Beyond Ordinary Kriging Master7

The Geostatistical Association of Australasia
PO Box 1719 West Perth WA Australia
Beyond Ordinary Kriging:

Non-Linear Geostatistical Methods in Practice
Proceedings of a 1 day Symposium held at Rydges Hotel, Perth CBD on Friday 30th October 1998
Major Sponsors
Mining & Resource Technology
MAJOR SPONSORS
Edith Cowan University
Contact: Dr Lyn Bloom l.bloom@eagle.fste.ac.cowan.edu.au
Mining & Resource Technology

Contact: Bill Shaw wjs@mrtconsulting.com.au
Resource Service Group

Contact: Dr Julian Barnes rsg_perth@iap.net.au
SATCHEL SPONSOR
Geoval Australias Geostatistical Experts.
Contact: John Vann johnvann@geoval.com.au
OTHER SPONSORS
Arne Berkmans, Mineral Resource Consulting CSIRO Division of Mathematical and Information Sciences Gemcom, Mining Software Snowden & Associates, Mining Consultants
Dedication
FOREWORD
This symposium, entitled Beyond Ordinary Kriging Non-Linear Geostatistical Methods in Practice, is the first such event organised by the Geostatistical Association of Australasia (GAA). The collected papers in this volume cover a range of applications of non-linear estimation techniques. This is the first technical meeting held in Australia dedicated to the topic of non-linear estimation. Non-linear estimation techniques are now becoming widely accepted within the Australian mining community, and are seeing some use in non-mining applications (for example, environmental hazard mapping). As such, the timing of this symposium is good, because it represents a chance to hear the approaches, pitfalls and suggestions of a wide-ranging group of mine-site practitioners, geological, geostatistical and statistical consultants and academics. Papers presented at this symposium deal mainly with mining examples, but the techniques presented have wider application. In particular we have several papers on the topic of modelling geological geometry and continuity using non-linear methods. The subject of conditional simulation, the other family of non-linear methods now operational, was excluded from this symposium, because the GAA plan to run a similar symposium in 1999 devoted to application of simulation techniques. As convenor, I would like to thank a number of people for their assistance in the hard work of organising this symposium. Most importantly, Meredith Gillespie (Geoval) put in most of the organisational leg-work and did a first class job. Thanks go to Geoval for their support in these administrative matters. Fellow executive committee member Bill Shaw (Mining and Resource Technology) and I shared the task of editing and reviewing papers, and GAA President, Louis Voortman, and Secretary, Ian Glacken, (both of Snowden Associates) also lent assistance. Sonja Neame of Resource Service Group helped with the financial and banking aspects. Thanks go to all of these individuals. To our major sponsors: Edith Cowan University, Mining and Resource Technology and Resource Service Group, we extend our thanks. The Australian Institute of Geoscientists (AIG) lent logistical assistance on a number of matters. Both the AIG and the Australian Institute of Mining and Metallurgy helped with promotions.
Symposium on Beyond Ordinary Kriging
Dedication
Thanks also to satchel sponsor, Geoval, and to those companies purchasing display booth space for the symposium: Gemcom, Global Mining Services, Geoval and Snowden Associates. Other sponsors were Arne Berckmans and the CSIRO Division of Mathematical and Information Sciences. Finally, thanks to authors, members of the GAA, and others, attending this first historic symposium.
John Vann GAA Executive Committee Perth, October 1998.
2003 Addendum This volume was re-edited to fit the standard GAA symposium volume format, for publication as a compact disc. Proof-reading of the volume was undertaken by Stella Searston, Roger Cooper and John Warner. John Vann designed the cover artwork, which was then produced by John Warner. The CD design, layout and manufacture were due to John Warner. Stella Searston and John Warner, May 2003.
Dedication
DEDICATION
This volume is dedicated to our colleagues:
Professor Michel David

May 10, 2000
Professor Georges Matheron

7 August, 2000
Program
PROGRAM
8.30 to 9.00 am Registration, Level 1
Session 1.
Chair: Dr John Henstridge (Data Analysis Australia) GAA Executive Committee 9.00 to 9.10 am 9:10 to 9:50 am Welcome and opening comments John Vann (Geoval) Symposium Convenor Keynote Address Beyond Ordinary Kriging a review of non-linear estimation John Vann, Daniel Guibal (Geoval) A practitioners implementation of Indicator Kriging Ian Glacken, Paul Blackney (Snowden Associates) Coffee Break
9.50 to 10.20 am 10.20 to 10.50 am
Session 2
Chair: Mr Bill Shaw (Mining and Resource Technology), GAA Executive Committee 10.50 to 11.20 am The application of Indicator Kriging in the modelling of geological data Brett Gossage (Resource Service Group) Non-linear modelling of geological continuity Dr John Henstridge (Data Analysis Australia) Comparison of Median and full Indicator Kriging in the analysis of gold mineralisation Donna Hill, Dr Lyn Bloom, Dr Ute Mueller (Edith Cowan University), Danny Kentwell (SRK Australia) Lunch
11.20 to 11.50 am 11.50 to 12.20 pm
12.20 to 1.20 pm
Program
Session 3
Chair: Dr Lyn Bloom (Edith Cowan University) GAA Executive Committee 1.20 to 1.50 pm Local recoverable resource estimation: a case study in Uniform Conditioning on the Wandoo project for Boddington gold mine Michael Humphreys (Geoval) A case study using Indicator Kriging: the Mount Morgan gold-copper deposit, Queensland Ivor Jones (WMC Resources) Practical application of Multiple Indicator Kriging to recoverable resource estimation for the Halleys lateritic nickel deposit Ian Lipton, Richard Gaze, John Horton (Mining and Resource Technology) Coffee Break
1.50 to 2.20 pm
2.20 to 2.50 pm
2.50 to 3.20 pm
Session 4
Chair: Daniel Guibal (Geoval) 3.20 to 3.50 pm 3.50 to 4.20 pm Median Indicator Kriging: a case study in iron ore Alison Keogh, Craig Moulton (Hamersley Iron) A proposed approach to change of support correction for Multiple Indicator Kriging, based on p-field simulation Sia Khosrowshahi, Richard Gaze, Bill Shaw (Mining and Resource Technology)
Program
Discussion Panel
Chair: John Vann (Geoval), GAA Executive Committee Panel: Dr Lyn Bloom (Edith Cowan University), GAA Executive Committee Richard Gaze (Mining and Resource Technology) Brett Gossage (Resource Service Group) Daniel Guibal (Geoval) Vivienne Snowden (Snowden Associates) 4.20 to 5.10 General discussion on the floor and with the panel
5.10 to 6 pm
Drinks on the Terrace
Vann and Guibal Keynote presentation
CONTENTS
Foreword.......................................................................................................................1 Dedication .....................................................................................................................1 Program ........................................................................................................................2 Beyond Ordinary Kriging An Overview of Non-linear Estimation ..................6 John Vann and Daniel Guibal Geoval.......................................................................6 A practitioners implementation of indicator kriging .............................................26 Ian Glacken and Paul Blackney Snowden Associates .............................................26 The Application of Indicator Kriging in the Modelling of Geological Data.........40 Brett Gossage Resource Service Group...................................................................40 Non-Linear Modelling of Geological Continuity ....................................................41 John Henstridge Data Analysis Australia................................................................41 Comparison of Median and Full Indicator Kriging in the Analysis of a Gold Mineralisation ............................................................................................................50 Donna Hill, Ute Mueller, Lyn Bloom Edith Cowan University..............................50 Local recoverable estimation: A case study in uniform conditioning on the Wandoo Project for Boddington Gold Mine ...........................................................63 Michael Humphreys Geoval ....................................................................................63 A case study using indicator kriging the Mount Morgan Gold-Copper Deposit, Queensland ..................................................................................................76 Ivor Jones WMC Resources .....................................................................................76 Practical application of multiple indicator kriging and conditional simulation to recoverable resource estimation for the Halleys lateritic nickel deposit. ............88 Ian Lipton, Richard Gaze, John Horton and Sia Khosrowshahi Mining and Resource Technology ...............................................................................................88 Median Indicator Kriging - A Case Study in Iron Ore ........................................106 Alison Keogh and Craig Moulton Hamersley Iron ................................................106 A proposed approach to change of support correction for multiple indicator kriging, based on p-field simulation .......................................................................121 Sia Khosrowshahi, Richard Gaze and Bill Shaw Mining and Resource Technology ................................................................................................................................121 Author Contact Details............................................................................................132
BEYOND ORDINARY KRIGING AN OVERVIEW OF NON-LINEAR ESTIMATION

John Vann and Daniel Guibal Geoval
Abstract
Many geostatistical variables have sample distributions that are highly positively skewed. Because of this, significant deskewing of the histogram and reduction of variance occurs when going from sample to block support, where blocks are of larger volume than samples. When making estimates in both mining and non-mining applications we often wish to map the spatial distribution on the basis of block support rather than sample support. The SMU or selective mining unit in mining geostatistics refers to the minimum support upon which decisions (traditionally: ore/waste allocation decisions) can be made. The SMU is usually significantly smaller than the sampling grid dimensions, in particular at exploration/feasibility stages. Linear estimation of such small blocks (for example by inverse distance weighting IDW or ordinary Kriging OK) results in very high estimation variances, i.e. the small block linear estimates have very low precision. A potentially serious consequence of the small block linear estimation approach is that the grade-tonnage curves are distorted i.e. prediction of the content of an attribute above a cut-off based on these estimates is quite different to that based on true block values. Assessment of project economics (or other critical decision making) based on such distorted grade-tonnage curves will be riskier than necessary. While estimation of very large blocks, say similar in dimensions to the sampling grid, will result in lower estimation variance, it also implies very low selectivity, which is often an unrealistic assumption. This paper presents an overview of the geostatistical approach to solving this problem: non-linear estimation. Linear estimation is compared to non-linear estimation, the motivations of nonlinear approaches are presented. A summary of the main geostatistical non-linear estimators is included. In a non-linear estimation we estimate, for each large block (by convention called a panel) the proportion of SMU-sized parcels above a cut-off grade or attribute threshold. A series of proportions above cut-off defines the SMU distribution. Use of such non-linear estimates reduces distortion of grade-tonnage curves and allows for better decision making. A partial bibliography of key references on this subject is included.
Vann and Guibal Keynote presentation Key Words: geostatistics, non-linear estimation, mining, environmental contamination, grade-tonnage curve, indicator, Gaussian transformation, lognormal distribution
Introduction
This paper uses mining terminology. This is convenient because: (1) this is the language of most of the geostatistical literature (even environmental and petroleum papers refer to grade tonnage relationships, nugget effects, selective mining units and so on) and (2) the main audience for this paper is likely to have a mining background. We hope non-miners are tolerant of this bias. This is a review paper and as such, an extensive bibliography is given. This bibliography goes beyond the papers cited in the text of our paper and is intended to give readers access to the primary literature.
Linear methods
First things first what is a linear interpolator?
Inverse Distance Weighting (IDW) interpolators are linear, as is Ordinary Kriging (OK). What do we mean by the term linear interpolator? A relatively non-mathematical understanding of linear weighted averaging can be gained from thinking about linear regression. In linear regression, the relationship between two variables, x and y, is considered to be a straight line (i.e. linear). The formula for this straight line is simple:
y = ax + b
Where a is the slope of the line and b is the value of y when x equals zero (i.e. the yintercept). If we specify a particular value of x we can therefore conveniently determine the expected y value corresponding to this x. It doesnt matter whether we specify an x value which is very small or very large, or anywhere in-between: the relationship between x and y is always the same the specified straight line. In other words, the formula used to estimate y does not alter as the magnitude of the x value changes. A linear interpolator has this property: the weights we assign to each of the N sample locations inside our estimation neighbourhood are independent of the specific data values at these locations. Think about the simplest kind of linear interpolator, IDW. An IDW estimate assigns the weight to a sample located within the estimation neighbourhood as: 1 d i
i =
d
j =1 j

Where are the weights, d are the distances from each sample location to the centroid of the block to be estimated and is the power1. Once the power to be used is specified, the ith sample is assigned a weight that depends solely upon its location (distance d i to the centroid). Whether the sample at this location had an average or extreme value does not have any impact whatsoever on the assignment of . OK is a more sophisticated linear interpolator proposed by Matheron (1962, 1963a, 1963b). OKs advantage over IDW as a linear estimator is that it ensures minimum estimation variance given: (1) A specified model spatial variability (i.e. variogram or other characterisation of spatial covariance/correlation), and (2) A specified data/block configuration (in other words, the geometry of the problem). The second criterion involves knowing the block dimensions and geometry, the location and support of the informing samples, and the search (or Kriging neighbourhood) employed. Minimum estimation variance simply means that the estimation error is minimised by OK. Given an appropriate variogram model, OK will outperform IDW because the estimate will be smoothed in a manner conditioned by the spatial variability of the data (known from the variogram). Now, contrast linear regression with non-linear regression. There are many types of non-linear relationships we can imagine between x and y, a simple example being:
y = ax 2 + b
This is a quadratic (or parabolic) regression, available in most modern spreadsheet software, for example. Note that the relationship between x and y is now clearly nonlinear the nature of the relationship between x and y is clearly dependent upon the particular x value considered. Non-linear geostatistical estimators therefore allocate weights to samples that are functions of the grades themselves and not solely dependent on the location of data.
Non-linear interpolators
Limitations of linear Interpolators
The fundamental limitations of linear estimation (of which OK provides the best solution) are straightforward:
1 The denominator of this fraction expresses the weight calculated as a proportion of the total weight allocated to all samples found within the search

1. We may be motivated to estimate the distribution rather than simply an expected value at some location (or over some area/volume, if we are talking about block estimation). Linear estimators cannot do this. The cases abound: recoverable ore reserves in a mine, the proportion of an area exceeding some threshold of contaminant content in an environmental mapping, etc. 2. We are dealing with a strongly skewed distribution, eg. a precious metal or uranium deposit, and simply estimating the mean by a linear estimator (for example by OK) is risky, the presence of extreme values making any linear estimate very unstable. We may require a knowledge of the distribution of grades in order to get a better estimate of the mean. This usually involves making assumptions about the distribution (for example, what is the shape of the tail of the distribution?) even in situations where we are ostensibly distribution free (for example using IK). 3. We may be studying a situation where the arithmetic mean (and therefore the linear estimator used to obtain it) is an inappropriate measure of the average, for example in situations of non-additivity like permeability for petroleum applications or soil strength for geological engineering applications. The specific problem of estimating recoverable resources was the origin of non-linear estimation and has been the main application. From a geostatistical viewpoint, non-linear interpolation is an attempt to estimate the conditional expectation, and further the conditional distribution of grade at a location, as opposed to simply predicting the grade itself. In such a case we wish to estimate the mean grade (expectation) at some location under the condition that we know certain nearby sample values (conditional expectation). This conditional expectation, with a few special exceptions (eg. under the Gaussian Model see later) is non-linear. In summary, non-linear geostatistical estimators are those that use non-linear functions of the data to obtain (or approximate) the conditional expectation. Obtaining this conditional expectation is possible, in particular through the probability distribution: Pr Z ( x o )| Z ( xi )
This reads: the probability of the grade at location xo given the known sampling information at locations Z(xi) (i.e. Z(x1), Z(x2) .Z(xN). This is the conditional distribution of grade at that location. Once we know (or approximate) this distribution, we can predict grade tonnage relationships (eg. how much of this block is above a cut-off ZC ?).
10
Available methods
There are many methods now available to make local (panel by panel) estimates of such distributions, some of which are:
Disjunctive Kriging DK (Matheron, 1976, Armstrong and Matheron, 1986a, 1986b); Indicator Kriging IK (Journel, 1982, 1988) and variants (Multiple Indicator Kriging; Median Indicator Kriging, etc.); Probability Kriging PK (Verly and Sullivan, 1985); Lognormal Kriging LK (Dowd, 1982); Multigaussian Kriging MK (Verly and Sullivan, 1985, Schofield, 1989a, 1989b); Uniform Conditioning UC (Rivoirard, 1994, Humphreys, 1998); Residual Indicator Kriging RIK (Rivoirard, 1989).
In a non-linear estimation we estimate, for each large block (by convention called a panel) the proportion of SMU-sized parcels above a cut-off grade or attribute threshold. A series of proportions above cut-off defines the SMU distribution. Note that there is a very long literature warning strongly against estimation of small blocks by linear methods (Armstrong and Champigny, 1989; David, 1972; David 1988; Journel, 1980, 1983, 1985; Journel and Huijbregts, 1978; Krige, 1994, 1996a, 1996b, 1997; Matheron, 1976, 1984; Ravenscroft and Armstrong, 1990; Rivoirard, 1994; Royle, 1979). By small blocks, we mean blocks that are considerably smaller than the average drilling grid (say appreciably less than half the size, although in higher nugget situations, blocks with dimensions of half the drill spacing may be very risky). The authors strongly reiterate this warning here. The prevalence in Australia of estimating blocks that are far too small is symptomatic of misunderstanding of basic geostatistics. Even estimating such small blocks directly by a non-linear estimator may be incorrect and risky. When using non-linear estimation for recoverable resources estimation in a mine, the panels should generally have dimensions approximately equal to the drill spacing, and only in rare circumstances (i.e. strong continuity) can significantly smaller panels be specified. Non-linear estimation provides the solution to the small block problem. We cannot precisely estimate small (SMU-sized) blocks by direct linear estimation. However, we can estimate the proportion of SMU-sized blocks above a specified cut-off, within a
Symposium on Beyond Ordinary Kriging 11

panel. Thus, the concept of change of support is critical in most practical applications of non-linear estimation.
Support effect
Definition
"Support" is a term used in geostatistics to denote the volume upon which average values may be computed or measured. Complete specification of support includes the shape, size and orientation of the volume. If the support of a sample is very small in relation to other supports considered, eg. drill hole sample upon which a gold assay has been made, it is sometimes assumed to correspond to "point support". Grades of mineralisation measured on a small support (eg. drill hole samples) can be much richer or poorer than grades measured on larger supports, say selective mining units (SMU) blocks. The grades on smaller supports are said to be more dispersed than grades on larger supports. Dispersion is usually measured by variance. Although the global mean grades measured (or estimated) on different supports (at zero cut-off) are the same, the variance of the smaller supports is higher, i.e. very high drill hole sample grades are possible, but large mining blocks have a smoother distribution of grades (fewer very high and very low grades). "Support effect" is this influence of the support on the distribution of grades.
The necessity for change of support

Change of support is vital for predicting recoverable reserves if we intend to selectively mine a deposit. Before committing the capital required to mine such a deposit, an economic decision must be made based only on the samples available from exploration drilling. Because mining does not proceed with a selection unit of comparable size to the samples, the difference in support between the samples and the proposed SMU must be accounted for in any estimate to obtain achievable results. When there is a large nugget effect, or an important short-scale structure apparent from the variography, then the impact of change of support will be pronounced. The histogram of drill hole samples will usually have a much longer "tail" than the histogram of mining blocks. Simplistic variance corrections, for example affine corrections, do not reflect the fact that, in addition to variance reduction, change of
12

support also involves symmetrisation of the histogram2. This is especially important in cases where the histogram of samples is highly skewed.
Recoverable resources
Recoverable resources are the portion of in-situ resources that are recovered during mining. The concept of recoverable resources involves both technical considerations, such as cut-off grade, SMU definition, machinery selection etc., and also economic/financial considerations such as site operating costs, commodity prices outlook, etc. In this paper, only technical factors are considered. Recoverable resources can be categorised as either global or local recoverable resources. Global recoverable resources are estimated for the whole field of interest; eg. estimation of recoverable resources for the entire orebody (or a large well-defined subset of the orebody like an entire bench)3. Local recoverable resources are estimated for a local subset of the orebody; eg. estimation of recoverable resources for a 25m x 50m x 5m panel.
A summary of main non-linear methods

Indicators
The use of indicators is a strategy for performing structural analysis with a view to characterising the spatial distribution of grades at different cut-offs. The transformed distribution is binary, and so by definition does not contain extreme values. Furthermore, the indicator variogram for a specified cut-off zc is physically interpretable as characterising the spatial continuity of samples with grades exceeding zc . Indicator transformations may thus be conceptually viewed as a digital contouring of the data. They give very valuable information on the geometry of the mineralisation. A good survey of the indicator approach can be found in the papers of Andre Journel (eg. 1983, 1987, 1989). An indicator random variable I ( x , zc ) is defined, at a location x , for the cut-off zc as the binary or step function that assumes the value 0 or 1 under the following conditions:
2 This symmetry can be demonstrated via the central limit theorem of classical statistics, which states that the means of repeated samplings of any distribution will have a distribution which is normal, regardless of the underlying distribution.. When we consider block support, the aggregation of points to form blocks will thus deskew the histogram. In the ultimate case, we have a single block, being the entire zone of stationarity and there is no skewness as such. 3 Global recoverable resources can be very useful as checks on local recoverable estimation, a good first pass valuation or can be used for checking the impact on grade-tonnage relationships of changing SMU, bench-height studies, etc. They are not specifically discussed in this paper. The interested reader is referred to Vann and Sans (1995).
13

I ( x, zc ) = 0 I ( x, zc ) = 1 if Z ( x ) zc if Z ( x ) > z c
The indicators thus form a binomial distribution, and we know the mean and variance of this distribution from classical statistics:
m= p
2 = p(1 p)
Where p is the proportion of 1s as defined above (for example, if the cut-off, zc is equal to the median of the grade distribution, p takes a value of 0.5, and the maximum variance is defined as 0.25).
After transforming the data, indicator variograms can be calculated easily by any program written to calculate an experimental variogram. An indicator variogram is simply the variogram of the indicator.
Indicator Kriging
Indicator Kriging is kriging of indicator transformed values using the appropriate indicator variogram as the structural function. In general the kriging employed is ordinary kriging. (OK). An IK estimate (i.e. kriging of a single indicator) must always lie in the interval [0,1], and can be interpreted either as 1. probabilities (the probability that the grade is above the specified indicator) or 2. as proportions (the proportion of the block above the specified cut-off on data support). In addition to its uses for indicator kriging (IK), multiple indicator kriging (MIK), probability kriging (PK) and allied techniques, the indicator variogram can be useful when making structural analysis to determine the average dimensions of mineralised pods at different cut-offs. Indicators are also useful for charactering the spatial variability of categorical variables (eg. presence or absence of a specific lithology, alteration, vein type, soil type, etc.). Henstridge (1998) presents examples of such applications for an iron deposit and Gossage (1998) give a more general overview of such applications of indicator kriging.
Multiple Indicator Kriging

Multiple indicator kriging (MIK) involves kriging of indicators at several cut-offs (see various publications by Andre Journel in the references to this paper as well as Hohn, 1988 and Cressie, 1993). MIK is an approach to recoverable resources estimation which is robust to extreme values and is practical to implement. Theoretically, MIK gives a worse approximation of the conditional expectation than disjunctive kriging
14

(DK), which can be shown to approximate a full co-kriging of the indicators at all cutoffs, but does not have the strict stationarity restriction of DK. The major difficulties with MIK can be summarised as: 1. Order relation problems: i.e. because indicator variogram models may be inconsistent from one cut-off to another we may estimate more recovered metal above a cut-off z c 2 than for a lower cut-off z c1 , where z c1 < zc 2 , which is clearly impossible in nature. While there is much emphasis on the triviality of order relation problems and the ease of their correction in the literature, the authors have observed quite severe difficulties in this regard with MIK. The theoretical solution is to account for the cross-correlation of indicators at different cut-offs in the estimation by co-kriging of indicators, but this is completely impractical from a computational and time point of view. In fact, the motivation for developing probability kriging (PK) was to approximate full indicator co-kriging (see below). 2. Change of support is not inherent in the method. In the authors experience, most practical applications of MIK involve using the affine correction, which assumes that the shape of the distribution of SMUs is identical to that of samples, the sole change in the distribution being variance reduction as predicted by Kriges Relationship. There are clear warnings in the literature (by Journel, Isaaks and Srivastava, Vann and Sans, and others) about the inherent deskewing of the distribution when going from samples to blocks. The affine correction is not suited to situations where there is a large decrease in variance (i.e. where the nugget is high and/or there is a pronounced short-scale structure in the variogram of grades). Other approaches can be utilised, e.g. lognormal corrections (very distribution dependent), or conditional simulation approaches (costly in time). A new proposal for change of support in MIK is given by Khosrowshahi et al. (1998).
Median Indicator Kriging

Median indicator kriging is an approximation of MIK which assumes that the spatial continuity of indicators at various cut-offs can be approximated by a single structural ~ , where m ~ is the median of the grade distribution. The function, that for z c = m indicator variogram at (or close to) the median is sometimes considered to be representative of the indicator variograms at other cut-offs. This may or may not be true, and needs to be checked. The clear advantage of median indicator kriging over MIK is one of time (both variogram modelling and estimation). The critical risk is in the adequacy of the implied approximation. If there are noticeable differences in the shape of indicator variograms at various cut-offs, one should be cautious about using median indicator kriging (Isaaks and Srivastava, 1989, pp 444). Hill et al (1998) and Keogh and Moulton (1998) present applications of the method.
Probability Kriging
Vann and Guibal Keynote presentation Probability kriging (PK) was introduced by Sullivan (1984) and a case study is given in Verly and Sullivan (1985). It represents an attempt to alleviate the order relationship problems associated with MIK, by considering the data themselves (actually their standardised rank transforms, distributed in [0,1]) in addition to the indicator values. Thus a PK estimate is a co-kriging between the indicator and the rank transform of the data U. When performed for n cut-offs, it requires the modelling of 2n+1 variograms: n indicator variograms, n cross-variograms between indicators and U, and finally the variogram of U. The hybrid nature of this estimate as well as the time-consuming complexity of the structural analysis makes it rather unpractical.
Indicator Co-kriging and Disjunctive Kriging

