http://dx.doi.org/10.

17159/2411-9717/2016/v116n7a6

When should uniform conditioning be

applied?
by K. Hansmann*

distributions. Additional studies have applied

&:<08>8 this approach to gold deposits (Assibey-
Blindly applying any methodology to estimate the recoverable resources of
Bonsu, 1998; Humphreys, 1998) and an iron
a mineral deposit without considering the suitability of the approach to the ore deposit (De-Vitry et al., 2007), indicating
deposit being evaluated can render misleading results. While ‘running the that the method is applicable on skew,
software’ provides an answer, one should, amongst numerous other lognormal distributions. While UC has been
considerations, understand the impact the underlying distributions and practically applied to different types of deposit,
assumptions have on the validity of the result. it is not known how well UC predicts actual
Uniform conditioning (UC) is a nonlinear estimation method that grades for the underlying grade distribution.
models the conditional distribution of smallest mining unit (SMU) block This paper discusses the UC estimation
grades within panels, and localized uniform conditioning (LUC) places method as well as the popular add-on, LUC,
these SMU at plausible locations within a panel. The localization process
presented by Abzalov (2006). A case study is
does not improve the accuracy of the UC result, but rather presents the
presented that compares UC and LUC estimates
result in a more practical format; particularly for use in mine planning.
A case study was carried out to compare the suitability of UC and LUC of two hypothetical data-sets, referred to
on two hypothetical data-sets. The data-sets are simulated realizations of herein as scenario 1 and scenario 2. Scenario 1
a normal grade distribution and a highly skewed lognormal grade distri- is a normally distributed grade distribution
bution which are akin to grade distributions found in mineral deposits. The and scenario 2 is a skew, lognormally
estimation methods were applied to both data-sets, and the results distributed grade distribution, both of which
compared with the actual grades of the simulated realizations. This paper are compared against the simulated
presents an overview of UC and LUC, with discussions around the case realizations that represent the actual grades.
study results. The conditions found in the two data-sets

;$<=58 are similar to those found in naturally
change of support, Gaussian anamorphosis, localized uniform occurring mineral deposits. The aim of this
conditioning, lognormal distribution, normal distribution, uniform investigation is to determine the underlying
conditioning. conditions of the grade distribution that
produces favourable results when applying UC,
and subsequently LUC, to such data-sets.

A:9=<5729><:
:>3<=1@2<:5>9><:>:4@$<=36<$
Uniform conditioning (UC) is a nonlinear
The following section describes a UC with LUC
estimation technique that estimates the
workflow, which follows the process outlined
conditional distribution of metal and tonnage
in Figure 1.
above cut-off within a mining panel. It does
not directly estimate grade, although grade is a ## ##" #" !"
typical outcome from the estimated metal-
This initial part of the UC workflow is to
tonnage distribution or the results produced by
prepare and carry out exploratory data
localized uniform conditioning (LUC). UC
analyses, including histograms, to understand
results are typically presented as a recoverable
the sample grade distribution and variability in
resource above multiple cut-off grades. The
the deposit. The data must be appropriately
advantage of UC is that it can be used on
widely spaced data, across domains that are
not strictly stationary, provided that there is
sufficient data for a conditionally unbiased
estimate of the panel mean grade (Rivoirard,
* University of the Witwatersrand, Johannesburg,
1994).
South Africa.
Previous studies where UC has been
© The Southern African Institute of Mining and
applied to porphyry copper deposits (Deraisme Metallurgy, 2016. ISSN 2225-6253. This paper
et al., 2008; Deraisme and Assibey-Bonsu, was first presented at, The Danie Krige
2011; Millad and Zammit, 2014) show the Geostatistical Conference 2015, 19–20 August
application of the method to normal grade 2015, Crown Plaza, Rosebank.
L


