Mining Geostatistics - Journel PDF
Mining Geostatistics - Journel PDF
Mining Geostatistics - Journel PDF
ACADEMIC PRESS GR -
d --s-c-
: -
Mininq Geostatist~cs
L
'." .'. //(
14. G. JOURNEL
Stanford University, USA
and
CH. J. HUIJBREGTS
Bureau de Recherches Geologiques et Minieres, France
AICACIEMIC PRESS
Harcourt Brace Jovanovich, Publishers
London .San Diego . New York . Boston
Sydney Tokyo Toronto
ACADEMIC PRESS LIMITED
24-28 Oval Road,
LONDON NW17DX
.dl
Copyright 01978 by The distribution of ore grades within a deposit is of mixed character, being
ACADEMIC PRESS LIMITED partly structured and partly random. On one hand, the mineralizing process
Reprinted with corrections 1981
Third Printing 1984 has an overall structure and follows certain laws, either geological or
Fourth Printing with corrections 1989 metallogenic: in particular, zones of rich and poor grades always exist, and
Fifth Printing 1991 this is possible only if the variability of grades possesses a certain degree of
continuity. Depending upon the type of ore deposit, this degree of continuity
will be more: or less marked, but it will always exist; mining engineers can
indeed be thankful for this fact because, otherwise, no local estimation and,
consequently, no selection would be possible. However, even though
mineralization is never so chaotic as to preclude all forms of forecasting, it is
All Rights Reserved never regular enough to allow the use of a deterministic forecasting tech-
No part of this book may be reproduced in any form by photostat, microfilm, nique. This is why a scientific (at least, simply realistic) estimation must
or any other means, without written permission from the publishers necessarily t.~keinto account both features -structure and randomness -
inherent in any deposit. Since geologists stress the first of these two aspects,
and statisticlans stress the second, I proposed, over 15 years ago, the name
geostatistics to designate the field which synthetizes these two features and
opens the way to the solution of problems of evaluation of mining deposits.
Practicians are well aware that such problems are numerous and often
very complex. They sometimes infer from this that the corresponding
Library of Congress Catalog Curd Number: '77-92823 mathematical theory is necessarily very complicated. Actually, they should
ISBN Hardback: 0-12-391050-1 set their minds at ease, for this is luckily not the case. The mathematical
ISBN Paperback: 0-12-391056-0
formalism necessary for geostatistics is simple and, in some cases, even trite.
The user of A. Journel and Ch. Huijbregts' book will come across nothing
more than variances and covariances, vectors and matrices; in short, only
linear calculations comparable to those which any computer user manipu-
lates every day. After the preliminary effort o f familiarizing themselves with
the notations, they will note, possibly to their surprise, that they can follow
Printed in Great Britain by S1 Edmundsbury Press Limited all of the authors' steps without difficulty.
Bury St Edmunds, Suffoik I might address a symmetric warning to statisticians. Those unfamiliar
with the coinplexity and specificity of mining problems might find this
vi FOREWORD FOREWORD vii
treatise banal and of little interest. I feel that they would be mistaken. A t this "We would like to know, as a function of x , the tonnage T(x)of the finally
point, a comparison with geophysics could be instructive (besides, the very selected blocks, and their mean grade m ( x ) ; this time, of course, we are
term "geostatistics" was formed after the model of "geophysi~cs" o r of referring to the mean value of the true mean grades of the selected blocks,
"geochemistry"). and no't of their estimated mean grades 2:. O n t o p of this, the evaluation,
If a theoretical physicist were t o read a treatise on geophysics, h e would based upon the rather poor information we have now (drilling on a 200-m
risk getting the disappointing impression that it I S but a collection of spacing), will inevitably be erroneous. So we shall also want to know the
particular cases of the theories of potential and of wave propagation: a olt of orders of magnitude of the errors committed in the evaluation of T ( x )and
gravitation and magnetism, quite a lot of electromagnetism, acoustics, erc., tn (x), for example through their estimation variances."
nothing very original, really. A n d yet, without special training, that very In the very same our Nobel prize physicist above admits being unable
physicist, be he a Nobel prize winner, would admit to be unable to Interpret to interpret the results of a seismic prospection campaign, I think that a
the results of a seismic prospection campaign. statist~ciannot versed in mining practice would honestly have to admit his
In the same way, a specialist in mathematical stat~stics,having paged inability to give a realistic solution to the above problem. Actually, this is a
through the present book, would not find much more than variances, rather simple and favoura'ble case: only one metal, a well-explored deposit
covariances, optimal linear estimators, etc., and would find no justificstion with a regular sampling grid, no missing data points nor preferential o r
in this for creating a separate field. However, let us submit to this specialist biased data, etc. O n e should also note that the mere statement of the
the following problem, which is quite a standard one for the mining engineer problem itself takes up about a page. A statistician who is not familiar with
(in geostatistical terminology, it is the problem of the estimation of recover- mlning may well be discouraged before he can even get a good idea of the
able reserves). problem at hand.
"We have, at a certain depth, an interesting copper deposit. T h e geolo- In fact, the difficulties are several. First of all, there is a violent contrast in
gists who have studied the deposit have arrived at certain conclusions about masses: even though the experimental data are numerous (5000 samples),
its genesis and structure, which you will find in this report. T h e available they represent but a minute part of the total mineralized mass (about 5 0
numerical data consist o f a square grid of drill-holes 200 m apart. T h e core tonnes out of hundreds of millions). Then there is a volume (or support)
sections were analysed metre by metre, and, thus, we have a total of some effect: the 20 m x 2 0 rn X 1 0 m blocks are not at all distributed (in the spatial
5 0 0 0 chemical analyses. We are considering open-pit selective mining. This sense as well as in the statistical sense) in the same way as the kilogram-size
means that, during the mining of our approximately 200 ~ r deposit, n ~ which core sections. Finally, there is the fact that the criteria to b e used for the
has been mentally divided into 5 0 0 0 0 blocks 20 m x 2 0 m X 1 0 m, we have future select~ono f blocks will not be based o n the true grades of the blocks,
the choice o f sending each newly excavated block either to waste, o r to the nor upon the data novv at hand, but on an intermediate level of information
processing plant. Taking account of an economic pre-feasibil~tystudy, we (the blast-holes). O n e can, thus, understand why t h e need to solve such
have decided that the blocks with a mean ore grade greater than x = 0.446 problems would instigate, over the years, the creation of a field as specialized
C u should be sent to the plant, and that the rest should be set aside with the as geophysics, with ills own terminology and special techniques-a field
waste. This cut-off grade x is subject to revision, since the market price of which carries the name of geostatistics.