In general, any practical function of the data can be expressed as a linear combination of indicators: f (Z ) = f n I (Z , z n )
n
Thus, estimating f ( Z ) amounts to estimating the various indicators. The best linear estimate of these indicators is their full co-kriging, which takes into account the existing correlations between indicators at various cut-offs. Full indicator co-kriging (also called Disjunctive Kriging, abbreviated to DK) theoretically ensures consistency of the estimates (reducing order relationships to a minimum or eliminating them altogether): this makes the technique very appealing, but there is a heavy price to pay: if n indicators are used, n2 variograms and cross-variograms need to be modelled, and this is unpractical as soon as n gets over 5 or 6, even with the use of modern automatic variogram modelling software. The various non-linear estimation methods can be considered as ways of simplifying the full indicator cokriging. Roughly speaking, there are three possible paths to follow (Rivoirard, 1994): 1. Ignore the correlations between indicators: this is the choice made by MIK already discussed. The authors consider this a fairly drastic choice. 2. Assume that there is intrinsic correlation, i.e.that all variograms and crossvariograms are multiples of one unique variogram. In that case, cokriging is strictly equivalent to kriging; this is the hypothesis underlying median IK. Needless to say, unfortunately, this very convenient assumption is rarely true in practice (see median IK, above). 3. Express the indicators as linear combinations of uncorrelated functions (orthogonal functions), which can be calculated from the data. Cokriging of the indicators is then equivalent to separate kriging of the orthogonal functions; this
Vann and Guibal Keynote presentation decomposition of the indicators is the basis of residual indicator kriging (RIK) and of isofactorial disjunctive kriging.
Residual Indicator Kriging

In this particular model, within the envelope defined by a low cut-off, the higher grades are randomly distributed. The proximity to the border of the envelope has no direct incidence on the grade, and this corresponds to some types of vein mineralisation, where there is little correlation between the geometry of the vein and the grades. The validity of the model is tested by calculating the ratios
ij (h) i ( h)
(cross-variograms of indicators over variograms of indicator) for the cut-offs zj higher than zi. If these ratios remain approximately constant, then the model is appropriate. Note that an alternative decreasing model exists where one compares the crossvariograms to the variogram associated with the highest cut - off (instead of the lowest ). residuals are defined from the indicator functions by I ( Z ( x), z i ) I ( Z ( x), z i 1 ) where Ti = E [I ( Z ( x), z i )] , i.e. the proportion of H i ( x) = Ti Ti 1 grades higher than the cut-off zi. i.e. H0 ( x) = 1 I ( Z ( x), z1 ) H 1 ( x) = 1 T1 ...... I ( Z ( x), z n ) I ( Z ( x), z n 1 ) H n ( x) = Tn Tn 1 The
The Hi ( x ) are uncorrelated and we have:

i I ( Z ( x), z i ) = H j ( x) Ti j =0
This means that the indicators can be factorised. In order to get a disjunctive estimate of f ( Z ( x)) , it is enough to krige separately each of the residuals Hi ( x ) . The Ti are simply estimated by the means of the indicators, I ( Z ( x), z i ) .
17
Vann and Guibal Keynote presentation In practice, the residuals are calculated at each data point, their variograms are then evaluated and independent krigings are performed. Another check of the model consists in directly looking at the cross variograms of the residuals: if they are flat, indicating no spatial correlation, the model works. Thus, essentially this model requires no more calculations than indicator kriging, while being more consistent when it is valid. The reader is referred to Rivoirard (1994, chapter 4) for a fuller explanation and a case study (chapter 13) of this approach. Residual indicators is one way to co-krige indicators by separately kriging independent combinations of them and recombining these to form the co-kriged estimate. Like MIK, this method involves working with many indicators and the same number of variograms. Thus, it can be time consuming.
Isofactorial Disjunctive Kriging

There are several versions of isofactorial DK, by far the most common is Gaussian DK. Gaussian DK is based on an underlying diffusion model (where, in general, grade tends to move from lower to higher values and vice versa in a relatively continuous way). The initial data are transformed into values with a Gaussian distribution, which can easily be factorised into independent factors called Hermite polynomials (see Rivoirard, 1994 for a full explanation and definition of Hermite polynomials and disjunctive kriging). In fact, any function of a Gaussian variable, including indicators, can be factorised into Hermite polynomials. These factors are then kriged separately and recombined to form the DK estimate. The major advantage of DK is that you only need to know the variogram of the Gaussian transformed values in order to perform all the krigings required. The basic hypothesis made is that the bivariate distribution of the transformed values is bigaussian, which is testable. Although order relationships can occur, they are very small and quite rare in general. A very powerful and consistent change of support model exists for DK: the discrete Gaussian model (see Vann and Sans, 1995). Gaussian disjunctive kriging has proved to be relatively sensitive to stationarity decisions, (in most cases simple kriging is used in the estimation of the polynomials). DK should thus only be applied to strictly homogeneous zones.
Uniform Conditioning
Uniform conditioning (UC) is a variation of Gaussian DK more adapted to situations where the stationarity is not very good.
18
Vann and Guibal Keynote presentation In order to ensure that the estimation is locally well constrained, a preliminary ordinary kriging of relatively large panels is made, and the proportions of ore per panel are conditional to that kriging value. UC is a relatively robust technique. However, it does depend heavily upon the quality of the kriging of the panels. As for DK, the discrete Gaussian model ensures consistent change of support. Humphreys (1998) gives a case study of application of UC to a gold deposit.
Lognormal Kriging
Lognormal kriging (LK) is not linked to an indicator approach and belongs to the conditional expectation estimates. If the data are truly lognormal, then it is possible, by taking the log, and assuming that the resulting values are multigaussian, to perform a lognormal kriging. The resulting estimate is the conditional expectation and is thus in theory the best possible estimate. This type of estimation has been used very successfully in South Africa. Unfortunately the lognormal hypothesis is very strict: any departure can result in completely biased estimates.
Multigaussian kriging
A generalisation of the lognormal transformation is the Gaussian transformation which applies to any reasonable initial distribution. Again, under the multigaussian hypothesis, the resulting estimate represents the conditional expectation and is thus optimal. This is a very powerful estimate much more largely applicable than lognormal kriging, but requires very good stationarity to be used with confidence. Compared to Gaussian DK, it is completely consistent, but based on stronger multigaussian assumptions and its application to block estimation is more complex.
Conclusions and recommendations

1. As we approach the end of this century, and nearly 40 years since Matherons pioneering formulation of the Theory of Regionalised Variables, there are a large number of operational non-linear estimators to choose from. Understanding the underlying assumptions and mathematics of these methods is critical to making informed choices when selecting a technique. 2. We join the tedious chorus of geostatisticians over many years and recommend that linear estimation of small blocks be consigned to the past, unless it can be explicitly proved through very simple and long known kriging tests that such estimation is adequate. It is our professional responsibility to change to culture of providing what is asked for regardless of the demonstrable and potentially serious financial risks of such approaches.
Vann and Guibal Keynote presentation Small block OK or IDW estimates should no longer be acceptable as inputs to important financial decisions. 3. A particular non-linear method is often applied by a given practitioner without considering, for the data set in hand, whether the main assumptions of that method are realistic. Some of these assumptions may be testable, for example assumptions about the cross correlations of indicators or assumptions about the nature of edge effects. Testing of such assumptions is rarely performed, in our experience. We therefore recommend that such tests be implemented (see Rivoirard, 1994). 4. The issue of change of support is critical in estimation of recoverable resources, and as such should remain a major topic in our field. The major criticisms of MIK, the most widely applied non-linear estimation method in Australia, have centred on change of support (as pointed out by Glacken and Blackney, 1998). The whole problem of recoverable reserves is the problem of change of support. We recommend that practitioners become highly familiar with the issue of change of support and bring a sophisticated appreciation of this problem to their practice. 5. Conditional simulation (another non-linear method) is now within the abilities of inexpensive desktop computing, leading to another possible future route to recoverable resources. For a given block, the average of a set of n conditional simulations is exactly equal to the kriged estimate of that block In fact, as the full conditional distribution of the block grade is accessible, any non linear estimate can be calculated. Now, it must be clear that simulations are based exactly on the same type of hypotheses as most non-linear estimation methods (stationarity, representativity of the variogram, etc.) and, from this viewpoint, they need to be assessed as critically as any estimation method. While multiple conditional simulation is an inefficient route to local recoverable resources in 1998, the exponential increase in computational speed witnessed in the past decade suggests that it will soon become viable. 6. We hope that geostatistics does not go down a proprietary route. By this we mean that the algorithmic basis of geostatistical methods should rightly be in the public domain (and thus debatable and open to cross-validation). The publication of GSLIB (Journel and Deutsch, 1998) sets standards in this regard. Publication of actual source codes is more debatable of course, and there is certainly no consensus about it.
References
Armstrong, M., 1989. Geostatistics. Proceedings of the 3rd International Geostatistical Congress, Avignon, France. September, 1988, Kluwer Academic Publishers (Dordrecht),(2 volumes).
20

Armstrong, M. and Champigny, N., 1989. A study on kriging small blocks. CIM Bulletin. Vol. 82, No. 923, pp.128-133. Armstrong, M., and Matheron, G., 1986a. Disjunctive kriging revisited, Part 1. Mathematical Geology, Vol. 18, No. 8, pp. 711-728. Armstrong, M., and Matheron, G., 1986b. Disjunctive kriging revisited, Part 2. Mathematical Geology, Vol. 18, No. 8, pp. 729-742. Baafi, E., and Schofield, N.A., 1997. Geostatistics Wollongong 96. Proceedings of the 5th International Geostatistical Congress, Wollongong, NSW, Australia, September, 1996, (2 volumes). Cressie, N., 1993. Statistics for spatial data. (revised edition). John Wiley and Son (New York), 900pp. David, M., 1972. Grade tonnage curve: use and misuse in ore reserve estimation. Trans. IMM, Sect. A., Vol. 81, pp.129-132. David, M., 1988. Handbook of applied advanced geostatistical ore reserve estimation. Developments in Geomathematics 6. Elsevier (Amsterdam), 216pp. Dowd, P.A., 1982. Lognormal kriging the general case. Mathematical Geology, Vol. 14, No. 5, pp. 475-489. Dowd, P.A., 1992. A review of recent developments in geostatistics. Computers and Geosciences, Vol. 17, No. 10, pp. 1481-1500. Glacken, I.M., and Blackney, P.A., 1998. A practitioners implementation of indicator kriging. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia. Goovaerts, P. Geostatistics for Natural Resources Evaluation. Oxford University Press. 484pp. Gossage, B., 1998. The application of indicator kriging in the modelling of geological data. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia. Guarascio, M., Pizzul, C. and Bologna, F., 1989. Forecasting of selectivity. in: Geostatistics Volume 2 (Proc. of the 3rd. Int. Geostatistical Congress at Avignon). (Ed. M. Armstrong), Kluwer Academic Publishers (Dordrecht), pp.901-909. Guibal, D., 1987. Recoverable reserves estimation at an Australian gold project. in: Matheron, G., and Armstrong, M., (Eds.), Geostatistical case studies. Reidel (Dordrecht), pp. 149-168.

Henstridge, J., 1998. Non-linear modelling of geological continuity. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia. Hill, D.L., Bloom, L.M., Mueller, U.A., and Kentwell, D.J., 1998. Comparison of median and full indicator kriging in the analysis of a gold mineralization. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia. Hohn, M.E., 1988. Geostatistics and petroleum geology. Van Nostrand Reinhold (New York). 264pp. Humphreys, M., 1998. Local recoverable estimation: A case study in uniform conditioning on the Wandoo Project for Boddington Gold Mine. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia. Isaaks, E.H., and Srivastava, R.M., 1989. Applied Geostatistics. Oxford University Press (New York), 561pp. Jones, I., 1998. A case study using indicator kriging the Mount Morgan goldcopper deposit, Queensland. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia. Journel, A.G., 1980. The lognormal approach to predicting local distributions or selective mining unit grades. Mathematical Geology, Vol. 12, No. 4, pp. 285303. Journel, A.G., 1983. Nonparametric estimation of spatial distributions. Mathematical Geology, Vol. 15, No. 3, pp. 445-468. Journel, A.G., 1982. The indicator approach to estimation of spatial data. Proceedings of the 17th APCOM, Port City Press (New York), pp. 793-806. Journel, A.G., 1985. Recoverable reserves the geostatistical approach. Mining Engineering, June 1985, pp. 563-568. Journel, A.G., 1987. Geostatistics for the environmental sciences. United States Environmental Protection Agency Report (Project CR 811893), U.S.E.P.A. (Las Vegas), 135pp. Journel, A.G., 1988. New distance measures the route toward truly non-Gaussian geostatistics. Mathematical Geology, Vol. 20, No. 4, pp. 459-475. Journel, A.G., 1989. Fundamentals of geostatistics in five lessons. Short Course in Geology: Volume 8. American Geophysical Union (Washington), 40pp.

Journel, A.G., and Deutsch, C.V., 1998.. GSLIB Geostatistical software library and users guide, Second Edition, Oxford University Press, New York Journel, A.G., and Huijbregts, Ch. J., 1978. Mining geostatistics. Academic Press (London), 600pp. Keogh, A.J., and Moulton, C., 1998. Median indicator kriging A case study in iron ore. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia. Khosrowshahi, S., Gaze, R., and Shaw, W., 1998. A new approach to change of support for multiple indicator kriging. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia. Kitanidis, P.K., 1997. Introduction to geostatistics Applications in hydrogeology. Cambridge University Press (Cambridge), 249pp. Kreyszig, E., 1983. Advanced engineering mathematics (Fifth Edition). John Wiley and Sons (New York), 988pp. Krige, D.G., 1994. An analysis of some essential basic tenets of geostatistics not always practised in ore valuations. Proceedings of the Regional APCOM, Slovenia. Krige, D.G., 1996a. A basic perspective on the roles of classical statistics, data search routines, conditional biases and information and smoothing effects in ore block valuations. Proceedings of the Regional APCOM, Slovenia. Krige, D.G., 1996b. A practical analysis of the effects of spatial structure and data available and accessed, on conditional biases in ordinary kriging. In: Geostatistics Wollongong 96. Proceedings of the 5th International Geostatistical Congress, Wollongong, NSW, Australia, September, 1996, pp. 799-810. Krige, D.G., 1997. Block kriging and the fallacy of endeavouring to reduce or eliminate smoothing. Proceedings of the Regional APCOM, Moscow. Lantujoul, Ch., 1988. On the importance of choosing a change of support model for global reserves estimation. Mathematical Geology, Vol. 20, No. 8, pp. 10011019. Lipton, I.T., Gaze, R.L., and Horton, J.A., 1998. Practical application of multiple indicator kriging to recoverable resource estimation for the Ravensthorpe lateritic nickel deposit. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia.
23

Marechal, A., 1978. Gaussian anamorphosis models. Fontainebleau Summer School Notes C-72, Centre de Morphologie Mathematique, (Fontainebleau).22pp. Matheron, G., 1962..Traite de geostatistique applique, Tome I. Memoires du Bureau de Recherches Geologiques et Minieres. No. 14. Editions Technip (Paris). Matheron, G., 1963a..Traite de geostatistique applique, Tome II: Le Krigeage. Memoires du Bureau de Recherches Geologiques et Minieres. No. 24. Editions Technip (Paris). Matheron, G., 1963b. Principles of geostatistics. Economic Geology. Vol. 58, pp.1246-1266. Matheron, G., 1973. Le krigage disjonctif. Internal note N-360. Centre de Geostatistique, Fontainebleau, 21pp. Matheron, G., 1976. A simple substitute for conditional expectation: The disjunctive kriging. In Guarascio, M., et. al. (Eds.), Advanced Geostatistics in the Mining Industry. Proceedings of NATO A.S.I.. Reidel (Dordrecht), pp. 221236. Matheron, G., 1984. Selectivity of the distributions and "the second principle of geostatistics" in: Verly, G., et al., (Eds.) Geostatistics for natural resources characterisation. Reidel Publishing Co. (Dordrecht), pp.421-433. Matheron, G., and Armstrong, M. (Eds.), 1987. Geostatistical case studies. Reidel (Dordrecht), 247pp. Myers, J.C., 1997. Geostatistical error management Quantifying uncertainty for environmental sampling and mapping. Van Nostrand Reinhold (New York), 571pp. Ravenscroft, P.J., and Armstrong, M., 1990. Kriging of block models the dangers re-emphasised. Proceedings of APCOM XXII, Berlin, September 17-21, 1990. pp.577-587. Rivoirard, J., 1987. Geostatistics for skew distributions. South African Short Course Notes, C-131, Centre de Morphology Mathematique (Fontainebleau) 31pp. Rivoirard, J., 1989. Models with orthogonal indicator Residuals. In: Armstrong, M., (Ed). Geostatistics. Proceedings of the 3rd International Geostatistical Congress, Avignon, France. September, 1988, (2 volumes). Rivoirard, J., 1994. Introduction to disjunctive kriging and non-linear geostatistics. Clarendon Press (Oxford) 180pp. Rouhani, S., Srivastava, R.M., Desbarats, A.J., Cromer, M.V., and Johnson, A.I., 1996. Geostatistics for environmental and geotechnical applications. ASTM. (West Conshohocken, PA), 280pp.

Royle, A.G., 1979. Estimating small blocks of ore, how to do it with confidence. World Mining, April 1979. Schofield, N.A., 1989a. Ore reserve estimation at the Enterprise gold mine, Pine Creek, Northern Territory, Australia. Part 1: structural and variogram analysis. C.I.M. Bulletin, Vol. 81. No. 909, pp.56-61. Schofield, N.A., 1989b. Ore reserve estimation at the Enterprise gold mine, Pine Creek, Northern Territory, Australia. Part 2: the multigaussian kriging model. C.I.M. Bulletin, Vol. 81. No. 909, pp.62-66. Soares, A., (Ed.) 1993. Geostatistics Troia 92. Proceedings of the 3rd International Geostatistical Congress, Troia, Portugal, Kluwer Academic Publishers (Dordrecht),(2 volumes). Sullivan, J., 1984. Conditional recovery estimation through probability kriging theory and practice. In: Geostatistics for natural resources characterisation, Part 1. Verly, G. et al. (Eds.), Reidel (Dordrecht), pp.365-384. Vann, J., and Sans, 1995. Global Resource Estimation and Change of Support at the Enterprise Gold Mine, Pine Creek, Northern TerritoryApplication of the Geostatistical Discrete Gaussian Model. Proceedings of APCOM XXV, AusIMM. (Parkville), pp.171-179. Verly, G., 1983. The multigaussian approach and its applications to the estimation of local reserves. Mathematical Geology, Vol. 15, No. 2, pp. 259-286. Verly, G., and Sullivan, J., 1985. Multigaussian and probability krigings an application to the Jerritt Canyon deposit. Mining Engineering, June, 1985, pp. 568-574. Wackernagel, H., 1995. multivariate geostatistics An introduction with applications. Springer (Berlin), 256pp.
25
Glacken and Blackney A practitioners implementation of Indicator Kriging
A PRACTITIONERS IMPLEMENTATION OF
INDICATOR KRIGING
Ian Glacken and Paul Blackney Snowden Associates

Abstract
Indicator Kriging (IK) was introduced by Journel in 1983, and since that time has grown to become one of the most widely-applied grade estimation techniques in the minerals industry. Its appeal lies in the fact that it makes no assumptions about the distribution underlying the sample data, and indeed that it can handle moderate mixing of diverse sample populations. However, despite the elegant and simple theoretical basis for IK, there are many practical implementation issues which affect its application and which require serious consideration. These include aspects of order relations and their correction, the change of support, issues associated with highly skewed data, and the treatment of the extremes of the sample distribution when deriving estimates. This paper discusses the theoretical and practical bases for these considerations, and illustrates through examples and case studies how the issues associated with the daily application of the IK algorithm are addressed. Finally, some less commonly-used IK applications are presented, and the limitations of IK are discussed, along with proposed alternatives.
Key Words: geostatistics, indicator kriging, minerals industry, categorical kriging, soft kriging.
Introduction
Indicator Kriging (IK) as a technique in resource estimation is over fifteen years old. Since its introduction in the geostatistical sphere by Journel in 1983, many authors have worked on the IK algorithm or its derivatives. The original intention of Journel, based on the work of Switzer (1977) and others, was the estimation of local uncertainty by the process of derivation of a local cumulative distribution function (cdf). The original appeal of IK was that it was non-parametric it did not rely upon the assumption of a particular distribution model for its results. From slow beginnings in the early eighties as a technique in mineral resource estimation, and in many other natural resource mapping applications, IK has grown to become one of the
26
Glacken and Blackney A practitioners implementation of Indicator Kriging most widely-used algorithms, despite the relative difficulty in its application. It is the prime non-linear geostatistical technique used today in the minerals industry. This paper presents an overview of the theory of IK, followed by some discussion of practical applications. A number of practical aspects concerning the implementation of the IK algorithm and its variants are then discussed, including various ways to overcome some of the shortcomings of the technique. Finally, some of the less common applications of the indicator approach are presented, and an approach which is the successor to IK is proffered.
Overview of theory of Indicator Kriging

The concept of indicator coding of data is not new to science, but has only been proposed in the estimation of spatial distributions since the work of Journel (1983). The essence of the indicator approach is the binomial coding of data into either 1 or 0 depending upon its relationship to a cut-off value, zk. For a given value z(x),
1 if i ( x; z k ) = 0 if z ( x) z k z ( x) < z k
This is a non-linear transformation of the data value, into either a 1 or a 0. Values which are much greater than a given cut-off, zk, will receive the same indicator value as those values which are only slightly greater than that cut-off. Thus indicator transformation of data is an effective way of limiting the effect of very high values. Simple or Ordinary kriging of a set of indicator-transformed values will provide a resultant value between 0 and 1 for each point estimate. This is in effect an estimate of the proportion of the values in the neighbourhood which are greater than the indicator or threshold value. Carried out over a larger area for a series of blocks (putting aside for a moment the change of support issue), IK has the potential to generate recoverable resources in other words, the proportion of a block theoretically available above a given cut-off grade (indicator threshold). The outcome of IK is a conditional cumulative distribution function (ccdf) in effect a distribution of local uncertainty or possible values conditional to data in the neighbourhood of the block to be estimated. This distribution of grades can be used for many purposes, in addition to simply deriving the average (or expected) value. Any relevant criteria may be used to derive the estimate required, not simply the arithmetic mean of the local distribution. The practice of IK involves calculating and modelling indicator variograms (that is, variograms of indicator-transformed data) at a range of cut-offs or thresholds which should cover the range of the input data. This approach is termed Multiple Indicator Kriging (MIK). Until recently, this has been a somewhat time-consuming exercise. One approximation is to simply infer the variogram for the median of the input data and to use this for all cut-offs. This so-called Median IK approach is very fast, since
Glacken and Blackney A practitioners implementation of Indicator Kriging the kriging weights do not depend on the cut-off being considered. Median IK also necessitates the solution of only one kriging system per block in contrast to the multiple systems required for MIK. However, Median IK has its own assumptions and drawbacks, as discussed below. Since IK generates at each point or block a cumulative distribution, this should be non-decreasing and valued between zero and one. These two requirements are sometimes not met, leading to so-called order relations violations. Many methods have been proposed to counteract the order relations issue the most commonly-used involves direct correction of the indicator values (eg. Deutsch and Journel, 1998, p82). Another approach, proposed by Dagbert and Dimitrakopoulos (1992), involves the use of nested indicator variables in other words, indicator variables which are defined by successively halving the data set to define the thresholds. This nested indicator kriging approach eliminates any problems associated with order relations, but suffers from data deficiency problems, especially at high thresholds. The IK algorithm has been extended to not only include the indicator transform of the data, but also the data itself. This approach, first postulated by Isaaks (1984) and termed probability kriging, is essentially indicator co-kriging between the indicatortransformed data and the uniform (01) transform of the sample data. As with most co-kriging, the downside is the calculation and modelling of the cross-variograms between the two data types in addition to the univariate indicator variograms. Rivoirard (1993) used the relationships between the cross-variograms of indicators at adjacent cut-offs to draw conclusions about the nature of the processes influencing the distribution of data values. This work led to the definition of the so-called mosaic and diffusion models, among others, which lead to a particular style of IK or other nonlinear estimation algorithm. Despite much theoretical development, in practice it is the straightforward implementation of MIK, using non-nested indicator transforms of data at multiple cutoffs leading to the definition of local distributions of grade, which has proved most popular and enduring. The next section examines the advantages and motivation behind the choice of IK as an estimation technique.
Why use IK?