   
VOLUME 116  
 645
When should uniform conditioning be applied?
#" !#!
The quality of the panel estimate determines the success of
the UC estimation (Rivoirard, 1994). A panel estimate should
be conditionally unbiased (Rivoirard, 1994; De-Vitry et al.,
2007), so that the UC conditional grade distribution will be
an accurate estimate of the actual grade distribution. The
panel estimate can be carried out using any linear estimator,
but conventionally ordinary kriging (OK) is used.
The panel size should be chosen relative to the spacing of
the sample data. De Vitry et al. (2007) suggested that the
panel should be as small as possible to ensure an accurate
estimate, but large enough for minimal conditional bias of the
estimate. The number of smallest mining units (SMU) within
the panel is linked to the resolution of the grade-tonnage
relationship, as the number of SMU discretizes the grade-
+>47=;@/* :>3<=1@2<:5>9><:>:4@$<=36<$#@$>9-@6<2?6>
?9><: tonnage curve of the panel (Harley and Assibey-Bonsu,
2007).

declustered, so that the normal score transforms of the grades

#!!#" #"#!!
are an accurate representation of the grade distribution. This The DGM can be used to derive the marginal histograms at
is important for the Gaussian anamorphosis function (to be different supports. The Gaussian anamorphosis function is
discussed later), and failure to correctly complete this will modelled by a set of Hermite polynomials, weighted with an
affect the results of the UC estimate. accompanying set of Hermite coefficients. A full description
of Hermite polynomials and how these may be calculated is
# !!
given by Rivoirard (1994).
As the discrete Gaussian model (DGM) for change of support A point and fitted model anamorphosis function for a
is used for UC, the data must be transformed to the normal and lognormal distribution are shown in Figure 3.
equivalent Gaussian (or normal score) values using The anamorphosis function for the lognormal distribution is
declustered weights. This is performed by transforming the constructed from the normal score data, by plotting pairs of
cumulative distribution frequency (CDF) of the original grade and Gaussian transformed grade values.
grades to a Gaussian probability CDF, on a percentile to
percentile basis for the entire data-set. Figure 2 shows the
CDF of a lognormal grade distribution (data from scenario 2)
on the left, with the green and red lines showing percentile
paired mapping of values to the equivalent normal score
values on the right.
The DGM relies upon the assumption of bivariate
Gaussianity of the transformed grades (Rivoirard, 1994).
Bivariate Gaussianity means that any linear combination of
the Gaussian transformed data is also Gaussian. Several tests
exist to determine if the transformed data conforms to such
conditions, and is therefore suitable for use with the DGM.
Schofield (1988), Rivoirard (1994) and Humphreys (1998) +>47=;@,*7176?9>.;@5>89=>)79><:@3=;7;:2@<3@4=?5;@.?67;8@9<
give practical examples on how these tests may be run. "?788>?:@9=?:83<=1;5@.?67;8@3<=@?@6<4:<=1?6@5>89=>)79><:

Gaussian Anamorphosi Gaussian Anamorphosi

— Data 1 — Model — Model — Data
30 60.0

25 50.0
Grade values [Z(x)]

Grade values [Z(x)]

20 40.0

15 30.0

10 20.0

5 10.0

0 0.0
-3 -2 -1 0 1 2 3 -3 -2 -1   0 1 2 3
Gaussian Values [Y(x)] Gaussian Values [Y(x)]

+>47=;@*(;39'@"?788>?:@?:?1<=0-<8>8@;0;=>1;:9?6@5?9?@?:5@1<5;6@3<=@82;:?=><@/@$>9-@?:@7:5;=6>:4@:<=1?6@5>89=>)79><:#@?:5@(=>4-9'@"?788>?:@?:?1<=%
0-<8>8@;0;=>1;:9?6@5?9?@?:5@1<5;6@3<=@82;:?=><@,@$>9-@?:@7:5;=6>:4@6<4:<=1?6@5>89=>)79><:
L

646  
 VOLUME 116


   

