See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/262463211

Transferring sampling errors into geostatistical modelling

Article in Journal- South African Institute of Mining and Metallurgy · December 2011

CITATIONS READS

2 100

3 authors:

Miguel A. Cuba Oy Leuangthong

MineSight ULC 37 PUBLICATIONS 326 CITATIONS
4 PUBLICATIONS 8 CITATIONS
SEE PROFILE
SEE PROFILE

Julián M. Ortiz
Queen's University
119 PUBLICATIONS 688 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Geologic Modelling View project

Mine Planning View project

All content following this page was uploaded by Julián M. Ortiz on 17 July 2014.

The user has requested enhancement of the downloaded file.

 Transferring sampling errors into
geostatistical modelling
by M. Cuba*, O. Leuangthong†, and J. Ortiz‡

Introduction
In mineral resources/reserves modelling, the
main and sometimes only source of
Synopsis
information is the exploratory drilling data-set.
Geostatistical modelling aims at providing unbiased estimates of the During different stages of the project
grades of elements of economic interest in mining operations, and
evaluation, various drilling campaigns are
assessing the associated uncertainty in these resources and
carried out in different periods of time and/or
reserves. Conventional practice consists of using the data as error-
free values and performing the typical steps of data analysis –
with different goals. As such, the exploratory
domaining, semivariogram analysis, and estimation/simulation. data-set is continually updated by each
However, in many mature deposits, information comes from campaign. The lifetime of an exploration
different drilling campaigns that were sometimes completed decades project can be several decades, and in this
ago, when little or no quality assurance and quality control context different technologies may be used
(QA/QC) procedures were available. Although this legacy data may during these drilling campaigns. Legacy data
have significant sampling errors, it provides valuable information from campaigns drilled at early stages of the
and should be combined with more recent data that has been subject project may not have been subjected to any
to strict QA/QC procedures. quality assurance and quality control (QA/QC)
In this paper we show that ignoring the errors associated with
procedures1. Even recent campaigns drilled
sample data considerably underestimates the uncertainty (and
will have different precision due to the drilling
consequently the economic risk) associated with a mining project.
We also provide a methodology to combine data with different technique used, for example, diamond or
sampling errors, thus preserving the relevant global and local reverse circulation drilling, or due to different
statistics. The method consists of constructing consistent simulated QA/QC standards.
sets of values at the sample locations, in order to reproduce the The effect of poor-quality data at different
error of each drilling campaign and the spatial correlation of the stages of a project has been widely discussed
grades. It is based on a Gibbs sampler, where at every sample in the literature2–4. The quantification of
location, the actual sample value (with error) is removed and a sampling error during sample collection and
conditional distribution is calculated from simulated values at preparation for chemical analysis is also well
nearby sample locations. A value is drawn from that distribution
documented; most operations perform routine
and kept only if it satisfies some statistical requirements—specif-
checks of the quality of their sampling
ically, the global relative error and local means and variances must
be reproduced. All sample locations are visited and simulated
procedures5–8.
sample values are generated iteratively, until the required statistics One outstanding problem in the evaluation
are satisfactorily attained over all sample locations. This generates of mineral resources and reserves is the use of
one realization of possible sample values, respecting the fact that data with inherent errors. The use of imprecise
the actual samples are known to carry an error given by the global data has been studied in the literature4,9–11,
relative error. Multiple realizations of simulated sample values can but it is uncommon to find that the errors have
be obtained by repeating the procedure. At the end of this been accounted for in the evaluation of the
procedure, at every sample location a set of simulated sample mineral inventory12,13.
values is available that accounts for the imprecision of the
information. Furthermore, within each realization, the simulated
sample values are consistent with each other, reproducing the
spatial continuity and local statistics. These simulated sets of
sample values can then be used as input to conventional simulation
on a full grid to assess the uncertainty in the final resources over
large volumes. The methodology is presented and demonstrated
using a synthetic data-set for clarity. * Department of Civil & Environmental Engineering,
University of Alberta, Edmonton, Canada.
Keywords † SRK Consulting (Canada) Inc, Toronto, Canada.
sampling error, realizations of simulated sample values, uncertainty ‡ Department of Mining Engineering, University of
in final resources. Chile, Santiago, Chile.
© The Southern African Institute of Mining and
Metallurgy, 2012.ISSN 2225-6253. Paper received
Mar. 2009; revised paper received Oct. 2012.
L