copper as well as the mining and processing costs can vary. s should be AndrC Journel and Charles Huijbregts, who have worked at the Centre de
considered as a parameter, and the evaluations should be made under G6ostatistique at Fontainebleau since its creation in 1968, are surely among
various hypotheses, i.e., with x = 0 . 3 % , 0.44b, O.j0/~,for example. the persons who have contributed the most to elevating this field to its
"Moreover, we must consider the fact that our block by block clecision is present level of practical and operational efficiency. They have personally
based, not on the unknown true block grade but on an estimator, this achieved studies and estimations upon more than 3 0 o r e deposits of the most
estimator being based on information which will be much richer than that diverse types. T h e vast experience which they thus accumulated, and the
now available (core drillings on a 200 m x 2 0 0 m grid): at tl~r: time of close and permanent contact they maintained with the everyday realities of
selection, we shall have the results of the chemical analyses of blast holes on mining, account for the primarily operational and concrete nature of this
a 5 m x 5 m grid. Based upon this final information, we shall build an book. Here is a practical treatise, written by practising mining engineers, for
estimator z: for each of our blocks, and we shall send this block to the plant the benefit of other practicians. T h e authors have deliberately - and justly
only if its estimated grade z: is greater or equal to the cut-off grade r. so - simplified to the utmost the purely mathematical part of their exposi,
viii FOREWORD
characterizes at least some of the aspects of a mineralization. The dis- words, the mining engineer must be involved, Outside the mining field,
tribution of altitude in a horizontal space characterizes a topographical geostatistics applied to forestry requires the ability to interpret*.@atial
surface. correlation between the densities of growth of different species of trees;
Let z ( x ) be the value of the variable z at the point x. The problem is to geostatistics applied to bathymetry must take into account all the problems
represent the variability of the function r ( x ) in space (when x varies). This related to the localization of the hydrographic survey ship at sea. Tb,us, a
representation will then be used to solve such problems as the estimation of book on gcostatistics cannot be dissociated from its practical context, in
the value z(x0) at a point xo at which no data are availat~le,or to estimate this case mining.
the proportion of values z(x) within a given field that are greater than a This instrumental nature of geostatistics both opposes and completes the
given limit (e.g., cut-off grade). explanatory aspects of the standard, deterministic, geological approach.
The geostatistical solution consists of interpreting each value z(x,) (z in For example, even if a metallogenic study is sometimes successful in
lower case) as a particular realization of a random variable Z(x,) (2in explaining the overall genesis of a given deposit, such a deterministic
upper case) at the point x,. The set of these auto-correlated random approach will b e clearly insufficient on a smaller scale; such as that of metal
variables {Z(x), when x varies throughout the deposit or domain D ) concentrations. In such a case, ~t is said that the grades are "randomly"
constitutes a random function. The problem of characterizing the spatial distributed within the defined genetic outline.
variability of z(x) then reduces to that of characterizing the correlations But noth~~ng is more preclse than this "random" distribution. It will differ
between the various random variables Z(x,),Z(x,) which constitute the from a goldl placer deposit to a porphyry-copper deposit, and the random
random function {Z(x), x E D ) .This fundamental interpretation is justified distribution of the grades within the deposit will condition its survey and its
a posteriori if it results in coherent and acceptable solutic~nsto the various selective mining, which represent considerable investments. On the smaller
problems encountered in practice. scale ol' mining, the probabilistic approach can relay the deterministic and
genetic approaches by which the deposit was delineated on a larger scale.
Our lumberman is not a botanist; similarly, the mining geostatistician
accepts the reality imposed by nature, and is content to study the
The adaptability of geosraristics
consequences of this reality on the mining project he must prepare. A
Geostatistics, and the probabilistic approach in general, is particularly I probabilistic approach can, of course, be guided by genetic considerations,
suited to the study of natural phenomena. Among the. many techniques and and the best mining geostatistician is the one who is able to combine his
mathematical models available to the practitioner of geomathemat~cs,a geological knowledge and his technical mining skill with a good use of the
particular technique o r approach will be chosen if, from experience, it is probabilistic language.
known to be best suited to the problem concerned. A Fourier analysis will
suit one problem, while, in other cases, geostatistical tools are best suited;
Plan of'the book
in some cases, a mathematical formulation may even be rejected for simple
qualitative experience. The acceptance of geostat~stics,and mining geostatistics in particular, rests
In the mining field, geostatistics provides a coherent set of probabil~stic on the coherence and effectiveness of the solutions it provides to the
techniques, which are available to each person (geologist, mining engineer, various problems encountered in practice. Thus, the primary aim of this
metallurgist) involved in the mining project; the proper use of these tech- book is to convince the varlous people involved in a mining project of the
niques depends on these people. Just as the efficiency of the axe and the practical interest of geostatistics. An engineer will not be convinced in the
saw depends on the arm of the lumberman, so the efficiency of any same manner as a mathematician or a researcher; he will first want to know
geostatistical study depends on the liaison between the various people and how the techniques have fared in practice. For this reason, before begin-
departments involved. The proper conclusions of an analysis of spatial ning any theoretical development,a practical presentation of geostatistics,
variability will be reached only if there is a close contact with the geologist by way of its mining applications, is given in Chapter I. This chapter can be
or, better still, if the analysis is carried out by the geologist himself. T h e read separately from the rest of the bbok, and is intended as a rksume' for
evaluation of recoverable reserves has no meaning unless technically readers who are pressed for time, and anxious to make a quick decision on
feasible processes of selection and mining are clearly defined; in other the usefulness of the proposed techniques. In the remainder of the book,
4 MINING GEOSTATISTICS INTRODUCTION 5
I
each new concept is introduced by a practical example drawn f r o m the quoted deposits and F O R T R A N programs. A t a first reading, the reader is
study of a real deposit. advised to o m ~ tthe various theoretical demonstrations. These demon-
T h e theoretical structure of the book begins with Chapter 11, which strations are necessary for the logical coherence of the book, but not for a
presents the theory of regionalized variables and the main tool of geo- preliminary overall understanding of geostatistics.
-
statistics t h e variogram - and the two variances of estimation and dis-
persion. First norations
Any geostatistical study begins with a structural analysis, which consists
of modelling t h e in situ spatial variabilities of the parameters studied, e.g., For the un~nitiated, one of the main difficulties of geostatistics is the
grades, thicknesses, accumulations. Chapters 111 and I V show how a struc- notation. These notar.ions are precisely defined as they enter into the work,
tural analysis is carried out in practice from the construction of variograms and are summarlzed in Index D. However, it may prove helpful at this
to their interpretation and fitting by models. Both chapters insist o n the stage to introduce the following basic notations.
"skilled" nature of the structural analysis. Even though many computer A deposlt is always a regionalized phenomenon in the three-dimen-
programs have been written for this purpose, a structural analysls must sional space. T o simplify notation, the coordinates (xu, x,, x,) of a point are
never be carried out blindly; it must integrate the qualitative experience of usually denoted x. Thus, the integral (extended over the three-dimensional
the geologist and carry it through t o the subsequent resource-reserves space) of the function f(x,, xu, x,) is wrltten as
estimation and the eventual planning of the mining project.
It is obvious that a massive porphyry-copper deposit cannot b e estimated
in the same way as a sedimentary phosphate deposit. T h e estimation
procedure must take into account both the structure of t h e spatial vari- T h e vector with three-dimensional coordinates (h,, h,, h,) is denoted h,
ability peculiar to each deposit, and the particular manner in which the with a modulus
deposit was surveyed. It is only by taking these peculiarities into account
that a confidence interval can be ascribed to each estimated value. Kriging
takes these factors into account, and is presented in Chapter V, along with and direction a. Note that a direction in a three-dimensional space is
the geostatistical approach to the estimation of the in situ resources. specified by two angles, its longitude and its latitude.