The primary motivation behind the use of IK in most earth science applications, and one of the main reasons for its introduction, is that it is non-parametric. Moreover, it is one of the few techniques that addresses mixed data populations. Figure 1 shows a typical scenario; a single, clearly-defined mineralogical/geological domain inside an orebody which is currently being mined, yet where the lead grades depicted show clearly mixed populations. In this instance, the geology and mineralogy of the orebody precludes further domaining. Since IK actually partitions the overall sample distribution by a number of thresholds, there is no need to fit or assume a particular analytically-derived distribution model for the data.
Glacken and Blackney A practitioners implementation of Indicator Kriging MIK requires the inference of a variogram model at each cut-off, and to a moderate degree, can handle different anisotropies at different cut-offs. Figures 2 and 3, from a gold project, show a single normal scores (Gaussian) variogram contour fan which suggests two directions of anisotropy, and two (non-adjacent) indicator variogram fans at a lower and an upper threshold, which demonstrate the partitioning of the anisotropy.
Figure 1. Histogram of lead data showing typical mixed populations
Figure 2.
Normal scores variogram contours in horizontal plane
showing two anisotropy directions
29
Figure 3. Equivalent indicator variogram fans for a lower (left) and an upper (right) threshold showing resolution of the two major anisotropy directions
Note that if the anisotropy changes too much between adjacent thresholds, the order relations violations become prohibitively large, but if the changes are gradual then the situation depicted can easily be handled. IK is a favoured option for highly-skewed data sets, as it offers a practical way of treating the upper tail of the distribution which does not depend entirely on an arbitrary upper cut value. This is discussed in more detail below, but IK allows the practitioner to use features of the actual data for defining any upper tail treatment. In many situations the level of knowledge of a deposit precludes the definition of clear geological or mineralogical domains. While unconstrained estimation is not advocated wherever there is the opportunity of defining constraining domains, if it is a necessity then MIK is an approach which can minimise the smoothing under certain conditions. One of the great benefits of the indicator transformation is that it allows common coding of diverse data types and their integration into the single process. Since all data is transformed to 01 space, other, secondary data types can easily be accommodated by the same coding scheme. This approach is discussed in some detail by Journel and Alabert (1989) and by Journel and Deutsch (1996). A practical example is given below. Finally, practitioners of IK are able to make use of one of the main motivations behind its development, that is the provision of an estimate of uncertainty at unsampled locations, via the inference of a distribution of values. This data can be used to derive an expected value, but also to yield risk-qualified outcomes, such as the
Glacken and Blackney A practitioners implementation of Indicator Kriging probability of exceeding a given grade the cut-off grade, or to map a given percentile of the distribution.
Practical implementation of IK
Treatment of upper and lower tails
An indicator kriging program will, in its most basic form, provide an estimate of the proportion of a model block above each of the indicator grades or thresholds assessed. To reduce this data into an estimate of mean block grade or grade above a cut-off, it is a requirement that each indicator class interval be assigned a grade. A number of sensitivities must be considered when undertaking the task of class interval grade assignment. If indicator grades have been carefully selected with adequate regard to the input grade distribution, then the distribution of grades within many classes will be nearly linear. The average grade of the input data or of the bounding indicator grades will normally suffice for the assignment of grade in these classes. The distribution of sample grades in the uppermost and lowermost grade classes of the distribution will not normally be linear and therefore require special treatment. In the case of a positively-skewed grade distribution, such as gold, the greatest estimation sensitivities relate to the grade assigned to the uppermost class. Distribution skew and grade outliers both influence the grade distribution in this class, which requires a more sophisticated method of mean grade selection if grade over-estimation or underestimation is to be avoided. Deutsch and Journel (1998) propose a modelling method based upon a hyperbolic distribution for representing the grade distribution above the uppermost indicator grade (Figure 4). The class mean grade calculated by this method is dependent upon the rate of decay of the hyperbolic function and the upper grade limit applied to the grade distribution. Both of these variables may be judged from the sample grade distribution. The choice of the decay rate parameter is generally less subjective than the selection of an upper grade limit for the class. If a slow decay rate is selected the class mean is sensitive to the magnitude of the class upper grade limit. If a high decay rate is chosen, the class mean is relatively insensitive to the upper grade limit.
31
Sample data and Hyperbolic model

100.0%
Cumulative Frequency
99.5%
99.0%
98.5%
Data Percentiles Hyperbolic Function
98.0% 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0
Grade
Figure 4. Example of a hyperbolic distribution fitted to the upper tail of a data set
In either case, the mean grade assigned to the uppermost class must be carefully considered if a good overall grade estimate is to result.
The data dilemma

When applying multiple indicator kriging to resource estimation it is necessary to settle on a finite number of thresholds that adequately represent the input data distribution shape. There is always a trade-off between the number of thresholds selected and the time available for the required analysis. In a typical gold mineralisation environment, the grade distribution of the mineralisation within a geological domain might, in the simplest case, be divided into decile classes based on grade or perhaps metal content. Additional indicators may be included to discriminate between components of mixed population distributions. The disadvantage of grade-based deciles is that many of the indicator grades will be concentrated at the lower grade end of the positively skewed grade distribution. Fewer indicator thresholds will represent the higher grades, which contain most of the metal. The advantage of metal-based indicator grade selection is the reduced emphasis on the lower grade end of the distribution and a better selection of indicators at grades associated with significant amounts of metal. Metal content indicators also provide the advantage of revealing the amount of metal that is associated with the higher grades. Metal indicators for gold projects often
Glacken and Blackney A practitioners implementation of Indicator Kriging reveal that 30 percent of the metal is derived from grades above the 90th grade percentile. In some cases, up to 70 percent of the metal may be above the 90th grade percentile. The recognition of the amount of metal that is dependent upon so few samples in an exploration drillhole database is quite often an eye-opener! The disadvantage of metal-based indicators is that, in practical terms, it is often impossible to define the continuity of grades for indicators set at or above the 90th grade percentile, due to the low numbers of high grade samples, and their sparse spatial distribution. The outcome is often that the spatial distribution of higher grade mineralisation, and thus the most significant part of the distribution (in terms of metal content) is estimated using poorly-defined or assumed grade continuity conditions. This in itself is not a fatal flaw, but requires some judgement and experience. This resource estimation problem is revealed by indicator analysis, but is not unique to the method. All resource estimation methods suffer from this weakness; the indicator approach only serves to highlight the issue. The only real answer to this dilemma is to collect more data. Enough close-spaced data needs to be collected from representative areas to allow the better definition of the high-grade continuity and the short-range continuity of lower-grade indicators. The other possibility that exists is to borrow continuity models from other sources such as in-pit grade control data located in similar geological environments with similar overall grade distributions. Unfortunately, a well-defined model of grade continuity is only part of the answer. If sampling density in the area being modelled is at typical exploration data levels, then the high grade continuity is likely to be shorter than the sample spacing. Unfortunately there is no easy solution to this dilemma, although conditional simulation is an approach which offers some promise.
Median IK
Median indicator kriging (Median IK) uses the median indicator variogram to define the continuity conditions for all indicators. This method is a simplified form of MIK that might be considered in the early stages of a resource project, when sample data is sparse and it is difficult or impossible to define grade continuity for a full range of indicators. The median indicator variogram is typically the most robust of all indicators, it tends to have the greatest range of continuity, and it is the easiest to define with some confidence from sparse data. The application of variograms from a single indicator to all thresholds reveals the main assumption associated with the median indicator method. This is that the direction and range of continuity does not vary with changing thresholds. Experience from full indicator variography studies shows that grade continuity almost always varies with indicator grade, and invariably declines with increasing indicator grade. This will tend to result in an overestimation of the quantiles of the upper grade classes with Median IK, resulting in a higher-than-normal expected grade. In practical terms,
Glacken and Blackney A practitioners implementation of Indicator Kriging Median IK is not a recommended technique where the data permits full estimation of a set of indicator variograms.
Use of a mineralisation indicator and choice of cut-offs

As mentioned above, in many geological estimation problems there may be no way to clearly define mineralogical or geological domains within which the spatial continuity and grade behaviour is more consistent than elsewhere. A typical scenario is an advanced exploration prospect with demonstrated geological and grade continuity, yet in which the detailed structures controlling and constraining mineralisation are unknown. Ideally in such a situation, the preference is to allocate some form of domain within which local controls on mineralisation may be applied, even if this is only a grade envelope. However, it is sometimes physically impossible to clearly separate ore-grade material from low-grade background material. In such a situation, descriptive statistics and geostatistics are often biased by the multitude of trace to low-grade assays, and this can also affect the choice of indicator cut-offs. One solution to this is to apply a mineralisation cut-off grade to the bottom end of the grade distribution. This will separate possibly-mineralised from clearly-background material. Values below this cut-off are rejected for statistical and geostatistical analysis. This leads to better-defined distributions, and assists with the selection of appropriate indicator cut-offs.
Change of support
The most trenchant criticism of IK by its detractors (for instance Dowd, 1992, or Matheron, 1982) is the issue of change of support that is, the generation of a distribution of grades for non-point (block, sensu stricto) data. Unlike ordinary kriging, inverse distance, and other linear estimation methods, the indicator transform is non-linear, as is the logarithmic transform, the uniform transform, and the normal scores transform. The consequence of this is that one cannot average indicators linearly, and thus cannot obtain a block distribution by averaging a series of point distributions derived at a smaller scale. The Ordinary Kriging (or inverse distance) corollary is to discretise a block by a series of points, estimate the grade at each point, and carry out the arithmetic mean to derive the block grade. However, if the statistic required from the IK distribution at each point (or sub-block) is known and fixed, such as the median or the mean value, there is nothing to stop the practitioner subdividing the block into a sub-blocks (actually points), estimating the ccdf by MIK at each point, deriving the desired statistic, and then averaging the result. This is shown diagrammatically in Figure 5. One downside of this approach is the extra computational effort. A short cut would be to use the same data configuration for each sub-block, thus yielding only one matrix inversion on the left-hand side of the kriging system; the difference between the sub-blocks would be in the different weights obtained due to
Glacken and Blackney A practitioners implementation of Indicator Kriging the differing positions of each sub-block in the parent block. Note that the objective of actually obtaining a distribution of uncertainty for each larger block (as shown in the top illustration of Figure 5) is a much more difficult task than simply correcting a series of mean values.
Block discretised by point distributions
Average block distribution
AVERAGING
Mean of each point distribution obtained
Mean of (unknown) block distribution obtained by averaging
Figure 5.
Diagrammatic illustration of permitted and non-
permitted changes of support during Multiple indicator kriging
The traditional approach to the change of support in MIK has been to apply a variance correction factor on a global basis to the point statistics, exactly as one might do with point kriged data. There are a number of common techniques for achieving this. Perhaps the most widely used is the affine correction, a simple factoring of variance from point to (theoretical) block variance. Another approach is to use the indirect lognormal correction as described in Isaaks and Srivastava (1989, p 472). The most elegant and theoretically correct solution to deriving block values is to move beyond the realm of estimation to that of sequential indicator simulation - this approach provides a truly local change of support, conditional only upon values in the neighbourhood. However, this approach is beyond the limits of the topic of indicator kriging addressed in this paper.
Variogram inference
Another criticism of indicator kriging, which paradoxically is a direct consequence of one of its great benefits, is the need to calculate a variogram and develop a model for each threshold value. For multiple (10 to 12) thresholds this was once a major exercise. However, advances in both computer speed and memory capacity and in software technology have produced new generations of fast and efficient variogram generation and modelling software. Such software allows the calculation of a full set of indicator variograms over all thresholds in one pass; this allows the practitioner to iterate between the various cut-offs and ensure a smooth variation in the variogram parameters. This is turn will serve to minimise the number of order relations
Glacken and Blackney A practitioners implementation of Indicator Kriging corrections required. The combination of modern software and hardware has all but eliminated the time penalty of variogram inference, and a full set of thresholds may now be generated and modelled in a few hours for a moderately-large data set, say up to a few thousand data. Other benefits of variogram calculation and modelling software include the provision of relative indicator variograms, scaled to a sill of one, which also eases the variogram modelling burden.
Other applications of IK
Categorical Kriging
The indicator transform also lends itself to the estimation of categorical data in other words, variables which are not continuous but which have discrete values. Some examples of categorical data are the presence or absence of a rock type (direct binary data requiring no transform), or a series of lithological or facies codes, or mineral sands hardness values. In this case, instead of kriging indicators at a set of thresholds, categorical IK will produce the probability of a given rock type or domain code occurring at a given location. Thus it is possible to produce probability maps for given lithologies based upon actual rock code data. This may be combined with indicator estimation or simulation of grade data, as described in Dowd (1996).
Soft IK
Soft indicator kriging, or soft kriging, is an area which holds great promise. Because of the common coding of all data types, both hard data values (i.e. those of the variable to be estimated) and soft data values (those correlated to the variable to be estimated) may both be used in the same indicator framework. The critical step is to calibrate the soft data from a scatter plot of soft and hard data. In a recent case study, the objective was to estimate cyanide-soluble copper values from a combination of few cyanide-soluble assays (the hard data) and many total copper assays (the soft data). Figure 6 shows a smoothed scatter plot between the cyanidesoluble copper data (on the Y axis) and the copper data (on the X axis), with shading which represents the probability. This was used to derive soft indicator values for the copper data which were valued anywhere between 0 and 1, not just 0 or 1 as with the hard cyanide-soluble copper values. These two data sets were then integrated into a multiple indicator kriging approach for the estimation of cyanidesoluble copper. The result was a significantly-improved definition of the cyanidesoluble copper than if only the limited hard data were used.
36
Figure 6.
Smoothed bivariate scatter plot between cyanide-
soluble copper (Y axis) and copper, used to derive soft indicator values for the copper samples.
Conclusions
Indicator kriging is now widely used in the mining industry as an estimation technique over a wide range of deposits and environments, because it offers practical solutions to some common estimation problems. In particular the issue of mixed or poorlydomained distributions, and the general trend away from the so-called parametric techniques (particularly after some spectacular disasters), has probably enhanced the acceptance of IK. The appeal of being able to generate (at least in theory) recoverable resources has undoubtedly contributed to the popularity of the approach. However, the use of IK is has its downside. Particular criticisms have been the relative difficulty of deriving true block distributions, the nuisance of order relations, and the sheer work involved in inferring variogram models at multiple thresholds. The practitioners in the industry, aided and abetted by academic research, have come up with solutions to almost all of the perceived problems, some more elegant than others. Notwithstanding these successful developments, the IK theory has now largely been superseded and improved by the conditional simulation paradigm. Conditional
Glacken and Blackney A practitioners implementation of Indicator Kriging simulation offers all of the advantages of IK and more. The single drawback still a major issue at most sites is the quantum leap in processing time and computing power required for the successful implementation of a simulation approach. However, even this is becoming a diminishing problem as computers exponentially increase in speed and memory capacity. It is predicted that within five years, indicator simulation and other simulation algorithms will be as commonplace in practical situations as indicator kriging is today. The extra dimension of simulation is hard to resist.
Acknowledgements
The management of Snowden Associates is thanked for its permission to publish this paper. The paper benefited from a review by Fleur Dyer.
References
Dimitrakopoulos, R, 1997. Indicator kriging course notes, W H Bryan Centre, University of Queensland Dimitrakopoulos, R, and Dagbert, M, 1992. Sequential modelling of relative indicator variables: dealing with multiple lithology types, in Geostatistics Troia 92, Soares, A Editor, Kluwer 1993. Dowd, P A, 1996. Structural controls in the geostatistical simulation of mineral deposits, in Baafi E Y, and Schofield, N A (Eds), Geostatistics Wollongong 96, pp 647-657. Dowd, P A, 1992. A review of recent developments in geostatistics, Computers and Geosciences Vol. 17, No. 10, p. 1481-1500. Isaaks E H, 1984. Risk qualified mappings for hazardous waste sites: a case study in distribution-free geostatistics. Masters thesis, Stanford University, Stanford, California, USA. Isaaks E H and Srivastava, R M, 1989. An introduction to applied geostatistics. Oxford University Press, New York. Journel , A G, 1983. Nonparametric estimation of spatial distributions, Math. Geol Vol 15, No. 3, p. 445-468. Journel A G, and Deutsch C V, 1996. Rank order geostatistics: a proposal for a unique coding and common processing of diverse data, in Baafi E Y, and Schofield, N A (Eds), Geostatistics Wollongong 96, pp 174-187. Journel A G, and Alabert F, 1989. Non-Gaussian data expansion in the earth sciences, Terra Nova 1: 123-34.
38

Journel, A G, and Deutsch, C V, 1998. GSLIB Geostatistical software library and users guide, Second Edition, Oxford University Press, New York. Matheron, G, 1982. La destructuration des hautes teneurs et le krigeage des indicatrices, Centre de Geostatistique et de Morphologie Mathematique, Note N-761, 33p. Rivoirard, J, 1992.. Relations between the indicators related to a regionalised variable, in Geostatistics Troia 92, Soares, A Editor, Kluwer 1993. Switzer, P, 1977. Estimation of distribution functions from correlated data. Bull. Inter. Stat. Inst., Vol 47, No. 2, p. 123-137.
39
Gossage Application of Indicator Kriging to modelling geological data
THE APPLICATION OF INDICATOR KRIGING IN THE MODELLING OF GEOLOGICAL DATA

Brett Gossage Resource Service Group
Abstract
Exhaustive geological databases are frequently compiled during the exploration and resource definition phases of mineral project development. A geological database is then often reviewed, plotted to section (or plan) to aid in the interpretation of a geological model before being forgotten. The geological or soft data in many cases provides critical information which warrants substantial further investigation to provide a better understanding of the geological controls on mineralisation, the possible distribution of important material types (eg: refractory ore) and often, appropriate bulk density assignment. Indicator Kriging (IK) can frequently provide a useful, cost effective means of validating and enhancing the more traditional geological interpretation. This paper reviews the application of IK in the development of both a geological and mineralised model for a mesothermal gold deposit which is interpreted as being both structurally complex and having multiple controls on the distribution of the gold mineralisation. The geological data investigated includes lithology, alteration and veining which have been reviewed both independently and in conjunction with geochemical drilling data. The results of the IK studies have been compared with the original geological/mineralisation model, and the relative merits of this assessment briefly discussed.
Key Words: Indicator kriging, probability, semivariogram, soft data.
40
Henstridge Non-linear modelling of geological continuity
NON-LINEAR MODELLING OF GEOLOGICAL CONTINUITY

John Henstridge Data Analysis Australia
Abstract
A frequent problem in the understanding of a mineral deposit is the classification of samples into rock types. While there may be several types of data used in this classification, one of statistical interest is multielement geochemical analysis. Provided there is a training dataset of samples where rock types have been manually classified and for which the geochemical information exists, it is possible to use discriminant analysis to classify subsequent rock samples into specific types. This is a logical extension of the first step in indicator kriging where samples are divided into two types above and below a certain grade. The continuity problem is that individual samples can be subject to significant random variation and thus a simple statistical classification is not smooth or does not reflect the geological continuity. One answer to this when there are just two rock types is that used in indicator kriging. Where there are more than two rock types the kriging approach is no longer satisfactory and a more probabilistic approach is necessary. This study investigates methods developed for regularly spaced data and used a combination of techniques developed by theoretical physicists and statisticians over the past twenty years. The example relates to an iron ore deposit.
Key Words: geostatistics, non-linear estimation, classification, Metropolis algorithms.
Introduction
One of the principal applications of geostatistics is the classification of blocks in a resource model into ore and waste. More generally there is a need to classify material into a number of different rock types. This is particularly the case where impurities must be considered as well as the ore grade itself. Examples are where iron ore mine plans must consider phosphorus levels or bauxite mines must consider clay content. However the classification cannot be properly done on a point-by-point or block-byblock basis since that ignores geological continuity. Continuity has two distinct though related roles in geostatistics. The first assumption of continuity is that despite the random variation in the data , the underlying structure
Henstridge Non-linear modelling of geological continuity at sampling points close to each other is likely to be similar. This provides the basis of the estimation of the variogram and the application of this to point or block estimates through kriging. This essentially sees kriging as an estimation method. The second assumption is that important aspects of the underlying geological structure being studied is continuous and hence the resulting output of the kriging process should reflect this. This leads to kriging as a smoothing process. Classical kriging assumes that the process being modelled is a continuous variable such as an ore grade. Furthermore the methods, being based on least squares and variances, are at best inefficient and at worst misleading when applied to data that is not normally distributed. However least squares methods are both computationally and theoretically the most tractable. While there is a need to consider models that do not depend upon normality, the penalty for such methods is usually greatly increased computation. Some such methods include disjunctive kriging that introduces some non-linearity (Matheron, 1976) or spatial generalised linear models that introduce the wider class of exponential family distributions (Diggle et al, 1998). The approach taken here relates to spatial models for categorical variables as occur with classification problems.
Discriminant analysis and classification

Discriminant analysis applied to geochemical data simultaneously considers a number of elements and models their joint distributions in each rock type. Typically with iron ore the measures are Fe, SiO2, Al2O3, P, TiO2 and loss on ignition (LOI). The simplest model assumes that these measures have a joint normal distribution in each rock type, perhaps after suitable transformation such as a power or logarithm. On the basis of these distributions it is possible to define a simple set of rules that can classify samples on the basis of their composition. This is illustrated by Figure 1 where the contours of the probability density functions for several rock types are illustrated in two dimensions (two elements). The normal distribution assumption implies that the contours are elliptical. On this basis the whole range of possible data values can be divided into several regions such that sample can be allocated to a rock type on the basis of the region they fall into. The boundaries between the regions are generally segments of ellipses. If an assumption is made that the shapes of the distributions are the same for all rock types, the boundaries become straight lines. While this might seem complex, it is quite simple to implement in practice. At the sample point the probability density function is calculated for the distributions corresponding to each of the rock types and the sample is assigned to the type which has the highest value of the density. The essence of this is the calculation of posterior probabilities p1, p2pk that give the likelihood of the sample coming from each of the k rock types and then choosing the rock type that is most likely. There are a number of other ways in which such a classification could be carried out it is an area of great activity in statistics. Of particular interest is modern work on non-parametric classification that no longer needs to assume a particular distribution
Henstridge Non-linear modelling of geological continuity but does generally require more data to be effective. McLachlan (1992) gives a somewhat mathematical summary. In what follows we simply assume that there is a classification method that works via posterior probabilities as above.
A B
Element 2
C
Element 1
Figure 1. A typical application in two dimensions where three rock types A, B and C have been identified according to a two element assay. The regions correspond to the classification functions.
Continuity Models
The standard geostatistical model assumes that the quantity being measured is defined continuously in the resource volume. Such a model does not readily carry over to the classification problem. Instead it is easier to use a lattice model where just a grid of points in three dimensional space is defined. A good review of lattice models is found in Cressie (1991). This can be thought of as a block model except that we just think of the centroid of each block. As a further simplification we just consider a regular lattice where the samples might be from a regular grid of drillholes and they have a regular vertical spacing. Continuity can now mean that each lattice point is likely to be similar to each of the adjacent points. The simplest model for adjacency is to take the points immediately to the north, south, east, west, above and below as neighbours. This is just one of the possibilities but with a rectangular grid it is effectively the minimum - it is illustrated in Figure 2. Of course via this nearest neighbour scheme every point in the lattice is generally related to every other point. However it assumes that the relationship works only by these nearest neighbours - a form of Markov property.
Henstridge Non-linear modelling of geological continuity The database is structured so that every point has a list of its neighbours. Most points in the interior have six such neighbours while boundary points or those along known discontinuities will have fewer. The database structure is usually such that an arbitrary number of neighbours can be managed which means that virtually any model for nearest neighbours can be used.
Figure 2.
The nearest neighbours of a lattice point.
The
neighbours of the shaded point are the points immediately to the north, south, east, west, above and below.
The basis of the continuity model is a statement about the probability that a point is of a particular rock type given the knowledge of the rock types at its nearest neighbours. This probability model can take a number of forms but the simplest with that we have had success has the prior probability of a point being of rock type i proportional to:
Exp ( (# NS neighbours of type i # NS neighbours not of type i) + (# E-W neighbours of type i # NS neighbours not of type i) + (# vertical neighbours of type i # vertical neighbours not of type i)).
Here the parameters , and measure the continuity in the north-south, east-west and vertical dimensions respectively. The posterior probability is then the product of this prior probability and the classification probability discussed earlier. The point is classified to the rock type with the greatest posterior probability. Notice that this probability model for continuity can be enhanced in many ways. We will give two examples:
In general the parameters , and could depend upon the rock type being considered. A typical application might be to enforce certain stratigraphic rules that may say that certain vertical sequences of rock types are impossible. It would also be possible to introduce knowledge that some rock types tend to be more extensive than others are.
44
The above model cannot allow for the main axis of variation to be rotated to the grid. This is a problem of the neighbourhood definition rather than the probability function itself. An appropriate neighbourhood model with more points as neighbours can readily correct this. However if each point is defined to have too many neighbours the calculations can get complex.
Algorithms
The fitting of the above models requires three steps, all of which require some care. The first and easiest is defining the basic point-by-point classification method. This is straightforward multivariate statistics. The second and third steps are the estimation of the parameters of the continuity model (the , and above) and the subsequent classification of the main body of the data. It is convenient to consider the final classification step first. The classification of each point requires knowledge of how the points around it have been classified but to classify them properly the first point must be classified. There is no direct solution to this and an iterative solution must be used. However the iterative solution is not straightforward since there tend to be many locally optimal solutions that must be avoided. The solution we have adopted is a version of the Metropolis-Hastings or annealing algorithm, a method first applied to such problems by Geman and Geman (1984). Initially the classification is carried out by adding perturbations to each of the posterior probabilities so that the classification is carried out with a controlled level of error. The starting point is the simple classification that ignores the continuity model. The points are then classified sequentially using the perturbed probability. This is repeated while slowly decreasing the scale of the perturbations to zero. Provided the perturbation process is scaled down slowly enough this method is sure to reach the global maximum likelihood solution even though it never evaluates the likelihood function. The impracticality of direct likelihood solutions to problems of this type has been a long-standing problem that first arose in statistical mechanics. The method also has many connections with the modern Markov Chain Monte Carlo (MCMC) approach that has revolutionised modern Bayesian statistics. If the method has one criticism apart from the computation involved it is that it is never possible to know whether the annealing or scaling down of the perturbations is applied slowly enough. However it is possible to make the process faster by constantly changing the order in which the points are classified at each step. The estimation of the continuity model parameters can form a second level of iteration. It is relatively straightforward for the example above and can be fitted by generalised linear model techniques (McCullagh and Nelder, 1989) in a package such as GLIM or S Plus.
An example
A small example of the application of this approach is presented here for an iron deposit that had been drilled on a 50 metre grid, with two metre composite samples taken vertically. A proportion was used as a training data set and manually classified into seven rock types ranging from ore through to the foot-wall underlying the ore body. Approximately 3500 samples were used in the example. Figures 3 and 5 show the result of the simple classification using a six element multivariate discrimination, after transforming the variables so that a normal distribution was a reasonable model. It can be seen that there is significant irregularity, which is hardly surprising since the element distributions for the different rock types overlapped somewhat. Figures 4 and 6 show the results after an annealing process with just one step, estimating the continuity parameters from the initial classification. The improvements that come from the continuity model are immediately obvious, giving a result that is physically much more reasonable, particularly for the vertical section. This could be improved even more with a more sophisticated continuity model, such as one that assumes that if a point has foot-wall above it then it must be foot-wall itself.
Height
140
160
180
200
2500
2600
2700 Easting
2800
2900
3000
Figure 3.
Vertical section of the deposit with the simple
classification before applying the continuity model. (+ = weathered layer, and = ore types, = foot-wall, = clay)
46
Height
140
160
180
200
2500
2600
2700 Easting
2800
2900
3000
Figure 4.
Vertical section of the deposit after applying the
continuity model.
97700
97800
97900
98000
North
98100
98200
98300
2500
2600
2700 Easting
2800
2900
3000
Figure 5.
Horizontal section of the deposit with the simple
classification before applying the continuity model.
47
97700
97800
97900
98000
North
98100
98200
98300
2500
2600
2700 Easting
2800
2900
3000
Figure 6.
Horizontal section of the deposit with the simple
classification after applying the continuity model.
Summary
The cost of these approaches is essentially greater computation for the annealing stage, something that made the methods impractical when first introduced in the early 1980s but which is now less of an issue. Their application is potentially quite wide since they are in some regards an extension to indicator kriging that divides the deposit into just two rock types. There is also a useful modularity to the model since it separates the continuity model from the model for what is happening at a specific point.
Acknowledgments
Cliff Robe River Mining provided the data for the initial development of these ideas. The diagrams were made with the assistance of Amanda Mellican of Data Analysis Australia.
References
Cressie, N.A.C., 1991. Statistics for Spatial Data, Wiley, New York. Diggle, P.J., Tawn, J.A. and Moyeed, R.A., 1998. Applied Statistics, Vol. 47, pp299-350.
Model-based geostatistics.
48