When should uniform conditioning be applied?
The Hermite polynomials are functions of the standard data spacing. A high R/r ratio is indicative of a strong
Gaussian distribution, and therefore they express correlation between SMU and panel grades, which indicates
probabilities and have the properties of a standard Gaussian good grade continuity in the deposit.
distribution (data is symmetrically distributed around the
mean value of zero, and has a unit variance). An additional ##"
property (see Table I) shows the variance of the grade Z(x) The result of a UC estimate is presented as a distribution of
expressed by the sum of Hermite coefficients, excluding the grades, shown as metal content and tonnages reported for a
0th, which describes the mean. series of cut-off grades. While this is insightful information
The number of Hermite coefficients used to fit the about the grade-tonnage distribution, it is not a particularly
anamorphosis model can vary, and the optimal number practical data format as the SMU location is not provided.
depends on the how well the polynomial set fits the Abzalov (2006) presents LUC as a simple extension to UC
underlying distribution. Neufeld (2005) recommends using that provides a practical solution for visualizing grades at the
less than 100 coefficients, although 20 to 30 coefficients are SMU level. A UC grade-tonnage distribution is decomposed to
usually sufficient.
The variance of grades depends on the support that the
grade represents, and a change-of-support model, like the
DGM, is used to predict the distribution of grade at different
supports. Grades at a point support have a higher variance
than grades of SMU, which similarly have a higher variance
than grades of panels. As the support of a grade increases, so
the grade values tend towards the population mean, have less
deviation from it, and are more symmetrical around it
(Figure 4).
There is a correlation between the distribution of grades
seen at a point support and the distribution of grades seen at
a SMU support, named the SMU change-of-support co- +>47=;@*&2-;1?9>2@8-<$>:4@8700<=9@;33;298@3<=@0<>:9@8?106;8#@&! #
efficient (r). Similarly, there is a correlation between the ?:5@0?:;68
distribution of grades seen at a point support and the distri-
bution of grades seen at a panel support, named the panel
change-of-support coefficients (R). The ratio R/r is the
correlation of SMU grades and panel grades. The R and r
change-of-support coefficients are determined by solving the
variance equations and the Gaussian anamorphosis
equations at SMU and panel supports, shown in Table I
(Rivoirard, 1994).

" """
The schematic in Figure 5 shows the relationship between
grades at a SMU support, grades at panel support, and how a
distribution of SMU grades is conditional on panel grade. A
low R/r ratio indicates a weak correlation between the SMU
and panel grades, which is caused by a high-nugget semivar- +>47=;@ *<:5>9><:?6@&! @4=?5;@5>89=>)79><:@( (''@3=<1@?@0?:;6@4=?5;
iogram and/or short semivariogram ranges relative to the ( (''@-;@2<==;6?9><:@<3@9-;@4=?5;8@>8@

Table I
(<0'@7?9><:8@3<=@2?6276?9>:4@.?=>?:2;@3=<1@;=1>9;@0<6:<1>?6@2<;33>2>;:98@0<>:9#@&! #@?:5@0?:;6@8700<=9
(<99<1'@7?9><:8@3<=@"?788>?:@?:?1<=0-<8>8@37:29><:@?9@0<>:9@8700<=9#@?:5@?5789;5@3<=@&! @?:5@0?:;6
8700<=9@39;=@>.<>=?=5@(/'

<>:9@8700<=9 &! @8700<=9 ?:;6@8700<=9

Variance equations

Gaussian anamorphosis equations

i Hermite coefficient r SMU change-of-support coefficient [0 ≤ r ≤ 1]

Hi Hermite polynomial evaluated for Gaussian value Z(V) Grade at panel support V
2
Dispersion variance Z(v) Grade at SMU support v
n Number of Hermite polynomial terms Z(x) Grade at point location x
R Panel change-of-support coefficient [0 ≤ R ≤ 1]
L