T h e next step is to evaluate the proportion of these in situ resources that Lower-case letters, z or y, are reserved for real values as, for example,
can b e recovered within a given technological and economical framework. the grade z(x) at point x or the density y(x) at t h e same point. T h e
This estimation of recoverable reserves must take the nature of the pro- corresponding upper-case letters, Z o r Y, denote the interpretation of this
posed selection and mining methods into account. Chapter V I presents the real value as a random variable or function, as, for example, the random
geostatistical formalization of these problems of selection and estimation function Z ( x ) with expectation E { Z ( x ) j . T h e real value z(x) is said to be a
of recoverable reserves. particular realization of the random variable or function Z ( x ) .
Since geostatistics interprets the regionalized variable z ( x ) as a partic- A n estimated value is distinguished from a real value, known o r
ular realization of a random function Z ( x ) , the possibility arises of drawing unknown, by an asterisk *. Thus, z*(x) represents the estimated value of
other realizations of this same random function, in much the same way as z(x), and Z * ( x ) represents the interpretation of this estimator as a random
casting a die several times. This new, simulated realization represents a variable or function.
numerical model of the real spatial variability z(x). Within certain li~nits,
t h e consequences of various planned mining methods can then b e
examined by simulating these methods o n the numerical model. T h e tech-
niques f o r simulating deposits and the practical uses of these sinlulations
a r e given in Chapter VII.
Each chapter begins with a summary of its contents, so as t o allow the
reader t o sort out items of particular interest. For this purpose, the reader
can also m a k e profitable use of t h e indexes of geostatistical concepts,
Geostatistics and Mining
Applications
SUMMARY
This chapter IS Intended to be both an introduction to, and an overview of,
geostatist~csas applied to the mining industry. Each successive problem encoun-
tered from the exploration phase to the routine operatmg phase is presented along
with ~ t geostatistical
s approach. The problems of characterizing v a r ~ a b ~ l i tstruc-
y
tures of variables of interest within the deposit are solved with the "variogram".
Optimal local estimates with corresponding confidence intervals are found by
"krigirg". T h e choice of various working techniques (machinery, need for stock-
p~lingor blend~ng)can be facilitated by simulating them on a geostatistically
determined numer~calmodel of thc deposit.
The main steps in proving a deposit are outlined schematically in section 1.A
along w ~ t hsome typ~calevaluation problems arising from such steps.
Section I B ~ntroducesthe geostatlstical terminology from a practical and intui-
tive point of view. Such key notions as regionalized variables, variogram, estima-
tion and d ~ ~ p e r s ~variances,
on are presented.
The geostat~st~cal approach to several of the problems encountered In section 1.A
is glten in s e c t ~ o nI C, which also o u t l ~ n e sthe sequence of steps in a geostat~stical
study.
Geological surveys resources, depend on the particular deposit and the type of ore considered.
And, finally, he knows that f o r a given o r e the variability of assays of core
A preliminary survey of a province o r favourable metallogenic e1.1viron- samples, cuttings or channel samples can b e markedly different.
ment, by such various means as geological mapping, geochemistry o r In this chapter, it will be shown how geostatistics takes these various
geophysics, will usually result in the location of a certain number of concepts into account by means of the variogram and structural analysis.
prospects. If one of these prospects locates a mineralized zone, the next
step is t o determine its value and extent by a more detailed survey in-
volving, for example, drilling, trenches o r even, sometimes, drives. T h e Sampling on a small grid
locations of these samples are generally not regular, but are based on Once the deposit has been deemed economically mineable, the next step is
geological considerations such as the need to verify the existence of the to define the technological and economical framework for its exploitation.
mineralization and to provide a qualitative description of it (e.g., T h e total in situ resources of a deposit are seldom entirely mineable,
mineralogy, tectonics, metallogenic type). T h e results of such sampling both for technological reasons (depth, accessibility) and for economic
may b e a proper basis on which to plan further reconnaissance, but they are reasons (mining cost, cut-off grade). T o define recoverable reserves within
not entirely suited for quantitative estimations of tonnage and yardage. a given technological and economic context, or, conversely, to define the ,
T h e first estimations will tend to b e more qualitative than quantitative, as type of mining which will result in maximum profit from the deposit it is
will be thegeologist's image of the deposit. necessary to have a detailed knowledge of the characteristics of this
deposit. These characteristics include t h e spatial distribution of rich- and
poor-grade zones, the variability o f thicknesses and overburden, the
Sys~ernaticsurvey on a large grid various grade correlations (economic metals and impurities).
T h e second stage of the survey usually consists of drllling on a large and T o determine these characteristics, it is usual to sample on T, small grid,
more or less regular grid. I n general, this grid will cover what is considered at least over the first production zone. T h e results of this sampling
to be the mineralized zone and should be sufficient to evaluate the overall compaign are then used to construct block estimations within this zone.
in situ resources (tonnages and mean grades). If promising results are These local estimatioris set the usual problem of determining estimators
obtained from this second stage, t h e survey may b e extended to some in-fill and evaluating the resulting estimation error. There will also b e problems
drilling o r by drilling in adjacent zones. A t this stage, the geologist and associated with the use o f these local estimations: the evaluation of
project engineer are faced with two fundamental quantitative problems. recoverable reserves d,epends o n whether t h e proposed mining method is
T h e first one is how to deduce t h e required global estimations from the block-caving with selection units of t h e order of 1 0 0 m x 1 0 0 m x 1 0 0 m, or
available information, which consists of qualitative geological hypotheses open-pit mining with selection units of the order of 1 0 m x 10 m x 10 m.
and quantitative data from core samples, trenches, pits, etc. In @eneral, From experience, the mining engineer knows that, after mining, the
estimations of a geometric nature such as mineralized surface and yardage actual product will differ from t h e estimations of the geologist. I n general,
will make more use of qualitative information than esrimations of a h e knows that tonnage is underestimated and, more importantly, quality is
"concentration" nature such as mean grades and densities. overestimated. In practice, this usually results in a more o r less empirical
T h e second problem is to express the confidence that can be given to "smudge factor" being applied t o t h e estimation of recoverable reserves,
these estimations. In practice, it is common t o classify the in situ resources which may reduce the estimated quantity of recovered metal from 5 to
in different categories such as, for example, those recommended by the US 20% This is not a ver!y satisfactory procedure, because it does not explain
Bureau of Mines: measured ore, indicated ore and inferred ore, T h e the reasons for these systematic biases between the geologist's estimations
boundaries between these categories d o not always appear clearly. T h e and actual production; in addition, it is far t o o imprecise when evaluating a
geologist knows from experience that the precision of an estimation, such marginally profitable new deposit.
as mean grade, does not only depend o n the number of the available assays Given the size of present day mining projects and the cost of their
but also o n the in situ variability of t h e grade in the deposit and the manner investments, the risks of such empirical methods are n o longer acceptable,
in which this variability has been sampled. H e also knows that the ccmcepts and it is imperative to set and rigorously solve such problems as those given
of continuity and zone of influence, which are used In the evaluation of below.