Geman, S. and Geman, D., 1984. Stochastic relaxation, Gibbs distributions and the Bayesian restoration of images, IEEE Transactions of Pattern Analysis and Machine Intelligence, Vol 6, pp 721-741. Matheron, G., 1976. A simple substitute for conditional expectation: the disjunctive kriging. In Advanced Geostatistics in the Mining Industry, Reidal, Dordrecht. McCullagh, P. and Nelder, J.A., 1989. Generalised linear Models, Chapman and Hall, London. McLachlan, G.J., 1992. Discriminant Analysis and Statistical Pattern Recognition, Wiley, New York.
49
Hill et al Comparison of Median and Full Indicator Kriging
COMPARISON OF MEDIAN AND FULL INDICATOR KRIGING IN THE ANALYSIS OF A GOLD MINERALISATION
Donna Hill, Ute Mueller, Lyn Bloom Edith Cowan University
Abstract
This paper presents a comparison of median indicator kriging and full indicator kriging in the analysis of gold mineralisation data from the Goodall gold mine in the Northern Territory of Australia. The sample data used in this study consist of exploration drilling data, suitably composited, taken from the 540mRL bench of the main open cut. The data set is sparse, irregularly spaced and has a highly skewed distribution. An enlarged data set, used for the variography, was obtained by including exploration data from 20m above and 20m below the bench. In addition, the blast hole data from the chosen bench were considered as a model of reality. Careful choice of the common semivariogram model for median indicator kriging was necessary. Grade tonnage curves from median indicator kriging and full indicator kriging were compared with one another and with the grade tonnage curve of the blast hole data. There was little difference in the grade tonnage curves for the two indicator kriging methods. In both cases the grade tonnage curves fell below that obtained from the blast hole data.
Key Words: indicator kriging, median indicator mineralisation, grade tonnage curve.
kriging,
gold
Introduction
In this study we compare median indicator kriging (mIK) and full indicator kriging (fIK) in the analysis of gold mineralisation data. mIK is sometimes chosen in preference to fIK merely because it is less time consuming since the same semivariogram model is used at all cut offs, even when the assumptions for mIK are only approximately satisfied . The question then arises as to how much, if any, detail is lost by this choice, particularly in the case of sparse, irregularly spaced and highly skewed data. The data used in this comparison come from the Goodall gold mine which is located in the Northern Territory of Australia, approximately 150km from
Hill et al Comparison of Median and Full Indicator Kriging Darwin. It falls within the Pine Creek 1:250 000 map sheet, at 8 525 000 mN: 750 000 mE. The mine produced 4.095 million tonnes of ore at a head grade of 1.99 g/t Au during openpit operations. Mining was completed in 1992 (Quick, 1994). For the Goodall mine both the exploration drilling data and the mining-stage blast hole data are available. Assuming the close-spaced blast holes approximate reality, predicted grade tonnages can be compared with actual grade tonnages to assess the performance of the two indicator kriging methods.
Data sets
The data used in this study are from the A-pod orebody of the main open pit. The first set comprises gold analyses from 21 inclined diamond and reverse circulation (RC) holes drilled during the exploration stage (see Figure 1).
Figure 1: Exploration data composites from within the bench.
The summary showed no substantial difference between the samples obtained from these two drilling types. We will refer to this data set as the exploration data set. The second set consists of gold analyses from 2.5m long blast holes drilled into the 540m RL bench. Both data sets were composited in such a way that samples represent 2.5 m vertical thickness at the same RL (Kentwell, 1997). The composited data were divided into mineralised and non-mineralised populations with a cut off of 0.5 g/t Au used to define the orebody. The part of the A-pod orebody analysed is located between mine co-ordinates 10 800 11 100N and 10 130 10 210E (see Figure 4 in Quick(1994)). Because of the small size of the exploration data set, an enlarged subset, referred to as the variography data set, containing additional drilling samples from up to 20m above and 20m below the 540m RL bench was used for variography. The total size of this set is 638 samples (see Figure 2). The location of the data set for IK in relation to the variography data set is shown in Figure 3.
Figure2.
Exploration data composites from up to 20m above and
20m below the bench.
560 550
Elevation
540 530 520 1013010140 10150 10160 1017010180 1019010200 11300 11200 11100 11000 10900 10800 10700 10600
Northing
Easting
Figure 2.
Exploration data composites data from the bench of
interest shown in dark.
Blast holes were drilled on a four metre by two metre grid. The size of the data set was reduced to represent 720 points on a four metre by four metre grid. The spatial distribution of grades is shown in Figure 4. Since only a single bench was used, this set was treated as two-dimensional (Kentwell, Bloom & Comber, 1997).
Figure 4.
Blast hole data from within the bench.
Summary statistics for the three data sets (see Table 1) exhibit the highly skewed nature of the distribution. There appears to be some difference between the distribution of the blast hole data and the variography data, which is not unexpected considering the different spatial extent and sample density of the two data sets.
Table 1: Summary statistics for the grades (g/t) of the exploration, variography and blast hole data. Data set Blast hole Exploration Variography N 720 21 638 mean 2.766 2.274 1.736 median 1.425 1.250 0.885 standard deviation 4.620 2.920 2.526 minimum 0.010 0.050 0.000 maximum 49.300 12.680 25.880 skewness 4.800 2.566 4.003
Variography
For the purposes of modelling the spatial structure and estimation all data will be treated as point data. Since the exploration data set is from a single bench the vertical bandwidth for calculating the experimental indicator semivariograms from the variography set was set to one. The parameters used in the calculation of each semivariogram are given in Table 2.
Table 2: Experimental semivariogram parameters Lag spacing Lag tolerance Number of lags
10m 5m 10
53

Vertical bandwidth Angular tolerance 1m 30
The semivariogram value for a lag was discarded if the number of pairs contributing to the semivariogram was less than 15 for that lag. Due to the long narrow shape of the mineralised zone, the experimental semivariograms in the East-West direction were only reliable for the first two to three lags. More emphasis was therefore placed on fitting a model for the North-South direction. Most semivariograms had a similar overall sill for both directions, so it was decided to fit models with purely geometric anisotropy at all cut offs. Indicator semivariograms were modelled for twelve cut offs, the deciles and the 95th, 97.5th and 99th percentiles. Semivariograms were standardised by dividing each semivariogram by the total sill, as suggested in Goovaerts (1997). Because of the sparseness of the exploration data set only five of these cut offs were used for indicator kriging (see Figure 5). A single spherical structure plus nugget was fitted in each case. Table 3 shows the standardised semivariogram model parameters for the five cut offs used and the 50% cut off.
Table 3: Standardised semivariogram model parameters.
Cut offs Percentage 20 40 50 60 80 95 Grade 0.30 0.66 0.88 1.23 2.39 6.05
Semivariogram parameters spherical model plus nugget Nugget Partial sill Anisotropy factor Maximum range 0.50 0.50 40 0.500 0.40 0.25 0.50 0.40 0.25 0.60 0.75 0.50 0.60 0.75 35 35 40 45 45 0.343 0.286 0.375 0.667 0.667
Indicator Kriging
The nugget to sill ratios for the semivariograms at the various cut offs are sufficiently similar to make the use of mIK an attractive option in this case. The initial estimation for mIK was performed with the 50% semivariogram. The range of the median semivariogram is only 35m in the direction of maximum continuity and 10m in the direction of minimum continuity, compared with 40m and 15m respectively for the 60% semivariogram. Because of the sparseness and irregular spacing of the exploration data set, for approximately 24% of all locations the distance between the closest sample datum and the point to be estimated was greater than the semivariogram range of the 50% semivariogram. At these locations all kriging
Hill et al Comparison of Median and Full Indicator Kriging weights were zero and the global cumulative distribution function was reproduced as the estimate. The slightly longer semivariogram range of the 60% semivariogram halved the number of locations at which this occurred. For this reason the 60% semivariogram was chosen as the common semivariogram model for mIK. Full indicator kriging of the exploration data set was performed for the cut offs at 20%, 40%, 60%, 80% and 95%. Non-estimation was less of a problem with fIK where ranges of the individual semivariograms varied from 35 m to 45 m (see Table 3).
1.5
1.5
gamma (h)
gamma (h)
0 20 40 60 80 100
0.5
0.5
0 0 20 40 60 80 100
h
1.5 1.5
gamma (h)
gamma (h)
0 20 40 60 80 100
0.5
0.5
0 0 20 40 60 80 100
h
1.5 1.5
gamma (h)
0.5
gamma (h)
0 20 40 60 80 100
0.5
0 0 20 40 60 80 100
h
1.5 1.5
gamma (h)
gamma (h)
0 20 40 60 80 100
0.5
0.5
0 0 20 40 60 80 100
55

1.5 1.5
gamma (h)
gamma (h)
0 20 40 60 80 100
0.5
0.5
0 0 20 40 60 80 100
Figure 5: Experimental and model semivariograms (top to bottom) for 20%, 40%, 60%, 80% and 95% cut offs, in North-South (left) and East-West (right) directions for exploration data.
Order relation deviations

Very few order relation deviations were encountered. There were 0.87% with an average magnitude of 0.0113 for fIK and none for mIK. These were corrected by averaging the cdf derived from an upward correction and the cdf derived from a downward correction as described in the GSLIB Users Guide (Deutsch and Journel, 1992). It was noted that slight changes to the semivariogram model affected both the amount and average magnitude of order relation deviations. For example, removing the nugget from the model at the 95th percentile caused a rise in the number of deviations to 3.06% with an average magnitude of 0.0451.
Results
Figures 6 and 7 show probability maps produced by mIK and fIK respectively. The smoothing nature of kriging is apparent, with neither method producing any probability estimates greater than 0.5 for the 20% cut off. Also fIK produced only one small region with probabilities lower than 0.5 at the 95% cut off, while mIK did not produce any. Hence there appears to be little difference between mIK and fIK, particularly at the lower cut offs. At the two upper cut offs mIK underestimates high grades slightly more than fIK does. Overall the location of regions of high and low probabilities matches the blast hole data in Figure 8.
56
57
Figure 6: mIK estimates for 20%, 40%, 60%, 80% and 95% cut offs.
58
Figure 7: fIK estimates for 20%, 40%, 60%, 80% and 95% cut offs.
59
Figure 8: Blast hole data indicators for 20%, 40%, 60%, 80% and 95% cut offs.
Grade tonnage curves

All data were treated as point data in two dimensional space rather than block data. The grade tonnage curves were calculated on the assumption that each value of the blast hole data and each estimate is representative of a block of size 4m by 4m by 2.5m centred on the location of the datum. No block support correction was applied because point estimates were compared with point support grade control data. To calculate the average grade above each cut off it was necessary to extrapolate above the highest cut off. As the distribution is highly skewed and a maximum possible value is not known, a hyperbolic extrapolation was used as given in the GSLIB Users Guide (Deutsch and Journel, 1992). It is given by
( z ) Hyp. = 1
z z > z K
where is a parameter greater than or equal to one, zK is the grade of the maximum cut off and is a constant such that (zK) = *(zK), the sample cumulative frequency. Figure 9 shows the grade tonnage curves produced from the mIK and fIK estimates using parameters = 1.5 and max = 0.995 together with the actual grade tonnage curve. .
60
Average Grade Tonnage Curves

18 16
blasthole
14
full IK
average grade above cutoff
12
median IK
10 8 6 4 2 0 0 10000 20000 30000 40000 50000 60000 70000 tonnage above cut off
Figure 9: Average grade tonnage curves for mIK, fIK and blast hole data.
Both grade tonnage curves produce tonnage estimates which underestimate the actual value at the chosen cut off. For each cut off chosen the estimates derived from mIK and fIK are almost identical, with fIK producing slightly higher values. However, even though mIK and fIK underestimate the average grades at the lower cut offs, they overestimate the average grades at the two highest cut offs.
Conclusions
This study reinforces the theory that little is lost by using the more time-efficient mIK rather than the more involved fIK. This is true even here where we are dealing with a highly skewed, sparse data set. However, as indicated earlier, care must be taken in choosing the cut-off value for the common semivariogram to be used in the mIK approach. Even though only one semivariogram is needed it may be worthwhile to model the semivariogram at several cut offs close to the median in order to ensure a sufficiently large range is used in the kriging procedure. For large data sets this may not be important, but for sparse data sets such as the one we used, a judicious choice for the cut off can help to minimise the number of locations at which only the global cumulative distribution function is used.
Acknowledgements
The authors would like to thank WMC Resources for making available the raw data (exploration and blast hole) from the Goodall mine. Thanks go also to D. J. Kentwell (a former Edith Cowan University graduate student now at SRK Consulting) for
61
Hill et al Comparison of Median and Full Indicator Kriging allowing us to use his composited data for this study and for making available the code for the bounding polygon used to delimit this irregularly shaped region.
References
Deutsch, C. V., & Journel, A. G. (1992). GSLIB: Geostatistical software library and users guide. New York: Oxford University Press. Goovaerts, P. (1997). Geostatistics for natural resources evaluation. New York: Oxford University Press. Kentwell, D. J. (1997). Fractal relationships and spatial distributions in ore body modelling. Unpublished masters thesis, Edith Cowan University, Perth, Western Australia. Kentwell, D. J., Bloom, L. M., & Comber, G. A., (1997). Improvements in grade tonnage curve prediction via sequential Gaussian fractal simulation. Mathematical Geology. (to appear). Quick, D., (1994). Exploration and geology of the Goodall gold mine. Proceedings of the AUSIMM annual conference, Darwin August 1994. 75-82.
62
Humphreys Case study in Uniform Conditioning, Wandoo project
LOCAL RECOVERABLE ESTIMATION: A CASE

STUDY IN UNIFORM CONDITIONING ON THE WANDOO PROJECT FOR BODDINGTON GOLD MINE
Michael Humphreys Geoval

Abstract
A practical case study for estimation of the large, low grade Wandoo deposit at Boddington Gold Mine is presented. This was broken down into seven zones, primarily by geology. One of these zones was a higher grade vein set that could be solid modelled and estimated separately. A geological mineralisation envelope was applied to constrain the estimation. This represented a broad, relatively continuous envelope in keeping with the geology of the orebody and was loosely based on a 0.1 g/t Au cut-off. Data was composited to 9m to reflect the intended mining bench height and the true variability expected from those benches. Waste dykes were only excluded from this compositing if they were considered large enough to be practically avoided when mining. Variograms of the composited data were not particularly clear for either Au or Cu, and a Gaussian transform was applied to help determine a model. Gold and copper do not exhibit the same anisotropy and there is significant variability at less than, or equal to ,the average drill spacing. Tests were conducted to determine the suitability of the Gaussian approach. This approach best represents a diffusion model of spatial continuity. These tests indicated that the Gaussian approach was suitable at Wandoo. A global evaluation of resources was carried out using the Discrete Gaussian Model. This was a first quick estimate of resources at cut-off that was useful as an order of magnitude check on the final local estimates. There was a requirement to represent selectivity in mining for a local resource estimate. It was unrealistic to try to achieve this by directly estimating such a small block size taking into consideration the average drillhole spacing. Therefore, the technique of Uniform Conditioning was applied to calculate the expected proportions above a cut-off. Previous experience has shown that this is a relatively robust method. Key Words: uniform conditioning, non-linear estimation, Discrete Gaussian Model, mining selectivity.
Introduction
Study data are from the Wandoo deposit at Boddington Gold Mine (BGM), located approximately 130 km southeast of Perth, Western Australia. The numbers present have been modified for confidentiality. Mineralisation is hosted by intermediate volcanic and intrusive rocks of the Archaean Saddleback Greenstone Belt. Unmineralised dolerite dykes transect the sequence. The estimation area has been broken up into seven distinct geological zones. One of these zones consists of two solid modelled, steeply-dipping actinolite veins a few metres wide. These are generally associated with higher grades. The estimation was confined to unoxidised host rocks.
Data
There are 2589 drillholes with over 118000 samples, mostly on 2m lengths, from diamond (DDH) and reverse circulation (RC) drilling (Figure 1). These were coded for rock type and zone. Exploration holes are variously spaced, but average a 25m x 25m spacing, with some zones drilled more sparsely. Inclinations vary from vertical to sub-horizontal. Hole azimuths also vary widely. A 25m x 25m x 9m block model was supplied, defining the geological zones, blocks to be estimated and rock type. This block size was chosen from the data spacing and mining considerations. Smaller blocks could not be used without possibly serious under-estimation of the variability (Vann and Guibal, 1998). An outer boundary to possible mineralisation was created by BGM geologists at a 0.1 g/t Au cut-off. This boundary was relatively insensitive to increases in cut-off up to approximately 0.5 g/t Au. An advantage of defining the outer boundary at a geological cut-off was that it allowed the application of different mining cut-offs within this boundary. Conversely, estimating with a high mining cut-off initially would probably require re-estimation if lower cut-offs were subsequently contemplated. It is very important in estimation to work with equal support (volume) samples. This is why the data were composited to equal lengths. A bench height of 9m was envisaged, therefore the samples were composited to that length along drillholes within the geological envelope (excluding defined, barren dolerite samples) to best represent real 9m bench variability. For this estimation, some dolerite was considered as unavoidably mined, and included. Dykes that were large enough to be easily excluded when mining, were excluded from compositing and estimation. Not to do so may bias the estimation.
64
Humphreys Case study in Uniform Conditioning, Wandoo project In gold, the effect of outlying values is usually significant and some approach must be taken to account for these. There is no single accepted method for determining upper cuts with theoretical justification available, and no strong argument to choose one method over another. A final cut of 40 g/t Au was employed.
Figure 1 Drillhole Location Map
Two tests for sensitivity of grade variability to cut-off were carried out (Figure 2) in each zone. These showed that (i) removing approximately the highest 10 values, and (ii) employing an upper cut of around 40 g/t Au, both reduced the outlier effect significantly.
65
Humphreys Case study in Uniform Conditioning, Wandoo project RC and DDH composites were compared in an area that was adequately covered by both datasets. Tests showed little difference in the statistical characteristics. This, along with the greater volume of RC drilling, led to the decision to keep both sets of data for the estimation. Combining data types should not be an automatic decision, but one consciously made with supporting results.
Mean Grade and CV by Number of High Grades Removed for Zone 1
2.5
Mean
"CV"
Grade (g/t), CV
2 1.5 1 0.5 0 0 20 40 60 80 100
Number of Highest Grades Removed
Mean Declustered Grade and CV by Upper Cut for Zone 1

2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 10 30 50 70 90 110 130
Mean
"CV"
Grade (g/t), CV
Upper Cut (g/t)
Figure 2 Sensitivity tests to cut-off grade
The spatial distribution of data is not uniform due to an irregular drilling grid, varied length and inclination of holes. There fore we used a Declustering (weighting) procedure so that statistics were not biased by preferential spatial position (eg many close-spaced holes in a high grade area). This does not decrease the number of composites used, but simply weights the histogram to produce an unbiased mean and variance. A simple Declustering usees the number of points in a block but we
Humphreys Case study in Uniform Conditioning, Wandoo project required the more accurate method using kriging weights which is far more time consuming. Table 1 shows the weighted means and variances compared to the original statistics. The weighted variances and means are generally lower than the unweighted statistics.
Au Number Mean Variance Weighted Mean Weighted Variance Cu Mean Variance Weighted Mean Weighted Variance
Zone 1 8460 1.14 2.82 1.07
Zone 2 4240 0.64 0.79 0.64
Zone 3 3339 0.88 4.06 0.66
Zone 4 1178 0.48 0.54 0.43
Zone 5 2078 1.21 4.14 0.91
Zone 6 776 1.20 11.22 0.87
Zone 7 129 10.17 168.7 8.78
2.15
Zone 1 1416.2 1546842 1472.1
1.30
Zone 2 1650.9 1572393 1599.7
2.02
Zone 3 1057.9 1256807 939.9
0.32
Zone 4 646.5 272742 645.1
2.21
Zone 5 503.1 318646 511.0
3.35
Zone 6 1556.9 1233720 1409.3
134.9
Zone 7 2234.6 3177538 1905.5
1536447 1592774 921811
297998
265296
1100511 2766932
Variograms were then calculated for Au and Cu in each of the zones on the 9m composites. These were not particularly clear and a Gaussian transform was employed to help define the underlying structure. Calculation of the Gaussian transform utilised the Declustering weights previously discussed. Variograms of the Gaussian transformed data presented clearer structures and were more easily modelled. Models were fitted in consultation with BGM geologists taking into account the known geological and mineralisation trends (see Figures 3 and 4). Models for the Gaussian variables were then transformed back to models on the raw data which will now reflect the declustered 3D spatial variability. The Gaussian transform is very powerful, and is part of the Discrete Gaussian method (diffusion methods). It is the only method having a built-in change of support to reflect a deskewing of the histogram for different volumes (eg samples versus blocks). Three tests were conducted to see whether the Discrete Gaussian model was applicable at Wandoo, as follows:
The Gaussian transform is very powerful and is part of the Discrete Gaussian Method (diffusion methods). It is the only method having a built-in change of support to reflect a deskewing of the histogram for different volumes (eg samples versus blocks). Three tests were conducted to see whether the Discrete Gaussian Model was applicable at Wandoo as follows: 1. checking indicator residuals (see Rivoirard, 1994, for example). If these show some spatial correlation (as seen in cross variograms) then a Gaussian approach is justified.
c:\msoffice\winword\template\normal.dot
Figure 3 Zone 1 Gaussian Au Variograms
68
Figure 4 Zone 1 Gaussian Cu Variograms
2. Using the Deutsch and Lewis normalised indicator approach. If the Gaussian reconstructed indicator variograms are the same as those calculated immediately from the data, then the Gaussian approach is justified (see Figure 5). 3. Checking ratios of indicator variograms. If the ratio of cross variogram to variogram increases with distance, then a diffusion or Discrete Gaussian Model is applicable (see Figure 6). All these conditions were satisfied (see Figures 5 and 6).
69
Theoretical indicator variogram for 50th percentile

0.50 0.40 0.30 0.20 0.10 0.00 0 50 100 150
Variogram
Major Semi-major Minor
Distance, m
Figure 5a Gaussian reconstructed indicator variograms (test 2)
Experimental variogram for 50th percentile