   
VOLUME 116  
 647
When should uniform conditioning be applied?
a series of grade values that reproduce the grade-tonnage Although both scenarios are equally sampled (at a
relationships. These plausible SMU grades are located into a density of 1%), the drill-hole spacing relative to the semivar-
SMU model based on the rank location of grades from a iogram ranges for scenario 1 is closer than for scenario 2. For
linear estimate of equally sized blocks. This results in a direct scenario 1, the samples are spaced at approximately one-third
grade model at the SMU resolution that respects the grade- of the semivariogram ranges (120 m) in X and Y, while for
tonnage distributions of the UC panels and attempts to reflect scenario 2 the samples are spaced at approximately the
the localized spatial grade distribution within the panel. maximum semivariogram range (40 m). Statistics of both
scenarios are compared, but there is no correlation between
?8;@8975 them as the simulations were run independently. Declustered
The objective of this case study is to assess the suitability of statistics are presented in Table II, histograms in Figure 8 and
UC and LUC for two data-sets with different grade distrib- Figure 9, and modelled semivariograms in Figure 10 and
utions, namely scenario 1 and scenario 2. The two scenarios Figure 11.
are distinctly different and represent two end-members of the In scenario 1, the grade distribution is symmetrical, with
range of grade distributions that may typically be seen in a comparatively low nugget effect (12%) and well-defined
mineral occurrences, being a symmetrical distribution and a continuity up to distances of 170–300 m. Slight anisotropy
positively skewed distribution. The grade distributions were was evident, and possibly some zonal anisotropy seen in the
synthetically generated and sampled to mimic how this would Y-direction where the variance does not reach the sill value.
be done in a mineral exploration project. The distribution is approximately normal, and is similar to
The UC with localization procedure described in this paper what one would find in a porphyry copper mineral
was followed for both data-sets, as outlined in Figure 1. occurrence. The first distribution has a smaller range of grade
values than the second (approximately half).
!"  #! ! ## In scenario 2, the grade distribution is approximately
Two sets of simulated data were generated, which were used lognormal, supported by the shape of the histogram. This
as base data for the assessment. A single realization was distribution has the characteristics of being asymmetrical,
simulated on a 2 m × 2 m × 2 m point grid, over a 800 m × strongly positively skewed, with a long tail. There is a higher
600 m area, with a thickness of 200 m, for each distribution. nugget effect (26%), with long-range continuity of approxi-
A plan view at surface through both simulations is shown in mately 60–90 m. The coefficient of variation (CoV) for
Figure 6 and Figure 7. scenario 2 is higher (1.3) than that of scenario 1 (0.4),
showing a wider spread and higher variability of grade
#!! #" ## values.
A spatially representative subset of data was taken from both
simulations, which makes up the sample database used for
#" !#!
this project. A total of 417 pseudo drill-holes, each containing Panel grade estimates, using OK, were produced for both
100 composites, was taken over the study area. The area is scenarios with the intent of minimizing conditional bias while
densely sampled, and this drilling grid would be consistent retaining some local variability. The block sizes were chosen
with that of a feasibility-stage project.

+>47=;@*6?:@.>;$@<3@9-;@8>176?9><:@<3@82;:?=><@,@(6<4:<=1?6@5>89=>%
+>47=;@*6?:@.>;$@<3@9-;@8>176?9><:@<3@82;:?=><@/@(:<=1?6@5>89=>)79><:' )79><:'

Table II

;82=>09>.;@89?9>89>28@3<=@82;:?=><@/@?:5@82;:?=><@,

71);=@<3@ 71);=@<3@ !>:@ !?@ !;?: ?=>?:2; &9?:5?=5@5;.>?9><:@ &;$:;88

7=9<8>8 <;33>2>;:9@<3@
8?106;8 )<=;-<6;8 49 49 49 49, 49 .?=>?9><:

Scenario 1 41 700 417 0.0 29.6 11.8 21.5 4.6 0.1 0.3 0.4
Scenario 2 41 700 417 0.0 59.3 4.1 30.3 5.5 3.7 19.5 1.3
L

648  
 VOLUME 116


   

When should uniform conditioning be applied?

