Quantitative Kriging Neighbourhood Analysis for the Mining

Geologist A Description of the Method With Worked Case

Examples
J Vann1, S Jackson2 and O Bertoli3

ABSTRACT The criteria for assessing the quality of kriging given a

Ordinary kriging and non-linear geostatistical estimators are now well
specified kriging neighbourhood (or neighbourhood) are well
accepted methods in mining grade control and mine resource estimation. established. However, outside of the specialist geostatistical
Kriging is also a necessary step in the most commonly used methods of literature (Armstrong, 1998; David, 1977; Rivoirard, 1987;
conditional simulation used in the mining industry. In both kriging and Chiles and Delfiner, 1999), there is little in the published domain
conditional simulation, the search volume or kriging neighbourhood is to describe QKNA or to guide geologists in implementation. In
defined by the user. The definition of this search can have a very this paper an attempt is made to redress this by setting out the
significant impact on the outcome of the kriging estimate or the quality calculations required for QKNA and defining some approaches
of the conditioning of a simulation. In particular, a neighbourhood that is to interpreting the results. Several practical worked mining case
too restrictive can result in serious conditional biases. The methodology
for quantitatively assessing the suitability of a kriging neighbourhood examples are also given. Finally, some comments are made on
involves some simple tests (which we call Quantified Kriging using the results of QKNA to assist with block size selection,
Neighbourhood Analysis or QKNA) that are well established in the choice of discretisation and mineral resource classification
geostatistical literature. The authors argue that QKNA is a mandatory decisions.
step in setting up any kriging estimate, including one used for This paper assumes the reader has a basic understanding of
conditioning a simulation. Kriging is commonly described as a linear geostatistics. Armstrong (1998), Chiles and Delfiner
minimum variance estimator but this is only true when the
neighbourhood is properly defined. Arbitrary decisions about searches (1999), Isaaks and Srivastava (1989) or Journel and Huijbregts
are highly risky, because the kriging weights are directly related to the (1978) can be referred to for the required background on
variogram model, data geometry and block/sample support involved in variograms and kriging.
the kriging. The criteria to look at when evaluating a particular kriging
neighbourhood are the following: MOTIVATION
1. the slope of the regression of the true block grade on the
estimated block grade; The motivation for QKNA
2. the weight of the mean for a simple kriging;
Ordinary kriging (OK) and non-linear geostatistical estimators,
3. the distribution of kriging weights themselves (including the including uniform conditioning and multiple indicator kriging,
proportion of negative weights); and
are now widespread and routine methods in mine resource
4. the kriging variance. estimation and grade control. In this paper kriging refers to OK,
Outside of the technical geostatistical literature, there is little in the unless otherwise indicated. Kriging (Matheron, 1962, 1963a,
published domain to describe the nature of QKNA and no practical 1963b; Journel and Huijbregts, 1978) is also a necessary step in
presentation of case examples. In this paper we attempt to redress this by the main methods of conditional simulation used in the mining
setting out the calculations required for QKNA and defining some
industry, eg sequential Gaussian simulation (SGS), turning bands
approaches to interpreting the results. Several practical worked mining
case examples are also given. Finally some comments are made on using (TB) and sequential indicator simulation (SIS). Conditional
the results of QKNA to assist with block size selection, choice of simulation (Journel, 1974; Lantuejoul, 2002) is now being
discretisation and mineral resource classification decisions. utilised by mining geologists in grade control, resource
estimation and risk analysis applications.
INTRODUCTION In both kriging and conditional simulation, the neighbourhood
is defined by the user (or at least it should be: in some cases a
This paper presents the methodology for quantitatively assessing black box approach may involve accepting default parameters).
the suitability of a kriging neighbourhood: ie the combination of Arbitrary specification of the neighbourhood is very risky
the search strategy and block definition used in a kriging. In this because the kriging weights are directly related to the variogram
paper kriging refers to ordinary kriging (OK), unless otherwise model, data geometry and block/sample support involved in the
indicated and the process of assessing a kriging neighbourhood kriging. Whilst kriging is commonly and correctly described as a
(for any kind of kriging) is referred to as Quantified Kriging minimum variance estimator this is only true when the
Neighbourhood Analysis or QKNA. The authors argue that neighbourhood is properly defined. This necessitates an objective
QKNA is a mandatory step in setting up any kriging estimate,
method to assess what constitutes an appropriate
including one used for conditioning a simulation.
neighbourhood.

1. FAusIMM, Principal Geologist Geostatistician, Quantitative Misconceptions about kriging searches

Geoscience Pty Ltd, PO Box 1304, Fremantle WA 6959. E-mail:
jv@quantitativegeoscience.com There is a widely held misconception that searching to the range
2. MAusIMM, Principal Geologist Geostatistician, Quantitative of the variogram is a good strategy for defining the
Geoscience Pty Ltd, PO Box 1304, Fremantle WA 6959. E-mail: neighbourhood. The choice of neighbourhood should be
sj@quantitativegeoscience.com influenced more by the slope of the variogram model at short
lags and the relative nugget effect (ie the ratio of the nugget
3. MAusIMM, Principal Mining Engineer Geostatistician, Quantitative
Geoscience Pty Ltd, PO Box 1304, Fremantle WA 6959. E-mail: variance to the total variance, expressed as a percentage) than by
ob@quantitativegeoscience.com the ranges per se.

