Objectives:

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

 Differentiate between experimental variogram and variogram model.

 Describe the different types of variogram models.
 Fit appropriate models to experimental variograms.
 Explain ways of testing how well a model fits data.
 Describe the relationship between semi-variogram and covariance functions.

8.1. Variogram Modelling

The experimental semi-variogram is only known at the points calculated for specific values
of h (discrete). In the kriging estimation process, we need γ(h) for any distance, whether a
multiple of h or not. This necessitates a continuous model that we can evaluate at any
required distance. This model will be a representation of the estimated variability or
continuity of the regionalised variable.

When we looked at classical statistics, we drew a histogram or probability plot of our data and
we then proposed some theoretical function for the distribution of values within the whole
population. We now have the equivalent of a spatial histogram for the sample data and need
to theorise about what this graph would look like if we had the whole population of all
possible pairs of values over the whole study area.

There are many possible models for the idealised `population' semi-variogram, γ(h). As with
probability distributions, there are mathematical restrictions on the models which can be
applied, mostly designed to ensure that we do not end up with answers involving, say,
negative variances. You can invent your own semi-variogram models if you wish, but
remember the restrictions.

66
The detail in regard to the shape and formulae for the different models are included
here. Generally semi-variogram models are split into two categories namely those without a
sill (unbounded) and those with sills (bounded).

8.2. Models without a sill

These include the Linear, the Generalised Linear (Power) and the De Wijsian models.

8.2.1. Linear Models

Linear models are basically a straight line. All you have to know is the slope of the line to
describe it Linear mode γ(h) = Co + ρ*h, where ρ = slope, Co = nugget effect and h = lag
distance. If points are rising, then simply fit a simple straight line using the eyeball technique
(line of best fit). The units of p (slope) are units2/m, e.g. (g/t)2 /m or MJ2/m, etc. The linear
model is not used much in the mining situations. It is used more often in geophysics where
you have closely packed data. Linear models describe perfectly occurring natural systems.

67
8.2.2. Generalised Linear Models (Power Model)

This model has the form γ(h) = C0 + ρ*hα, where α can be 2>α>0 but never equals to or greater
than 2; ρ = slope of the line. Generalised linear models are rarely used as actual models. They
are popular with theoretical geostatisticians because they are easy to use without being trivial.
They are useful teaching models.

If α = 2, it becomes a quadratic function. Usually an indicator of trend in data and which

is problematic in variography.

8.2.3. De Wijsian model

The De Wijsian model has the form γ (h) = Co + 3α*loge(h). Where α is the
dispersion coefficient (variance divided by mean) and 3α is essentially the slope of the line.
Dispersion measures how variance increases as distance increases. This was developed as an
idealised model for lognormal data. If you assume log-normality, then theoretically the
variance gets bigger as the area gets bigger. The reverse is not true though. To use the de
Wijsian model you must have lognormal data.

68
8.3. Models with a sill

8.3.1. Spherical Model

This is a model first proposed by Matheron. The formula is a cubic one since it represents
volumes, and relies on two parameters: the range of influence (radius of the sphere) and the
sill (plateau) which the graph reaches at the range. In addition to these, there may be a
positive intercept on the γ axis --- the "nugget effect" described above. The formula and shape
for this model are shown on the right hand side. The mathematical description of the model
is summarised below:

where γ is the semi-variogram and h the distance between the two points of interest. The
parameter a represents the range of influence of the semi-variogram. We generally interpret
the range of influence as that distance beyond which pairs of sample values are unrelated. C is
the sill of the Spherical component and C₀ the nugget effect on the γ axis. The final height of
the semi-variogram model is C₀+C.

69
Unlike the previous two models, there are modifications which you can make to the standard
Spherical model. There are often cases where the semi-variogram graph reaches a definite
"sill' but does not quite match the shape of a single Spherical model. In this case you may mix
Spherical components (nested structures) with different ranges of influence and/or sill values
in order to achieve the correct shape. Remember that, if you have (say) three component
Sphericals (structures), the final height of the graph will be C₀+C₁+C₂+C₃. The formula for such
a model is simply the combination of ordinary Spherical models, remembering to stop each
one as it reaches its range of influence.