+>47=;@*&?106;@->89<4=?1@?:5@27176?9>.;@3=;7;:2@->89<4=?1@3<=@82;:?=><@/

+>47=;@*&?106;@->89<4=?1@?:5@27176?9>.;@3=;7;:2@->89<4=?1@3<=@82;:?=><@,

relative to the average sample spacing, at 50 m × 50 m ×

20 m. Ten discretization points were chosen in the X and Y
directions, based on a quantitative kriging neighbourhood
analysis (QKNA) optimization of the block variance.
Discretization points in the Z direction were chosen to be
equal to the compositing length. A sufficiently large search
neighbourhood was chosen for the panel estimate to ensure
high slopes of regression, without the introduction of too
many (>5%) negative kriging weights.
The panel model estimation for scenario 1 had a mean
slope of regression of 0.97; while that of the scenario 2
panel model estimation was 0.72. Figure 12 and Figure 13
+>47=;@/ *0;=>1;:9?6@8;1>.?=><4=?1#@1<5;6@8;1>.?=><4=?1#@?:5 show plan views of the respective scenarios, at surface
0?=?1;9;=8@3<=@82;:?=><@/ elevation. Comparing these to the simulated data (Figure 6
and Figure 7), the grade smoothing effect and reduction of
variance from the OK is evident.

#"  !
Sample data was converted to normal scores and tested for
bivariate Gaussianity. For both scenarios, the test results
were consistent with bivariate Gaussian conditions of the
Gaussian transformed data. Change of support was carried
out for each scenario using the DGM, with parameters shown
in Figure 3.
In scenario 1 the SMU change-of-support coefficient
indicates a strong correlation between point and SMU grades.
For scenario 2, the SMU change-of-support coefficient implies
a weak correlation between point and SMU grades and is a
result of the high nugget effect.
+>47=;@//*0;=>1;:9?6@8;1>.?=><4=?1#@1<5;6@8;1>.?=><4=?1#@?:5 Panel change-of-support coefficients are measured from
0?=?1;9;=8@3<=@82;:?=><@, the direct variance of estimated panel grades. Well-informed
L


   
VOLUME 116  
 649
When should uniform conditioning be applied?

+>47=;@/,*&7=3?2;@06?:@.>;$@<3@<=5>:?=@=>4;5@0?:;6@;89>1?9;@?:5 +>47=;@/*&7=3?2;@06?:@.>;$@<3@<=5>:?=@=>4;5@0?:;6@;89>1?9;@?:5
6<2?9><:@<3@08;75<@5=>66%-<6;8@3<=@82;:?=><@/@ 6<2?9><:@<3@08;75<@5=>66%-<6;8@3<=@82;:?=><@,

Table III
&! @?:5@0?:;6@5>80;=8><:@.?=>?:2;8@?:5@=;8769>:4@2-?:4;@%<3@%8700<=9@2<;33>2>;:98@?:4;5@.?67;8@?=;@3<=
1>:>1?@?:5@1?>1?@<3@>8<.?=>?:2;@4=<70>:48

&2;:?=><@/ &2;:?=><@,
2
Sample variance (sill) 21.7 g/t 30.6 g/t2
Dispersion variance in SMU (theoretical value) 4.22 g/t2 24.45 g/t2
SMU change of support coefficient, r 0.903 0.503
Dispersion variance in panel (measured values) 6.17 g/t2–13.39 g/t2 0.91 g/t2–1.04 g/t2
Panel change of support coefficient, R 0.537–0.791 0.230–0.215
Change of support correlation R/r 0.595–0.876 0.427–0.457