0.50
Variogram
0.40 0.30 0.20 0.10 0.00 2 60 120 180
Major Semi-major Minor
Distance, m
Figure 5b Indicator variograms (test 2)
70
Figure 6 Indicator variogram ratio test for Zone 3 (test 3)
Knowing that the Discrete Gaussian Model was applicable a global estimate was made (see also Vann and Sans, 1995 or Guibal, 1987). By modelling the composite histogram and knowing the Gaussian transform function, the histogram of any size block that we want to consider for estimation can be obtained. This gives a prediction of the global tonnes and grade above a cut-off and is a good first pass approximation to a local result. The last step before local estimation is to test the estimation/neighbourhood parameters. This is important and is all too rarely performed. A consequence of not testing and using too small a neighbourhood would be a biased, poor quality and badly representative estimate. Many different configurations were tested and results compared for estimation variance, bias (slope of the regression of true value with the estimated value) and weight of the mean (a measure of the need for closer spaced and/or more data in the neighbourhood). For further discussion see Armstrong and Champigny (1989), Krige (1994, 1996a, 1996b and 1997), Ravenscroft and Armstrong (1990) and Royle (1979). Examining the kriging weights can also help determine if large negative weights or other possible problems exists. Table 2 shows some results. It is desirable that the
71
Humphreys Case study in Uniform Conditioning, Wandoo project weight of the mean is below 10%, the slope of the regression is close to 1.0 (above 0.9 is preferable) and that estimation variance is minimised.
Table 2 Kriging Neighbourhood Test Results for Zone 1
Sample grid No. of informing composites Estimated block size Ordinary kriging result Estimation variance Slope of the regression Z/ZE Simple kriging result Weight assigned to the mean Sample grid
25 x 25 25 x 25 25 x 25 25 x 25 25 x 25 25 x 25 25 x 25 25 x 25 x9 x9 x9 x9 x9 x9 x9 x9 3x3x5 3x3x7 3x3x9 3x3x 3x5x5 3x5x7 3x5x9 3x5x 11 11 25 x 25 25 x 25 25 x 25 25 x 25 25 x 25 25 x 25 25 x 25 25 x 25 x9 x9 x9 x9 x9 x9 x9 x9 0.0971 0.9029 0.0911 0.9382 0.0886 0.9580 0.0874 0.9701 0.0937 0.9524 0.0893 0.9745 0.0874 0.9862 0.0866 0.9931
0.1179
0.0773
0.0539
0.0393
0.0647
0.0361
0.0201
0.0104
No. of informing composites Estimated block size
25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 9 9 9 9 9 9 9 5x5x5 5x5x7 5x5x9 5 x 5 x5 x 7 x5 x 7 x9 x 9 x 11 11 13 13 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 25 x 6 6 6 6 6 6 6 0.0884 0.9989 0.0018 0.0869 1.0042 -0.0072 0.0862 1.0071 -0.0125 0.0862 1.0090 -0.0188 0.0859 1.0093 -0.0203 0.0858 1.0024 -0.0122
Ordinary kriging result 0.0920 Estimation variance Slope of the regression Z/ZE 0.9877 Simple kriging result Weight assigned to the mean 0.0190
The Gaussian model, the variograms and the results of neighbourhood testing provide the parameters necessary for the kriging estimation and the non-linear local estimation by Uniform Conditioning. Ordinary kriging was performed but this could not be used to give a resource reflecting the real mining selectivity. Kriging of smaller blocks would seriously understate the true variability. Therefore, Uniform Conditioning was applied to obtain a more realistic resource estimate corresponding to the intended mining selectivity. A selective mining unit (SMU) of 8.3 x 8.3 x 9m was used as the minimum basis for determining ore or waste parcels. Uniform Conditioning (Rivoirard, 1994) takes the locally estimated ordinary kriging result and applies a change of support to calculate the expected histogram of grades for that 25 x 25 x 9m block based on an SMU of 8.3 x 8.3 x 9m. Results are then reported for each large block as the proportion of the block above cut-off and the
Humphreys Case study in Uniform Conditioning, Wandoo project grade above cut-off from that expected SMU histogram, knowing the estimated grade of the entire block. Histograms comparing composites, kriged 25 x 25 x 9m blocks and SMU results are given in Figure 7. These show the expected deskewing effect of larger block sizes. The Discrete Gaussian Method is one of the few approaches that takes this important deskewing into account. Affine corrections do not for example.
Figure 7 Comparing Histograms for Different Supports
The global results obtained previously can be used as an order of magnitude check for the Uniform Conditioning local results. In this study, agreement between global and local results was good. If alternative cut-offs are required then it is simply a matter of
Humphreys Case study in Uniform Conditioning, Wandoo project running only the Uniform Conditioning step with the new cut-offs no other work need be re-done.
Acknowledgements
The help, assistance and permission of the Boddington joint venture comprising Normandy Mining, Acacia Resources and Newcrest Mining is gratefully acknowledged. Thanks to David Burton and Simon Williams who provided assistance on the project concerned. Thanks also to Olivier Bertoli, Henri Sanguinetti and Daniel Guibal of Geoval who provided much valuable assistance and input.
References
Armstrong, M. And Champigny, N., 1989. A study on kriging small blocks. CIM Bulletin. Vol. 82, No. 923, pp. 128-133. Deutsch, C.V. and Lewis, R.W., Advances in the Practical Implementation of Indicator Geostatistics: Appendix A: A Test for the Validity of Parametric Methods. 23rd APCOM Proceedings Guibal, D., 1987. Recoverable Reserves Estimation at an Australian gold project. Geostatistical Case Studies, G. Matheron and M. Armstrong (eds.), pp149168. Kluwer Academic Publishers. Krige, D.G., 1994. An analysis of some essential basic tenets of geostatistics not always practised in ore valuations. Proceedings of the Regional APCOM, Slovenia. Krige, D.G., 1996a. A basic perspective on the roles of classical statistics, data search routines, conditional biases and information smoothing effects in ore block valuations. Proceedings of the Regional APCOM, Slovenia. Krige, D.G., 1996b. A practical analysis of the effects of spatial structure and data available and accessed, on conditional biases in ordinary kriging. In: Geostatistics Wollongong 96. Proceedings of the International Geostatistical Congress, Wollongong, NSW, Australia, September, 1996, pp. 799-810. Krige, D.G., 1997. Block kriging and the fallacy of endeavouring to reduce or eliminate smoothing. Proceedings of the Regional APCOM, Moscow.
74

Ravenscroft, P.J., and Armstrong, M., 1990. Kriging of block models the dangers re-emphasised. Proceedings of APCOM XXII, Berlin, September 1721, 1990, pp. 577-587. Rivoirard, J., 1994. Introduction to disjunctive kriging and non-linear geostatistics. Clarendon Press (Oxford), 181pp. Royle, A.G., 1979. Estimating small blocks of ore, how to do it with confidence. World Mining, April 1979. Vann J. and Guibal D., 1998. Beyond Ordinary Kriging An overview of nonlinear estimation. Beyond Ordinary Kriging Seminar, Geostatistical Association of Australasia, Perth, WA (this volume). Vann J. and Sans H., 1995. Global resource estimation and change of support at the Enterprise Gold Mine, Pine Creek, Northern Territory Application of the geostatistical Discrete Gaussian model. Proceedings of the APCOM XXV 1995 Conference (Brisbane), Aus. I.M.M., PP. 171-180.
75
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit
A CASE STUDY USING INDICATOR KRIGING THE MOUNT MORGAN GOLD-COPPER DEPOSIT, QUEENSLAND
Ivor Jones WMC Resources
Abstract
In August 1882, the Morgan brothers recognised a mineral deposit, now known as the Mount Morgan Gold-Copper Deposit. The final production figures for the mine were 250 tonnes of gold and 360,000 tonnes of copper from 50 million tonnes of ore, making the average grades 4.99g/t gold and 0.72% copper. A three dimensional grade model was made of the pre-mined gold distribution within the Main Pipe mineralisation between the Slide Fault (to the west) and the Andesite Dyke (to the east), and bound to the south by the South Dyke. Indicator kriging provided a method for estimating the grade in the strongly skewed gold distribution, without the problems of smearing of the high grades as seen in linear techniques. The application of indicator kriging using grade thresholds based on the declustered sample decile values was shown to be a poor application of indicator kriging, but was greatly improved by modifying grade thresholds above the median so that the amount of contained metal was evenly distributed between these classes. The pre-mined resource estimate for this portion of the Main Pipe mineralisation using a 2g/t lower selection limit was 3,526,800 tonnes with an average grade of 11.98g/t, equivalent to 42.25 tonnes gold.
Key Words: geostatistics, non-linear grade estimation, indicator kriging.
Introduction
The Mount Morgan Mine in central Queensland was for many years described as the greatest gold mine on earth with grades as high as 2000 ounces per tonne. The aim of the study was to prepare a detailed three dimensional computer model of the grade distribution within a select area of the Mount Morgan deposit (pre-mining), and to compare some different estimation techniques. Indicator kriging was used to estimate the grades in this paper.
76
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit The study area is a section of the Main Pipe mineralisation between the Andesite Dyke and the Slide Fault, and north of the South Dyke (Figure 1). The northern, lower and upper boundaries are those of the known data.
Slide Fault
South Dyke
Andesite Dyke
Figure 1 Location of the Study Area in relation to the major geological features within the Mt Morgan deposit.
The Main Pipe mineralisation was a concentrically zoned sub-horizontal pipe-like orebody (Jones and Golding, 1994), with the highest grades in the centre of the pipe. At the western end of the study area, there is a narrow vertical high grade gold shoot. The aim of the study was to prepare a detailed three dimensional computer model of the grade distribution within a select area of the Mount Morgan deposit (pre-mining). The model itself was used to investigate the three dimensional grade distribution of the deposit relating it to the recorded geology using computer visualisation. In order to do this, a three dimensional computer model of the distribution of metal within the study area was prepared, and indicator kriging was used for grade estimation. A comparative study using two different methods of determining grade thresholds in the study was also performed. Indicator kriging, as proposed by Journel (1982), has been a well accepted technique by the geostatistical community for dealing with skewed distributions and extreme values. It is a non-parametric estimation procedure, is not based on the data fitting a particular statistical distribution, is resistant to the influence of outliers, and is based on the knowledge that different parts of a mineralisation can have different spatial
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit characteristics. High grade mineralisation may be limited in spatial extent, and in strongly skewed distributions, the contained metal can contribute a significant proportion of the ore reserves (Journel and Arik, 1988), although not all high grade occurrences may have been sampled. Alternatively, low grade mineralisation may be pervasive, and spatially extensive. It was therefore a suitable procedure for gold grade estimation in the Mount Morgan Deposit.
Data analysis
Underground mining at Mount Morgan was primarily by square set stoping, and large open stopes or chambers. A face sample was taken for each square set and assayed for gold, copper and in select areas silica, the assays being recorded on level plans. The square sets, were approximately 5 feet by 6 feet square by 7 feet 9 inches high (Patterson and Thomas, 1910). The assays for each square set from every fourth level were digitised as a point representing the centre of the square set location. The square set stope data was the basis of this study.
Moving window statistics

Moving Window Statistics were calculated for the study area as well as the rest of the Main Pipe mineralisation. They were used as a tool for examination of the data in a spatial context, and as a tool for determining if high grade zones could be separated from the remainder of the deposit for modelling purposes. The contoured plans of the moving window statistics (Figure 1) and visual observations of the digitised data indicate that the high grade shoot had gradational boundaries. It was therefore not split into a domain separate to the main data.
Statistical summary of data distribution

The histogram of the Au data (Figure 2) shows a distribution that is positively skewed, the histogram of the logarithm of Au is roughly symmetrical, and the logprobability plot is relatively smooth and roughly straight with a high grade population of data above around 80g/t Au.
Preparation of indicator variables for Au

The principle of indicator kriging involves the transformation of the original random variable into indicator data instead of using the variable itself. The thresholds are chosen from the histogram of the data such as z1, z2, z3, z4 and z5, respectively the 10th, 30th, 50th, 70th and the 90th percentiles (Isaaks and Srivastava, 1989). Isaaks and Srivastava (1989) also comment that it is important to include more thresholds where the information is required, for instance in a precious metal deposit there might be more thresholds in the higher grades.
78
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit In the first Indicator Kriging study, the indicators I1, I2, I3, ..., I9 had thresholds of the 10th, 20th, 30th, ..., 90th, percentiles.
79
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit
Figure 2, the de-clustered histogram (top), log-histogram (centre) and log-probability plot (bottom) of the Au data.
Calculation of the contained metal represented by samples above the threshold grade was by spread sheet. The samples, all being the same length, were multiplied by their de-clustered weight and grade in order to get the weighted value for the histogram from the de-clustered data. The samples were then sorted by grade, and the cumulative totals of the weight*grade determined. The cumulative total at each of the sample grades is then a relative representation of the contained metal below that grade within the study area when compared to the cumulative total for the highest grade. The contained metal represented by samples above that grade are then the cumulative total for the highest grade minus the cumulative total at the threshold grade. In the second Indicator Kriging study, the thresholds were modified to balance the contained metal in the high threshold classes. More than 80 percent of the contained metal was represented by the data above the 50 percentile (10.73 g/t Au), and more than 50 percent of the data was above the 80 percentile (22.48 g/t Au). 35 percent of the contained metal was represented in the top decile of the data. The disproportionate amount of metal occurring in the top classes was unsatisfactory with decile thresholds (not enough thresholds to define the distribution and control the amount of contained metal), and in the second Indicator Kriging study, the data threshold approach was modified using the quantity of contained metal. The criteria for selection of the indicator thresholds in the second Indicator Kriging study was as follows:
Thresholds below the median grade (as defined by the de-clustered data) were selected on every tenth percentile (the decile values). Thresholds above the median grade were selected on the basis of the cumulative metal content above the median grade. The classes were divided up so that each class represented approximately 10% of the contained metal in the study, as represented by the samples.
There were 13 thresholds in the secondary indicator kriging study, including 8 above the median.
Indicator variograms for the Au variable

Indicator variograms of the Au variable were prepared in this study for grade threshold values based on the Au variable grade thresholds. Full directional variography studies were prepared by firstly determining any axes of anisotropy of the spatial continuity for the indicator variables at each threshold, calculating experimental indicator variograms for each of the axes of anisotropy determined for
80
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit the indicator variables at each threshold, and finally preparing a mathematical model for each of the variograms. Determination of the principal axes of anisotropy by planimetric variogram maps. Experimental indicator variograms were not calculated for threshold grades below the 20 percentile (3.17 g/t Au) or above the 82 percentile (23.72 g/t Au) as there is not enough data in the class above or below to prepare stable and valid variograms suitable for interpretation. Variograms applied to lower or higher thresholds were those of the nearest threshold with an experimental variogram. In relation to the mineralisation, the determined orientation of the major axis of the mineralisation is approximately that of the core of the Main Pipe mineralisation. Variogram models were prepared for directional variograms for each grade threshold and the model parameters recorded for grade estimation. Apparent zonal anisotropy between the directions 1 and 2, and direction 3 (as for lower grade thresholds) is caused by a mixture of similarity in grades in the direction of principal anisotropy, and the influence of hole effects across the mineralisation in directions 2 and 3. The directions of the Zc = q0.50 or median variogram are very similar to those of the pair-wise relative variograms for the Au variable that were interpreted separately. The median indicator experimental variograms are presented in figure 3 with fitted models.
Figure 3 Experimental indicator variograms for Zc = q0.50 The variogram for the principal axis of anisotropy is variogram (1), the variogram for the semi-major axis is variogram (2) and the variogram for the minor axis is variogram (3).
Grade estimation of Au by indicator kriging

Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit The Au variable in this study has a distribution that is positively skewed and characterised by mixed sample populations in the data. Indicator kriging has been used for estimating the grade distribution in this study. As discussed already, two estimates by indicator kriging were prepared. The first indicator kriged estimate (IK1) used the nine decile thresholds, a power function to estimate the distribution in the lower class and a hyperbolic function for the top class. The second indicator kriged estimate (IK2) used threshold values determined using the decile values for the first five grade thresholds (up to the median), the balance of metal to determine the higher thresholds (8 classes representing 10% contained metal in each class), a power function to estimate the distribution in the lower class and a hyperbolic function for the top class. The required result from the indicator kriging was an expected average value of each block (an E-type estimate). The E-type estimate for the block is first determined using the probabilities and the mean grade of each class for each cell.
Estimation of the lower tail distribution

The lower class (the class below the lowest threshold of the indicators) was estimated using a power model (Deutsch and Journel, 1992). The power model was used for the lowest class estimate because, even though the class is the lowest class, the grades are still of economic interest. The calculated model was plotted against the de-clustered data on a scatter plot (percentile versus value) to ensure the fit of the power model to the de-clustered data. A p value of 2.1 was determined to provide a relatively good fit to the data. The power model was the same for both the IK1 and IK2 indicator kriged estimates, as there was no change to the lowest class threshold between these estimates.
Estimation of the upper tail distribution

The top class (above the highest grade threshold of the indicators) was estimated using a hyperbolic model (Deutsch and Journel, 1992). The hyperbolic models for the top class in the IK1 and IK2 estimates were determined using an excel spreadsheet. The calculated hyperbolic models were plotted against the de-clustered data to ensure the models matched the data.
Order relations corrections

Order relation corrections were calculated as an average of the upper and lower values. There was a relatively small number of order relation corrections in the lower classes is because the variograms and their orientations are roughly consistent between neighbouring thresholds. The largest number of order relation corrections is in the
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit grade thresholds near the middle of the distribution, but the average correction was in the order of 1%.
Correction for the volume-variance relationship.

In order to get block estimates for recoverable reserves from indicator kriging, a volume-variance correction is often used as indicator kriging produces point estimates. Individual recoverable reserves in blocks were not required for this study, and so no volume-variance correction was applied.
Validation of indicator kriged estimates

After checking the order relation corrections, the indicator kriged estimates were checked by examination of the mean against the declustered data mean, visual examination of the estimates in relation to the data, and examination of scatter plots of moving window statistics for raw data against the average of all the block estimates in the same window. The simplest validation is to look at the statistics of the data, and compare the declustered mean of the raw data to the mean of the estimates. Table 1 shows that the estimates have a lower mean value than the de-clustered raw data, and that the IK1 estimate had a relatively low maximum value.
Table 1 Comparison of statistics of gold for de-clustered data with indicator kriged estimates.
mean value Median Range
De-clustered data Au g/t 12.41 10.73 787
Indicator kriged IK1 estimate g/t 10.98 8.51 86.31
Indicator kriged IK2 estimate g/t 11.40 8.59 125.17
A factor in the lower mean grade is the inability of the estimates to reproduce the extreme values. Removal of the four top high grade values from the de-clustered data reduced the mean of the de-clustered data by approximately 0.2 g/t. Another factor in the lower estimate is the zone around the upper, lower and northern limits of the workings which were a part of the block model, and which were not sampled. This part of the deposit is the low grade boundary where the orebody graded into the weakly mineralised host rock (uneconomic for mining at the time of underground mining). Inclusion of the blocks in these peripheral parts of the deposit in the calculation of the statistics give a lower mean grade for the estimates than for the de-clustered data, as the de-clustered data does not include the low-grade halo. By restricting the blocks to those closer to the raw data (removing some of the more
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit peripheral blocks in an arbitrary manner), the mean of the indicator kriged Au estimates was 12.12 g/t (using the decile thresholds) and 12.39 g/t (using the modified thresholds), as opposed to the declustered mean for the Au data (12.41 g/t). Visual inspection was also an important part in the validation of these estimates. The estimations were checked in cross-sectional and plan views against the raw data. The large amount of raw data made this difficult, and summaries using coloured grade ranges were used to compare the two data sets. Further examination of the estimates was made using a comparison between the logprobability plots of the de-clustered raw data and the estimates. The most obvious features of this comparison are the minimum low grade estimate, and that the higher grade estimates are lower than those of the raw data. There is an apparent artificial maximum estimate in the indicator kriged estimate using the decile grade thresholds, but this is not a feature of the estimate using the modified thresholds. The higher grade estimates using modified thresholds appear to represent the de-clustered histogram better than the indicator kriged estimate using the decile thresholds. Examination of the basic Au statistics, the visual inspections of the indicator kriged estimates with the raw data and the comparison of the log-probability plot of the declustered Au data and the indicator kriged Au estimates indicate the indicator kriged estimates using the decile thresholds did not reproduce the high grade end of the distribution well. The low grade estimates were also not well reproduced with a lack of definition of the distribution at the low end, and an artificial minimum. An attempt to review some of the estimates on a local basis was made by preparing moving window statistics of the Au and comparing the average value (if more than 30 data points) with the average of the indicator kriged estimates from within the same window in the model. Comparison between the average estimate in the block, and the average of the sample data in the block was not entirely valid as the estimates used data outside of the block as well to estimate the grade in the block model, and the average of the data was no more than a simple average (a polygonal estimate). Outliers were a problem in each of the comparisons. Although a point on the scatter plot indicated that the estimates were much lower than the mean of the data, the mean data value was controlled by an extreme value / outlier, and the estimates did not reproduce this extreme value. A small spread of points around the 45 degree line on the scatter plot is a result of a lack of precision in local estimates. This is because the estimates are conditioned by the class the data belongs to, not the actual data value. The indicator kriged Au estimates using the modified thresholds are more consistent with the raw data, and from the review of the estimates and the data, there is no reason to assume this model is not valid.
Grade-tonnage results
In a global sense, the grade tonnage figures for the various estimates are compared. The high grade blocks are crucial with a large proportion of the metal in the deposit being represented by only a small amount of high value data and consequently only a
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit few blocks in the model. An important feature of these global numbers is the relatively low amount of contained metal in the indicator kriged estimate using decile thresholds for the 50 g/t cut-off.
Table 2 Tonnes, grade and contained gold above cut-off gold grades by the IK1 and IK2 estimates.
Cut-off IK1 IK2
2 g/t Au 3,526,800t 11.98 g/t Au 42,251 kg Au 3,526,920t 12.24 g/t Au 43,170 kg Au
25 g/t Au 339,240t 34.31 g/t Au 11,639 kg Au 339,480t 35.90 g/t Au 12,187 kg Au
50 g/t Au 11,040t 52.37 g/t Au 578 kg Au 29,760t 57.00 g/t Au 1,696 kg Au
The IK2 reports slightly higher tonnages and grade at the 25 g/t Au and 50 g/t Au cutoff reports. Examination of the amount of contained metal above each cut-off for the IK1 and IK2 estimates showed that for each of the cut-offs, there was more contained metal reported by the IK2 estimate than by the IK1 estimate. An interesting feature of the IK1 estimate is the low amount of contained metal above the 50 g/t Au cut-off. The IK2 estimates report 1,100 Kg of gold more than the IK1 estimates in the 50 g/t Au cut-off report. In this case, if a resource estimate was applied using the decile thresholds only (as was the IK1 estimate), it is probable that the resource will be under-stated by the estimates and the grade-tonnage figures.
Observations from visual comparison of the estimates

The indicator kriged estimates had good continuity of the high grade estimates. The indicator kriged estimates also had minor near vertical shoots of high grade estimates that were not as well defined as the main shoot. These high grade shoots were too small to have been separated from the remainder of the data by building them into separate geological domains during the data review phase. The lowest estimates for the indicator kriged estimates were higher than the low end of the de-clustered data distribution. This is because the indicator kriged estimates are conditioned on the class the sample value belongs to, rather than the actual sample value as in the linear estimates. An obvious overall trend was that in the indicator kriged estimates the effect of smearing of extreme values was restricted.
Discussion
In this study, the available data was sufficiently dense throughout the study area to allow a fairly detailed examination of how different techniques perform. It is not
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit possible to examine them against reality because the exhaustive dataset is not known, and local production figures are not available. Global production figures are not suitable for comparison either as the deposit included several different geological domains including an oxidised zone, supergene enrichment zone (gold) and depletion zone (copper - Mundic Zone) and pyrrhotite rich ore-bodies near the western margins of the deposit. Two indicator kriged estimates were prepared, one using decile thresholds (the IK1 estimate) and another using modified thresholds (the IK2 estimate). The IK1 estimate, which used grade thresholds based on the declustered sample deciles, was a poor estimator. It was particularly poor at preserving the contained gold in the gradetonnage figures and in nearly all the grade-tonnage estimates (Table 2), the IK1 estimate reported nearly 1 tonne less of contained gold than the IK2 estimate. The IK1 estimate resulted in only 578 kg Au reported above 50 g/t Au , approximately 1,100 kg Au less than was reported for the IK2 estimate. The biggest problem with the IK1 estimate appears to have been the use of grade thresholds based on declustered sample decile values, and the consequent poor definition and use of the high grade values. This resulted in significant under estimation of the resource represented by the high grade values, and subsequent under estimation of the total resource. Clearly this degree of under estimation of a resource would be reason for concern in a feasibility study to determine the economics of mining a mineralisation with a skewed distribution. The IK2 estimate, which used grade thresholds based on the balance of metal for thresholds above the median value, was much better than the IK1 estimate. The extreme values were better accounted for by the increased number of classes. This effectively controlled the influence of high grade values on surrounding blocks. Indicator kriging dealt with the outliers in the most effective manner, but lacked the precision for local estimates. The indicator kriging did however provide the best global estimate because of the control on the high grade values. This study supports the earlier reports that the orebody was concentrically zoned in respect to the gold grades. The peripheral parts of the Main Pipe were also shown to be gradational, but were not defined adequately by grade control samples and therefore not accurately defined by the modelling. The high grade shoots, also defined by the model appeared more prolific than were previously thought, and appeared to be evenly distributed along the central axis of the deposit. It was further recognised that the high grade shoots had irregular and gradational boundaries, and were more irregular in shape than previous reports have suggested.
Acknowledgements
I would like to express my sincere thanks to Professor Roussos Dimitrakopoulos for his tireless guidance in helping me complete this study. I would also like to acknowledge the support of Mount Morgan Ltd, Perilya Mines N.L., CRA
Jones Case study Indicator Kriging and the Mount Morgan goldcopper deposit Exploration Pty Ltd and the University of Central Queensland who provided access to company data.
References
Deutsch, C.V., and Journel, A.G., 1992. GSLIB Geostatistical Software Library and Users Guide, Oxford University Press, New York, 340p. Isaaks, E.H., and Srivastava, R.M., 1989. Applied Geostatistics. Oxford University Press (New York), 561pp. Jones, I.W.O., and Golding, S.D., 1994. The Mount Morgan Mine Central Queensland: Three dimensional relationships of the mineralization in the Mount Morgan Gold-Copper Deposit, in Proceedings of the 1994 Field Conference, Central Coastal Queensland, pp. 64-79 (Geological Society of Australia, Brisbane). Journel, A.G., 1982. The Indicator Approach to Estimation of Spatial Distributions, in Proceedings of the 17th APCOM (International Symposium on the Application of Computers and Mathematics in the Mineral Industry), SMEAIME, Golden, Colorado, USA, pp. 793-806. Journel, A.G. and Arik, A., 1988. Dealing with Outlier High Grade Data in Precious Metals Deposits, in Proceedings of Computer Application in the Mineral Industry, (Fytas, Collins and Singhal, eds.), Canada, Balkema, Rotterdam, pp. 161-171. Patterson, B.G., and Thomas, G.A., 1910. Mount Morgan practice in recording and estimating Ore Tonnages and Values, Transactions of the Australian Institute of Mining Engineers, 15:301-334.
87
Lipton et al Indicator Kriging, conditional simulation and the Halleys deposit
PRACTICAL APPLICATION OF MULTIPLE