70
71
8.3.2. Exponential Model

This is a model developed to represent the notion of exponential decay of "influence"

between two samples. It relies on two major parameters: the range of influence (a scaling
parameter) and the sill (plateau) which the graph tends towards at large distances. There is
also a possible "nugget effect". The formula and shape for the Exponential model are shown
below:

0 𝑖𝑓 ℎ = 0
𝛾 (ℎ) = { −ℎ
𝐶0 + 𝐶1 [1 − exp⁡( ) ] 𝑖𝑓 0 < ℎ ≤ 𝑎1
𝑎1

C = C1 + Co

C1

Co

Where γ is the semi-variogram and h the distance between the two points of interest. The
parameter a represents the so-called "range of influence" of the semi-variogram, C the sill of
the Exponential component and C₀ the nugget effect on the γ axis. You will note that the
asymptotic height of the semi-variogram model is C₀ + C1. You may also note that, although
parameter a is referred to as the range of influence, it is not possible to interpret is in the
same way as for the Spherical model. This distance a is not the distance at which samples
become "independent" of one another. In fact, the exponential model reaches about two-
thirds of its height at a distance a, and must go to four or five times this distance to come close
to its asymptotic "sill". You can also expand the formula above to indicate a nested exponential
model as previously shown for the spherical model.

72
8.3.3. Gaussian Model

This is a model developed to represent the notion of exponential decay of "influence"

between two samples. It relies on two major parameters: the range of influence (a scaling
parameter) and the sill (plateau) which the graph tends towards at large distances. There is
also a possible "nugget effect". The formula and shape for the Exponential model are shown
below:
0 𝑖𝑓 ℎ = 0
2
𝛾 (ℎ) = { −ℎ
𝐶0 + 𝐶1 [1 − exp⁡( 2 ) ] 𝑖𝑓 0 < ℎ ≤ 𝑎1
𝑎1

The formula for this curve is similar to that for a cumulative Normal distribution --- hence the
name Gaussian --- it does not imply that the sample values must be Normal.

Where γ is the semi-variogram and h the distance between the two points of interest. The
parameter a represents the so-called "range of influence" of the semi-variogram, C the sill of
the Gaussian component and C₀ the nugget effect on the γ axis. You will note that the
asymptotic height of the semi-variogram model is C₀ + C1. You may also note that, although
parameter a is referred to as the range of influence, it is not possible to interpret is in the
same way as for the Spherical model. This distance a is not the distance at which samples
become "independent" of one another. In fact, the Gaussian model reaches about two-thirds
of its height at a distance √a, and must go to four or five times this distance to come close to
its asymptotic "sill".

73
8.3.4. Hole Effect Model

This is a model developed to represent a cyclic or periodic relationship between two samples.
It relies on two major parameters: the cycle distance (a full cycle of the periodicity) and the
sill (plateau) which the graph tends to oscillate around. There is also a possible "nugget effect".
In many cases, a fourth parameter is added to these three, known as a "damping" or decay
parameter. Without this parameter the cyclic effect would continue on to infinity. In practical
circumstances, the relationship usually tends to trail off and this may be reflected by including
the damping parameter. The formula for the Hole Effect model with damping is:

Where the formula for this model would be:

Where γ is the semi-variogram and h the distance between the two points of interest. The
parameter ξ is the cycle interval (distance), ζ represents the so-called "decay" or damping

74
parameter of the semi-variogram, C the sill of the Hole Effect component and C₀ the nugget
effect on the γ axis. You will note that the asymptotic height of the semi-variogram model
is C₀ + C. The damping on the model is inverse to the magnitude of the parameter --- the bigger
the ζ the less the damping effect. It is also scaled by the distance h, and so will be relative to
the scale at which we calculate the graph.

Cautionary note: the hole effect model is not mathematically stable and can lead to some
weird results if used without care.

8.3.5. Paddington Mix model

This model is included mostly as an illustration of how you can reflect the geology or structure
of the measurements by combining components of various different shapes. In this case we
have a fairly continuous phenomenon which has a weaker cyclic component present. The
model was first used (by us) in an Australian gold deposit which was "shear enhanced". That
is, the gold was present throughout the mining area but values were higher close to a quartz
shear. Since shears in rock tend to occur with great regularity, the obvious Spherical structure
for gold grade was modified by a `ripple' effect due to the presence of shears. The shape of
the graph changed according to whether the direction was parallel to or across the direction
of the quartz shears.