The Journal of The Southern African Institute of Mining and Metallurgy VOLUME 112 NOVEMBER 2012 971
Transferring sampling errors into geostatistical modelling
Regardless of the type of sample obtained during drilling Diversions from this goal may be due to the characteristics of
campaigns, this data is the starting point for evaluating the the material, the equipment used for increment extraction,
quantity and quality of the various elements of interest. The the handling of increments after collection, and finally the
quality of the final resource evaluation and the uncertainty analytical process itself.
quantification performed for risk assessment depend strongly In the case of samples from drilling campaigns, the lot
on the data quality. Accounting for data quality from corresponds to the drilled core of a given length or the
different campaigns has been addressed through very simple detritus obtained in reverse circulation drilling, representing
approximations such as adding an independent random error the extracted cylinder. From these lots, and after a series of
to each sample value3. stages of division of the total mass and reduction of the
The present article deals with a methodology to particles size through crushing and pulverizing, a sample,
incorporate sampling errors from different drilling campaigns usually of a few grams, is obtained.
into the geostatistical modelling process. The aim is to The goal of sampling is to obtain a selected sample that
propagate the uncertainty associated with the data quality to represents correctly some properties of the lot. Particularly,
all subsequent steps of the project evaluation, including risk we are interested in the concentration (grade) of some
related to the mine plan, classification of resources and elements that have economic interest. Inevitably, when
reserves, transfer of geological risk into the financial selecting a subset of the lot, the sample will have properties
performance of a project, and so forth. slightly different from those of the lot. If sampling rules are
We first review some fundamentals of sampling theory strictly followed, there will be no bias, but some fluctuation
and of the geostatistical approach This is followed by an or error around the true value should be expected.
overview of the methodology and some details related to the The theory of sampling (ToS) has developed a systematic
simulation of multiple sample data-sets, in order to account taxonomy of errors and principles for minimizing or
for their precision. We provide some implementation details; eliminating sampling errors. The combination of errors from
and finally, we show an application to a synthetic data-set. these sources identified by Gy5,7,14 is termed the total
We conclude with a discussion, recommendations, and future sampling error (TSE). The relevant individual errors for this
work. paper are shown in Table I.
Regardless of their type, the errors generate a difference
between the sample properties and those of the lot. As long
Sampling error as each particle of the lot has an equal probability of
When dealing with particulate materials, sampling theory belonging to the sample, there should be no bias; that is, in
provides a set of rules to ensure that some basic principles expected value the sample will have the same properties as
are followed when extracting a subset of the original volume the lot. However, some dispersion around the true value
in a representative manner. The original volume of material should be expected. A relative error quantifies this
that is being characterized through the sampling process is dispersion.
called a lot5,7. A sample is a small quantity of material When accounting for all the sample locations, a global
extracted from the lot in such a manner that the sample relative error (GRE) can be quantified:
represents the essential characteristics of the lot. Sampling as
a process, however, aims at ensuring that the integrity of the [1]
sample is preserved.
The sampling process involves a stepwise reduction of the
mass and the fragment size from the lot to the sample. The where n is the total number of sample locations, Zs (xi) is
number of steps involved is a function of material character- the sample value at location xi, and ZL (xi) is the true value of
istics and the requirements of the analytical procedure. the lot sampled at location xi.

Table I

Taxonomy of sampling errors in regard to material characteristics, plant processes, sampling equipment, and
analytical procedures7,14,15

Sources of error Nature of error

Material characterization (sampling uncertainties, known, Fundamental sampling error (FSE), related to compositional heterogeneity due
but never eliminated, minimized by strict adherence to principles) to the material properties of the lot
Grouping and segregation error (GSE) related to distributional heterogeneity due
to grouping and segregation in the lot
Sampling equipment and materials handling (sampling errors, Increment delimitation error (IDE), geometry of outlined increment is not
unknown but eliminated by strict adherence to ToS principles) completely recovered; can be completely eliminated
Increment extraction error (IEE), material extracted does not coincide with th
delineated increment; can be completely eliminated
Increment preparation error (IPE), all sources of non-stochastic variation after
extraction of the material; error should always be nil
Analytical processes (mainly sampling errors) Analytical error (AE), all sources of error associated with materials handling and
processes in the laboratory
L