INDICATOR KRIGING AND CONDITIONAL SIMULATION TO RECOVERABLE RESOURCE ESTIMATION FOR THE HALLEYS LATERITIC NICKEL DEPOSIT.
Ian Lipton, Richard Gaze, John Horton and Sia Khosrowshahi Mining and Resource Technology
Abstract
The renewed interest in the evaluation and development of lateritic nickel deposits in Australia comes at a time when multiple indicator kriging (MIK) has finally achieved a level of acceptance in the gold mining industry. The application of MIK to lateritic nickel deposits presents some theoretical and practical problems not normally encountered in gold deposit modelling. These are highlighted in a case history of resource and reserve estimation for the Halleys lateritic nickel deposit. Nickel and cobalt are spatially correlated and were modelled together using nickel as the mother variable and carrying cobalt. MIK and E-type estimates for nickel and cobalt provided both selective and non-selective mining estimates. The spatial distribution of magnesium in the deposit is significantly different to that of nickel and cobalt and therefore an independent model was necessary for magnesium. Ordinary kriging (OK) was used to estimate magnesium grades and estimates were constrained by an interpreted boundary separating high and low magnesium domains (the magnesium discontinuity). Conditional simulation was used to simulate the local variability of the grades and validate the resource model. The simulations showed that the MIK model was a very good predictor of nickel grades and volumes. They also showed that the interpreted position of the magnesium discontinuity was critical to the correct estimation of global magnesium grades and that the initial interpretation tended to place the boundary too low in the profile. The resource model was subsequently adjusted to reflect this. For the Halleys deposit the MIK and E-type estimates were similar. This reflects the good continuity of the nickel grades. The E-type estimates have been used for mine planning work since they are much easier to process and present. The MIK estimates show the higher grades that are
Lipton et al Indicator Kriging, conditional simulation and the Halleys deposit expected to be achieved by selective mining. The composite MIK/OK model has proved to be a workable solution that provides a practical basis for mine design and scheduling.
Key Words: geostatistics, non-linear estimation, mining, nickel, multiple indicator kriging, cobalt, magnesium, ordinary kriging, conditional simulation.
Introduction
Estimation of recoverable resources and reserves is one of the most critical aspects of modern mining geology. The accurate assessment of the tonnage and grade of run of mine ore, inclusive of ore loss and dilution may be the difference between a healthy profitable operation and an expensive early mine closure. Geostatistics offers several methods for recoverable resource estimation and in recent years these techniques have been put to effective use, particularly in the gold mining industry. Multiple Indicator Kriging (MIK) and conditional simulation are achieving growing acceptance in Australia as practical and cost effective methods for resource estimation and grade control. Most implementations so far seem to have been for estimation of a single economic variable, such as gold. This paper explores the practical application of these methods to the evaluation of lateritic nickel deposits. Several variables need to be considered and this introduces some theoretical and practical problems. These problems and some alternative methods of resolving them are discussed with particular reference to the lateritic nickel deposits at Ravensthorpe in Western Australia. The Ravensthorpe Nickel Project is located close to the south coast of Western Australia and is controlled by Comet Resources NL. It is currently the subject of a study into the feasibility of development as an open pit mine, with nickel-cobalt ore to be treated by pressure acid leach (PAL) technology.
Geology
The Ravensthorpe Nickel Project is located close to the southern margin of the Archaean Yilgarn Craton, within the Ravensthorpe Greenstone Belt. The project is centred on the Halleys deposit, a single tabular lateritic nickel body some 3km long and 1km wide. The deposit occurs within the Bandalup Ultramafics, a north-northwest striking, serpentinised komatiite suite with rare interflow sedimentary units (Sampson, 1998). A thick lateritic regolith is interpreted to have developed on these rocks during the Tertiary Period, and is partially preserved at the Halleys deposit. The regolith profile has been subdivided into five principal units (Table 1).
89
Table 1 Summary of major subdivisions of the weathering profile at Halleys
Zone
Description
Lateritic residuum Strongly indurated, ferruginous and siliceous duricrust and leached siliceous pedolith Upper ferralite Lower ferralite Saprolite Fresh ultramafic Weakly to strongly indurated porous cellular quartzgoethite (-minor clay) rock above Co-Ni-Mn redox Weakly to strongly indurated porous cellular quartzgoethite (-minor clay) rock below Co-Ni-Mn redox Clay-serpentine-goethite-carbonate weathered ultramafic with local silica veining Komatiitic olivine cumulate
The geochemical, mineralogical and textural characteristics of these units reflect weathering processes dominated by:

leaching of magnesium, residual enrichment of immobile elements such as nickel, iron, aluminium and chromium, dissolution and reprecipitation of silica, residual and supergene enrichment of elements such as cobalt and manganese.
These processes have resulted in a geochemical profile that displays similarities to those developed in other lateritic nickel deposits, and in particular the Cawse deposit near Kalgoorlie (Brand et al, 1996). It must be emphasised, however, that each lateritic nickel deposit has its own set of geochemical, mineralogical and structural characteristics that may require different approaches to resource estimation. Fresh, unweathered komatiitic rocks are interpreted to underlie the profile but have rarely been intersected by evaluation drilling. Most drill holes terminated in weakly weathered serpentinite in which there is no significant accumulation of nickel.
90
Lipton et al Indicator Kriging, conditional simulation and the Halleys deposit Magnesium grades are typically greater than 20% and nickel grades are less than 0.3%. Fresh bedrock is overlain by moderately weathered saprolite in which there has been partial leaching of mobile cations such as magnesium and minor residual enrichment of nickel. Nickel grades increase rapidly above a 0.3% threshold. The top of the saprolite is commonly marked by a rapid decrease in magnesium content and increase in iron content. This geochemical boundary, termed the magnesium discontinuity, is a redox front that marks the contact with an overlying ferralite zone. It ranges from very sharp (e.g. from less than 1% magnesium to greater than 10% magnesium across one or two sample intervals) to gradational over several metres. The decrease in magnesium grades is mirrored by a rise in Fe grades since the boundary is conformable with the base of limonite and goethite formation. The magnesium discontinuity tends to rise at the margins of the ultramafic unit probably due to the lower solubility of the marginal facies of the ultramafic unit compared to the central facies. The ferralite zone consists dominantly of iron oxyhydroxides, silica and minor clays. The ferralite zone is highly porous due to the relatively isovolumetric leaching of almost all of the magnesium. It is generally enriched in nickel, locally enriched in cobalt and hosts the majority of the Halleys deposit. The ferralite zone is overlain by leached siliceous pedolith and lateritic residuum that are largely depleted in nickel, cobalt and magnesium but enriched in silica, iron and aluminium. The nickel enrichment zone forms a single gently undulating slab. Approximately two thirds of the zone occurs within the ferralite horizon and the remaining third occurs within the saprolite. Within the enrichment zone higher nickel grades tend to occur at one or two horizons. The most continuous of the high grade zones occurs towards the top of the ferralite and is associated with the main zone of cobalt enrichment, although the peak cobalt grades typically occur 2m to 4m higher in the profile than the peak nickel grades. The cobalt mineralisation is strongly associated with manganese accumulation and is interpreted to be controlled by a redox front. A second deeper zone of higher grade nickel mineralisation is also often present and is commonly associated with the magnesium discontinuity.
Drill hole data.

The Halleys deposit has been tested by over one thousand Reverse Circulation drill holes drilled on a regular 50m by 40m grid. All holes were vertical and were sampled at 2m intervals. Samples were riffle split and were assayed using a four acid digest and ICPOES finish. Sampling and assaying quality were monitored by field re-splits, inter laboratory check analyses, assessment of standard sample assay results and twinned core drill holes. All results were satisfactory.
91
Selection of resource modelling method

The Halleys deposit is being considered for development as an open pit mine, with the ore to be treated by pressure acid leach (PAL). There are three principal elements of interest; nickel, cobalt and magnesium. Nickel is the main ore element but cobalt also occurs in potentially economic quantities. Magnesium is the most important gangue component in the ore because it is the principal consumer of sulphuric acid in the PAL circuit. There is therefore a direct relationship between magnesium grade and operating cost. Ideally, the resource model should honour the local grade variability whilst also reflecting any strong geological boundaries within the deposit. Constraining the resource estimates within boundaries defined by economic cut-off grades would produce unrealistically sharp divisions between ore and waste, since there are no natural nickel grade boundaries within the enrichment zone. Such an approach would also effectively remove the flexibility to evaluate the resource model at different cutoff grades. This flexibility is important because optimisation of the life of mine schedule may require a variable cut-off grade strategy and, in any case, the optimum cut-off grade is unlikely to be known at the time of the resource estimation. Examination of the drill hole sections indicated that for the estimation of nickel, cobalt and magnesium grades the most significant, and strongest, boundaries correspond to the major changes in geochemical gradients within the weathering profile:

The top of nickel enrichment, as defined by a sharp increase in nickel grades above a nominal 0.3 % nickel cut-off. The Co-Ni-Mn redox zone, The base of nickel enrichment, as defined by a sharp decrease in nickel grades to background levels below a nominal 0.3 % nickel cut-off. The magnesium discontinuity.
These boundaries were interpreted using the drillhole assay data supported by the geological logging. Although other elements are also either enriched or leached within the weathering profile they generally display enrichment or depletion characteristics largely consistent with the five zones defined above. The boundaries were wireframed and used to constrain statistical analysis, geostatistical analysis and resource estimation. The effectiveness of these boundaries in subdividing the data into homogeneous geological domains can be seen in the distributions of the subdivided data. Figure 1 shows log-probability plots of the 2m drill hole samples. The plots show the separation of distinct enriched and unenriched nickel populations. Small proportions of enriched material assigned to the interpreted footwall are interpreted to be due to
Lipton et al Indicator Kriging, conditional simulation and the Halleys deposit narrow, structurally controlled zones of deeper weathering. Conversely, the small proportions of unenriched material assigned to the enrichment zone are interpreted to be due to isolated core stones of unenriched rock preserved within the ferralite zone. The interpreted magnesium discontinuity divides the data into an upper low grade zone and a lower high grade zone within the saprolite. These zones are important for the control of the magnesium grades in the resource estimate. Mixed populations are still evident in the log-probability plots (Figure 2). This is due to a combination of the locally gradational nature of the magnesium discontinuity and the presence of corestones of saprolitic material within the ferralite zone.
93
94
95
Lipton et al Indicator Kriging, conditional simulation and the Halleys deposit Multiple Indicator Kriging (MIK) was initially selected as the preferred method of resource estimation since:

It is less prone to over-smoothing of grades and conditional bias than Ordinary Kriging (OK) and other linear interpolation methods, It allows direct estimation of recoverable resources inclusive of dilution and ore loss. It is non-parametric and does not depend on prior assumptions about the shape of the distributions, unlike parametric methods, such as Multigaussian Kriging and Disjunctive Kriging.
Ideally, for economic evaluation of the deposit, all the elements of interest should be evaluated simultaneously. That is, the nickel, cobalt and magnesium grades of a given portion of the deposit at a given cut-off grade should be assessable. For methods that produce whole block estimates, this ideal presents no problem but for MIK some difficulties arise. MIK is a probabilistic method that defines the distribution of the grades of samples within each search window, providing a discrete approximation to the Conditional Cumulative Distribution Function (CCDF) for each block. When considering multiple variables the spatial relationships between the variables must be retained in the representations of the individual CCDFs. That is, the model should provide estimates of each variable for the same portion of the CCDF. For MIK interpolation, the local CCDFs are usually based on one variable, for which multiple indicator parameters are modelled. Multiple Indicator Co-kriging systems require the co-regionalisation of the two variables to be modelled. However, such an approach only has a significant advantage if one of the variables is undersampled relative to the other. A further practical consideration is the additional calculation and modelling of indicator cross-variograms that are required with co-kriging methods. An alternative approach to co-kriging is to establish one variable as the mother variable and carry the grades of other variables with it. The mother variable and its search and interpolation parameters control the estimation of the other variable(s). In this way, the CCDFs of the other variables are based on exactly the same statistical and spatial support as the mother variable. For this to be a valid approach the variables must be spatially correlated and have similar patterns of continuity. In order to determine whether MIK could be applied in this way to the multi-element resource model for the Halleys deposit, it was necessary to examine the spatial relationships between the elements. This was carried out through simple statistical and geostatistical analysis of the drill hole data.
Bivariate statistical analysis of the data

The relationships observed between nickel, cobalt and magnesium were confirmed by generating scatter plots and correlation coefficients. These showed:
Moderate to strong positive correlation between nickel and cobalt, particularly at the Co-Ni-Mn redox front, No apparent statistical correlation between magnesium and nickel.
The statistical relationship between nickel and cobalt indicated that simultaneous MIK modelling of nickel and cobalt grades might be appropriate but the case for magnesium was less clear. In geological terms there is an overall negative correlation between nickel and magnesium due to the leaching of magnesium and residual accumulation of nickel in the weathered profile. Superimposed on this trend is the accumulation of higher grade nickel mineralisation near the CoNiMn redox and near the magnesium discontinuity, which is a steep geochemical gradient rather than a homogeneous domain. The distribution of magnesium grades in the nickel enrichment zone is also strongly bimodal. For these reasons it was difficult to establish any statistical correlations between magnesium and nickel.
Variography
Directional variography was carried out using grades and indicators. Since the nickel enriched material forms a gently undulating zone and the zone is not constrained by internal grade boundaries, unfolding techniques were used to honour the continuity of mineralisation. The digitised CoNiMn redox surface was used to unfold samples within the ferralite zone and ensure that high grade nickel and cobalt intercepts were correctly correlated from hole to hole. Beneath the magnesium discontinuity the nickel, cobalt and magnesium data were unfolded to the magnesium discontinuity. Ferralite and saprolite data were examined separately. When unfolding techniques are used, the inferences of dip orientations are not relevant, as the variogram calculations are restricted to the two-dimensional plane of the unfolding surface. All the grade variograms were well structured. The variograms showed some regional anisotropy but for practical grade estimation purposes were largely isotropic in the plane of the unfolding. Ranges from 65m to 130m were observed and modelled. Down hole variograms provided clear structures with low nugget variance and modelled ranges of 10m to 13m in the minor axis direction. The variograms for nickel and cobalt were very similar and displayed structures with ranges of about 20m and 100m, as illustrated in Figure 3. The variograms for magnesium showed a single structure with a range of about 60m. Interestingly, the magnesium variograms changed little when the data were unfolded to the CoNiMn redox rather than the magnesium discontinuity. For the Halleys resource model the objective was to build local CCDFs based on nickel and determine the grades of elements associated with nickel, above various nickel cut-off grades. In MIK modelling, the grade variogram parameters are not used directly in the interpolation process. Since nickel is the variable of greatest importance and elements such as cobalt and magnesium have similar anisotropies it was initially decided to use nickel as the mother variable and carry the grades of the other
Lipton et al Indicator Kriging, conditional simulation and the Halleys deposit elements with it during MIK. Using this approach, only the indicator variograms for nickel were required.
98
The cumulative log-probability plots of nickel grades were used to establish cut-off grades at regularly spaced percentiles across the distribution. Directional variograms were generated at each cut-off using the same lags and tolerances as were used for the grade variography. All the nickel indicator variograms were well structured and displayed continuity similar to the grade variograms. The indicator variograms for the ferralite zone had lower nugget variance and slightly longer ranges that those from the saprolite zone, indicating that continuity of nickel mineralisation is better in the ferralite zone.
Grade estimation
Multiple Indicator Kriging (MIK) works on a probabilistic basis to define the distribution of the grades of samples within each search window, providing a discrete approximation to the CCDF for each block. As this distribution is based on the samples found within the search window centred on any given point, it changes from block to block to reflect local grade variability. Several cut-off classes were selected for each model block, so that an equal number of samples occurred in each class. The cut-off classes varied from block to block to reflect the local variability of the samples within the search volume. For example, in a low grade area, the cut-offs were generally low, but for blocks in a high grade area, higher cut-offs were selected. This provided a better approximation of the CCDF of the local samples than could be achieved with a fixed set of cut-offs for the whole data set. Whilst this is ideal for modelling the distribution of grades within the deposit, it is impractical for the purposes of reporting the resource or mine planning. Therefore, after the CCDFs were estimated at each block centroid, the model was further processed to provide probabilities and grades within blocks, for a range of fixed cut-offs that covered the range of potential mining cut-off grades. Each of the spatial domains defined by the geological interpretation were interpolated separately. The samples in each domain were unfolded to the relevant unfolding surface in that domain. The MIK method allows resource estimates to be reported in several ways, including: the average block grade, (E type estimate); or
the recovered tonnes and grade (MIK estimate) defined by a selective mining unit (SMU).
The E-type estimate is simply the average grade of the block derived by weighting the grade in each cut-off class by the probability for that class. When assessing the
Lipton et al Indicator Kriging, conditional simulation and the Halleys deposit resource in terms of the E-type estimates, it is assumed that the entire block has a single grade and that selective mining is not possible. Once a cut-off grade is applied, the E-type estimate therefore incorporates high dilution and high ore loss. The CCDF distribution at a block centroid initially reflects the variance of the grades of the samples within the search window. When estimating recoverable resources we are more interested in the variance of grades based on SMU support, as opposed to sample support. The SMU block size represents the smallest practically mineable volume that can be extracted, given the still imperfect resolution of sampling at the grade control stage. The variance of the SMU grades will be smaller than the variance of the samples. In order to convert the prior CCDFs based on sample support to posterior CCDFs reflecting SMU support, a variance correction is required. The volume-variance relationship, or Krige's relationship, states that the variance of samples within a domain of interest is equal to the variance of samples within SMU sized blocks plus the variance of SMU sized blocks within the domain of interest. This relationship allows local recoveries based on SMU support to be estimated. There are several possible approaches to change of support variance correction including the affine correction and indirect lognormal correction. For the Halleys deposit, an indirect lognormal correction of variance was used. This is a method that borrows the transformation that would have been used if the original sample support distribution and the transformed block support distribution were lognormal. The idea behind it is that while skewed distributions may differ in important respects from a true lognormal distribution, change of support may affect them in a manner similar to that described by two lognormal distributions with the same mean but different variances (Isaacs and Srivastava, 1989). The correction operates by squashing the CCDF at each centroid. The mean of the distribution is maintained but the variance is reduced. The indirect lognormal correction results in some reduction of skewness. For example, for positively skewed data, the tail of very high values is squashed in towards the mean grade more than are the low grade values. Although the mean of the CCDF will be unchanged, the mean grade above a cut-off grade will change. The change of support correction factors can be derived theoretically using the grade variogram models. However, the grade variograms derived in this particular study were of such large range and low nugget variance that the correction factors estimated from the models, assuming an SMU size of about 5m by 5m by 2m, would result in minimal change of support. In practice, factors such as undulations in the enrichment zone, movement during blasting and irregularities in bench floors mean that theoretical selectivity will not be achieved. The variance correction used for the Halleys resource model was adjusted to reflect these practical conditions.
Conditional Simulation
Kriging, in common with other grade interpolation methods, results in smoothing of the true dispersion (the variability) of the grade. This leads to overestimation of low grades and underestimation of high grades. The probabilistic estimates produced by
Lipton et al Indicator Kriging, conditional simulation and the Halleys deposit MIK methods often reduce this effect but the local grade variability may still be incorrectly characterised. This failure may result in incorrect estimation of recoverable resources. Since recoverable resource estimates were the aim of the modelling for Halleys deposit, the MIK results were validated using conditional simulation. Conditional simulation is a series of methods for producing models on a detailed scale to simulate the spatial and statistical characteristics of a deposit. The simulations are conditioned to the known data points and honour the spatial continuity as modelled by the variogram, but are not smoothed. They therefore preserve the local variability of the deposit. A sequential Gaussian technique was used for the Halleys deposit. In simple terms, a typical sequential Gaussian conditional simulation involves:

transforming the data values to a normal distribution using a Gaussian transform function; creating a dense grid of points, or nodes, for which the grades will be simulated; selecting a node at random and allocating a value at random to that node, conditioned to the conditional moments (simple kriging mean and variance) defined at each location from the real data points; repeating the process by selecting further nodes at random but now using the values at previously simulated nodes as well as the real data, until all nodes are simulated; and finally, back-transforming the simulated Gaussian values.
Each simulation run produces an image or realisation of the deposit that correctly reflects the statistical and spatial variability of the real data. The true data values are retained wherever they correspond with a node. Multiple realisations are required to enable the impact of local variability to be assessed. Collectively these realisations are referred to as the conditional simulation model. Two trial areas were drilled in the Halleys deposit, one towards the eastern margin of the ultramafic unit where the magnesium discontinuity was gradational and a second area in the middle of the deposit where the discontinuity appeared to be much sharper. Holes were drilled at 10m spacings along a series of intersecting lines, running parallel, orthogonal and diagonal to the regular drilling grid. The pattern was chosen to provide detailed coverage in several directions (azimuths), whilst minimising the total amount of drilling. The holes in the detailed drilling pattern were sampled on 1 m intervals.
The original and close spaced drilling data were used to generate a conditional simulation model comprising 50 realisations that reflect the spatial and statistical variability of the data. Three variables were simulated:

the nickel grade in the ferralite zone and saprolite zone; the elevation of the magnesium discontinuity; the magnesium grade in the ferralite zone and saprolite zone.
In the case of the grade variables, the 2m composite assays from both the regional RC drilling (40m by 50m grid) and the close spaced RC drilling (10m by 10m grid) were used as the conditioning data. For the elevation of the magnesium discontinuity, the data points on the digitised surface were used for conditioning. A mixed simulation model using nodes on a 1m by 1m by 2m grid was then developed for each realisation as follows:

all the nickel grades were used from both the ferralite zone and the saprolite zone without constraint; the elevation of the magnesium discontinuity was applied to the simulation to define the ferralite and saprolite zones; magnesium grades from the ferralite zone and saprolite zone were modelled independently.
For the purpose of comparison of volumes and grades of the simulation model with the recoverable resource model, the simulation model was regularised to a cell size of 5m by 5m by 4m, the nominal size of the SMU. In addition, three other MIK resource models were generated for the trial area to examine the effects of using additional drilling on the interpretation of the magnesium discontinuity and on the grade estimation. These models used various combinations of regular and detailed drilling data and they were compared to the best case, worst case and most likely result conditional simulation models. The most likely result was defined as the probability weighted mean grade above a nominated cut-off for all 50 cases. From these comparisons it was apparent that:

the MIK model was producing good estimates of the volume and grade of nickel mineralisation above various cut-off grades; the variability of the elevation of the magnesium discontinuity was high, and accordingly difficult to predict locally;
102
the variability of the magnesium discontinuity significantly affected the average magnesium grades within individual domains, the MIK estimates for magnesium were lower than the simulated grades.
Since the magnesium content of the ore is directly related to ore processing costs it was considered undesirable to run the risk of underestimating magnesium grades in the resource model. In order to reduce this risk a revised, more conservative approach to modelling the magnesium grades was taken. The simulation indicated that the interpreted elevation of the discontinuity tended to be too low, so the interpreted discontinuity was raised by 2m to create a new constraining surface. Samples were flagged according to the original interpretation but then interpolated into ferralite and saprolite domains defined by the revised surface. The simulation models suggested that the MIK estimates based on a nickel mother variable were underestimating magnesium grade. This was probably due to the weakness of the correlation between magnesium and nickel grades and showed that the initial conclusion that magnesium grades could be carried with the nickel grades was invalid. For the revised model, magnesium grades were interpolated using ordinary kriging. The revised modelling resulted in a local increase in magnesium grades immediately above the magnesium discontinuity, a small increase in the magnesium grade of the resource overall, and a much better correlation with the simulations.
Mine Planning
The MIK, E-type and OK resource estimates for the Halleys deposit are quite similar (Table 2) even at higher cut-off grades.
Table 2.. Comparison of resource estimates at a 0.5% and 1.0% nickel cut-off grades
Cut-off Estimate OK E-type MIK Mtonnes 63.2 64.7 59.4
0.5% Ni Ni (%) 0.87 0.86 0.90 Co(%) 0.04 0.04 0.04 Mtonnes 16.6 15.6 16.9
1.0% Ni Ni (%) 1.25 1.25 1.31 Co (%) 0.05 0.05 0.05
This reflects the good continuity of the nickel grades and effective unfolding of the data during grade estimation. Larger variations would be expected in deposits where short scale variability of grades is higher and in these deposits evaluation of the recoverable portion of the resource becomes more crucial.
Whole block estimates are much simpler to work with for mine planning. Not only are they computationally much less onerous, but they also avoid the difficulties of combining probabilistic estimates of variables such as nickel and cobalt with smoothed, whole block estimates of other variables such as magnesium and bulk density. Importantly, whole block estimates are also much easier to present to parties such as engineers and financiers, who may have no familiarity with geological or geostatistical concepts or who may only be familiar with more empirical approaches to modelling ore loss and dilution. At the feasibility study and financing stages of a project these may be relevant considerations. For the mining studies that form the core of the reserve estimation and mine planning for the Ravensthorpe Nickel Project the E-type estimates for nickel and cobalt were used in preference to the MIK estimates. Mine planning was therefore based on marginally lower nickel grades than are expected to be achieved by selective mining. The weak negative correlation between nickel and magnesium indicates that selective mining by nickel cut-off grades may also achieve slightly lower magnesium grades. The E-type and OK estimates provided simple robust means of designing and scheduling the mine and the MIK estimates can be used to test the sensitivity of project cash flow to the better recoveries anticipated from selective mining.
Conclusion
The studies completed for the Ravensthorpe Nickel Project provide an illustration of some of the difficulties that emerge when estimating recoverable resources for nickel laterite deposits. The challenge is to provide estimates for several variables that honour the statistical and spatial characteristics of the individual variables but also honour the spatial relationships between those variables. Multiple Indicator Kriging with nickel as the mother variable carrying cobalt was shown to be a viable approach to the estimation of recoverable nickel and cobalt resources in the Halleys deposit. In deposits where cobalt is of sufficient economic importance to warrant inclusion in cut-off grade criteria, further adaptation of this method by use of equivalent grades may be necessary. This case history also highlights the value of conditional simulation as a tool for validating both grade estimates and the position of grade boundaries within the deposit. As well as detecting a deficiency in the original estimates of magnesium grade the conditional simulation also provided a view of short range variability in the deposit and an additional guide to resource classification. For mine planning purposes, the choice between MIK estimates for multiple variables rather than E-type or OK estimates depends on the sensitivity of the project to selective mining estimates and the practicalities of handling and presenting data for multiple CCDFs. For the Halleys deposit, the strong continuity of grades resulted in relatively small differences between the MIK and E-type estimates so a combination
Lipton et al Indicator Kriging, conditional simulation and the Halleys deposit of E-type and OK estimates was adequate for most mine planning purposes. For projects in which variability is greater and a higher degree of selectivity is required during mining, more complex, computationally intense approaches are required to evaluate multi-element MIK models.
Acknowledgements
The authors would like to thank Comet Resources NL for permission to publish this paper.
References
Brand N.W., Butt C.R.M., and Hellsten K.J., 1996. Structural and Lithological Controls in the Formation of the Cawse Nickel Laterite Deposits, Western Australia Implications for Supergene Ore Formation and Exploration in Deeply Weathered Terrains. In Grimsey E. J. and Neuss I. (eds), Nickel 96, Mineral to Market. The AusIMM Publication Series No. 6/96, 185-190. Isaacs E.H. and Srivastava R.M., 1989. An Introduction to Applied Geostatistics. Oxford University Press. Sampson D., Geology of the Ravensthorpe Nickel Project. Comet Resources NL unpublished report.
105
Keogh and Moulton Median Indicator Kriging and iron ore case study
MEDIAN INDICATOR KRIGING - A CASE STUDY IN IRON ORE