10 MINING GEOSTATISTICS I. GEOSTATISTICS AND MINING APPLICATIONS 11
(i) The definition of the most precise local estimator and its confidence The dircct study of the mathematical function z(x) is excluded because
interval, taking account of the peculiarities of the deposit and the its spatial variability is usually extremely erratic, with all kinds of dis-
ore concerned. It is obvious that a gold nugget deposit cannot be continuities and anisotropies. It is equally unacceptable to interpret-all the
evaluated in the same way as a sedimentary phosphate deposit. numerical values z(x), z(xl) as independent realizations of the same
(ii) The distinction between in sinu resources and mineable reserves. It random variable 2, because this interpretation takes no account of the
should be possible to model rigorously the effect of the mining spatial auto-correlation between two neighbouring values z ( x ) and z(x +
method on the recovery of the in situ resources. h). Both the random and structured aspects of the regionalized variable are
(iii) The definition of the amount of data (drill cores, exploration drives) expressed by the probabilistic language of "random functions".
necessary for a reliable estimation of these recoverable reserves. A random function Z ( x ) can be seen as a set of random variables Z(X,)
defined at each point x, of the deposit D: Z(x)={Z(x,), Vxi ED).The
Mine planning random variables Z ( x , ) are correlated and this correlation depends on both
the vector h (modulus and direction) separating twp points x, and x, + h
Once the estimations of resources and recoverable reserves have been and the na,ture of the variable considered. At any point x,, the true grade
made, the next step is to plan the mining project itself. This involves z(x,), analysed over a core sample, for example, is interpreted as one
choosing mining appliances, type of haulage, requirements of stockpiling particular realization of the random variable Z(x,). Similarly, the set of
and blending, mill planning, etc. Some of these choices are conditional, in , E D } defining the deposit or the zone D is inter-
true grades { ~ ( x , )Vx,
all or part, upon the structures of the variabilities of the ore characteristics, preted as one particular realization of the random function {Z(xi),V X E~
either in situ or at any of the various stages of mining, haulage, stockpilmg Dl.
I
and mill feeding. It will be shown later how geostatistics can be used to The next section will deal with the operational aspect of this formaliza-
provide representations o f these variabilities by means of numerical models tion by describing how the spatial variability of the regionalized variable
of the deposit. The effects of various extraction and blending processes can z(x) can be characterized.
then be tested by simulating them on these numerical models.
I.B.2. The variogram
1.B. THE GEOSTATISTICAL LANGUAGE Consider two numerical values z (x) and z (x + I t ) , at two points x and x + h
separated by the vector h. The variability between these two quantities is
I.B.l Regionalized variable a n d r a n d o m function characterized by the variogram function 2y(x, h), which is defined as the
expectation of the random variable [ Z ( x ) - Z ( x + h ) j 2 , i.e.,
A mineralized phenomenon can be characterized by the spatial distribution
of a certain number of measurable quantities called "regionalized vari- 2y(x, h ) = E{[Z(x)-Z(x + h)12). (1.1)
ables". Examples are the distribution of grades in the three-dimensional
space, the distribution of vertical thicknesses of a sedimentary bed in the In all generality, this variogram 2 y ( x , h ) is a function of both the point x
horizontal space, and the distribution of market price of a metal in time. and the vector h. Thus, the estimation of this variogram requires several
Geostatistical theory is based on the observation that the variabilities of .
realizations, [zk(x), zk(x + h)], [zk (x), zk(x + h)], . . , [z~,.(x),zk-(.x -+ h)],
all regionalized variables have a particular structure, T h e grades z(x) and of the pair of random variables [ Z ( x ) ,Z ( x + h)]. Now, in practice, at least
+
z(x h ) of a given metal at points ~c and x + h are auto-correlated; this in mining applications, only one such realization [z(x), r ( x + h ) ] is avail-
auto-correlation depends on both the vector h (in modulus and direction) able and this is the actual measured couple of values at points x and x +h.
which separates the two points, and on the particular mineralization T o overcome this problem, the intrinsic hypothesis is introduced. This hy-
considered. The variability of gold grades in a placer deposit will differ pothesis is that the variograrn junction 2 y ( x , h ) depends only on the
from that of gold grades in a massive deposit. The independence of the two separation vector h (modulus and direction) and not on the location x. It is
grades z(x) and z(x + h ) beyond a certain distance h is simply a particular then possible to estimate the variogram 2y(h) from the available data: an
case of auto-correlation and is treated in the same way. I estimator '2-y*(h) is the arithmetic mean of the squared differences between
., * ,*
1.8.3. The estimation variance be known. In mining applications however, the standard 95%
Gaussian confidence interval [ * 2 a E ] is used, where ~riis the estimation
The variogram defined in (1.1) can be viewed as an estimation variance, the
variance of the error committed when the grade at point x is estimated by variance. ..
' ..,*
.
the grade at point x + h. From this elementary estimation variance, 2 y ( h )= Experimental observation of a large number of experimental histograms
E { [ z ( x ) - ~ ( x +h)I2), geostatistical techniques can be used to deduce the of estimatuon errors has shown that the Gaussian distribution tends to
variance of estimation of a mean grade Zv by another mean grade Z,. This underestimate the proportion of low errors (within *u&) and of high
estimation variance is expressed as errors (outside the range f 3 ~ ~ However, ) . the standard Gaussian
confidence interval [ f 2 a E ] in general correctly estimates the 95%
experimental confidence interval. Figure 11.17 shows the histogram of
The mean grades Z, and ZV can be defined on any supports, e.g., V may estimation errors of 397 mean block grades from the Chuquicarnata
represent a mining block centred on the point y, and v the set of N drill deposit in Chile. The arithmetic mean of the errors is *O.OlO/o Cu for a
cores centred on points {x,, i = 1 to N ) . The mean grades are then defined mean grade of 2% Cu. The experimental variance. of the 397 errors is
equal to 0.273 (% CU)? which is close to the geostatistically predicted
by
value of 0.261. The predicted 95% confidence interval is, according to the
1 1 "
zv --
= Z(x)d;r and Z,=- 7 iT(xi).
N iz,
Gaussian standard, [*2J(0.261)] = [&1.02% Cu]. In fact, 96% of the
observed errors fall within this interval.
The notation y(V, v ) , for example, represents the mean value of the
elementary semi-variogram function y ( h ) when the end-point x of the
vector h = (x - x l ) describes the support V and the other end-point x '
independently describes the support v , i.e., The elementary variogram 2 y ( h ) , estimated from experimental data, is
defined on the support of these data, e.g., core samples of given cross-
section and length. By means of formula (I.3), the variograrn 2 y v ( i r )
defined on any other support V can be deduced from the variogram 2 y ( h ) ,
e.g., V may be the support of the mining block. In fact,
The general formula (I.3), shows that the estimation variance, i.e., the
quality of the estimation, depends on all four of the following.
(i) The relative distances between the block V to be estimated and the appears as the variance of estimation of the mean grade of the block V
, information v used to estimate it. This is embodied in the term centred on point x by the mean grade of the block V centred on the point
T(V, t.1. x + h. This formalism for the change of support has great practical appli-
(ii) The size and geometry of the block V to be estimated. This is cations. Since future mining will be carried out on the basis of blocks and
embodied in the term y(V, V). not drill cores, i t is of great benefit to be able to evaluate the spatial
(iii) The quantity and spatial arrangement of the information u, which is variability of the mean grades of these blocks by means of the variogram
embodied in the term y(v, u ) . 2 y v ( h ) defined on the block support.
(iv) The degree of continuity of the phenomenon under study, which is
conveyed by its characteristic semi-variogram y ( h :I.
I.B.4. The dispersion variance
Thus, the theoretical formula (1.3) expresses the various concepts that
There are two dispersion phenomena well known to the mining engineer.
we intuitively know should determine the quality of an estimation.