5th International Mining Geology Conference Bendigo, Vic, 17 - 19 November 2003 215
J VANN, S JACKSON, and O BERTOLI

In fact, as the range of a variogram approaches zero (pure resource estimation (Krige, 1951), and marked the beginning of
nugget effect) it can be shown that the neighbourhood required modern approaches to resource estimation, leading to
for good estimation will progressively get larger. In the case of geostatistics. The interpolation method called kriging is simply a
pure nugget, correlation between any two points in a domain is linear regression solution to the grade interpolation problem.
zero. Therefore, samples located within any limited search The awkward phrase conditional unbiasedness has been used
neighbourhood will be uncorrelated to the true grade of the historically by geostatisticians to describe the following property
block! In other words, local estimation is risky and will be of kriging: blocks estimated to have a certain grade Z*V will, on
increasingly riskier as we define progressively smaller average, have that grade (David, 1977).
neighbourhoods. In the case of pure nugget, the most reliable Figure 1 shows the well-known regression between estimated
estimate will be made with the largest number of samples. In fact block grades Z*V and true block grades ZV (see Journel and
in this case searching the whole domain will be the minimum Huijbregts, 1978 for more detailed discussion). The correlation
estimation variance solution. between true and estimated grades is always imperfect in any
On the other hand, when the relative nugget effect approximates practical situation, ie any non-exhaustive drill sampling, and the
zero and the range is very long relative to the block dimensions, regression line will usually be flatter than Y = X (ZV = Z*V).
the closest samples are highly correlated to the true block grade. This implies that using any data honouring estimator, like
Therefore, only nearby samples will be required to ensure that a nearest neighbour (or for that matter inverse distance estimators,
good estimate is made when kriging and relatively restricted which as the power increases, approximate nearest neighbour),
searching can produce a minimum estimation variance solution. will result in conditional biases, specifically:
When no QKNA is performed, choice of search is sometimes
1. estimated grades Z*V greater than the mean grade will, on
made on the basis of comparing the output results from several
average, be too high; and
estimations with different searches. In the authors experience the
danger in this approach is that the most financially attractive 2. estimated grades Z*V less than the mean grade will, on
result will often be selected. Since the most financially attractive average, be too low.
estimate is generally a product of the most restricted search, the The lack of perfect or exhaustive information implies that
risk is that the most conditionally biased result tends to be the correlation between estimates and true block grades will be
selected. imperfect. This in turn implies overstatement of high grade and
understatement of low grades by local estimators, on average. As
Consequences of overly-restricted the spatial correlation of grades deteriorates (ie high relative
neighbourhoods: conditional bias nugget and/or presence of a short range structure) the slope of
the regression gets progressively less than one. Because it is
The neighbourhood we choose has a very significant impact on important (for example in any material classification) that an
the outcome of the kriging estimate. In particular, a estimator results in the correct average grade being estimated for
neighbourhood that is too restrictive will result in serious blocks in various grade classes, smoothing is required to ensure
conditional biases (Krige, 1994, 1996a, 1996b). Understanding that the regression slope is as close as practicable to one. Note:
that conditional biases can be eliminated by a regression the four quadrants marked on Figure 1 denote the correct and
approach was the primary original contribution of D G Krige to incorrect classification of material for a given cut-off grade.

= *

FIG 1 - Information effect diagram showing linear regression. Quadrants II and IV are correct classification of waste and ore, respectively;
while quadrants I and II are corresponding incorrect classifications.

216 Bendigo, Vic, 17 - 19 November 2003 5th International Mining Geology Conference
 QUANTITATIVE KRIGING NEIGHBOURHOOD ANALYSIS FOR THE MINING GEOLOGIST

