Geostatistics - Geostatistical

simulation: examples
Julián M. Ortiz - https://julianmortiz.com/

October 5, 2024

Summary
Sequential Gaussian and sequential indicator simulation are
commonly used to build models of continuous and categor-
ical variables. In this chapter, we present examples of im-
plementation of both methods and show how these are inte-
grated to account for joint geological and grade uncertainty,
in the context of a resource modeling example in mining.

1 Example

1.1 Categorical simulation

In this example, we start by considering the rock types as a
categorical variable. For simplicity, we model the presence
and absence of unit 20. In order to prepare the data, we

1
code the rock type to discriminate between samples with
rock type 20 vs all the rest. The new codes are:

ˆ Rock type 20: 200

ˆ All other rock types: 100

Sequential indicator simulation (for categorical variables)

requires the definition of the number of categories (2 in this
case), the categories codes (200 and 100 in this case), the
global proportions, the indicator variograms and a set of
parameters for the search.

Global proportions

Categorical variables, as continuous variables, may be sam-

pled in a non-representative way. The high grade domain
may be preferentially sampled, thus, when computing the
global proportions, we need to correct for this spatial bias
via declustering.
The same parameters used for declustering the grades
are used to decluster the proportions. A cell of size 35 by
35 by 12 m3 is used to determine the weights assigned to
each sample. Statistics before and after declustering are
presented in Figure 1. It can be seen that the declustered
proportions are as follows:

ˆ Code 100 (all other rock types): 65.108 %

ˆ Code 200 (rock type 20): 34.892%

2
Figure 1: Raw and declustered statistics of the proportions of the coded rock types.

Indicator variograms

Since the sample locations are the same as for the grades,
the search parameters for inferring indicator variograms are
kept as before. They are summarized in Table 1.
Notice that data are coded as indicators for the two cases
considered: code 100 and code 200. In each case, the codes
are transformed to a binary indicator. It is obvious from look-
ing at the formula of the experimental indicator variogram
that in the binary case, the output is the same if the indica-
tors are interchanged, as only transitions from one category

3
 Lag Azimuth Dip
Direction Lag Tolerance Azimuth Tolerance Dip Tolerance Bandwidth
[m] [m] [°] [°] [°] [°] [m]
N30°E 35.0 17.5 30 22.5 0 15.0 30.0
N120°E 35.0 17.5 120 22.5 0 15.0 30.0
Vertical 12.0 6.0 0 180.0 -90 15.0 20.0

Table 1: Parameters for indicator variogram calculation

to any other matters for the computation of the indicator

variograms.
The resulting experimental indicator variograms are then
modeled with a nugget effect and two nested structures.
Notice the non-stationarity evident in the experimental re-
sults (Figure 2). Despite this, we model the indicator vari-
ograms as stationary random functions (Table 2), with a sill
(that depends on the proportions). The plots are standard-
ized to a unit sill to ease the modeling.

Notice that in most cases, the categories are not prop-

erly modelled with a stationary random function, since
they tend to present very clear structure (trends) in
space, rather than being spread all over the domain
in a regular fashion. Therefore, categorical simulation
methods always struggle to represent properly the spa-
tial distribution of categories.

Type Sill Angle 1 Angle 2 Angle 3 Range Y” Range X” Range Z”

Nugget 0.01
Spherical 0.29 30 0 0 170 140 85
Spherical 0.70 30 0 0 170 140 ∞

Table 2: Parameters of the indicator variogram model

4
Figure 2: Experimental indicator variogram and fitted model in the three principal
directions.

Search plan

A point support grid (consistent with the point support that

is later used for simulating the grades) is considered (Table
3). The search plan consists of 24 samples and previously
simulated nodes within a 70m search radius. Since the grid
is quite dense, each sample is assigned to the closest node,
requiring a single search for conditioning information and
making the simulation faster. Multiple grids are used with
three levels.

1.2 Results of categorical simulation

The results of sequential indicator simulation are 20 dense
realizations conditioned by the sample data. The E-type es-

5
 Coordinate Nodes Min Coordinate (m) Spacing (m)
East (X) 400 24450.5 1.0
North (Y) 600 25050.5 1.0
Elevation (Z) 12 3826.0 12.0
Total 2880000

Table 3: Definition of the point grid for simulation.

timate provides a map of the likelihood of finding each rock

type. The map of conditional variances, reflects the areas
where the category is uncertain. Four realizations and the
e-type and conditional variance maps are shown in Figure
3.
From the maps of the realizations it is clear the non-
stationary nature of the variable. The Tourmaline breccia
unit (rock type 20, coded as 200) is mainly concentrated at
the center of the domain. Some occurrences appear in the
simulations at the edges of the domain, but these happen
because we use simple indicator kriging, so where condi-
tioning is weak, simulated categories are drawn from the
global (prior) proportions.
The e-type map clearly shows where each category should
be found. All the green areas are mixed cases, where some
simulations output a value of 100 and others a value of 200.
These are the areas close to the unknown boundaries be-
tween the main unit and the others. This is again seen in
the conditional variance map, where a clear countour of the
boundary is evident in red. Further to this, areas with large
uncertainty are seen where conditioning is scarce.