panels, as defined by the number of samples in the

# #""# #!!!!"
neighbourhood and semivariogram ranges, will have better The grade-tonnage relationship for the normally distributed
estimation confidence. To account for this, different panel scenario is shown in Figure 14. The global UC prediction of
change-of-support correlations are used for panels with tons and grades is very close to the actual grades, and shows
similar confidence in the estimation. Three such isovariance a slight improvement on the OK grade-tonnage curve.
groupings were used for each scenario, and the ranges of In the case of the normally distributed grades (scenario
these results are shown in Table III. Relatively higher 1), where there is good data coverage (relative to the
variances of estimated panel grade groupings are typically semivariogram ranges), OK performs well for determining
from better-informed areas where there is better grade recoverable resources. Slopes of regression for the OK model
continuity that results in a greater spread of estimated panel were, on average, close to unity, indicating a very low
grades. Conversely, relatively lower estimated panel grade conditional bias (which is reflected in the OK estimates being
variance groupings are for worse-informed areas, where there close to the actual values). However, UC marginally
is more evidence of smoothing and panel estimates are closer outperforms OK in terms of estimating a recoverable
to the mean, giving rise to less variance resource, as it more closely predicts the grade and tonnage of
the simulated reality. For low cut-off grades, there is slightly
""" " """ #" ##" less tonnage than predicted for the both the OK and UC
UC was carried out for both data-sets using the panel model model, but the UC estimates are closer to the actual values.
and DGM. After completion of UC, the model was localized Where the grade data has an underlying lognormal distri-
using a local SMU model, which was estimated using smaller bution (scenario 2) with relatively poor data coverage, the
kriging neighbourhoods to reflect local variability. simulated model shows a decline in tonnage (or volume) as
the cut-off grade increases (Figure 15). OK generates a
moderate estimation of the grade and tonnage extractable for
>82788><:@<:@7:>3<=1@2<:5>9><:>:4@
any cut-off grade. This lack of adherence to the grade-
The performance of UC may be assessed by how closely the tonnage curve can be explained by a grade smoothing, which
UC grade-tonnage estimate conforms to the actual simulated was expected, as the slopes of regression of the panel
model and an OK model, as a benchmark for a linear estimate were, on average, poor. UC gives a better result than
estimator. This comparison was made globally, to OK, but the resultant estimation of grades and tonnage does
demonstrate the effects of incorrectly predicting the not closely conform to the actual values. As the selectivity
extractable tonnage of the deposit, and locally, to increases (i.e. high-grade areas are targeted), the average
demonstrate the effects of getting individual panel grades grade of the actual material will be higher than the OK model
right/wrong. predicts.
L

650  
 VOLUME 116


   

When should uniform conditioning be applied?

+>47=;@/*"6<)?6@4=?5;%9<::?4;@27=.;8@3<=@82;:?=><@/

+>47=;@/ *"6<)?6@4=?5;%9<::?4;@27=.;8@3<=@82;:?=><@,

At low cut-off grades, a linear estimated model frequently # #""# #!!!!"
shows an overestimation of volume or payable ground. This
An assessment was done to compare the grade-tonnage
is referred to as the ‘vanishing tons’ problem as described by
results of panels that are well estimated and did not contain a
David (1977), which is seen when mining commences and
conditional bias (as determined by the slope of regression)
less material is recovered than was predicted. This is caused
against poorly estimated panels. Panels chosen for this
by a conditional bias and/or smoothing in the estimate,
assessment are shown in Figure 16, where values with the
which are reflected respectively by low slopes of regression
better slopes of regression fall on or close to the 1:1
and/or higher estimation variance in the estimated result.
regression line.
This phenomenon is amplified by a high nugget effect and
For the normally distributed data, if the mean panel grade
small block sizes used for estimation.
estimate is correct, the UC accurately predicts the distribution
In order to resolve a conditional bias, one can estimate
of grades and tonnage (Figure 17). For the lognormally
grades into larger blocks. However, estimating into larger
blocks can produce an over-smoothed histogram, or too much distributed data, UC predicts the grade-tonnage relationship
average material, and does not provide the accuracy required (Figure 19 and Figure 20) fairly well. For both distributions,
to select blocks for mining. This is the ‘kriging oxymoron’ if the mean panel grade is wrong, the distribution of SMU
(Isaaks, 2004), which states that a kriged estimate cannot be grades will not necessarily match the simulated distribution
conditionally unbiased and accurate at the same time. UC (Figure 18 and Figure 21). It appears that, in addition to an
uses the ‘conditionally unbiased’ large block estimator to unbiased panel estimate, UC performs slightly better on an
condition the average of a distribution of small blocks, individual panel basis when the underlying distribution is
thereby maintaining the correct grade-tonnage curves and normal.
applying a conditional distribution to obtain an accurate
histogram of small block (SMU) grades. This attempts to ##" #"
" #!!!!"
satisfy the apparent contradiction embodied in the kriging The localization of the UC result places individual SMU
oxymoron. grades (derived from the SMU grade-tonnage curve within
L


   
VOLUME 116  
 651
When should uniform conditioning be applied?