In looking at Figure 1, it is obvious that in practice we never The weight of the mean in a simple kriging is not so easy to
know the true grades of blocks. However, we can infer the obtain because it is not offered as an option in some mine
relationship between true block values and estimated block planning software.
values under certain assumptions. A critical assumption is that
the variogram is modelled reliably and represents the domain of Slope of the regression
interest adequately, ie the assumption of intrinsic stationarity (see
Journel and Huijbregts, 1978, Chapter II) is reasonable. The Under the assumptions stated previously (that the variogram is
assumption of intrinsic stationarity enables the calculation of valid and the regression is linear), it is possible to calculate the
covariances, thus correlations, between any specified supports, main parameters of the regression between estimated and true
for example between true and estimated blocks. We must also block grades. We repeat that the actual scatter plot (as illustrated
assume that a linear regression can capture the essence of the and summarised by the ellipse in Figure 1) cannot be plotted
relationship between true and estimated block grades. because we dont know individual true block grades! What can
Under the assumptions stated previously (the variogram is be calculated is the covariance (thus correlation coefficient)
valid and the regression is linear), it is possible to calculate the between estimated and true block grades. The slope is given in
main parameters of the regression between estimated and true terms of this covariance and the variance of the estimated blocks
block grades, given a certain variogram model, informing data by the expression:
set, block size and neighbourhood. Note that the actual scatter
plot (as illustrated and summarised by the ellipse in Figure 1) Cov( Zv , Zv* )
a=
cannot be plotted because we dont know individual true block Var( Zv* )
grades! What can be known, however, are the slope of the
regression and the covariance (thus correlation coefficient) where:
between estimated and true block grades. a is the slope of the regression
The process of QKNA is, in essence, one of adjusting the
neighbourhood to arrive at good regression statistics in order to Zv is the true block grade
reduce or eliminate conditional bias. It is important to appreciate Zv* is the estimated block grade
that this will necessitate smoothing. This necessity to smooth is a
consequence of the information effect (Journel and Huijbregts, Note that the value of a is often directly given by kriging
1978) which implies that the estimated block values will have a programs. However, we include in a technical appendix more
lower variance than the true block values. The only way to obtain details on the derivation of a for interested readers.
block estimates that are equivalent to the real (true) block Ideally, the slope of the regression a should be very close to 1.0
grades is to base the estimate on exhaustive data: ie mine the and thus imply conditional unbiasedness. In these circumstances,
orebody out with a drill rig. In all realistic cases, some the true grade of a set of blocks should be approximately equal to
smoothing is necessary and the task is to minimise conditional the grade predicted by the kriged estimation. The slope and its
bias. In other words: smoothing is the price we must pay for interpretation are discussed more fully by Krige (1994; 1996a) and
non-exhaustive information. Rivoirard (1987).
The question is: how much do we smooth to ensure The slope should be 1.0 for conditional unbiasedness. A
conditional unbiasedness? The answer to this question requires rewriting of the expression for the slope in terms of correlation
analysis (QKNA), not guesswork. coefficient is possible:
Z V
EVALUATION CRITERIA a =
Z *
V

The criteria to look at when evaluating a kriging

where:
The objective of QKNA is to determine the combination of
search neighbourhood and block size that will result in a is the slope of the linear regression
conditional unbiasedness. The criteria to consider when is the linear (Pearson) correlation coefficient
evaluating a particular kriging neighbourhood, in the order of
priority used in practice by the authors, are: Z V is the standard deviation of true block grades

1. the slope of the regression of the true block grade on the * is the standard deviation of estimated block grades
V

estimated block grade; From the above expression we can see that even for slope
2. the weight of the mean for a simple kriging; equalling one, the correlation may be less than one (because the
smoothing effect of kriging necessitates that the variability of
3. the distribution of kriging weights themselves (including estimates is lower than that of true blocks). We discuss the
the proportion of negative weights); and interpretation of slope further in the section below on resource
4. the kriging variance. classification, but note here that no statistical criteria should ever
be the sole basis of resource categorisation.
Calculation of required parameters
Weight of the mean for simple kriging
Any useful kriging program will provide kriging variance for
each estimated block. The most important criterion, the slope of Instead of performing an ordinary kriging, where the sum of the
the regression, can usually be obtained from kriging programs. It weights is set to one, we can run simple kriging (SK) where the
may be necessary to use a debug option, where blocks are sum of the weights is not constrained to add up to one. The
kriged individually and the weights and various other statistics remaining weight is allocated to the mean grade of the domain
are reported to an output file. In some commercial mining (the weight of the mean) and is an inversely proportional index
software, the output requires further post-processing to obtain the of screen effect. A sample is said to be screened if another
slope of regression. sample lies between it and the point being estimated, in which
If the slope of the regression cannot be obtained in your case the estimation weight of the screened sample is reduced. If
software, we recommend that you liaise with the developers to the variogram indicates high continuity, screen effect will be
have this simple enhancement made available. pronounced; conversely, high nugget effect (or significant short

5th International Mining Geology Conference Bendigo, Vic, 17 - 19 November 2003 217
J VANN, S JACKSON, and O BERTOLI