Alison Keogh and Craig Moulton Hamersley Iron
Abstract
The iron ore deposits of the Pilbara region in Western Australia have some unique characteristics requiring advanced geological block modelling solutions. Iron ore grade estimation commonly involves a suite of chemical variables. With Hamersley Irons strong focus on across-site blending to precise customer requirements, it is important that all grade variables are estimated accurately. In this case study, the 84 East deposit is examined at the Channar mine site, east of Paraburdoo. This deposit was originally estimated by inverse distance techniques. After two years of mining, the deposit has recently been remodelled using ordinary kriging and median indicator kriging. The grade distributions of Fe and SiO2 observed in the 84 East deposit are bimodal, as a consequence of the geological nature of iron ore mineralisation. The distributions of the contaminants Al2O3, P and LOI are positively skewed. Median indicator kriging was applied, using two medians to represent the ore and BIF populations, to try to improve the estimates of Fe and SiO2 in the geological block model. A comparison is made between the results of inverse distance, ordinary kriged and median indicator kriged grade estimates. Each method has a different effect on geological block model grade distributions and local grade estimates.
Key Words: iron ore, geological block model, geostatistics, grade distributions, non-linear estimation, mining, median indicator kriging, ordinary kriging, bimodal.
106
Introduction
Hamersley Iron Pty. Ltd. (Hamersley) currently operates six open cut iron ore mines in the Pilbara region of Western Australia. With the exception of Hamersleys newest development, Yandicoogina, the other five mines plan as one mine to produce blended ore. These include the Mount Tom Price, Paraburdoo, Channar, Marandoo and Brockman mining operations. Two products, lump and fines, are sold to customers at the port of Dampier. Each product must be blended to stringent grade specifications to meet customer requirements. The quality of each product is controlled on six chemical variables - Fe, SiO2, Al2O3, P, Mn and Loss On Ignition (LOI) - and a number of physical ore characteristics. The geological block models for each deposit are fundamental to good mine planning and sound resource estimation. For the purposes of this paper, the term geological block model refers to a computer-generated block model of an ore deposit, in which each block contains information about the geology, grade, tonnage, density and dimensions of that block in space. The aim of these geological block models is to provide estimates of grade and tonnage for mine planning and mineral resource reporting purposes. These estimates are used for different purposes: global resource estimates, strategic mine planning, whole-of-life mine plans, long-term and short-term mine planning and grade control. It is important to use an unbiased estimation method which can provide a cost-effective, fit-for-purpose grade and tonnage estimate. By providing the best estimate, mining costs can be decreased by minimising the need for unplanned digging. In this case study, the 84 East deposit is examined at the Channar mine site, east of Paraburdoo (Figures 1 and 2). This deposit was originally estimated by inverse distance grade techniques. After two years of mining, the geology was reinterpreted and the orebody remodelled using ordinary kriging and median indicator kriging. To test the relative success of the three grade estimation methods, a comparison is made between the results of inverse distance, ordinary kriged and median indicator kriged grade estimates. This includes a comparison of input composite grade distributions with the resulting model grade distributions, and local grade estimates in each model with respect to the drillholes. These results are interpreted with a focus on the geological features of the deposit.
Geology
The Channar mining area is situated approximately 20 km southeast of the town of Paraburdoo. The 84 East deposit is one of several deposits in the Channar mining area (Figure 2) with significant identified mineral resources (Table 1).
Table 1: Mineral Resource Statement Channar mining area deposits as at 31st December 1997. Million tonnes Fe % SiO2 % Al2O3 % P%
Keogh and Moulton Median Indicator Kriging and iron ore case study 265 63.0 3.8 2.1 0.098
7800000 mN 7800000 mN
400000 mE
500000 mE Cape Lambert
600000 mE
900000 mE 900000 mE
700000 mE
800000 mE
Port Hedland
Dampier
7700000 mN
Karratha
Roebourne
Pannawonica
7600000 mN
ROBE RIVER
Wittenoom
BROCKMAN 2 DETRITALS
7500000 mN
MARANDOO
MT TOM PRICE
Tom Price
YANDI (BHP) YANDICOOGINA (HI)
Brockman Iron Formation Iron Ore Mines
Paraburdoo
0 25 50
Kilometres
PARABURDOO CHANNAR
MT WHALEBACK Newman
JIMBLEBAR
Figure 1. Location map: Pilbara iron ore mines.
108
Figure 2. Location map: Paraburdoo and Channar deposits.
These bedded iron ore deposits are hosted in the Brockman Iron Formation of the Hamersley Group within Banded Iron Formation (BIF; Figure 3). The Hamersley Group is a sequence of BIF, shale, dolomite and acid volcanics up to 2500 m thick, intruded by dolerite dykes and sills. Geological Members contain consistent thicknesses of layered BIF or ore with continuous marker shale bands. A detailed description of the geology of the Hamersley Province and iron ores is provided in Harmsworth et al. (1990) and Blockley (1990). For further background information refer to discussions on iron ore geostatistics by Marechal and Srivastava (1977) and iron ore resources by Pal (1993). Each deposit has unique grade issues and grade variability related to geology. Iron ore mineralisation occurs where BIF has become relatively enriched in iron and depleted in silica. The transition from ore to unmineralised BIF varies from a sharp contact to being gradational over tens of metres. Within the orebodies, shale bands, dolerite dykes and sills, and near-surface clays and cavities cause local increases in contaminants. Mineralisation at the 84 East deposit is near-surface. Martite-hematite and martitegoethite ore types are predominant and the ore-BIF boundaries are gradational. Relative to Hamersley Irons blended ore product grades, the ore is moderate to high in Fe, slightly higher in P and Mn, and moderate in Al2O3. Drilling density is 60m x 60m, with 1.5m downhole samples, composited to 1.75m. The majority of the mineable ore occurs in the Joffre Member, with lesser amounts in the Dales Gorge Member and Colonial Chert Member (Figure 3). Local bedding in these geological Members dips gently (20o) to the south with open east-west trending
Keogh and Moulton Median Indicator Kriging and iron ore case study fold axes. The dip of the orebodies in 84 East generally reflects the dip of the beds. Minor high-angle faults and dolerites cross-cut these strands. Each geological Member is subdivided into strands, distinguished by their relative shale content (Figure 3). Mineralisation is typically of higher purity (higher Fe with lower contaminants such as Al2O3, P and LOI) in those strands with lower shale content. The Joffre Member contains six strands (J1 to J6), and the Dales Gorge Member three strands (DG1 to DG3; Figure 3). The thickness of each strand is interpreted from drillhole gamma traces (Figure 3) and sample grades. In the 84 East deposit, the 430 drillholes were reinterpreted and section interpretations were compiled from drillhole and mapping information. This provided good geological control for the geological block model.
110
Figure 3.
Stratigraphic Columns - Hamersley Group, Joffre
Member and Dales Gorge Member.
111
Grade distributions
In the 84 East deposit, the distributions of Fe and SiO2 are strongly bimodal. This is depicted in Figure 4 in the grade distribution histograms of drillhole composites from the Joffre Member. This bimodality is due to the geological nature of iron ore mineralisation: BIF contains low Fe and high SiO2 values, whereas ore contains moderate to high Fe and low to moderate SiO2 values. These two major populations also result in a strong inverse relationship of Fe to SiO2. In contrast to the bimodal distributions of Fe and SiO2, contaminants such as Al2O3 and P display a single, positively skewed grade distribution (Figure 4). High Al2O3 and P values occur in shaly strands and clay-rich cavities. The distribution of LOI also depends on local geological controls, as LOI increases in shaly strands and in near-surface hydrated ore zones. It is important to use an estimation method which deals with the nature of these distributions appropriately. Since none of the grade distributions are normal or lognormal, a non-linear estimator is ideal (Isaaks and Srivastava, 1989).
Variography
Variography was completed using the composites from each geological Member, because the consistent layering results in good continuity within individual members. No further domaining was applied. Grades were estimated into each strand, using only composites from that strand. Variograms employed in kriging within each Member were based on data within the relevant Member. This enabled shaly strands to be estimated separately from non-shaly strands, minimising mixing of populations. Variograms from these Members display no appreciable drift. Figure 5 shows the Joffre Member Fe variograms for 84 East. For the purposes of the ordinary kriged estimate, one variogram was completed in each direction; downhole (minor axis of continuity), direction 1 (major axis) and direction 2 (semi-major axis). For the purposes of the median indicator kriged estimate, two variograms were calculated in each direction for both Fe and SiO2. For Fe, the medians of each population were set at 38% Fe and 62% Fe, with a population boundary (close to an inflection point) at 46% Fe (Figure 6). For SiO2, the two medians were set at 5% SiO2 and 38% SiO2, with a population boundary (close to an inflection point) at 14% SiO2 (Figure 6). Indicator variograms were calculated for each median. Means were calculated for each indicator class (by decile, with extra classes at the 2.5th, 5th and 95th percentiles). The deciles were then separated into two groups representing the populations of ore above the inflection point and BIF below. The appropriate variogram was used when estimating these deciles. For further discussion on the
Keogh and Moulton Median Indicator Kriging and iron ore case study method of median indicator kriging, refer to Journel (1982), Journel and Huijbregts (1978), Glacken and Blackney (1998) and Vann and Guibal (1998).
113
Figure 4. 84 East deposit histograms of Fe, LOI, SiO2, Al2O3 and P in drillhole composites, inverse distance (ID) model, ordinary kriged (OK) model and median indicator kriged (IK) model.
Figure 5. composites.
84 East deposit Joffre Member Variograms of Fe
114
Figure 6.
84 East deposit Fe and SiO2 composite grade
distributions showing population boundaries and medians for Fe and SiO2 indicator variograms.
In the Joffre Member at 84 East, the nugget effect is typically 5 - 20 % of the population variance, reflecting the consistent layering and low variability at short ranges. In the maximum continuity directions, ranges are greater than 130m for most elements, and are up to 400m. In less consistent strands, the nugget effect can be up to 60 % of the population variance, although this is quite rare. The two median indicator variograms (Figure 5) display similarly shaped structures, but different ranges and directions, suggesting that the two populations have different continuity. The absolute variograms show similar structures, ranges and directions to the BIF (38%Fe) indicator variograms. Typical variography shows low nugget effect and long ranges, reflecting good continuity. This drill spacing of 60m is less than half of the minimum range of 130m. While the drill spacing is well within the maximum ranges, there are shorter-scale structures (for example; Figure 5 shows short-scale Fe structures at approximately 75m) which constitute a high proportion of the population variance. Therefore drilling is required at this spacing or closer spacing to define these structures. This suggests that the drill spacing is just adequate for the purposes of grade estimation for long term planning purposes.
Grade estimation
The individual characteristics of each deposit are reviewed before deciding the best approach to producing each geological block model. The method used to build the geological block model and estimate grades is dependent on two major factors: the purpose of the model, mining constraints and the specific characteristics of the geology and grades. Statistics and variography are used to help choose optimal grade estimation parameters. The method used to estimate the ore-BIF transition will impact on local estimates, resource and reserve figures. Although the estimation of high grade (>60% Fe) areas has the most impact on resource and reserve figures, it is important to estimate transitional categories (for example, 55-60% Fe and 50-55% Fe) well because of the blending nature of the mining and potential for upgrading lower grade ore. Some of the estimation techniques which may be used include:

inverse distance (ID); ordinary kriging (OK); inverse distance estimation of separate ore and BIF domains; ordinary kriging within separate ore and BIF domains; median indicator kriging (IK); full or multiple indicator kriging (FIK).
An improvement was required in the grade estimation of the 84 East deposit, for the purposes of medium to long-term mine planning. The previous geological block model was estimated by inverse distance (ID) methods. This model did not provide adequate local estimates for accurate mine planning. Records showed an inverse distance cubed weighting was applied, with a large search distance of 300 metres. No ore-BIF boundary was used. Kriging techniques were used to try to improve the inverse distance estimated geological block model. Two models were produced for the 84 East deposit; an ordinary kriged estimate (OK) for all elements and a median indicator kriged estimate (IK) for Fe and SiO2. The ordinary kriged model and median indicator kriged model differed only in the variogram parameters and estimation method used. All other parameters, such as search distance, search orientation and discretisation, were the same. No ore-BIF boundary was used for either model. Ordinary kriging was used to estimate Al2O3, P, Mn and LOI in both the OK model and IK model. Fe and SiO2 were estimated by ordinary kriging in the OK model. In the median indicator kriged (IK) model, median indicator kriging methods (using medians to reflect the two populations of BIF and ore) were applied to estimate Fe and SiO2. The aim was to produce a model which honoured the gradational nature of the ore-BIF boundary, and dealt with the complexity of a bimodal distribution without the constraints of an ore-BIF boundary.
In other deposits, median indicator kriging has been used for some contaminant elements where their distribution is highly skewed, to minimise the impact of outlying values on the estimate. This has been applied where extreme values appear to be less continuous geologically, such as sporadic occurrences of Mn along fault and joint planes. Because ordinary kriging does not perform well for strongly skewed distributions, where data may be sparse, it is possible that the use of ordinary kriging in such cases may assign incorrect weights to key areas in some elements. Playkrig software was used to help determine the optimal discretisation and search parameters using the modelled variogram structures. Post-indicator kriging steps included dealing with order relation problems and back-calculating grades using the means for each decile weighted by the proportion of samples within each indicator class.
Model comparisons
The median indicator kriged model (IK) and ordinary kriged model (OK) were compared with the previous inverse distance estimated model (ID) for 84 East. Comparison of local estimates showed an improvement from ID to the kriged models. Specifically, the kriged Fe estimates show ore-BIF grade transitions which reflect the geology observed in drillholes and mine faces. In contrast, the ID Fe estimates display abrupt, vertical changes between drillholes. The ID model was estimated using inverse distance cubed methods with a search distance of 300 metres. Although local estimates generally closely reflect samples from the nearest drillholes, where two adjacent drillholes differ substantially (eg; ore vs. BIF) the transition is abrupt and in many cases does not reflect the observed geology (Figure 7). Effectively, these grade estimation parameters cause a similar result to that of a polygonal estimation. Local estimates differ only slightly between the OK model and IK model. Subtle differences are apparent in the ore-BIF transitions, where blending and low grade category blocks occur. The median indicator kriged model also appears to show more continuous high grade (HG) blocks between drillholes which display HG composite values (Figure 7). Estimates appear to be geologically sensible and the block estimates better reflect the input composite values. The histograms of the ID block model do not reflect the input (composite) data well. Whilst a bimodal distribution is apparent, the shape and range of block values are substantially different. The highest SiO2 and lowest Fe composite values are nonexistent in the block model, resulting in a sharp drop-off at this end of each distribution. The ID method does not adequately reflect the gradational nature of oreBIF boundaries seen in mine faces. Smoothing has not occurred to an appropriate degree.
Keogh and Moulton Median Indicator Kriging and iron ore case study The histograms of the kriged geological block models show bimodal Fe and SiO2 and positively skewed Al2O3, P, Mn and LOI distributions (Figure 4). In each case, the block model minimums have increased and maximums have decreased as expected due to smoothing (volume-variance effect). The local estimates of the kriged geological block models reflect the drillholes well, but have also incorporated a degree of smoothing which reflects the observed geology. The Fe and SiO2 block model distributions display a higher proportion of intermediate values. This suggests an addition of the two smoothed distributions and appears to be a reasonable outcome of smoothing a bimodal distribution.
Figure 7. 84 East deposit model slices: Section 30 of Median Indicator Kriged Model and Section 30 of Inverse Distance Estimated Model.
The OK block model Fe and SiO2 histograms are similar to the IK block model histograms (Figure 4). While they both show similar shaped distributions, the very low Fe and very high SiO2 values are only present in the ordinary kriged blocks. This suggests that in the IK model the most extreme values have less influence in the estimation, because the outside percentile bins are treated separately. Outlying values are not generally considered to be part of the same population as the majority of data, as they often reflect different geological genesis: for example, remobilisation of elements along faults or near-surface. Therefore, median indicator kriging is
Keogh and Moulton Median Indicator Kriging and iron ore case study considered to provide a more valid method to estimate an element with a very skewed or bimodal distribution. To fully test the results of the various grade estimation methods in 84 East, reconciliations must be performed between the models and blasthole production data. Reconciliations are currently in progress. Multiple indicator kriging may serve to define the various distributions even better, and is therefore the next logical step to improve the grade estimation of skewed or bimodal elements in iron ore. The most obvious constraints in using this method are the time required to produce all of the indicator variograms and computing time. Once the results of median indicator kriging are tested, the benefits of applying this method will be examined.
Conclusions
This case study compares the results of inverse distance, ordinary kriging and median indicator kriging estimation in the 84 East iron ore deposit east of Paraburdoo mine. The comparison includes examination of local estimates and data distributions. The kriging methods have improved the estimation compared with the inverse distance method, in the block data distributions and the pattern of local estimates. Median indicator kriging has been tested on the bimodal distributions of Fe and SiO2 in iron ore. A decision was made to separate the two main populations of BIF and ore from these distributions and estimate medians for each. The differences between ordinary kriging and median indicator kriging of Fe and SiO2 are slight. Both methods provide geologically sensible local estimates and an appropriate degree of smoothing. However, the median indicator kriging method is considered to provide a better way of estimating an element with a bimodal or strongly skewed distribution. Outlying assay values have less influence when using the median indicator kriging method. Detailed reconciliation work is required to test how well each of these kriging estimates compare with inverse distance estimates for mine planning purposes. Optimal methods of grade estimation will continue to be explored.
Acknowledgements
The authors gratefully acknowledge the permission of Hamersley Iron Pty. Limited and Joint Venture partners to publish and present this paper. The authors wish to thank Klaus Gosman and Thanh Phan for their time and support in preparation of figures and James Farquhar, Louis Voortman, Michelle Franks, Jacqui Coombes and Fiona Caristo for their editing and technical suggestions.
References
Blockley, J.G., 1990. Iron ore. In: Geology and Mineral Resources of Western Australia. Geological Survey of Western Australia Memoir 3. pp.679-692. Glacken, I.M., and Blackney, P.A., 1998. A practitioners implementation of indicator kriging. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia. Harmsworth, R.A., Kneeshaw, M., Morris, R.C., Robinson, C.J and Srivastava, P.K., 1990. BIF-derived iron ores of the Hamersley Province. In: Geology of the Mineral Deposits of Australia and Papua New Guinea. Hughes, F.E. (Editor), The Australasian Institute of Mining and Metallurgy (Melbourne), pp.617-642. Isaaks, E.H., and Srivastava, R.M., 1989. Applied geostatistics. Oxford University Press (New York), 561pp. Journel, A.G., 1982. The indicator approach to estimation of spatial data. Proceedings of the 17th APCOM, Port City Press (New York), pp. 793-806. Journel, A.G., and Huijbregts, Ch. J., 1978. Mining geostatistics. Academic Press (London), 600pp. Marechal, A and Srivastava, P., 1977. Geostatistical study of a Lower Proterozoic iron orebody in the Pilbara region of Western Australia. 15th APCOM Symposium, Brisbane, Australia. pp.617-642. Pal, M., 1993. Iron ores of Australia. International Seminar, Iron Ore 2000 and Beyond, Society of Geoscientists and Allied Technologists, Bhubaneswar, India, Proc., pp.38-60. Vann, J. and Guibal, D., 1998. Beyond ordinary kriging - an overview of non-linear estimation. Proceedings of a one day symposium: Beyond Ordinary Kriging. October 30th, 1998, Perth Western Australia. Geostatistical Association of Australasia.
120
Khosrowshahi et al Change of support correction based on p-field simulation
A PROPOSED APPROACH TO CHANGE OF

SUPPORT CORRECTION FOR MULTIPLE INDICATOR KRIGING, BASED ON P-FIELD SIMULATION
Sia Khosrowshahi, Richard Gaze and Bill Shaw Mining and Resource Technology
Abstract
While there have been significant improvements in estimation techniques, ways of addressing the difficult problem of support effect have languished. For geostatisticians working in mineral resource estimation this issue is critical. There is now widespread industry acceptance that mining optimisation studies are best built on resource models that predict the tonnes and grade that will be identified during mining. Ore loss and dilution vary with the scale of mining, notionally represented by the size of the selective mining unit (SMU). Recoverable resource models, i.e. those incorporating change of support correction, are based on implicit assumptions. The various approaches used in disjunctive kriging (DK), multiple indicator kriging (MIK), probability kriging (PK) and multigaussian approach (MG) are discussed. A proposed alternative to the adoption of parametric change of support correction factors is presented where the local correction factor is deduced empirically. The approach is based on characterisation of the short-scale variability within the domain of interest. The simulation concept is borrowed from MG but with the conditioning moments replaced by the MIK cdf. The distributions of the expected SMU sized values (grades) are thus derived empirically. This proposed approach, referred to here as p-field correction for multiple indicator kriging (pMIK), is expected to allow the local impact of selective mining or bulk mining to be characterised.
Key Words: geostatistics, non-linear estimation, multiple indicator kriging, change of support.
121
Introduction
Point support and block support
There have been significant improvements in estimation techniques with the introduction first of linear methods (e.g. simple kriging or ordinary kriging) and subsequently nonlinear methods (e.g. indicator kriging). These methods have proved to be significant advances over the classical methods of weighted averaging to interpolate values at some locations between known data. From the outset, the primary advantages of kriging were recognised as being that it is an unbiased estimator and that it minimises the error of estimation. This is achieved by combining information about the directional correlation of the data with the configuration of the data around the unknown location. The variogram is a popular tool for quantifying this spatial relationship. Using a kriging plan, the appropriate weights are allocated to the surrounding data to enable the unknown grade to be interpolated. The work originally done by D. G. Krige in South Africa (Krige, 1951, 1962) was primarily concerned with selective mining of gold to a cut-off grade and changes to the distribution of values as their support (size) increased from samples to ore blocks. The impact of correlation and regression on the grade-tonnage curve was recognised by Krige and has become embodied in expressions referred to as Kriges relationship. This may be summarised as the total variance is equal to the sum of the within block variance and the between block variance (David, 1977, p. 98). This relationship provides a theoretical basis for the application of a change of support correction, so that estimation based on a point support (e.g. 1 m composited drill sample assays) can be used to represent information on a larger support (e.g. ore blocks).
Resource estimation
One of the major difficulties for classical methods of resource estimation is this issue of support. Polygonal estimates can be biased if the drill hole samples (the point support data) are each allocated a specific volume of influence. No matter how this volume is assigned, the assumption that the point estimates correctly represent grades for such volumes is demonstrably and intuitively incorrect. The variance of sample grades and volume (block) grades would be the same, contrary to Kriges relationship. The differences in grade expected for a series of large blocks (e.g. stoping blocks or mining benches) will be much less than that seen in the drill samples. Similarly the maximum grade of large blocks will be much less than the maximum grade of the samples. At the resource estimation stage the sampling data set is relatively sparse compared to that finally available from grade control. Krige noticed that estimates made using such
Khosrowshahi et al Change of support correction based on p-field simulation sparse data were conditionally biased due to the smoothing effect of estimation. Nevertheless the objective in resource estimation is to predict what will happen during mining. We are unable to accurately define the positions of the smaller ore parcels that will be defined by grade control. Frequently we only have the sparse data. Let us assume that the grade control data will have the same distribution characteristics as the resource sampling data, but is dispersed (or lost) within the deposit. Models can then be constructed to indicate the probability of values above a threshold that will be identified (or found) assuming a nominal selective mining unit (SMU) size. Thus a two step process is required: estimation of the local distribution and then correction of that distribution to represent the distribution of SMU blocks. Non-linear geostatistics addresses estimation of the point support local distributions. There are various strategies that must then be used to adjust the point support distribution to that expected for blocks of a nominated volume. These strategies are collectively referred to here as change of support corrections (Isaaks and Srivastava, 1989, pp. 458-488). Such corrections, e.g. adjusting one distribution to another while keeping the means the same, involve making assumptions which are examined further in this paper.
Recoverable resource models

