differences can be re-written as:

The variance of the difference gives us the estimation variance ( ), which is the

variogram . That is:

The above variogram formula gives us the spatial variability at a particular distance h and

orientation ϴ. It is convention to use the semi-variogram instead of the

variogram (please note that in literature the two terms are used interchangeably
referring to the semi-variogram. As such, the two terms will also be used interchangeably
from here forth in these notes). The semi variogram can therefore be represented as
follows:

Sometimes, instead of using notations for the sample values, z(x) and z(x +
h) are used respectively. The subsequent equation will then look like this:

20
The semi-variogram provides a descriptive statistic (the variance of the differences) of a
small number of samples in the form of a distance function.

4.3. Example: Calculating the standard experimental semi-variogram

Consider the 25 samples discussed in the sections above. Calorific values of each sample is
given. Calculate the experimental semi-variogram value at 50m West-East (W-E). Also,
calculate the average of the differences and explain what the value describes about the data.
Additionally, assuming that the samples come from a normal distribution, estimate the
population variance. Compare the estimated population variance and the semi-variogram
value calculated.

Example: Solution

21
 = 0 MJ/Kg, meaning there is no trend in data.

< , which makes sense since as spatially variability increases, it will eventually reach the
population variance (To expand on this latter).

22
Check Your Understanding: Question

Given 25 samples with associated calorific values measured. Calculate the experimental semi-
variogram value at h =100m (S-N). Additionally, calculate the mean of the differences and
estimate population variance. Comment of the final answers. Assume normally distributed
data.

A single point on a standard variogram is calculated as follows:

• Gather all pairs of sample points, in all directions, with the same separation distance.

• Calculate the difference between each pair of points.

• Square the difference (this will remove any negative values).

• Add all the squared differences.

• Divide by twice the number of pairs of data used.

23
Check Your Understanding: Solution

Although there is trend in the data, our assumption of no trend will still apply (at-least for
the techniques we are using in this course). The reason being that we will only model the
part which has no trend in our variogram. However, there are other techniques wherein we
can incorporate trend such "kriging with a trend" method, etc.

24