+>47=;@/*297?6@!! 
@0?:;6@82?99;=@06<9#@8-<$>:4@)6<28@3<=@ @?:?68>8@3<=@82;:?=><@/@(6;39'@?:5@82;:?=><@,@(=>4-9'

+>47=;@/*"=?5;%9<::?4;@27=.;@3<=@0?:;6@/,,@(:<=1?6@5>89=>)79><:@82;:?=><@/'

+>47=;@/*"=?5;%9<::?4;@27=.;@3<=@0?:;6@/@(:<=1?6@5>89=>)79><:@82;:?=><@/'

the panel) at specific locations within the panel, based on the UC in preference to linear estimates. If the OK SMU estimate
estimated grades of the OK SMU model. The success of the provides a good spatial representation of the local grades,
localization is verified by visual comparison and by statis- then the location of the UC grades within the panel will be
tically comparing the rank values of the actual versus the UC more accurate. This confirms Abzalov’s (2006) findings that
ranking (Figure 22). the localization success is dependent on available data
The success of localization depends entirely on the (among other factors).
reliability of the OK SMU estimate. However, the smoothing If the data is closely spaced enough to provide accurate
and inaccuracy of this estimate is the prime motivation to use localization, then it is also likely that the data is sufficiently
L

652  
 VOLUME 116


   

When should uniform conditioning be applied?

+>47=;@/*"=?5;%9<::?4;@27=.;@3<=@0?:;6@/@(6<4:<=1?6@5>89=>)79><:@82;:?=><@,'

+>47=;@, *"=?5;%9<::?4;@27=.;@3<=@0?:;6@/,/@(6<4:<=1?6@5>89=>)79><:@82;:?=><@,'

+>47=;@,/*"=?5;%9<::?4;@27=.;@3<=@0?:;6@,@(6<4:<=1?6@5>89=>)79><:@82;:?=><@,'

+>47=;@,,*<10?=>8<:@<3@&! @$>9->:%0?:;6@=?:>:4@3<=@5>33;=;:9@5?9?@5>89=>)79><:8
L


   
VOLUME 116  
 653
When should uniform conditioning be applied?
closely spaced for a linear estimation to accurately predict the of Iron Ore 2007, Perth, Australia. Australasian Institute of Mining and
model grade value. In this circumstance, the benefit of using Metallurgy, Melbourne. pp. 67–77.
a nonlinear UC estimator over a linear estimator is not as
significant as the benefit seen with widely spaced data. This
DERAISME, J. AND ASSIBEY-BONSU, W. 2011. Localised uniform conditioning in the
is evident in the grade-tonnage predictions for the well-
multivariate case - An application to a porphyry copper gold deposit.
estimated data, where the OK and UC results are similar; and
Proceedings of the 35th International Symposium on Application of
the predictions for the poorly estimated data, where the UC
results show a significant improvement over the OK results. Computers and Operations Research in the Minerals Industry (APCOM),

Although LUC is a useful addition to UC, it does not Wollongong, Australia, 24–30 September 2011. Baafi, E.Y., Kininmonth,
improve the accuracy of the UC estimate and the localization R.J., and Porter, I. (eds.). Australasian Institute of Mining and Metallurgy,
algorithm cannot predict the placements of SMU beyond the Melbourne.
available data. This is the main problem: one cannot simulta-
neously know the local mean and the local variability from DERAISME, J., RIVOIRARD, J., and CARRASCO CASTELLI, P. 2008. Multivariate uniform
limited local data. The single largest contribution of the
conditioning and block simulations with discrete gaussian model:
localization approach is to present a UC model in a more
Application to the Chuquicamata deposit, Proceedings of the VIII
accessible and immediately useful format for mine planning.
International Geostatistics Congress (Geostats 2008), Santiago, Chile, 1–5
December 2008. Gecamin, Santiago. pp. 69–78.
<:2678><:8
UC performs well in terms of estimating grades and tonnages HARLEY, M. and ASSIBEY-BONSU, W. 2007. Localised uniform conditioning: How
when there is a normal underlying grade distribution and
good are the local estimates? Proceedings of the 33rd International
good sample coverage relative to the variogram ranges, which
Symposium on Application of Computers and Operations Research in the
result in low conditional biases. In such circumstances a
linear estimator can also closely predict recoverable resources Minerals Industry (APCOM), Santiago, Chile, 24–27 April 2007. Magri,