The first is that the dispersion arpund their mean value of a set of data
collected within a domain V increases with the dimension of V. This is a
logical con~sequenceof the existence bf spatial correlations: the smaller V,
The estimation variance itself is not enough to establish a confidence the closer the data and, thus, the closer their values. The second is that the
interval for the proposed estimate, the distribution of the errors must also dispersion within a fixed domain V decreases as the support u on'which
MINING GEOSTATISTICS I. GEOSTATISTICS AND MINING APPLICATIONS 17
each datum is defined increases: the mean grades of mining bloclts a r e less I.B.6. S i m u l a t i o n s of d e p o s i t s
dispersed than the mean grades of core samples. T h e regionalized variable z(x), which is the subject of the geostatistical
These two phenomena a r e expressed in the geostatistical concept of structural analysis, is interpreted as one possible realization of the random
dispersion variance. Let V b e a domain consisting of hr units with t h e same function Z(x). It is possible to generate other realizations z,(x) of this
support 0. If the N grades of these units are known, their variance can be random function Z ( x ) by the process of simulation. This simulated real-
calculated. T h e dispersion variance of the grades of units v within V, ization r,(x), will differ from point to point from the real realization -.(x),
written D ~ ( v / v ) , is simply the probable value of this experimental vari- but, statistically, will show the same structure of variability (same histo-
ance and is calculated by means of the elementary variogram 2 ~ ( h ) grams and var~ograms);moreover, at the experimental data locations n-,, the
through the formula simulated values willl b e equal to the true values, i.e., z,(x,)= z(x,), Vx, E
0 2 ( u / V) = y ( v , V ) - y(v, v). (1.4) data set. T h e simulated and actual realization will, thus, have the same
concentrations of rich and poor values a t the same locations. T h e two
Taking the generally increasing character of the variogram into account, real~zationscan be viewed as two possible variants of the same genetic
it can be seen that D'(v/v) increases with the dimension of V a n d - phenomenon characterized by the random function Z ( x ) . T h e simulated
decreases with the dimensions of v . realization, however, has the advantage of being known at all points and
This formula can be used, for example, t o calculate the dispersion not only at the data points x,.
variance of the mean grades of production units when the size of the units Various methods of extraction, stoping, blending, stockpiling, etc., can
( u ) varies or when the interval of time considered varies (i.e., V varies). then be simulated o n this numerical model {z,(x), x E D } of the deposit o r
zone D. Using feedback, the effect of these simulated processes on recov-
ery, fluctuations o f o r e characteristics, and on the run3ing of the mirle and
I.B.5. Coregionalizations
T h e concepts of variogram and various variances, which so fa] have been
defined for the regionalization of o n e variable, can be generalized to spatial
coregionalizations of several variables. In a lead-zinc deposit, for example,
t h e regionalizations of the grades in lead, Zl(x), and in zinc Z2(x), are
characterized by their respective variograms 2 y l ( h ) and 2y?(h). However,
these two variabilities are not independent of each other and we can define
a cross-variogram for lead a n d zinc:
~ Y I ~ ( ~ ) = E { [ Z I ( X ) - Z+I ~( X
) I ~ Z ~ ( X ) - Z+
Z h)I}.
(X (1.5)
O n c e the cross-variogram is calculated, we can form the matrix
mill in general can be used to modify certain parameters of the mining Case study, after J. P. Delhotnme a n d P. Delfiner (1973)
project.
This was a 'study of the regionalization of water heights (measured-in rain
The techniques for the simulation of a regionalization z ( x ) can be
gauges) after rain showers, but ~t could equally well apply to the regional-
extended to simulations of coregionalizations, i.e., to the: simulations of the
ization of the vertical thickness of a sedimentary bed.
simultaneous spatial distributions of the various charxtcristic variables
Figure 1.2 shows the location of rain gauges in the Kadjemeur basin in
(thicknesses, grades, density) of a mineralization.
Tchad (Central Africa). After one particular rainfall, the height of water in
each rain gauge was measured. These data were then used to construct an
experimental semi-variogram to ~ h i c ha linear model was fitted. Figure 1.3
shows lhree simulated rainfalls, each of which has the same linear vario-
gram as the real rainfall and the same water heights at the actual rain gauge
locatiol~s.Pit points other than these rain gauge locations, the three simu-
lations can differ markedly from each other. This is iloticeably evident in
the extreme right-hand corner of the basin, where there are no rain
gauges
and negative drill-holes. Mineralized thickness o r mean grade may be (ii) T h e variance o f estimation u: = E { [ z ~ - z $ ] ~ is
) minimal. This
estimated by the'arithmetic mean of the positive data o n a pseudo-regular minimization of the estimation variance amounts to choosing the
grid, or by any of the traditional methods of extension of profiles or (linear) estim,ltor with the smallest standard Gaussian 95%
cross-sections. However, through the estimation variance, geostatistics can confidence interval (k2uE).
quantify the reliability of each of these traditional methods and so provide Consider the example shown o n Fig. 1.4, in which structural isotropy is
a criterion for choice between them. T h e estimation variance ciin also be assumed, i.e., the variability is the same in all directions. Provided that all
used to objectively classify resources and to predict the number, location the N = 8 data are of the same nature, and all defined o n the same support,
and distribution of additional information required to improve the quality kriging gives the following m o r e o r less intuitive results:
of estimations, e.g., to move a given tonnage from the "possible" to the
"proved" category, cf. A. G. Royle (1977). (i) T h e symmetry relations, A2 = A 3 , A , = A 6 ;
(ii) The inequalities, A l a A , , V i Z 1 ; A 7 G A 4 , i.e., t h e datum Z4 screens
I.C.2. Local estimation the influence of 2 7 ; A e 3 A 4 , Le., part of the influence of Z4 is
transferred to the neighbouring data Z5, z6and Z,.
O n c e the deposit has been deemed globally mineable, the next phase is
local, block by block estimation. This local estimation gives an idea of the
spatial distribution of the in situ resources, which is necessary for the
evaluation' o f the recoverable reserves. This local estimation is u*jually
carried out on the basis of samples obtained from a smaller sampling grid,
e.g., through in-fill drilling. T h e estimation variance of a give11 panel, cf.
formula (I.3), can be calculated as a function of the grid size and, thus,
indicates the optimal sampling density to b e used for t h e local estimation.
Once this second, denser sampling campaign is completed, the next
problem is to determine the best possible estimate of each block. This
amounts t o determining the appropriate weight to b e assigned to each
datum value, whether inside o r outside the panel. T h e determination of
these weights must take into account the nature (core samples, channel
samples) and spatial location of each datum with respect to the block and
the other data. It must also take into account the degree of spatial
continuity o f the variable concerned, expressed in the various features of
the variogram, Zy(jhl, a). FIG. 1.4. Kriging configuration of block V,
F o r example, suppose that t h e mean grade Zv of the block V is to be
estimated from a given configuration of N data Z,, as shown on Fig, 1.4.
T h e geostatistical procedure of "kriging" considers as estimator Z $ of this I.C.3. Selection a n d recoverable reserves
mean grade, a linear combination of the N data values:
T h e terms "selection" a n d "recoverable reserves" have little meaning
unless they are used within the context of precisely defined technological
conditions for recovery. From experience, the mining engineer knows that
T h e kriging system then determines the N weights {A,, i = 1 to N } such the quantity and quality of recovery improves as:
that the following hold.