scale structure) implies that weights will be spread far away from ability to specify a maximum number of data to used in an
a block to reduce conditional bias and minimise estimation estimate (Nmax) allows the search to be relaxed
variance. automatically.
SK is also called kriging with known mean, and is based on the Another practical implementation question of interest is the
assumption that the global mean grade is known and equal to m: use of quadrant or octant searches. The kriging estimate does
perform a degree of declustering, but such strategies are often
ZV* SK m = SK
i ( Z( x i ) m) still very important.
where:
APPLICATIONS
m = 1 SK
i is the weight assigned to the global mean grade
or weight of the mean
Application of QKNA to kriging estimation
The weight of the mean for a given neighbourhood, denoted
m, gives a clear idea of the quality of kriging because it is a The optimal solution would be to tune the neighbourhood for
measure of the weakness of the screen effect. The larger m is, every block estimated but this is obviously not practical.
the weaker we expect the screen effect to be. Consequently, all Therefore the objective is to find best compromise solution: a
things being equal, it is better to choose a larger kriging neighbourhood is defined for a domain that will be adequate for
neighbourhood as m increases, (Rivoirard, 1987). As a general as many blocks as possible. If differences between optimal
rule, we prefer the weight of the mean to be close to zero. The neighbourhoods and this compromise neighbourhood are too
objective in QKNA is to obtain the combination of the best slope big (as can happen when data density is variable across a
with a minimised weight of the mean. domain) then there may be an argument to further divide
Note that the use of SK here is solely for QKNA and that, in estimation area and perform multiple runs.
general, the stationarity assumptions of SK are not suited to The kriging weights depend upon the data in the sense that the
mining grade estimation. It can be shown that OK is exactly the variogram model we choose is intimately linked to the histogram
same as SK when m is replaced by its kriged estimation and spatial continuity of the samples themselves; however, the
(Armstrong, 1998). kriging equations contain no direct reference to the data values
themselves. This means the set of weights obtained for a given
Kriging weights (and negative weights) sampling/block geometry and a specified variogram model will
always be the same, regardless of the sample grades. Because of
If we can expand the neighbourhood and still assign a this property, we need only test a range of typical data
meaningful positive weight to the incremental samples thus configurations in order to determine the optimal search (we dont
obtained, then the search is too restrictive. At the margins of an have to test every block). The authors recommend testing a range
optimised search, kriging weights should be very small of blocks as follows:
(trivial), or even slightly negative. 1. well informed blocks (amidst plenty of sampling
Most mining grade variables are certainly not spatially smooth, information, especially with samples internal to the block);
ie there is at least some nugget effect. In these circumstances, a
screen effect can be expected and at some distance negative 2. less well informed blocks but still with data surrounding the
weights will be observed. The distance we need to search before block; and
negative weights are encountered progressively increases as the 3. poorly informed blocks, including:
effective nugget effect increases. In the case of pure nugget
every sample found gets equal weight (1/N) no matter how far
blocks with no internal samples; and
we search. Negative weights are not problematic if they represent blocks with no samples in certain directions (ie blocks
a small proportion (say <5 per cent) of total weight. at the edges of domains).
The authors advise against modified kriging algorithms that
adjust negative weights (eg Deutsch, 1996) or set them to zero, Block size determination
since such approaches will assure conditional bias. It is important to understand that the block size is critical in all
Kriging variance cases where a cut-off will be applied to an estimate (ie unless the
The minimum variance solution to the ordinary kriging estimate is to be used only in a global sense and without a cut-off).
equations results in an estimation variance, also called the There is a long bibliography of warnings against estimation of
kriging variance (KV). Refer to the appendix for the expression small blocks (for example Armstrong and Champigny, 1989;
for KV. Ravenscroft and Armstrong, 1990; Royle, 1979; Vann and Guibal,
A mapping of KV gives an idea of the relative quality of 2000). The question is, how small is too small?
estimation (mainly in terms of data density and geometry), but By running QKNA for a range of block sizes, in relatively
maps of regression slope may be more useful. well- and poorly-informed instances, quantitative determination
of appropriate block sizes is possible. The results of such an
Determining minimum and maximum numbers of analysis will usually show that the slope of the regression and
data weight of the mean rapidly become unacceptable as the block
size reduces, except for those blocks actually containing samples.
The minimum number of data used in an estimate (Nmin) can As a general summary, the block size needs to increase as the
be studied as a variable in a QKNA exercise. However, as a rule nugget (and other short-scale discontinuities) increases. It is
of thumb using less than ten - 12 samples is not recommended, unusual for blocks appreciably smaller than half the drilling grid
especially in the presence of any appreciable short-scale structure dimensions to yield acceptable QKNA results, unless the grade
or nugget effect. The default value for Nmin in software systems continuity is very high (ie very low nugget and long ranges).
may be as low as 1 or 2: this is effectively a nearest a neighbour
interpolation and in most cases this cannot be technically
defended.
Discretising the block in kriging
There are cases where the search will be set larger than strictly The discretisation in block kriging is used to calculate the
necessary to ensure that an adequate number of data are used in point-block average values of variogram (or covariance)
less sampled areas of a domain (eg at the edges). In these cases, functions, ie ( x i , V ) or C ( x i , V ). The general process for

218 Bendigo, Vic, 17 - 19 November 2003 5th International Mining Geology Conference
 QUANTITATIVE KRIGING NEIGHBOURHOOD ANALYSIS FOR THE MINING GEOLOGIST

determining discretisation of a block involves iterative

10000.00
calculation of (V , V ) or equivalently C (V , V ) with a range of
discretisations. Ideally, we calculate for each discretisation
1000.00
several times, moving the origin of the discretisation grid each
time. Stable results indicate that the discretisation is adequate.
100.00
The number of discretised points should be compatible with the
dimension of the block in units of composite-length in the 10.00
direction approximately parallel to the drilling, ie when a
composite can no longer be reasonably considered as a point in 1.00
space but rather a regularised variable (its dimension is no