Other applications for this type of model include:

 potholes in platinum reefs

 diamonds occurring on the sea-bed (genuine ripples)

 plant or tree yields where fields have been trenched

 occurrences of species which show `competition' effects

75
where the full formula would be:

Where γ is the semi-variogram and h the distance between the two points of interest. The
parameter a represents the range of influence of the semi-variogram, C the sill of the Spherical
component. The parameter ξ is the cycle interval (distance), ζ represents the so-called "decay"
or damping parameter of the semi-variogram, Chef the sill of the Hole Effect component
and C₀ the nugget effect on the γ axis. You will note that the asymptotic height of the semi-
variogram model is C₀ + C + Chef.

8.4. How well does the model fit the data?

Noel Cressie came up with a very nice "goodness of fit" statistic in the late 1980s which goes
a long way to measuring how well the model fits the data. However, it preferable to use a

76
combination of statistic and visual assessment in assessing how well the model fits the
experimental data. The Cressie goodness of fit statistic, is calculated as follows. For each point
on the semi-variogram graph, calculate the Cressie statistics and sum the terms.

This is analogous to but not the same as the χ² goodness of fit test, even allowing for the
weighting by the number of pairs. This statistic allows bigger deviations between the
experimental and model semi-variogram as the model becomes higher. It demands closer
fitting at the lower levels -- usually the lower distances -- than at higher ones. It also demands
better fitting where you have more pairs of samples in a point. The fit is directly weighted by
the number of pairs of samples. The actual magnitude of the statistic depends on the total
number of pairs of samples found during the calculation. Now, not all samples are paired the
same number of times in all different directions. This means that a Cressie statistic of a certain
size in one direction is not necessarily equivalent to the same value in another direction. To
adjust for this, Clark suggested a modification to remove the scaling by total number of pairs.
Simply dividing through by the total number of pairs of samples in that particular semi-
variogram fit.

Another useful statistic for testing how well the model fits the data is using Normalised Means
Square Error as shown below:

77
There are many sentiments amongst geostatisticians that cross validation is the better option
for evaluating the performance of a variogram model in estimating a regionalised
variable. Cross validation compares alternative variogram models based on the results of the
final objective, which most often is some kind of estimation. The exercise consists of
estimating locations with known sample values, and comparing the estimated and the
sampled value; the alternative variogram models would result in different sets of estimates;
the better variogram model would be the one that yields a lower average error. The are two
ways that this exercise can be performed: (1) a spatial, but otherwise classical, leave-one out
re-estimation, whereby one sample at a time is removed from the data set and re-estimated
from the remaining data; to avoid undue influence of the nearest datum in a drill hole, one or
more of the closest samples are typically not used in there-estimation; and (2) a subset of the
data (say, 40 or 50% of the total) is removed completely from the data set, and re estimated
using the remaining data.

8.5. Covariance Function

Matheron showed that, if the semi-variogram model has a sill or final asymptote, then the
final height of the semi-variogram is theoretically equal to the population variance of
measured values. That is, as h→∞, then γ→σ². If the semi-variogram has a sill, this tends to
support the assumption that the samples come from a `stationary' distribution with a fixed
mean and standard deviation. In this case, some workers prefer to use the covariance function
rather than the semi-variogram function. The relationship between the two is:

γ(h) = σ² - covh

Where covh is used here as shorthand for the covariance between sample values at distance
(and direction) h, and is estimated by:

78
There does not seem to be any indication in the geostatistical literature that Nh -1 might be
more appropriate here, since we have to use g bar or some other estimator to estimate the
population average, μ. This shortcoming in the literature has been pointed out rather
forcefully by Dr Jan Merks in some moderately intemperate articles.

8.6. Example: Variogram Models

An omni-directional experimental variogram for %Fe spatial distribution calculated Unit 6 is

shown below. Answer the questions that follow:

(a) Fit a model that best fits the data. What model will best fit the data?

(b) Substitute the values of C0, C1 and a in the appropriate model algorithm.

(c) Determine the γ (h) at the following values of h: 1 m, 2 m, 3 m, 4 m, 5 m, 6 m.

79
Example: Solution

80
Check Your Understanding: Question

The experimental semi-variogram {γ* (h)} constructed in Unit 6 is shown below. C, C1 , Co and a have been interpreted by the resource
geologist for two possible variogram models. But there is a disagreement in the office on whether the line of best fit shows an exponential
model {γe (h)} or a spherical model {γs (h)}. Calculate the variogram model values for both a spherical {γs (h)}) and exponential model {γe (h)}.
Compare the spherical and exponential models' values to the original semi-variogram values. Use lags starting at h = 100 m until h = 1900
m increasing at a lag separation of 100m. Which model best fits the data?

81
Check Your Understanding: Solution

82