972 NOVEMBER 2012 VOLUME 112 The Journal of The Southern African Institute of Mining and Metallurgy
Transferring sampling errors into geostatistical modelling
Sampling theory allows us to quantify some of the sample error results in an inflation of the data variance and
components of this global relative error, particularly, the consequently, an increase in the variogram above the
fundamental sampling error (FSE). Other components cannot intended sill. Note further that while the variogram clearly
be directly quantified, but can be minimized through best rises above the sill, this should not be misinterpreted as a
practices, as they are related to the FSE and to the trend in the data. A trend demonstrated via the experimental
homogeneity of the lot, the number of increments to be used variogram tends to show a consistently increasing slope
to compose the sample, proper sampling equipment design, beyond the range of correlation16. As such, the rise in the
etc. apparent sill above the standard sill is believed to be due to
the sample error.
Geostatistical approach In conventional practice, the exploratory data-set of the
mineral deposit is assumed to be error-free, and the experi-
In geostatistical applications, a probabilistic approach based
mental semivariogram of the data-set with its sampling error
on the concept of random variables is used to account for our
(from now on referred to as the available data-set) is used
lack of information at unsampled locations. Similarly, we can
directly both for estimation and simulation. Since these
consider that at those locations where we have a sample, but
techniques reproduce exactly the conditioning data, the
whose values are uncertain, random variables can represent
conditional variance at such locations is zero, reflecting a
these uncertainties. Specifically, we can assign a probability
misleading certainty about these values and underestimating
distribution to the attribute of interest, which in our context
the propagated uncertainty due to the actual imprecision at
is the mineral or metal grade. It should be emphasized that
sample locations. This practice leads to an understatement of
locally the relative error should be of equal magnitude as the
the actual uncertainty related to the resources of the deposit.
global relative error, but since the grade changes from one
Based only on the available data-set and its respective
location to another, the dispersion is proportional to the
sampling error (i.e. GRE), inference of the actual mean and
grade value. This is, of course, a model of the spread of error
variance is difficult; means and variances of the possible
around the true value. Other approaches could be
data-sets with sampling error could be smaller than, similar
considered4.
to, or larger than that of the data-set with no sampling error.
The distribution of the samples from a lot is frequently
Since we do not know the actual values without error, we
assumed to be Gaussian. However, experimental results have
must centre our statistical inference on the available data-set,
shown that this is not always the case, especially when
which includes errors.
dealing with precious metals with high nugget effect and
highly skewed distributions7.
From a geostatistical point of view, the lack of precision
in the sample data can be accounted for by performing
estimation or simulation conditional to data values that are
independently drawn from the distribution at each sample
location. This approach has been used before3. However,
when significant relative errors are considered, one should
also ensure that the spatial correlation between these
simulated values at sample locations is preserved. If the
values are simulated independently at every location, the
spatial correlation will be partially lost and this will be
reflected in the estimated or simulated models at subsequent
stages. In the following section, we propose a methodology to
handle data with different errors, imposing their spatial
correlation and accounting for their precision.

Dealing with the sampling error

As mentioned, the common approach to accounting for
sampling errors is to draw at each sample location a value
from a distribution, usually Gaussian, centred at the sample
value and with a variance proportional to the relative error
associated. Since this process is done independently at each
sample location, some spatial correlation is lost, and the
simulated sample values show a higher nugget effect than
the original values (with error), and the underlying actual
values, which are unknown. This increase may not be
relevant when the sampling error is low. However, when
dealing with large relative errors, it may hide the actual
Figure 1—Experimental semivariograms of synthetic data-set (black
continuity of the variable. This is illustrated in Figure 1.
dots), 100 simulated data-sets with sampling relative error of 30% (grey
One other consequence of adding an independent sample lines), average of the experimental semivariograms of the 100 simulated
error to the data values is an explicit addition to the variance data-sets with error (dashed line), and semivariogram model of the true
of the data. As the sill of the variogram should be the data-set (solid line) for 0º azimuth (top) and 90º azimuth (bottom). All the
variance of the data, this approach to explicitly account for values were transformed to normal score units
L