2x2x2

3x3x2

4x4x2

5x5x2

6x6x2

7x7x2

8x8x2

9x9x2

10 x 10 x 2

11 x 11 x 2

12 x 12 x 2
longer 0 at the scale of the block). For example, in an open pit
situation with steeply inclined drilling, a block 10 m high would
be discretised to 2 in the Z (vertical) when using 5 m composites,
but to 5 in the vertical when using 2 m composites or 1 in the
vertical for 10 m composites. FIG 2 - Visualisation of example of sensitivity of C (V ,V ) to
discretisation. Note that the variance for results, plotted on the
In general, higher discretisations are better, with the only
Y-axis does not show significant continuing downward trend
disadvantage being some computing speed consequences. Note
beyond about 6 6 2. In this case 6 6 2 or 7 7 2
that the speed penalty for higher discretisation when kriging is
would be acceptable.
not as severe as it is in IDW interpolation, where discretisation to
n points requires n estimates. In kriging the discretisation is used
solely for the calculation of C (V , V ) and C ( x i , V ). substantially by the adequacy of the neighbourhood used. The
Table 1 shows results of a sensitivity analysis of discretisation authors note that in more automated approaches to simulation
on C (V , V ) (for the same variogram and block size used in case sometimes employed in grade control the definition of the
Study A, later). Figure 2 shows a graph of the impact of neighbourhood may not be transparent to the user.
discretisation on the resulting range of C (V , V ) values.
Application of QKNA to multivariate situations
Application of QKNA to conditional simulation
QKNA can be applied to the multivariate situation as well. It
Conditional simulation is widely applied by mining geologists requires a consistent model of co-regionalisation to allow for the
for grade control, resource estimation and risk analysis. Kriging modelling of spatial correlation intra-variables (say gold with
is also a necessary step in the methods of conditional simulation gold, copper with copper) and inter-variables (say gold with
most commonly used in the mining industry: sequential Gaussian copper). The assumption of stationarity inherent in the use of
simulation (SGS) and turning bands (TB) and sequential such a model is more restrictive because it needs to apply to
indicator simulation (SIS). Refer to Chiles and Delfiner (1999), domains that ought to display an acceptable level of statistical
Goovaerts (1997) and Lantuejoul (2002) for details on simulation homogeneity for all variables and their spatial correlations.
algorithms. Depending on the nature of the spatial relationship between the
In sequential algorithms, such as SGS and SIS, the kriging step variables under consideration, different models of correlation
is at the core of the algorithm and conditioning of the simulation may be used (intrinsic correlation, linear model of
is by construction. In TB a non-conditional simulation is co-regionalisation, complex model of co-regionalisation). For
performed, followed by an explicit conditioning step by kriging. more details on multivariate applications, refer to Wackernagel
In both cases, the quality of the kriging is determined (1995) and Wackernagel and Grzebyk (1994, 1995).

TABLE 1
Example of sensitivity ofC (V ,V ) to discretisation. NB. same variogram model for Case Study A: for each discretisation,
13 calculations ofC (V ,V ) were made with randomly changed grid origin.

Discretisation 2 2 2 332 442 552 662 772 882 9 9 2 10 10 2 11 11 2 12 12 2

Cvv test1 1.31 1.26 1.24 1.22 1.23 1.24 1.23 1.23 1.23 1.24 1.23
Cvv test2 1.35 1.23 1.29 1.27 1.24 1.23 1.23 1.23 1.23 1.24 1.23
Cvv test3 1.22 1.26 1.24 1.25 1.24 1.25 1.24 1.23 1.24 1.23 1.24
Cvv test4 1.11 1.25 1.27 1.22 1.22 1.23 1.23 1.23 1.23 1.23 1.23
Cvv test5 1.28 1.26 1.28 1.24 1.25 1.25 1.23 1.23 1.23 1.24 1.24
Cvv test6 1.48 1.24 1.26 1.23 1.25 1.23 1.23 1.23 1.23 1.23 1.24
Cvv test7 1.21 1.30 1.24 1.25 1.23 1.23 1.24 1.24 1.23 1.23 1.24
Cvv test8 1.27 1.20 1.27 1.23 1.23 1.23 1.24 1.23 1.24 1.24 1.24
Cvv test9 1.17 1.25 1.26 1.23 1.23 1.24 1.22 1.23 1.23 1.23 1.23
Cvv test10 1.41 1.31 1.26 1.24 1.25 1.24 1.23 1.23 1.24 1.24 1.23
Cvv test11 1.29 1.28 1.22 1.24 1.25 1.23 1.24 1.23 1.23 1.23 1.23
Cvv test12 1.28 1.30 1.26 1.23 1.24 1.24 1.23 1.24 1.23 1.24 1.23
Cvv test13 1.37 1.25 1.27 1.26 1.23 1.24 1.23 1.23 1.24 1.23 1.24
Variance 9807.10 976.01 363.44 170.93 59.07 44.95 34.52 9.65 15.70 5.35 7.74

5th International Mining Geology Conference Bendigo, Vic, 17 - 19 November 2003 219
J VANN, S JACKSON, and O BERTOLI

Some further comments on resource

classification
One of the factors which can impact on resource classification
that is often not adequately considered is the scale of 200.
classification, ie the volume of mineralisation that is being
0.29
classified (Stephenson and Vann, 2000). It is unwise, and 0.97 0.95
2.50
generally unnecessary, to classify a Resource or Reserve estimate 4.70 0.93
0.53 7.12
at the scale of a block (especially when small blocks have been 0.59 9.03
1.03