6
Figure 3: Four realizations and the e-type and conditional variance maps for a repre-
sentative bench.

7
1.3 Continuous simulation
The next step is to simulate the spatial distribution of grades,
but this has to be done for each domain separately. Since
we have two categories, grades are simulated independently
in each domain (we will assume hard boundaries between
the units).
The procedure is:

ˆ Decluster the data: this can be done jointly or by cate-

gory
ˆ Transform to normal scores: the grades within each
category are transformed independently to a standard
Gaussian distribution.
ˆ Calculate and model variogram of normal scores of grades:
this is done independently for each category.
ˆ Perform simulation: grades are simulated with sequen-
tial Gaussian simulation independently for each cate-
gory. Notice that the simulation is not yet constrained
to the extent of the domain (which is determined in
a stochastic manner with sequential indicator simula-
tion). Therefore, the entire domain is simulated.

Once simulations are completed, the models can be merged

with the indicator simulations such that:

ˆ For realization 1 to L:

– Read the value of each node in the indicator simu-

lation realization.
– Assign the simulated grade for the corresponding
category to that location.

8
Declustering and transformation

The declustered distribution and corresponding normal score

transform is shown as an example, for unit 100 in Figure 4.

Figure 4: Declustered and transformed histogram for unit 100.

Variograms of normal scores

The variograms for unit 200 were shown when presenting

multiGaussian kriging. Here, we show the resulting vari-
ograms of normal scores for unit 100 (Figure 5).

Search plan

The parameters for Gaussian simulation are the same as

for indicator simulation in terms of the grid definition and
search plan.

9
 Figure 5: Variogram of normal scores for unit 100.

1.4 Results of continuous simulation

The results of the simulation of grades are presented in Fig-
ures 6 and 7.
The realizations capture the spatial correlation and aniso-
tropy and show the true variability between points. The e-
type follows closely what was seen before with kriging. The
conditional variance highlights the proportional effect and
the uncertainty in areas of scarce sampling.
Notice for unit 100 the smooth behavior at the center of
the deposit, where scarce conditioning exists (since this is
where unit 200 is). The estimate tends to the global mean
and the variance is high. Also, it can be seen that for the
simulation of grades in unit 200, in the borders, conditioning
is weak generating a high variance.

10
Figure 6: Four realizations, the e-type estimate and the conditional variance for grades
in unit 100.

11
Figure 7: Four realizations, the e-type estimate and the conditional variance for grades
in unit 200.

12
1.5 Merging the models
The final step is to merge the categorical simulations rep-
resenting the extent of the domains, and the continuous
simulations representing the uncertainty in the grade distri-
bution.
This is done by taking a very simple approach: the simu-
lated grade for realization  at location  is assigned based
on the simulated category ssm () at that location in real-
ization  of the set of categorical simulations. This means if
category ssm () = sk then, the grade assigned at location 
is that of the simulations for category sk at that location, for
that same realization.
The results of the assembled model are presented in Fig-
ure 8.
This set of realizations captures the joint uncertainty in
the extent of the geological units and in the grades. It can
be processed to represent the expected uncertainty of the
true deposit, considering the samples available.

1.6 Final comments

In this example, we showed how to combine the categorical
and continuous simulation models to account for the uncer-
tainty in both the distribution of geological domains and the
grade distribution within domains.
The resulting set of realizations can be used to determine
the response of the deposit to any transfer function, by pro-
cessing each realization through the transfer function and
obtaining the statistical distribution of the response.
It should be noticed that by integrating the categorical

13
and continuous models, the hard boundaries between cate-
gories, end up “diluted” in the set of models. This reflects
the fact that we are uncertain about the location of the
boundary between categories (although we may be sure the
the boundary is hard).

14
Figure 8: Four realizations, the e-type estimate and the conditional variance for the
grades in the final assembled model.

15
Index
boundaries, 6
conditional variances, 6
declustered proportions, 2
declustering, 2
E-type estimate, 6
global proportions, 2, 6
hard boundaries, 8
indicator variogram, 3
joint uncertainty, 1, 13
multiGaussian kriging, 9
non-stationary, 6
proportional effect, 10
sequential Gaussian simula-
tion, 1, 8
sequential indicator simula-
tion, 1, 8
trends, 4

16