The Journal of The Southern African Institute of Mining and Metallurgy VOLUME 112 NOVEMBER 2012 973
Transferring sampling errors into geostatistical modelling
This can be illustrated with a simple example (see The Gibbs Sampler17,18 works by selecting realizations
Figure 2). Let us assume the original data-set with no from a second-order stationary random function, so that at
sampling error is accessible, and 100 realizations are the conditioning data locations they reproduce the GRE with
sampled, assuming the distribution of samples with error for respect to the available data-set; the variability due to the
each data location is Gaussian, with a constant coefficient of sampling error is transferred to the simulated model in terms
variation in order to mimic a sampling process with error. of local means and variances (Figure 3).
Notice that means and variances of the simulated data-sets
with sampling error can be smaller, similar or larger than Gibbs sampler algorithm for simulating additional
those of the original data-set. Once we have a data-set with
data-sets
sampling error, we cannot infer if its statistical parameters
are different than those of the error-free distribution, as this Since the experimental semivariogram has to be reproduced,
is dictated purely by chance. Since the error is added a Gibbs sampling algorithm is used to generate the
independently at every location, on average the distribution alternative data-sets, imposing the reproduction of the spatial
will show a higher global variance, as depicted in Figure 2 structure and of global and local statistics. The sampling
(right). This requires assuming the spatial correlation strategy splits the algorithm in two parts to speed up the
inferred from the data is not significantly affected by the processing time; however, this strategy is not rigid and many
sampling error, as we do not have access to the true other alternatives based on the Gibbs sampler algorithm
underlying variogram. could be proposed to obtain the same results described here.
The proposed approach considers simulating at the Basically the aims of each part are as follows:
sampling locations, honouring the sample values in expected ® Reproduce the global mean, experimental semivar-
terms, and reproducing globally the GRE. Furthermore, the iogram in normal score units, and the GRE values in
simulated values must retain the spatial correlation inferred the simulated subsets of values at sample locations; all
from the available information. Notice that these constraints of them within given tolerances. Each subset represents
do not require the distribution at sample locations to be a drilling campaign and can have different GRE values
Gaussian-shaped. Each realization will represent a plausible ® From the selected data-sets in the first part, adjust
set of samples obtained at the sample locations with the locally the local means to match the sample values; the
known precision provided by the GRE. reproduction of the experimental semivariogram in
normal score units, and the GRE values are still
preserved
® For the Gibbs sampler algorithm, the back-transfor-
mations are carried out using the transformation table
of the declustered available data-set. The sampling
process is performed in Gaussian units since the
conditioned distributions are parametric and the
parameters can be easily and correctly calculated using
the simple kriging (SK) approach. The experimental
semivariograms are calculated from the normal score
transformed values, since it is this semivariogram that
is reproduced in simulation
® The respective targets of the two parts of the Gibbs
sampler algorithm are approached within some

Figure 3—Sketch of the proposed approach compared to the conven-

tional methodology for a 1D second-order stationary RF; the dashed
Figure 2—Histograms of global means (top) and global variances lines represent some confidence intervals that denote the local
(bottom) of 100 realizations where each location was simulated variances along the second-order stationary RF and the solid line in the
independently assuming a local Gaussian distribution of samples with centre the conditioned mean. Notice the conditioned local means are
error and local and constant coefficients of variation. The black dots at the same for the two approaches and the local variances are different.
the bottom of each histogram represent the reference values of the The outer confidence limits are of the proposed approach and the inner
original data-set with no sampling error confidence limits of the conventional approach
L

974 NOVEMBER 2012 VOLUME 112 The Journal of The Southern African Institute of Mining and Metallurgy
Transferring sampling errors into geostatistical modelling

Figure 4—Sketch of the process of approaching one of the Gibbs sampler conditions to its respective target. By using tolerances the Gibbs sampler
conditions do not have to be approached to their targets indefinitely; this is helpful when the Gibbs sampler algorithm has multiple conditions

tolerance intervals. The use of tolerances helps to simulated values and calculate the respective GRE
reduce the calculation time; they act as bandwidths of values for each subset, each representing a drilling
acceptable solutions with respect to the target values campaign with different sampling error
(see Figure 4). Conditions that have reached their 2. Visit all the sample locations following a random path.
respective targets within the tolerances do not continue The term iteration is used for one round when all the
trying to approach the target but move freely within the sample locations have been visited
tolerance limits. This allows the algorithm to a. Temporarily remove the value of the visited
concentrate on the conditions that have yet to be met location and calculate the parameters of the
and not be stressed by conditions that are already conditioned distribution by simple kriging of the
acceptable, thus expediting the entire process
® For the first part of the algorithm, a semivariogram
model is fitted to the experimental semivariogram
values of the available data-set in normal score units.
Using this semivariogram model, several data-sets are
simulated unconditionally to initialize the algorithm.
Since the domain of the samples is finite, a smaller
variance is obtained on average in the simulated data-
sets, since it represents the dispersion variance over
the domain, rather than the theoretical variance. Since
this bias affects equally high and low values in
Gaussian units after back-transformation to original
units, this may carry a bias in the global mean,
depending on the skewness of the distribution. This
bias is illustrated in Figure 5. When the starting
realization has a significantly lower variance than the
target, convergence to target statistics is hard to obtain.
Therefore, for optimization purposes these realizations
are discarded after a maximum number of iterations is
performed without satisfactorily converging to the
target statistics within their allowable tolerances.
The result of the first part of the algorithm (see Figure 6)
is a set of simulated data-sets that honour the experimental
semivariogram of the available data-set in normal score
units, the GRE values for all the subgroups in the available
data-set (each representing a drilling campaign), and the
global means. Basically, this part of the algorithm filters out
those unconditional realizations that are hard to solve, and
delivers data-sets that are easy to solve for local correction to
the second part of the algorithm. The workflow of the first
part of the algorithm is summarized as follows:
Figure 5—Distributions of global means (top) and global variances
1. Initialize the algorithm simulating an unconditional
(bottom) in original scale units of 100 unconditional simulated data-sets
realization using the semivariogram model of the at the available data-set locations. The black dots below the two plots
available data-set. The unconditional realization is represent the referential global mean and global variance of the
simulated at sample locations. Back-transform the available data-set