Z (m)
used relative to the sample spacing). The classification process is 180. 1.74 9.55
2.50 8.43
usually best approached at the scale of domains or significant 2.93 6.24 0.56
3.87 0.22
subsets of domains (Domain A above a certain RL, for example). 2.78 1.97 0.18
2.15 0.86 0.47
However, the various statistics resulting from QKNA are very 1.30
useful as inputs to resource classification along with geology, 0.48
-0.07
appraisal of drill spacing and so on. -0.15
160.
Associations of slope and estimation quality statistics with the
Joint Ore Reserves Committee Code classification of Mineral
Resources into Measured, Indicated and Inferred (JORC, 1999)
49820. 49840. 49860. 49880.
require expert consideration of Competent Persons
knowledgeable of the geological, geostatistical and other X (m)
Isatis
specifics of the deposit at hand. In particular, output of QKNA
are complementary to other inputs to classification (data quality, FIG 3 - Cross-sectional view of the estimated block, sample
geology confidence, etc).
locations, kriging weights and search ellipse selected for the
The reader is also referred to the JORC Code (JORC, 1999) well-informed case in Case Study A. Note, some samples inside
and Stephenson and Vann (2000) for more guidance. the ellipse do not have weights because the maximum number of
samples criterion has been met.
CASE STUDIES
We present here results of two short case studies to illustrate the
outcome of QKNA. In each case a table is given showing the TABLE 3
results of QKNA tests for three situations in a deposit: QKNA statistics for Case Study A.
1. a well informed block, ie one with information on all Test result Well Reasonably Poorly
sides and samples in (or very near to) the estimated block, informed informed informed
in other words, an ideal case;
Slope of regression > 0.93 < 0.8 < 0.6
2. a reasonably informed block, ie one with less information Correlation coefficient < 0.9 < 0.7 < 0.5
than a well informed block but still moderately well
Weight of the mean < 15% > 25% > 50%
informed; and
Cumulative sum of <0.5% 0% 0%
3. a poorly informed block, ie one with less than adequate negative weights
information, for example no samples in or close to the
Kriging standard deviation ~ 0.5 ~ 0.7 ~ 0.9
block, or a lack of samples on one side, etc.
The definition of these three cases usually requires testing of a
range of possibilities, ie a number of well informed situations, Archaean gold deposit: a well informed, reasonably informed
etc. Results for typical cases are given in our examples. Note and poorly informed block.
that the case studies could also be presented to show how the The results in Table 3 show that a well informed block has a
slope of regression improves by varying the size of the search
low weight of the mean (13 per cent) and a slope of regression of
from a restricted one to a reasonable one, but space
considerations preclude this here. 0.93, indicating that the estimation would not be much improved
by searching further. Consequently, a trivial amount of negative
Case A gold grade estimation in a gold deposit weight is allocated, and this is not a problem at all! The kriging
standard deviation (ie square root of the kriging variance) is
(3D)
significantly lower than for the other two cases. Also, the
Table 2 shows the variogram model for Case Study A and Figure correlation coefficient indicates that we can expect quite good
3 shows a cross sectional view of a well informed block. Table 3 (but not perfect) reconciliation between estimated and true block
shows the results of QKNA tests for three situations in an grades.

TABLE 2
Variogram model for Case Study A.

Nugget Structure 1 Structure 2 Structure 3

(Spherical) (Spherical) (Spherical)
Variance of structure 2.03 1.67 0.07 0.18
Proportion of total sill 51% 42% 2% 5%
Range D1 (m) (Az 330, Dip 0) 22 28 16
Range D2 (m) (Az 240, Dip 60) 24 70 16
Range D3 (m) (Az 060, Dip 30) 200 70 100

220 Bendigo, Vic, 17 - 19 November 2003 5th International Mining Geology Conference
 QUANTITATIVE KRIGING NEIGHBOURHOOD ANALYSIS FOR THE MINING GEOLOGIST

The poorly informed block, on the other hand, is only

TABLE 4
marginally acceptable for Inferred Resources, in the opinion of
Variogram model for Case Study B.
the authors, because the slope is now close to 0.5 and the weight
of the mean is significantly higher than for a well informed Nugget Structure 1 (Spherical)
block, at 50 per cent of total weight. The correlation expected
Variance of structure 25 10
between estimated and true blocks is now quite poor (less than
0.5). The problem here is that to improve these statistics, we Proportion of total sill 71% 29%
would need to search much further, making the estimate almost Isotropic range (m) 750
global in nature and thus of little local utility. Areas comprised
of blocks with statistics worse than these should not really be
classified as Resources, in the JORC sense, if any degree of
selective mining is envisaged.

Case B ash content estimation for a coal

deposit (2D)
7247500. 0.60
0.46
0.430.50
Estimation in two dimensions is appropriate for deposits having a 0.51
0.37
0.350.60
layer or vein-like geometry, eg narrow veins and coal seams (see 0.52
0.410.45
Bertoli et al, 2003, for further details on 2D estimation). This 0.51
0.44
0.46
0.50 0.71
case study is for ash content in a coal deposit. The ash content 0.68
0.51
0.48
estimate is used in a decision to send or by-pass a wash plant 0.16
-0.01
facility, and is thus the basis of material classification. 2.32
2.66 1.74