(i) T h e mining selection unit becomes smaller (although this intuition is
(i) T h e estimation is unbiased. This means that, on average, the error true only for highly selective mining);
will b e zero for a large number of such estimations, i.e., (ii) The information available at the time of selection improves, to allow
E{Zv - Z*v}= 0. better discrimination between rich and poor units.
MINING GEOSTATISTICS I. GEOSTA'TISTICS A N D MINING APPLICATIONS 23
These two essential concepts are formalized by geostatistics in the the time of' selection. Selection is made on the basis of the estimated grades
following way. of these unlts, which will inevitably result in a recovery different from
. .. . the
optimum recovery based on the true grades.
Figure I.5(b) shows the histogram of the true grades of the units V
1. Recovery depends o n the support of the selection unil superimposed on the histogram of their estimated values. As actual selec-
Consider the two histograms shown on Fig. IS(a). The first one represents tion will bc made on the basis of these estimated values, the proportion of
the distribution of the mean grades of 5 m x 5 m x 5 m units u in the deposit recoverable units V must be read from the histogram of the estimated
D. These units may represent, for example, selective underground mining grades (dotted area on Fig. I.5(b)) and not from the histogram of the true
by small blocks. The second histogram represents the distribution of the grades (hatched area on Fig. I.S(a)).
mean grades of 50 mX 50 m x 50 m units V in the same deposit D, which If the estimated grades are obtained by kriging, then the dispersion
may represent, for example, mining by block-caving. It is obvious that the varlance ,D2(V*/D) of the histogram of the estimated grades can be cal-
first histogram ( v ) will be much more dispersed than the second one (V), culated from a relation known as the "smoothing effect" due to kriging:
increase. This behaviour is characteristic of the vertical variability of nickel In practically all deposits, a characteristic behaviour o r structure of the
grades in the lateritic type deposits of New Caledonia, (cf. C h . Huijbregts spatial variability of the R e v under study can b e discerned behind a locally
and A. Journel, 1972). erratic aspect. In almiost all deposits, there exist zones which are richer than
others; samples taken in a rich zone will be, o n average, richer than those
taken in a poorer zone, i.e., the value of the R e v f(x) depends on the
spatial position x, seic Fig. 11.2.
Thus a R e v possesses two apparently contradictory characteristics:
(i) a local, random, erratic aspect which calls to mind the notion of
random variable;
(ii) a general (or average) structured aspect which requires a certain
functional representation.
A proper formulation nus st take this double aspect of randomness and
structure into account in such a way as to provide a simple representation
of the spatial variability and lead to a consistent and operational approach
to the solution of problems. O n e such formulation is the probabilistic
interpretation as provided by random functions.
general, independent but are related by a correlation expressing the The sel of 2111 these distribution functions, for all positive integers k and for
spatial structure of the initial regionalized variable z ( x ) . every possillte choice of support points in R", constitutes the "spatial law"
of the l i F Z k x ) .
Statistical inference In minin,g applications, the entire spatial law is never required, mainly
because the first two moments of the law are sufficient to provide an
The probabilistic interpretation of a R e v z ( x ) as a particular realization of acceptable approximate solution to most of the problems encountered.
a certain RF Z ( x ) has an operative sense only when it is possible to infer all Moreover, the amount of data generally available is insufficient to infer the
or part of the probability law which defines this R F in its entirety. entire spatial law. In geostatistics (more precisely, linear geostatistics), only
Obviously, it is not rigorously possible to infer the probability law of a the first two moments of the RF are used; in other words, no distinction is
R F Z ( x ) from a single realization z ( x ) which is, in addition, limited to a made between two RF's Z , ( x ) and Z 2 ( x )which have the same first- and
finite number of sample points x,. It is impossible, for instance, to deter- second-order moments and both functions are considered as comprising
mine the law of the R V "the result of casting the die" from the single the same model.
numerical result 5 resulting from a single cast of the dit:, and, in particular,
it is not possible to determine whether or not the die is biased. Many casts
of the die are necessary to answer such a question. Simidarly, many realiza- Mathematical expectation, or first-order moment
tions r l ( x ) , z ~ ( x ) ,. . . , zk(.r) of the R F Z ( x ) are required in order to infer Consider a RV Z ( x ) at point x. If the distribution function of Z ( x )has an
the probability law of Z ( x ) . Since, in practice we shall be limited to a single expectation (and we shall suppose that it has), then this expectation is
realization { r ( x , ) }of the R F at the positions x,, we appear to have come to a generally a function of x, and is written
dead end and, in order to solve it, certain assumptions are necessary. These
assumptions involve various degrees of spatial homogeneity and are intro- E { Z ( x ) }= m ( x ) .
duced under the general heading of the hypothesis of stationarity.
In practice, even if only over a certain region, the phenomenon under Second -ordrermoments
study can very often be considered as homogeneous, the R e v repeating
itself in space. This homogeneity or repetition provides the equivalent of The three second-order moments considered in geostatistics are as follows.
many realizations of the same RF Z ( x ) and permits a certain amount of (i) The variance, or more precisely the "a priori" variance of Z ( x ) .
statistica1 inference. Two experimental values z ( x o ) and z(xo-t-h ) at two When this variance exists, it is defined as the second-order moment
different points xo and xo+h can, thus, be considered as two different about the expectation m ( x ) of the R V Z ( x ) , i.e.,
realizations of the same R V Z ( X O )This . type of approach is not peculiar to
geostatistics, it is used to infer the distribution law of the RV Z ( x ) from a Var { Z ( x ) )= E { [ Z ( x ) -tn(x)lZ}.
histogram of data values { z ( x , ) }or , more simply, to infer the mathematical As with the expectation m ( x ) ,the variance is generally a function of
expectation E { Z ( x ) }from the arithmetic mean of the data. X.
(ii) The covariance. It can be shown that if the two RV's Z ( x 1 ) and
Z ( x z )haye variances at the points x l and x:, then they also have a
ll.A.2. Moments and stationarity covariance which is a function of the two locations X I and xz and is
Consider the R F Z ( x ) . For every set of k points in R" (n-dimensional writt.en as
space), X I ~, 2 , .. . , xk, called support points, there corresponds a k - C b l ,~ 2 =)E { [ Z ( x 1 ) - m(x1)1[Z(x2>-m(xz)l).
component vectorial random variable
(iii) T h e variogram. The variogram function is defined as the variance of
the increment [ Z ( x l ) - Z ( x z ) ]?nd
, is written as
2 r ( x l , X Z ) = Var { Z ( x l ) - Z ( x d ) .
The function y ( x l , x 2 ) is then the "semi-variogram".
32 MINING GEOSTATISTICS , ,
11. THE THEORY OF REGIONALIZED VARIABLES 33
The hypothesis of starionarity Relation (11.4) indicates that, under the hypothesis of second-order
stationarity, the covariance and the variogram are two equivalent tools for
It appears from the definitions that the covariance and variogram functions
characterizing the autlo-correlations between two variables Z ( x + h) and
depend simultaneously on the two support points xl and x ~If. this is indeed
Z ( x ) separated by a distance h. W e can also define a third tool, the
the case, then many realizations of the pair of RV's {Z(xl), Z ( x 2 ) )must be
correloaram :
available for any statistical inference to be possible.