975
L

The Journal of The Southern African Institute of Mining and Metallurgy VOLUME 112 NOVEMBER 2012
Transferring sampling errors into geostatistical modelling

Figure 6—Flow diagram of the first part of the algorithm to simulate data-sets that reproduce GRE, semivariogram, and global mean

normal scores using the nearby values at sample The second part of the algorithm (see Figure 7) takes the
locations. These are the values simulated simulated values from the first part and corrects the
unconditionally in the previous step reproduction of the local means so that they are be equal to
b. Draw a simulated value from the conditioned the available data-set within some tolerance intervals. The
distribution experimental semivariogram and the GRE were imposed in
c. With the sampled value both in normal score and the first part of the algorithm and now move freely within the
in original units respectively check if: tolerance limits. Correcting the local means generates the
® The GRE value of each subset approaches its respective following effects:
target ® The GRE reproduction is better for all subsets, as it gets
® The normal scores experimental semivariogram of the closer to the target values with less dispersion
simulated data-set approaches the reference semivar- ® The global means and global variances match the
iogram. This is checked for several lags and directions targets more closely
® The global mean of the simulated data-set approaches ® The nugget effect of the experimental semivariograms
the reference mean, in original units. of the simulated data-sets increases; however, this is
d. The term attempt is used for each time a sample controlled by the tolerance in the reproduction of the
is drawn from the conditioned distribution. For experimental semivariograms.
implementation, a limit in the number of The summary of the second part of the algorithm is:
attempts is restricted to a maximum value. If the 1. For each realization, take the simulated values from
sampled value satisfies all the conditions of step the first part of the algorithm
2c accept the sampled value, keep it as part of 2. Visit all the locations following a random path
the simulated data-set; otherwise reject it and go a. Temporarily remove the value of the visited
to step 2b. location and calculate the parameters of the
3. The number of allowable iterations is restricted to a conditioned distribution by simple kriging of the
maximum value. When this value is reached the normal scores of the nearby values at sample
algorithm discards the simulated data-set and restarts locations
from step 1. If the maximum number of iterations has b. Draw a simulated value from the conditioned
not been reached, go to step 2. This part of the distribution
algorithm stops when all required realizations are c. With the sampled value both in normal score and
completed. in original scale units check if:
L

976 NOVEMBER 2012 VOLUME 112 The Journal of The Southern African Institute of Mining and Metallurgy
Transferring sampling errors into geostatistical modelling

Figure 7—Flow diagram of the second part of the algorithm that corrects the local means of the data-sets generated in the first part

® The GRE values approach to their respective targets and how it would allow combining information from two
® The experimental semivariogram of the simulated data- different campaigns with different sampling errors, the
set approaches to the experimental semivariogram of following synthetic example has been prepared.
the available data-set The goal is to account for these sampling errors in the
® Over all realizations, the mean of the simulated values subsequent step of simulation. The procedure is:
at that particular sample location approaches the actual ® Generate 100 realizations of the values at sample
sample value (with error), which is assumed as the
locations, honouring the GRE, local values, and
correct mean, since we do not have access to the error-
semivariogram in normal score units, using the
free value.
proposed Gibbs sampler methodology
d. For implementation, the number of attempts is
® Simulate 100 realizations of the full grid conditioned
restricted to a maximum value. If the sampled
on each case to one of the simulated sample data-sets
value satisfies all the conditions of step 2c
® Compare the results with the uncertainty assessed
accept the sampled value, keep it as part of the
when disregarding the sampling errors.
simulated data-set; otherwise reject it and go to
step 2b. Let us consider an available data-set with 400 sample
3. The algorithm runs until the local means over the locations placed over a regular grid of 500×500 units of
realizations at all sample locations are similar to the distance. The available data-set was split into two sub groups
available data-set values within the given tolerances. with a different sampling error associated with them. This
The resulting simulated data-sets now honour the GRE was done in order to mimic the integration of two sampling
for all the subsets representing different drilling campaigns, campaigns in the study (see Figure 9). This available data-
the experimental semivariogram in normal score units, and set was built sampling one realization over a regular grid and
the local means in original scale units. adding randomly a categorical code to each sample. The
The data uncertainty due to sampling errors captured in sampling errors were assigned according to categorical codes,
the previous steps must now be transferred into the next step GRE equal to 20% for code 1 and 40% for code 2.
of geostatistical modelling. Each set of simulated data values The Gibbs sampler algorithm runs in two parts. The first
is used as input in a three-dimensional conditional part tends to reduce the global means and global variances;
simulation model of the grades (see Figure 8). however, this effect is corrected in the second part, where the
dispersion is also reduced. Even when the global mean is
Example implementation fairly reproduced, the global variances are greater than the
To illustrate the practical implementation of this algorithm target of the available data-set (see Figure 10).
L