Y (m)
7.31 10.35
13.117.41
Table 4 shows the experimental variogram model for ash 7245000. 10.04
1 1.30
9.94
accumulation and Figure 4 shows a plan view of a well informed 1.801.99
2.00
1.30
0.11
-0.22
block. The key characteristics of the model fitted to this 0.09
-0.00 0.37
0.38
0.43
variogram are: a high nugget effect (~70 per cent) and a spatial 0.49
0.62
0.48
0.45
0.42
structure that is relatively short (500 m) given the data spacing 0.73
0.73
(250 250 m).
As a consequence of the poor spatial continuity of ash 7242500.
(evidenced by the variogram model) the weight of the mean
tends to be elevated. This motivates searching progressively
further away in order to weight a larger number of samples with
the objective of reducing conditional bias. 198500. 200500.
Table 5 shows the results of QKNA tests for ash accumulation X (m)
(ash content density thickness) for three situations in: a well Isatis
informed, a reasonably informed (ie surrounded by information
but no central sample) and a poorly informed block. Each block FIG 4 - Plan view of the estimated block, sample locations, kriging
has dimensions of 125 250 m (with the long dimension along weights and search ellipse selected for the well-informed case in
strike). Case Study B. Note, some samples inside the ellipse do not have
The results in Table 5 were obtained using a search radius of weights because the maximum number of samples criteria has
1500 m down dip and 3500 m along the strike of the coal seam been met.
with up to 80 samples for the best informed blocks. The best
informed blocks are well estimated (slope of 0.9, correlation on
average of 0.70 and up to 0.80) but as soon as the density of TABLE 5
information decreases (eg when the block does not contain any QKNA statistics for Case Study B.
samples) the quality of estimation deteriorates rapidly. The
conclusion is that the ability to produce reliable local estimates is Test result Well Reasonably Poorly
seriously compromised in such situations. Any attempt to informed informed informed
estimate smaller blocks or to use smaller neighbourhoods (in a Slope of regression ~ 0.90 < 0.70 < 0.30
data honouring strategy) will have costly consequences in terms
of material misclassification. Correlation coefficient < 0.70 < 0.50 < 0.20
Weight of the mean > 30% > 50% > 80%
CONCLUSIONS Cumulative sum of < 1% 0% 0%
negative weights
Kriging is commonly described as a minimum variance
Kriging standard deviation ~ 2.0 ~ 2.5 ~ 3.0
estimator but this is only true when the block size and
neighbourhood are properly defined. The methodology for
quantitatively assessing the suitability of a kriging
neighbourhood QKNA involves some straightforward tests unrealistic expectations of the grade above an elevated cut-off
that are described in detail in this paper. These tests can be grade may be raised by estimates based on overly restricted
implemented in most good mine planning software systems. The searching.
definition of the search in both kriging and conditional The authors reiterate that QKNA is, in our view, a mandatory
simulation can have a very significant impact on the outcome of step in setting up any kriging estimate, including one used for
the kriging estimate or the quality of the conditioning of a conditioning a simulation. In the end, the results of a kriged
simulation. In particular, a neighbourhood that is too restrictive estimate or conditional simulation can be very sensitive to the
can result in serious conditional biases. QKNA at a relatively neighbourhood and a quantitative approach to selection of the
early stage (wide spaced drilling) can be very important because neighbourhood is thus essential.

5th International Mining Geology Conference Bendigo, Vic, 17 - 19 November 2003 221
J VANN, S JACKSON, and O BERTOLI