On the other hand, if these functions depend only on the distance
-
between the two support points (i.e., on the vector h = X I x2 separating x l
and xz), then statistical inference becomes possible: each pair of data The hypothesis of second-order stationarity assumes the existence of a
{ z (xk),t (xkr)} separated by the distance (xi, - xk ), equal to the vector h, can covariance and, thus, of a finite a priori variance, Var {Z(x)} = C(0). Now,
be considered as a different realization of the pair of RV's {Z(x ,}, Z(x2)}. the existence of the variogram function represents a weaker hypothesis
It is intuitively clear that, in a zone of homogeneous mineralization, the than the existence of the covariance; moreover, there are many physical
correlation that exists between two data values z(xk) and ~ ( x k does ) not phenomena and RF's which have an infinite capacity for dispersion (cf.
depend on their particular positions within the zone but rather on the section II.A.5, Case Study I), i.e., which have neither an a priori variance
distance which separates them. nor a covariance, but for which a variogram can be defined. As a
consequence, the second-order stationarity hypothesis can be slightly
Strict stationariry A R F is said to be stationary, in the strict sense, when reduced when assuming only the existence and stationarity of the vario-
its spatial law is invariant under translation. More precisely, the two gram.
k-component vectorial RV's { Z X ). . . , Z x } and {Z(xl +
.
h), . . , Z ( x k + h ) } are identical in law (have the same k-variable dis- Intrinsic hypothesis A. RF .Z(x) is said to be intrinsic when:
tribution law) whatever the translation vector h.
However, in linear geostatistics, as we are only interested in the pre- (i) the mathematical expectation exists and does not depend on the
viously defined first two moments, it will be enough to assume first that these support point x,
moments exist, and then to limit the stationarity assumptions to them. E{Z(x)} = m, Vx;
Stationarity of order 2 A R F is said to be stationary of order 2, ,when: (ii) for all vectors h the increment [ Z ( x + h)-Z(x)] has a finite variance
(i) the mathematical expectation E{Z(x)} exists and does not depend which does not depend on x,
on the support point x ; thus, Var { Z ( x + h ) - Z ( x ) } = E { [ Z ( x + h ) - 2 ( x ) 1 2 ) = 2 y ( h ) , Vx. (11.6)
E{Z(x)}= m , Vx; (11.1) Thus, second-order stationarity implies the intrinsic hypothesis but the
converse is not true: the intrinsic hypothesis can also be seen as the
(ii) for each pair.of RV's {Z(x), Z ( x +h)} the covariance exists and limitation of the second-order stationarity to the increments of the RF
depends on the separation distance h, (x 1.
C ( ~ ) = E { Z ( X + ~ ) ~ Z ( X ) ) - ~VX,
~, (11~2)
Quasi-starionariry In practice, the structural function, covariance or
where h represents a vector of coordinates (h,, h,, h,) in the three- variogram, is only used for limited distances lhl s b. The limit b represents,
dimensional space. for example, the diameter of the neighbourhood of estimation (i.e., the
zone which contains the information to be used, cf. Fig. 11.3). In other
The stationarity of the covariance implies the stationarity of the variance cases, b may be the extent of an homogeneous zone and two variables
and the variogram. The following relations are immediately evident: Z ( x k ) and Z ( x k+ h ) cannot be considered as coming from the same
homogeneous minera1iz:ation if Ihl> b.
In such cases, we can, and indeed we must, be content with a structural
~ ( h ) = % [ z ( x + h)-Z(x)12}= C(0)-C(h), Vx. (11.4) function C(x, x i - h ) or sy(x, x + I T ) , which is no more than locally stationary
34 MINING GEOSTATISTICS 11. THE THEORY OF REGIONALIZED VARIABLES 35
(for distances Ihl less than the limit b ) . This limitation of the hypothesis of Their proof is o f no interest in the application of geostatistics; we shall
second-order stationarity (or the intrinsic hypothesis if only the variogram simply state, the properties and comment on them.
is assumed) to only those distances Ih]S b corresponds to an hypothesis of
quasi-stationarity (or to a quasi-intrinsic hypothesis).
In practical terms, we can define sliding neighbourhocrds inside of which Positive definite conditions
the expectation and the covariance can be considered as stationary and the Let Z ( x ) be a stationary RF of expectation m and covariance C ( h ) or
data are sufficient to make statistical inference possible. semi-variogram y ( h ) . Let Y be any finite linear combination of the type
The hypothesis of quasi-stationarity is really a compromise between the n
scale of homogeneity of the phenomenon and the amount of available data. Y = I1
= I AIZ(x,)
(11.7)
In fact, it is always possible to produce stationarity by considerably reduc-
ing the dimension b of the zones of quasi-stationarity, but then most of for any weights A,.
these zones would not have any data in them, and, thus, it would be This linear combination is a RV and its variance must never be negative,
impossible to infer the quasi-stationary moments in these zones. Var { Y ) a0. Explicitly, this variance is written
From a data grid of 1 to 3 m, a 200-m-long hiIlside would appear as a
quasi-stationary model, but would require a non-stationary model
(because of the increasing altitude) for a grid of 20 to 30 m, cf. Fig. 11.3. If Var {Y}= I C,I AIA,C(x,- x l ) 3 0. (11.8)
the data grid is of the order of 200 rn, then a certain stationarity again
becomes evident: at the scale of the peneplanar basin the entire hillside
appears as a local variability of the stationary RF "altitude". This example The covariance function must be such that it ensures that the previous
shows, incidentally, that it is not possible to devise a theclretical test for the variance w~llalways be positive or zero. By definition then, the function
a priori verification of the hypothesis of stationarity using a single realiza- C ( h ) is said to be "positive definite".
tion of the RF. It is always possible to find a particular realization of a Thus, not any function g ( h ) can be considered as the covariance of a
stationary RF which would display a systematic and continuous change in stationary RF: it is imperative that it be positive definite.
value (apparent non-stationarity) over a limited distance, so that a Taking account of relation (II.4), expression (11.8) can be written in
theoretical test could never refute the hypothesis of stationarity. terms of the semi-variogram:
In the case where the variance C(0) does not exist, only the intrinsic
hypothesis is assumed, and the variance of Y is defined on the condition
that
- - - - - -- - -
200 rn
X
* and the term C(0) is thus eliminated, leaving
When only the intrinsic hypothesis is assumed, y ( h ) exists b u t C ( h ) may T h e definition of the variogram as the variance of increments entails the
not exist; the only linear combinations that have a finite variance a r e those following properties:
verifying the weights condition C,A, = 0.Thus, if only these authorized
linear combinations are considered (which is the case in linear glzostatis-
tics), there is no need to calculate the a priori variance C(O), nor t o know In general, but not always, as h increases, the mean quadratic deviation
the expectation rn = E { Z ( x ) }o r the covariance C ( h ) ;it is enough to know between the two variables Z ( x ) and Z ( x + h ) lends to increase and so y ( h )
the semi-variogram. increases from its initial zero value.
Note that, for the authorized linear combinations, formula (11.10) is
simply deduced from formula (11.8) by substituting (- y ) for C . This rule
for passing from a formula written in terms of covariance to the cor-
responding formula written in terms of semi-variogram is general in linear
geostatistics.