The Journal of The Southern African Institute of Mining and Metallurgy VOLUME 112 NOVEMBER 2012 977
Transferring sampling errors into geostatistical modelling

Figure 8—Sketch of uncertainty transfer due to sampling error into the simulated models. Each simulated data-set is the input of a simulated realization;
when summarizing the realizations, the sampling error features are reproduced in the simulated maps on average at the conditioning data locations as
well as the available data-set values

Figure 9—Locations of the two sub-data-sets with different sampling errors of GRE 20% (empty dots) and GRE 40% (solid dots)

The variability in the reproduction of the GRE values for In the first part of the algorithm only the global
the two subsets is also reduced by the second part of the conditions are targeted; the second part targets the
algorithm. In the first part, the GRE values are reproduced reproduction of the local features. Although the first part of
within the tolerance limits, and in the second part the the algorithm starts with unconditional realizations, the GRE
dispersion is reduced (see Figure 11). converges to the target values for each subset with
L

978 NOVEMBER 2012 VOLUME 112 The Journal of The Southern African Institute of Mining and Metallurgy
Transferring sampling errors into geostatistical modelling

Figure 10—Reproduction of the global means (blue line) and global variances (red line) for the first part (top) and the second part (bottom) of the algorithm;
for two GRE targets of 20% and 40% with 5% tolerance, global mean tolerance 10%, and semivariogram tolerance 5%

Figure 11—Reproduction of the two groups of GRE 40% (blue line) and GRE 20% (red line) for the first (top) and the second part (bottom) of the algorithm;
for two GRE targets of 20% and 40% with 5% tolerance, global mean tolerance 10%, and semivariogram tolerance 5%
L

The Journal of The Southern African Institute of Mining and Metallurgy VOLUME 112 NOVEMBER 2012 979
Transferring sampling errors into geostatistical modelling
subsequent iterations. This implies that the simulated values
tend locally toward the measured sample values (with error),
as shown in Figure 12 (top). When considering the average
of the simulated values at a sample location, the second part
of the algorithm allows a much better reproduction of the
target values, that is, the values of the available data-set
(Figure 12, bottom).
Figure 13 shows the reproduction of the semivariogram
while accounting for the sampling error. Further, we note
that the simulated data-sets also reproduce the semivar-
iogram of the available data following the first part of the
algorithm. It can be observed that the projection of the nugget
effect is lower than in the case of adding an independent
error to every location (Figure 1).
After the second part of the algorithm, the semivariogram
reproduction improves and the dispersion between
realizations is slightly lower (Figure 14).

Discussion
Recall that in the conventional approach, resource modelling
depends on the available data-set, which is known to have
errors. It is common to ignore the sample error and/or
assume that an increased nugget effect in the semivariogram
adequately accounts for sample errors in the final model. Figure 13—Experimental semivariogram reproduction in normal score
Conventional simulation modelling of resources then units of the 100 simulated data sets (grey lines), average of experi-
proceeds by (1) transforming the conditioning data mental semivariograms (dashed lines) and experimental semivariogram
(disregarding the sampling error) to normal score units, (2) of the available data set (black dots) for the first part of the Gibbs
fitting a semivariogram model, (3) simulating many sampler algorithm for 0º azimuth (top) and 90º azimuth (bottom)
realizations, (4) back-transforming to original units, and (5)
summarizing the realizations in order to assess uncertainty
(Figure 15 shows an example of one such summary).

Figure 14—Experimental semivariogram reproduction in normal score

units of the 100 simulated data sets independently transformed to
normal score units from the simulated values in original units (grey
lines), average of experimental semivariograms (dashed lines) and
Figure 12—Correlation plots of the local mean values versus the values experimental semivariogram of the available data set (black dots) for
of the available data set with sampling error for the first (top) and the second part of the Gibbs sampler algorithm for 0º azimuth (top) and
second part (bottom) of the algorithm 90º azimuth (bottom)
L