There has been a significant rise in the sophistication of resource estimates being carried out on orebodies, with the successful application of kriging to many deposits previously believed intractable. Recent published examples include the use of multiple indicator kriging for gold deposits (Gaze, et al, 1997), iron ore deposits (Collings, et al, 1997), and nickel (Lipton et al, 1998). Based on the successful performance of probabilistic resource models at a number of operations, the mining industry is gaining confidence in such computer intensive methods. By comparison with the sophistication of kriging techniques being applied, ways of addressing the support effect have languished. The mining resolution is notionally represented by the size of the selective mining unit (SMU). This may be regarded as the average minimum dimensions for a discrete ore block. In many cases ore blocks exceed these dimensions. Where this happens, assumptions that depend on the SMU block size have less impact. As the mining block size increases, the mining scale changes from selective mining to bulk mining. The low costs provided by economies of scale must be balanced against the lower mining head grade as dilution increases. Mines that have difficulty with dilution are frequently found to have chosen too small an SMU size in building the resource model. In striving to demonstrate a high head grade, such models overestimate the ability of the proposed mining fleet to mine selectively and achieve tonnage targets.
123
Khosrowshahi et al Change of support correction based on p-field simulation The precision and resolution of the grade control sampling (the information effect) may be regarded as part of the mining selectivity process for the subsequent discussion.
Current practices
Change of support correction
Linear estimation (e.g. polygonal, inverse distance, and kriging) methods produce a single outcome for each location. The estimation error depends on data density. Estimation of average grades for block sizes approximating the nominal SMU size would require a comparable sampling data density (resolution). Such information is usually only available at the grade control stage. Although we are unable to accurately define the positions of smaller units, it is possible to forecast on a probabilistic basis, over an area suited to the exploration data density, the proportion of SMU size blocks that will be selected above a nominated cut-off grade. Non-linear estimation methods were designed to address this support issue. A brief survey follows of how the change of support is currently addressed in recoverable resource estimation. Emphasis here is on declaring the implicit assumptions rather than on a mathematical exposition of the various methods, which are provided in the references. The approaches used in disjunctive kriging, multiple indicator kriging, probability kriging and the multigaussian approach are discussed. More detailed comparative reviews of these methods are provided by Marechal (1984); Journel (1985); and Knudsen and Baafi (1987). This discussion provides the context for a proposed new approach to change of support correction.
Disjunctive kriging (DK)

The DK method (Matheron, 1976) was developed to predict the recovery of material at different cut-off grades from the distribution of sample values. It does this by modelling the density function of the variable of interest (e.g. grade) from all the available data. The sample data set is transformed to the standard normal distribution using a limited expansion of Hermite polynomials (see Journel and Huijbregts, 1978, p.573-580). The solution to these equations is derived using numerical integration and the model is checked against the original data set for goodness of fit. Following the normalisation, structural analysis (variography) is performed and the process of disjunctive kriging is carried out. Using DK there are a number of approaches available for change of support correction. For example in the discrete Gaussian model, the process of back transformation (anamorphosis) may be accompanied by a change of support coefficient r. This is determined by assuming that the means are the same, the variance of the points is derived
124
Khosrowshahi et al Change of support correction based on p-field simulation from the variogram and the points are randomly distributed inside the blocks (Rivoirard, 1994, p.81). The covariances of the points (xx) and the blocks (vv) are related by:
x x = r 2 v v
i j i
The mathematics of DK is complex and the numerical integration process may suffer from unsuccessful convergence specifically related to the definition of the extreme tails of the experimental distribution. Rivoirard (1994, p.90) describes it as a neat and yet mathematically consistent model, applicable under various distribution assumptions to local estimation and simulation. Deutsch and Journel (1992, p83) note that the DK method is a slight generalisation of the bivariate Gaussian model and due to the problem of which transfer function to retain one may be better off using the most congenial and parsimonious of all RF models, the multivariate Gaussian model.
Multiple indicator kriging (MIK)

The application of an indicator function as a threshold value (e.g. a cut-off grade) to the sample data set enables the data values to be transformed to a set of indicators, so that the data set is now a set of 0 and 1 values (e.g. the indicator 1 is assigned if the sample data value is equal to or above the cut-off grade, otherwise the indicator 0 is assigned). This is a non-linear transformation in that linear estimates (whether arithmetic averaging or kriging) carried out on the transformed values do not retain the original sample data relationships. Kriging the indicators for a given location produces a result between 0 and 1 which cannot be transformed back to a sample grade. Rather this kriged indicator value must be regarded as a pseudo-probability (p-value) of exceeding the threshold (cut-off grade) at that location. Multiple indicator kriging (MIK) involves the determination of indicators at a number of thresholds so that the cdf model of p-values is available for each location. To define the distribution of blocks of a nominal SMU size, the variance of the cdf models must be reduced. This post-processing of the MIK inventory is frequently done using either the affine correction or the indirect lognormal correction. The affine correction (Isaacs and Srivastava, 1989, p.471) is the more simple approach. The sample distribution is symmetrically squeezed about the mean value to reduce the variance. This reduces the high grades but increases the low grades to preserve the shape of the distribution. For gold deposits with strong positive skew, or iron deposits with a negative skew and a maximum grade (e.g. 69.4 % Fe for hematite), this may not be appropriate. Rossi and Parker (1993) note additional difficulties caused by a high nugget to sill ratio for variograms.
125
Khosrowshahi et al Change of support correction based on p-field simulation The indirect lognormal correction (Isaacs and Srivastava, 1989, p.472) assumes for transformation that the point distributions and the inferred block distributions are both lognormal. The values are transformed to reduce the variance and then they are rescaled to correct the mean. The extent of the rescaling required depends on how well the assumption of lognormality holds for the point distributions. This results in asymmetric squeezing of the distribution about the mean, reducing the skewness of the distribution. The post-processing of the MIK point support model is carried out to define a block support model that predicts the mining resolution expected using a proposed SMU size. After obtaining the SMU distribution, application of a cut-off grade allows the probability of exceeding that cut-off to be predicted. The model can be interrogated to define the probability (or its inverse, the risk) of achieving a tonnage and grade. Probabilities of exceeding thresholds may be contoured to demonstrate the spatial relationships in the resource model, and the grade can be reported above nominated cut-off grades. Journel (1985, p.565) commends this method for its simplicity and robustness, noting that it is distribution free. Some significant points to be aware of are:

Reporting the tonnes and grades from such models requires a clear understanding of the p-values and associated grade estimates. The resulting model is assumed to define the recoverable resource by including planned dilution, i.e. the impact of the effective SMU size. This relates to the grade control data support (both precision and resolution) and the mining practices. There is no theoretical or parametric demonstration that the change of support correction has correctly modelled these effects. This can be demonstrated using conditional simulation and is strongly recommended.
Probability kriging (PK)

Probability or proportion kriging (PK) is an enhancement of indicator kriging (IK) described by Sullivan (1984). The IK method uses only the indicator data whereas PK uses the grade information in addition to the indicator values through a co-kriging process. Since the grade values and the indicator values are of different orders of magnitude, use of both in the same system may give numerical problems. Consequently the data values are transformed to a uniform distribution, sorting them by increasing grade and assigning them their cumulative frequency (i.e. a rank order transform). The cdf is then estimated by solving a co-kriging system involving both the indicators and the rank order transformed data. The PK approach requires calculation of the indicator variograms, the rank order transformed (grade) variograms and the cross variograms between these two variables. Journel (1985) notes: The PK approach retains the essential characteristics of simplicity and robustness of the IK approach. However, despite the additional complexity and
Khosrowshahi et al Change of support correction based on p-field simulation information retained, probability kriging does not alter the requirement for a change of support correction. Again the affine correction or the indirect lognormal correction may be applied to the point support distributions.
The multigaussian approach (MG)

Unlike PK and IK, which are nearly free of major probabilistic interpretations or hypotheses, the multigaussian approach (MG) of Verly (1984) is highly dependent on the probabilistic interpretation of the grade values and distribution hypotheses. The sample values are transformed to a standard normal distribution. Variograms are produced and kriging is carried out followed by a Monte Carlo simulation to solve the integration required for the change of support. An underlying assumption is that the normal score transforms are multivariate normally distributed, i.e. that all the interrelationships between the variable of interest, at all thresholds, are normally distributed. As with lognormal kriging (a special case of MG where the lognormal transform is used instead of the Gaussian transform) local errors in the estimation of the variance of the cdf may be enhanced. This can have the unfortunate consequence of creating ore in regions of consistent low grade where the data set is sparse. Journel (1985) notes: The MG approach stands out by its extreme rigour and consistency. A single, albeit strong, hypothesis is made in the beginning from which all subsequent results are deduced without any further hypothesis or approximation. Also, the MG approach is presently the only approach available that can handle the difficult problem of support effect with strict consistency. This does not necessarily mean that it would provide the most accurate solution.
Alternative methods
An implicit assumption in the approaches outlined previously for change of support correction is that the blocks are selected as ore or waste independently. This ignores the spatial relationship inherent in mining data. In addition each of the methods makes specific assumptions regarding distributions and their relationships through transforms. Change of support correction for the non-linear approaches discussed (DK, MIK, PK, and MG) all suffer from a subjective assumption of the appropriateness of a parametric distribution model. An alternative to the adoption of parametric change of support correction factors is required. It can be safely assumed that ore grades average arithmetically, the mean grade of points and blocks does not change, and that the variance of block grades reduces; however these assumptions are not enough (Rossi and Parker, 1993). A number of pragmatic solutions were presented by Rossi, Parker and Roditis (1993) including using conditional simulation to provide a training model to define the change of support correction, modelling of grade control sampling, modifying the kriging method using grade
Khosrowshahi et al Change of support correction based on p-field simulation thresholds and/or adjusting the kriging plan to produce block estimates with the theoretically appropriate dispersion variance. These methods are all made possible or enhanced by the availability of some short spaced data for the orebody. Using conditional simulation to provide a training model to deduce the change of support correction factor has the advantage that the result is deduced empirically rather than parametrically. However the application of a local training set to the global resource model raises questions of representativity and stationarity. Conditional simulation approaches attempt to determine the local SMU cdf empirically. Given good quality data (always the ultimate limitation) the empirical method frequently provides more realistic solutions than parametric modelling to geological and mining problems. There is now an expectation amongst geostatisticians that comprehensive conditional simulation models can be produced for recoverable resource estimation. Unfortunately, with many methods of simulation the computation time increases exponentially as the volume to be simulated increases. Add to this the massive data storage requirements for a large number of realisations (or the definition of many ccdf models) and other more efficient solutions still look attractive.
The proposed approach

By using the Gaussian kriged values as conditional moments for a simulation process, the MG model suggests a different approach to Journels difficult problem of support correction. However the MG approach is parametric with a strong multi-normal hypothesis. The proposed approach is to borrow the simulation concept from MG but with the conditioning moments replaced by the MIK cdf. This has been developed and trialled following successful experience with probability field simulation (p-field simulation) models to define variability (Khosrowshahi and Shaw, 1997). The p-field approach (Srivastava, 1992) creates non-conditional simulated realisations which reflect the deposit variability, and are then conditioned to the local cdf estimated by MIK. This p-field correction for multiple indicator kriging (pMIK) allows the impact of selective mining and bulk mining to be assessed for the deposit of interest. The method has been found to be sensitive to the number of simulation realisations and the resolution of discretisation used for the simulation model. However the speed of the p-field algorithm, and the requirement to only retain the local ccdf models for the nominated SMU size, means that a large number of realisations are manageable. The use of local conditioning overcomes another implicit assumption in all previous change of support corrections. The problem of adaptive geometry (Rivoirard, 1994; Rossi, Parker and Roditis, 1993) occurs where the nominal SMU size bears no
Khosrowshahi et al Change of support correction based on p-field simulation relationship to the real spatial continuity of mined ore blocks. Many ore zones of contiguous mineralisation are much bigger than the nominal SMU and the assumption that SMU sized blocks are all independent is just not true in practice. The pMIK approach compensates for this by producing contiguous SMU blocks with similar values where there is good continuity. In areas where the MIK model defines a cdf with high variance, the effect of local dilution incurred for a nominal SMU size is significantly increased. It must be remembered that there may be causes for not achieving the predicted recoverable grade that are not addressed by this proposed method. Examples that lead to additional imprecision during mining include:

Imprecision or bias of the grade control assay results at the cut-off grade, Vertical heave and lateral movements due to blasting, Insufficient survey control of sample locations relative to dig lines, and Inappropriate application of visual control for selection of ore.
The impact of grade control procedures on dilution, unseen ore loss and mining recovery has been examined in other forums (e.g. Shaw and Khosrowshahi, 1997). Despite the many difficulties in predicting the recoverable resource estimate, significant improvements can be achieved by addressing issues one at a time. The first step is to use good estimation methods that include an appropriate change of support correction based on local data. The proposed pMIK method is expected to provide an effective way to achieve empirical correction and to be an improvement on parametric based methods. This approach is currently still under development.
References
Collings, P. S., Khosrowshahi, S. and Ness, P. K., 1997. Geological modelling and geostatistical resource estimation of the Hope North Deposit. Ironmaking Resources and Reserves Estimation, Perth, Australasian Institute of Mining and Metallurgy, pp.105 - 116 David, M., 1977. Geostatistical ore reserve estimation. Pub. Elsevier, Amsterdam, 364pp. Deutsch, C. V. and Journel, A. G., 1992. GSLIB geostatistical software library and users guide. Oxford University Press, Inc, New York, 340 pp. Gaze, R. L., Khosrowshahi, S., Gibbs, D., and Grove, A., 1997. Geological and geostatistical resource estimation of the Cleo Deposit, Sunrise Dam a balanced approach. Gold & Nickel Ore Reserve Estimation Practice, Towards 2000
Khosrowshahi et al Change of support correction based on p-field simulation Australasian Institute of Mining and Metallurgy Mineral Resources And Ore Reserves Seminars, Australasian Institute of Mining and Metallurgy, Kalgoorlie Branch, WA, 21 October, pp121-139.
Isaacs, E. H. and Srivastava, R. M., 1989. Applied geostatistics, Oxford University Press, 561 pp. Journel, A. G., and Huijbregts, C. J., 1978. Mining geostatistics, Academic Press, London, 600pp. Journel, A.G., 1985. Recoverable reserves estimation - the geostatistical approach. Mining Engineering, pp.563-568, June. Khosrowshahi, S. and Shaw, W. J., 1997. Conditional simulation for resource estimation and grade control principles and practice. Proceedings of the World Gold 97 Conference, Singapore, Australasian Institute of Mining and Metallurgy, pp.275-282. Knudsen, H. P. and Baafi, E.Y., 1987. Indicator kriging and other new geostatistical tools. Proceedings of the Pacific Rim Congress, Australasian Institute of Mining and Metallurgy, pp. 859-864. Krige, D. G., 1951. A statistical approach to some basic mine valuation problems on the Witwatersrand. Journal of Chemical, Metallurgical and Mining Engineering Society of South Africa, December, pp.119-139 (and following discussion during 1952). Krige, D. G., 1962. Effective pay limits for selective mining. Journal of the South African Institute of Mining and Metallurgy, pp.345-363, January. Lipton, I. T., Gaze, R. L., Horton, J. A. and Khosrowshahi, S., 1998. Practical application of multiple indicator kriging and conditional simulation to recoverable resource estimation for the Halleys lateritic nickel deposit. Beyond Ordinary Kriging: Non-Linear Geostatistical Methods In Practice, Perth, Geostatistical Association of Australasia, (this volume). Marechal, A., 1984. Recovery estimation: a review of models and methods. Verly, G. (Ed) Proceedings of 2nd Geostatistical Congress, California, NATO ASI series, Reidel, Dordrecht (Holland), pp.385-420. Matheron, G., 1976. A simple substitute for conditional expectation: disjunctive kriging. Proceedings of the 1st Geostatistics Congress, Rome, NATO ASI series, Reidel, Dordrecht (Holland), pp.221-236.
130

Rivoirard, J., 1994. Introduction to disjunctive kriging and non-linear geostatistics, Clarendon Press, Oxford, 180pp. Rossi, M.E. and Parker, H. M., 1993. Estimating recoverable reserves: is it hopeless? Dimitrakopoulos, R., (Ed.), Geostatistics For The Next Century, Kluwer Academic, Boston, pp.259-276. Rossi, M.E., Parker, H. M., and Roditis, Y. S., 1993. Evaluation of existing geostatistical models and new approaches in estimating recoverable reserves. Proceedings of the XXIV APCOM, Montreal, Canada. Shaw, W. J. and Khosrowshahi, S., 1997. Grade control sampling and ore blocking: optimisation based on conditional simulation. Third International Mining Geology Conference Proceedings, Launceston Tasmania, Australasian Institute of Mining and Metallurgy and Australian Institute of Geoscientists, pp.131-134. Srivastava, R. M., 1992. Reservoir characterisation with probability field simulation. Society of Petroleum Engineers Annual Conference, Washington, D.C., pp. 927938. Sullivan, J., 1984. Conditional recovery estimation through probability kriging. Theory and practice. Verly, G. (Ed) Proceedings of 2nd Geostatistical Congress, California, NATO ASI series, Reidel, Dordrecht (Holland), pp.365-384. Verly, G., 1985. The block distribution given a point multivariate normal distribution. Verly, G. (Ed) Proceedings of 2nd Geostatistical Congress, California, NATO ASI series, Reidel, Dordrecht (Holland), pp.495-515. Verly, G. and Sullivan, J., 1985. Multigaussian and probability krigings - application to the Jerritt Canyon deposit. Mining Engineering, pp.568-574, June.
131
Author contact details
AUTHOR CONTACT DETAILS

Mr Paul Blackney Principal Consultant, Snowden Associates 87 Colin St, West Perth, WA 6005 ph: (08) 9481 6690 fax: (08) 9322 2576 email: pblackney@snowden.com Senior Lecturer Department of Mathematics Edith Cowan University 100 Joondalup Drive, Joondalup, WA 6027 ph: (08) 9400 5883 fax: (08) 9400 5811 email: l.bloom@eagle.fste.ac.cowan.edu.au Geostatistician, Mining and Resource Technology Level 3, Kirin Centre, 15 Ogilvie Road, Mt Pleasant, WA 6153 ph: (08) 9316 2710 fax: (08) 9316 1791 email: rlg@mrtconsulting.com.au Manager, Resource Division, Snowden Associates 87 Colin St, West Perth, WA 6005 ph: (08) 9481 6690 fax: (08) 9322 2576 email: iglacken@snowden.com Senior Geologist, Resource Services Group 35 Ventnor Avenue, West Perth, WA 6005 ph: (08) 9481 0885 fax: (08) 9481 0526 email: rsg.perth@rsg.com.au
Dr Lyn Bloom
Mr Richard Gaze
Mr Ian Glacken
Mr Brett Gossage
132
Mr Daniel Guibal
Technical Manager, Geoval Level 13, 256 Adelaide Terrace, East Perth, WA 6000 ph: (08) 9221 0600 fax: (08) 9221 4985 email: dguibal@geoval.com.au Manager, Data Analysis Australia Unit 8, 154 Hampden Road, Nedlands, WA 6009 ph: (08) 9386 3304 fax: (08) 9386 3202 email: john@daa.com.au Dept of Mathematics, Edith Cowan University 2 Bradford St, Mount Lawley, WA 6050 ph: (08) 9370 6360 fax: (08) 9370 6179 email: dhill@cown.edu.au Senior Geostatistician, Geoval Level 13, 256 Adelaide Terrace, East Perth, WA 6000 ph: (08) 9221 0600 fax: (08) 9221 4985 email: michaelh@geoval.com.au Principal Geologist, Mining and Resource Technology Level 3, Kirin Centre, 15 Ogilvie Road, Mt Pleasant, WA 6153 ph: (08) 9316 2710 fax: (08) 9316 1791 email: jah@mrtconsulting.com.au
Dr John Henstridge
Ms Donna Hill
Mr Michael Humphreys
Mr John Horton
133
Mr Ivor Jones
Senior Mine Geologist, WMC Resources PO Box 285, Leinster, WA 6437 ph: (08) 9026 5267 fax: (08) 9026 5442 email: ivor.jones@wmc.com.au Acting Specialist Geologist Resources, Hamersley Iron PO Box 114, Paraburdoo, WA 6754 ph: (08) 9143 4109 fax: (08) 9143 4264 email: alison.keogh@hi.riotinto.com.au Director, Mining and Resource Technology Level 3, Kirin Centre, 15 Ogilvie Road, Mt Pleasant, WA 6153 ph: (08) 9316 2710 fax: (08) 9316 1791 email: sia@mrtconsulting.com.au Principal Geologist, Mining and Resource Technology Level 3, Kirin Centre, 15 Ogilvie Road, Mt Pleasant, WA 6153 ph: (08) 9316 2710 fax: (08) 9316 1791 email: itl@mrtconsulting.com.au Specialist Resource Geologist, Hamersley Iron PO Box 114, Paraburdoo, WA 6754 ph: (08) 9143 4109 fax: (08) 9143 4264 email: craig.moulton@hi.riotinto.com.au Department of Mathematics, Edith Cowan University 2 Bradford St, Mount Lawley, WA 6050 ph: (08) 9370 6360 fax: (08) 9370 6179 email: u.mueller@cown.edu.au
134
Ms Alison Keogh
Dr Sia Khosrowshahi
Mr Ian Lipton
Mr Craig Moulton
Dr Ute Mueller
Mr William Shaw
Director, Mining and Resource Technology Level 3, Kirin Centre, 15 Ogilvie Road, Mt Pleasant, WA 6153 ph: (08) 9316 2710 fax: (08) 9316 1791 email: wrt@mrtconsulting.com.au General Manager, Geoval Level 13, 256 Adelaide Terrace, East Perth, WA 6000 ph: (08) 9221 0600 fax: (08) 9221 4985 email: johnvann@geoval.com.au
Mr John Vann
135

Beyond Ordinary Kriging Master7

Încărcat de

Informații document

Descriere originală:

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Beyond Ordinary Kriging Master7

Încărcat de

Drepturi de autor:

Formate disponibile

The Geostatistical Association of Australasia

PO Box 1719 West Perth WA Australia

Beyond Ordinary Kriging:

Mining & Resource Technology

Mining & Resource Technology

Resource Service Group

Symposium on Beyond Ordinary Kriging

John Vann GAA Executive Committee Perth, October 1998.

Symposium on Beyond Ordinary Kriging

Professor Michel David

Professor Georges Matheron

Symposium on Beyond Ordinary Kriging

9.50 to 10.20 am 10.20 to 10.50 am

11.20 to 11.50 am 11.50 to 12.20 pm

Symposium on Beyond Ordinary Kriging

Symposium on Beyond Ordinary Kriging

Drinks on the Terrace

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

BEYOND ORDINARY KRIGING AN OVERVIEW OF NON-LINEAR ESTIMATION

Symposium on Beyond Ordinary Kriging

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

Vann and Guibal Keynote presentation

The necessity for change of support

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

A summary of main non-linear methods

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

Multiple Indicator Kriging

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

Median Indicator Kriging

Indicator Co-kriging and Disjunctive Kriging

Residual Indicator Kriging

The Hi ( x ) are uncorrelated and we have:

Symposium on Beyond Ordinary Kriging

Isofactorial Disjunctive Kriging

Symposium on Beyond Ordinary Kriging

Conclusions and recommendations

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

Vann and Guibal Keynote presentation

Vann and Guibal Keynote presentation

Symposium on Beyond Ordinary Kriging

Vann and Guibal Keynote presentation

Vann and Guibal Keynote presentation

Symposium on Beyond Ordinary Kriging

Glacken and Blackney A practitioners implementation of Indicator Kriging

Ian Glacken and Paul Blackney Snowden Associates

Symposium on Beyond Ordinary Kriging

Overview of theory of Indicator Kriging

Why use IK?

Figure 1. Histogram of lead data showing typical mixed populations