ACKNOWLEDGEMENTS Stephenson, P R and Vann, J, 2000. Commonsense and good

communication in mineral resource and ore reserve estimation, in
The authors would like to acknowledge the improvement made Mineral Resource and Ore Reserve Estimation The AusIMM Guide
possible by the helpful comments on an early version of this to Good Practice, Monograph 23 (Ed: A C Edwards) pp 13-20 (The
paper by Rick Adams and Ted Coupland of Cube Consulting, Australasian Institute of Mining and Metallurgy: Melbourne).
Phil Jankowski of Sons of Gwalia, Mike Stewart of MPI Mines Vann, J and Guibal, D, 2000. Beyond ordinary kriging an overview of
and Aaron Tomsett of Quantitative Geoscience. The comments non-linear estimation, in Mineral Resource and Ore Reserve
Estimation The AusIMM Guide to Good Practice, Monograph 23
of two anonymous reviewers were appreciated, leading to useful (Ed: A C Edwards) pp 249-256 (The Australasian Institute of Mining
improvement of the arguments presented. and Metallurgy: Melbourne).
Wackernagel, H, 1995. Multivariate Geostatistics An Introduction With
REFERENCES Applications (Springer: Berlin).
Wackernagel, H and Grzebyk, M, 1994. Multivariate analysis and
Armstrong, M, 1998. Basic Linear Geostatistics, pp 94 - 95; pp 96 - 97 spatial/temporal scales: real and complex models, in Proceedings
(Springer: Berlin). XVIIth International Biometric Conference, Hamilton, Ontario,
Armstrong, M and Champigny, N, 1989. A study on kriging small Canada, Volume 1, pp 19-33.
blocks, CIM Bulletin, 82:128-133. Wackernagel, H and Grzebyk, M, 1995. Linear models for spatial or
Bertoli, O, Job, M, Vann, J and Dunham, S, 2003. Two-dimensional temporal multivariate data, in Proceedings Sixth International
geostatistical methods theory, practice and a case study from the Meeting on Statistical Climatology, Galway, Ireland, pp 427-429.
1A shoot nickel deposit, Leinster, Western Australia, in Proceedings
Fifth International Mining Geology Conference, pp 189-195 (The
Australasian Institute of Mining and Metallurgy: Melbourne). APPENDIX
Chiles, J P and Delfiner, P, 1999. Geostatistics: Modelling Spatial The slope of the regression a is given in terms of the covariance
Uncertainty (Wiley Inter-Science: New York).
between estimated and true block grades and the variance of the
David, M, 1977. Geostatistical Ore Reserve Estimation (Developments in
Geomathematics 2), (Elsevier: Amsterdam).
estimated blocks:
Deutsch, C V, 1996. Correcting for negative weights in ordinary kriging, Cov( Zv , Zv* )
Computers and Geosciences, 22:765-773. a=
Var( Zv* )
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
(Oxford University Press: New York). where:
Isaaks, E H and Srivastava, R M, 1989. Applied Geostatistics, (Oxford
University Press: New York). a is the slope of the regression
JORC, 1999. Australasian Code for Reporting of Mineral Resources and Zv is the true block grade
Ore Reserves, The Joint Ore Reserves Committee of The
Australasian Institute of Mining and Metallurgy, Australian Institute Zv* is the estimated block grade
of Geoscientists and Minerals Council of Australia.
The variance of the estimates can be calculated from linear
Journel, A G, 1974. Geostatistics for conditional simulation of ore geostatistics.
bodies, Economic Geology, 69:673-687.
Journel, A G and Huijbregts, Ch J, 1978. Mining Geostatistics, pp
306 - 307 (Academic Press: London). Variance of block estimates
Krige, D G, 1951. A statistical approach to some basic mine valuation Block grade estimates are written:
problems on the Witwatersrand, J Chem Metall Min Soc S Afr,
N
52:119-139.
Krige, D G, 1994. An analysis of some essential basic tenets of
Zv* = i Z( x i )
i=l
geostatistics not always practised in ore valuations, in Proceedings
Regional APCOM: Computer Applications and Operations Research The variance of estimated block grades can be calculated from
in the Minerals Industries, Slovenia, pp 15-18. the following relationship:
Krige, D G, 1996a. A basic perspective on the roles of classical statistics,
N N
data search routines, conditional biases and information and
smoothing effects in ore block valuations, in Proceedings Var( ZV* ) = i jC ( x i , x j )
i=1 j=1
Conference on Mining Geostatistics, Kruger National Park, South
Africa, pp 1-10 (Geostatistical Association of South Africa).
where:
Krige, D G, 1996b. A practical analysis of the effects of spatial structure
and data available and accessed, on conditional biases in ordinary C(xi,xj) is the value of the covariance function between sample
kriging, in Proceedings Fifth International Geostatistical Congress, location xi and sample location xj and the kriging weights
Geostatistics Wollongong 96 (Eds: E Y Baafi and N A Schofield) associated to these samples are denoted i and j.
pp 799-810.
Lantuejoul, C, 2002. Geostatistical Simulation: Models and Algorithms,
(Springer: New York). Covariance between true and estimated blocks
Matheron, G, 1962. Traite de geostatistique applique, tome I, Memoires The covariance between true and estimated blocks Cov( Zv , Zv* )
du Bureau de Recherches Geologiques et Minieres, No. 14, (Editions can be calculated (see for example, Armstrong, 1998):
Technip: Paris).
N
Matheron, G, 1963a. Traite de geostatistique applique, tome II: Le Cov( ZV , ZV* ) = iC ( x i , V )
krigeage, Memoires du Bureau de Recherches Geologiques et
i=1
Minieres, No. 24. (Editions Technip: Paris).
Matheron, G, 1963b. Principles of geostatistics, Economic Geology, where:
58:1246-1266.
Ravenscroft, P J and Armstrong, M, 1990. Kriging of block models the
C ( x i , V ) is the mean value of the covariance function between
dangers re-emphasised, in Proceedings APCOM XXII, Berlin, each sample location xi and the block to be estimated V.
17 - 21 September 1990, pp 577-587. Note that given the kriging equations which state for any I:
Rivoirard, J, 1987. Two key parameters when choosing the kriging
neighbourhood, J Math Geol, 19:851-856. N
Royle, A G, 1979. Estimating small blocks of ore, how to do it with
confidence, World Mining, April.
Cov( x , x
j=1
j i j ) = C (xi,V )

222 Bendigo, Vic, 17 - 19 November 2003 5th International Mining Geology Conference
 QUANTITATIVE KRIGING NEIGHBOURHOOD ANALYSIS FOR THE MINING GEOLOGIST

we have: Kriging variance

* *
Cov( ZV , Z ) = Var( Z )
V V
The expression for the minimum estimation variance (obtained
by kriging), also called the kriging variance (KV), is:
so that: N

2OK = Var( zV* zV ) = i ( x i , V ) (V , V ) +

a =1 i=1
Var( Zv* )
All the terms in this expression are defined previously.
where:
is the Lagrange multiplier used in solving the kriging
system written with the covariance function.
Note also that the kriging system and all kriging equations can
be expressed in terms of the semi-variogram function ((h)) only.
This is particularly interesting for intrinsic random functions for
which the covariance is not defined (see Journel and Huijbregts,
1978).

5th International Mining Geology Conference Bendigo, Vic, 17 - 19 November 2003 223