980 NOVEMBER 2012 VOLUME 112 The Journal of The Southern African Institute of Mining and Metallurgy
Transferring sampling errors into geostatistical modelling
indirectly imposed by the global sampling error, but without
local control (Figure 17, right).
When comparing the distributions of the global means
and variances of the conventional approach (see Figure 18)
and of the proposed approach that reproduces the GRE at the
sample locations (Figure 19), it can be seen that global
means and global variances spread over a larger range in the
proposed approach.
The global means for the proposed approach are centred
on a value slightly lower than in the conventional case.
However, global variances appear centred at a very similar
value, with more dispersion in the case of the proposed
approach. This can be explained by the fact that every
realization is conditioned to a different set of samples and,
therefore, the reference histogram changes. Depending on the
distribution of high and low samples, accounting for the
sampling error yet preserving the spatial correlation allows
the global distribution to fluctuate. This is the result of
transforming each simulated data-set independently and
back-transforming the simulated maps using individual
transformation tables.

Conclusions
Most widely used geostatistical techniques in the mining
industry for building resource/reserve models assume that
the conditioning information is error-free; conditioning data

Figure 15—Average of 100 simulated values (top) and conditional

variances (bottom) using conventional approach

This paper proposes a methodology that accounts directly

for sample error in the available data at the sample locations.
The result of this proposed approach is a set of realizations of
sample data at the sample locations. Each simulated data-set
acts as the conditioning sample data used to generate each
simulated realization. Resource modelling in this context
requires that for each simulated data-set, we (1) transform
the data to normal score units, (2) simulate one realization,
and (3) back-transform the realization to original units.
Semivariogram modelling is performed only once, on the
original data-set. The uncertainty from this approach can be
assessed once all realizations are generated. Figure 16 shows
a similar summary of local means and variances as those
shown in Figure 15, but the results in Figure 16 are based on
the modelling approach outlined here. Notice that the
variability at the sample locations is no longer zero but in
fact reflects the uncertainty related to sampling quality.
The average value at a sample location, over the set of
realizations generated with the Gibbs Sampler algorithm,
should be equal to the sample value (with error), since this is
the only hard data available. However, since some tolerances
were used in the second part of the algorithm, some fluctu-
ations occur, as illustrated in Figure 17 (left). Globally, the
allowable departure can be controlled at the expense of Figure 16—Average of 100 simulated values (top) and conditional
computation time for convergence (if more strict tolerances variances (bottom) using transfer variability for data sets with sampling
are imposed). In the case of the local variances, these are error
L

The Journal of The Southern African Institute of Mining and Metallurgy VOLUME 112 NOVEMBER 2012 981
Transferring sampling errors into geostatistical modelling

Figure 19—Global means (top) and global variance (bottom) of 100

simulated maps using data sets with sampling error
Figure 17—Scatter plots of local means (top) and global variances
(bottom) of the sets of simulated maps. For the two scatter plots the
conventional approach versus the proposed approach are compared
are perfectly reproduced at their respective locations in the
geostatistical model. The uncertainty in the conditioning
information cannot be accounted by using these techniques;
therefore, it is not transferred into the geostatistical model.
Since the exploratory data-set in mining consists of many
different subsets corresponding to different drilling
campaigns (e.g. reverse circulation drilling, diamond core
drilling, or sampled information taken using technologies
from different time periods or with different QA/QC
standards), the presence of errors in the exploratory data-set
is unavoidable. This should be accounted for in the charac-
terization of the mineral deposit.
The present paper deals with errors in the data-set that
arise from the sampling process, that is, the fundamental
error. In practice there are additional sources of errors, such
as errors in laboratory analysis or errors in manipulation of
the sample, which cannot be easily quantified. A total error
could be used with the proposed methodology, instead of
only the fundamental error, as presented in the example.
The proposed approach is able to deal with many
different subsets of information that comprise the exploratory
data-set, each subset with a different sampling error. The
general procedure consists of simulating alternative data-sets
such that global statistics are preserved and the sampling
error is imposed by considering that, at sample locations, the
sample value is an outcome of a random variable with a
mean given by the actually sampled value, which has an
error, and a variance controlled by the amount of global
Figure 18—Global means (top) and global variances (bottom) of 100 relative error. These random variables are not independent,
simulated maps using conventional approach therefore the values are drawn using a Gibbs sampler in
L