and provide a spatially representative grade model, although E.J. (ed.). Gecamin, Santiago. pp. 105–112.
the UC estimate of tons and grades is slightly better.
When there is an underlying lognormal distribution and HUMPHREYS, M. 1998. Local recoverable estimation: A case study in uniform
poor sample coverage relative to variogram ranges, conditioning on the Wandoo Project for Boddington Gold Mine.
conditional biases of a linear panel estimate will occur. This Proceedings of a One-day Symposium: Beyond Ordinary Kriging, Perth, 30
results in UC providing a more accurate global estimate of
October 1998. Geostatistical Association of Australia. pp. 63–75.
grades and tonnage than a linear estimate. The individual
panel results predict the actual grade and tonnage distri-
ISAAKS, E. 2004. The kriging oxymoron: A conditionally unbiased and accurate
bution when there is no evidence for conditional bias for that
panel. predictor (2nd edn.). Proceedings of Geostatistics Banff 2004.

LUC results are favourable when there is sufficient closely Leuangthong, O. and Deutsch, C.V. (eds). Springer. pp. 363–374.
spaced data, in which case it is likely that a linear estimation
could also accurately predict the model grade values. NEUFELD, C.T. 2005. Guide to recoverable reserves with uniform conditioning.
Therefore, the benefit of using a nonlinear UC estimator over Guidebook Series, volume 4. Centre for Computational Geostatistics (CCG),
a linear estimator is more significant when the data is widely University of Alberta, Canada.
spaced.

MILLAD, M.G. and ZAMMIT, K.M. 2014. Implementation of localised uniform

2:<$6;541;:98 conditioning for recoverable resource estimation at the Kipoi Copper
My thanks to those who reviewed this work, and partic- Project, DRC. Proceedings of the Ninth International Mining Geology
ularly to Michael Harley of Anglo American for his advice. Conference. Australasian Institute of Mining and Metallurgy, Melnourne.
pp. 207-214.
;3;=;:2;8
ABZALOV, M.Z. 2006. Localised uniform conditioning (LUC): a new approach for RIVOIRARD, J. 1994. Introduction to Disjunctive Kriging and Nonlinear

direct block modelling. Mathematical Geology, vol. 38, no. 4. pp. 393–411. Geostatistics. Centre de Geostatistique, Ecole des Mines, Paris, France.

ASSIBEY-BONSU W. 1998. Use of uniform conditioning technique for recoverable SCHOFIELD, N. 1988. Ore reserve estimation at the Enterprise gold mine, Pine

resource/reserve estimation in a gold deposit. Proceedings of Geocongress Creek, Northern Territory, Australia. Part 1: structural and variogram
1998. Geological Society of South Africa, Johannesburg. pp. 68–74. analysis. CIM Bulletin, vol. 81, no. 909. pp. 56-66.

DAVID, M. 1977. Geostatistical Ore Reserve Estimation (4th impression). VERLY, G. 1986. Multigaussian kriging - a complete case study. Proceedings of
Elsevier, New York. pp. 301-320. the 19th International Symposium on Application of Computers and
Operations Research in the Minerals Industry (APCOM), Littleton, CO.
DE-VITRY, C., VANN, J., and ARVIDSON, H. 2007. A guide to selecting the optimal Ramani, R.V. (ed.). Society of Mining Engineers, Littleton, CO.
method of resource estimation for multivariate iron deposits. Proceedings pp. 283–298. N
L

654  
 VOLUME 116