982 NOVEMBER 2012 VOLUME 112 The Journal of The Southern African Institute of Mining and Metallurgy
 Transferring sampling errors into geostatistical modelling
order to preserve the spatial correlation. Once the alternative 7. PITARD, F.F. Pierre Gy’s Sampling Theory and Sampling Practice -
Heterogeneity, Sampling Correctness and Statistical Process Control. 2nd
data-sets have been simulated, each one feeds a simulated edn. CRC Press, Boca Raton, Florida, 1993. pp. 488.
model constructed in the conventional way. The result is that
8. MAGRI, E.J. Some experiences in heterogeneity tests and mine sampling.
overall uncertainty is increased when the sampling error is Proceedings of the Third World Conference on Sampling and Blending,
considered, demonstrating that assuming the samples to be Porto Alegre, Brazil, 23–25 October 2007. Costa, J.F.C.L. and Koppe, J.C.
(eds.), Fundacao Luiz Englert, 2007. pp. 329–348.
error-free, when in fact they are not, leads to an
understatement of the actual uncertainty in the resources and 9. FREULON, X. Conditioning a Gaussian model with inequalities. Proceedings
of the Fourth International Geostatistics Congress—Geostatistics Tróia’92.
reserves. Soares, A. (ed.). Kluwer Academic, Dordrecht, 1993, vol. 1. pp. 201–212.
10. FREULON, X. Conditional simulation of a Gaussian random vector with
Acknowledgments nonlinear and/or noisy observations. Geostatistical Simulations.
Armstrong, M. and Dowd, P.A. (ed.). Kluwer Academic, Dordrecht, 1994.
The authors wish to acknowledge the three anonymous pp. 57–71.
reviewers for their constructive comments that improved this 11. ZHU, H. and JOURNEL, A.G. Formatting and integrating soft data: stochastic
paper. imaging via the Markov-Bayes algorithm. Proceedings of the Fourth
International Geostatistics Congress—Geostatistics Tróia’92, Soares, A.
(ed.). Kluwer Academic, Dordrecht, 1993, vol. 1. pp. 1–12.
References
12. SINCLAIR, A.J. and BLACKWELL, G.H. Applied Mineral Inventory Estimation.
1. WAKEFIELD, T. Legacy Data quality and its impact on mineral resource Cambridge University Press, Cambridge, 2002. pp. 381.
estimation. MININ 2008 – III International Conference on Mining
13. ROLLEY, P.J. Geologic uncertainty in the development of an open pit mine: a
Innovation. Arias, J., Castro, R., and Golosinski, T. (eds.). Gecamin Ltda.
risk analysis study. Unpublished M.Sc. thesis, University of Queensland,
Santiago, Chile, 2008. pp. 3-10.
Australia, 2000. pp. 164.
2. VALLEE, M., AND SINCLAIR, A.J. Quality control of resource/reserve
estimation—where do we go from here? CIM Bulletin, vol. 1022, 1998. 14. GY, P. Sampling of discrete materials – a new introduction to the theory of
pp. 55–57. sampling. I. Qualitative approach. Chemometrics and Intelligent
3. MAGRI, E. and ORTIZ, J. Estimation of economic losses due to poor blast Laboratory Systems, vol. 74, no. 1, 2004. pp. 7–24.
hole sampling in open pits. Geostatistics 2000, Proceedings of the 6th 15. GY, P. Sampling of Heterogeneous and Dynamic Material Systems:
International Geostatistics Congress, Cape Town, South Africa, 2000. Theories of Heterogeneity, Sampling and Homogenizing. Elsevier,
Kleingeld, W.J. and Krige, D.G. (eds.), vol. 2, pp. 732–741, 10–14. Amsterdam, 1992. pp. 653.
4. EMERY, X., BERTINI, J.P., and ORTIZ, J.M. Resource and reserve evaluation in 16. LEUANGTHONG, O. and DEUTSCH, C.V. Transformation of residuals to avoid
the presence of imprecise data. CIM Bulletin, vol. 98, no 1089, 2005. artifacts in geostatistical modelling with a trend. Mathematical Geology,
pp. 2000 vol. 36, no. 3, 2004. pp. 287-305.
5. GY, P. Sampling of Particulate Materials—Theory and Practice. 2nd edn.
17. GEMAN, S. and GEMAN, D. Stochastic relaxation, Gibbs distribution and the
Elsevier, Amsterdam, 1982. pp. 431.
Bayesian restoration of images. IEEE Transactions on Pattern Analysis
6. FRANÇOIS-BONGARÇON, D. Geostatistical tools for the determination of
and Machine Intelligence, vol. 6, no. 6, 1984. pp. 721–741.
fundamental sampling variances and minimum sample masses.
Proceedings of the Fourth International Geostatistics Congress— 18. EMERY, X. and ORTIZ, J.M. Histogram and variogram inference in the
Geostatistics Tróia’92. Soares, A. (ed.). Kluwer Academic, Dordrecht, multigaussian model. Stochastic Environmental Research and Risk
1993, vol. 2. pp. 989-1000. Assessment, vol. 19, no. 1, 2005. pp. 48–58. N

The Journal of The Southern African Institute of Mining and Metallurgy VOLUME 112 NOVEMBER 2012 983

View publication stats