IC(h)j G C(O), Schwarz's inequality. 1 Transition phenomena Very often, in practice, the semi-variogram stops
increasing beyond a certain distance and becomes m o r e o r less stable
Since the degree of correlation between two variables Z ( x ) and Z ( x + h ) around a limit value y ( ~called
) a "sill" value, which is simply the a priori
generally decreases as the distance h between them increases, so, in variance of the R F
general, does the covariance function, which decreases from its value at the ?,(co) = Var { Z ( x ) }= C ( 0 ) . (11.13)
origin C(0). Correspondingly, the semivariogram y ( h ) = C(0)- C ( h ) In such cases, the a priori variance exists as does the covariance. Such
increases from its value at the origin, y(O)= 0, see Fig. 11.4, variograms, which are characterized by a sill value and a range, are called
"transition" models, and correspond to a R F which is not only intrinsic but
Absence of correl~tion also second-order stationary, cf. Fig. 11.4.
Very often, in practice, the correlation between two variables Z ( x ) and Zone of influence In a transition phenomenon, any data value z ( x )will be
Z ( x + h ) disappears when t h e distance h becomes too large: correlated w ~ t hany other value falling within a radius a of x. This cor-
relation and, hence, the ir~fluenceof one value on the other will decrease as
C ( h ) + 0 , when lhl+ co,
the distance between the two points increases. Thus, the range corresponds
and, in practice, we can put C ( h )= 0 , once Ihl2 a. T h e distance a beyond to the intuitive idea of a z o n e of influence of a RV: beyond the distance
which C ( h )can be considered t o b e equal to zero is called the "range" and lhj = a, the RV's Z ( x ) and Z ( x + h ) are no longer correlated.
it represents t h e transition from the state in which a spatial correlation
exists (Ih I < a ) to the state in which there is absence of correlation (lh I > a ) , Anisorropies There is n o reason to expect that the mineralization will
exhibit the same behaviolur in every direction, i.e., that the mineralization
cf. Fig. 11.4.
-
38 MINING GEOSTATISTICS 11. THE THEORY OF REGIONALIZED VARIABLES 39
A s a consequence, an experimental variogram which increases at least as 2. A cross-semi-varijogram ykfk(h) can take negative values, whereas a
rapidly as lhlZ for large distances h is incompatible with the intrinsic direct semi-variogram is always positive. A negative value of the cross-
hypothesis. Such an increase in the variogram most often indicates the semi-variogram ind~catesthat a positive increase in one of the variables
presence of a trend o r drift,? i.e., a non-stationary mathematical expec- (k') corresponds, on average, to a decrease in the other (k). This could be
tation: explained, for example, by a phenomenon in which element k is substi-
tuted for element k'.
E I Z ( x ) } = m(x),
3. Under the second-order stationary hypothesis, the existence of the
where m (x) depends on x. cross-covariances entails that of the cross-variograms and these two groups
of structural tools are related by the expression
ll.A.4. Caregionalization
A regionalized phenomenon can b e represented by several inter-correlated 4. T h e cross-variogram is symmetric in (k', k ) and (h, - h ) while this is
variables and, in certain cases, it may b e useful to study them sirnul- not necessarily so for the cross-covariance:
taneously. Y A ,(h)
~ = Y W ( ~ ) and Y Y (~h ) = ~ k z (-h
k 1,
In a lead-zinc deposit, for example, the two mineralizations are closely (11.18)
Ckks(h)= C ~ , k ( - h ) and C ~ X ( - h )f Ckzk(h).
related and it may b e of interest to know if, o n average, zones that are rich
in P b are equally rich in Zn. In other polymetallic deposits, it would be A lag effect in one variable in relation to another (e.g., rich P b grades
useful to b e able to use the information provided by one variable to make lagging behind rich Zn grades in a given direction) is represented by a
u p for missing information o n another variable. dissymetr~ccross-covariance in (h, -h). Such a lag effect would not be
T h e probabilistic approach to a coregionalization is similar to that of the visible on the cross-variogram, and, where it is suspected, the cross-
regionalization of a single variable. covariance should be used. Otherwise, for reason of consistency with the
The simultaneous regionalizations (or the coregionalization) of K ReV's study of the direct variograms, it is preferable l o use cross-variograms.
.
{zl(x), , . ,ZK(X)} are interpreted as a particular realization of the set of K 5 . Correlariotz coeficrent. 'The point to point correlation coefficient
inter-correlated RF's {Zl(x), . . . , ZK(x)}. between two variables & (x) and Zk(x) located at the same point x is
U n d e r the hypothesis of second-order stationarity we define: written
(i) for each R F Zk(x), t h e mathematical expectation
If the coregionalization {Zl(x), . . . , ZK(x)} verifies the intrinsic hy- The experimental curve of Fig. 11.7 can be approximated by a straight
pothesis, then the R F F(x, A k ) is also intrinsic anld has the following line representing the uninterrupted logarithmic growth of the variance with
stationary semi-variogram: the increase in field size V: ..
a 2 ( v )= a log (V/v),
where u = 0.0075 ft2 is the area of cross-section of the analysed core
and this single direct semi-variogram T ( h ) will be used in subsequent samples, and V is the horizontal surface area of the field. The a priori
operations. In particular, the economic function f(x, A k ) can be kriged variance of the logarithms of the gold grades should be the limit of a2(v)
(estimated optimally). as the field size V tends towards infinity. Experimentally, such a limit does
not exist and, as a consequence, a second-order stationary model of RF
cannot be used. It then becomes necessary to use a model that is only
ll.A.5. Case studies intrinsic, having a variogram but no covariance.
1. Infirrite a priori variance (Witwatersrand gold prouitzce)
Under the hypothesis of second-order stationarity, both a covariance and,
thus, a finite a priori variance, Var { Z ( x ) =
} C(Of, are assumed. But there
are theoretical models of RF's and also physical phenomena that display an
unlimited capacity for dispersion andbfor which neither an a priori variance
nor a covariance can be defined.
Brownian motion is an example of such a phenomenon and is well
known to physicists. It can be described by the Wiener-Livy process,
which has an infinite variance and, thus, no covariitnce. However, the
increments of this process do have a finite variance:
Var { Z ( x + h ) - Z ( x ) } = 2 y ( h ) = A . (hi.
Log oreo
The variogram of the Wiener-LCvy process is, thus, diefined and it can be FIG. 11.7. Variance versus area of dispersion.
shown that it increases linearly with the distance Jhl.
The dispersion of gold grades in the Witwatersrand provinces of South
Africa provides a more practical mining example. South African authors 2. Quasi-stationarity (on a copper deposit)
(in particular D.G. Krige, 1951, 1976) have shown that the experimental
A quasi-stationary type of mineralization is illustrated by a massive copper
variance a Z ( v )of gold grades within a field of size V increases continu-
deposit explored by vertical bore-holes. The copper grade decreases
ously with the dimensions of the field.
regularly with the depth, which indicates the presence of a trend (non-
The essence of Krige's original results is shown in Fig. 11.7. The variance
stationarity) in the vertical direction.
of the logarithms o f the gold grades increases without interruption as a The experimental variogram of the copper grades calculated using all the
function of the logarithm of the area o f the field V over which the variance
vertical drill cores is shown on Fig. 11.8. The characteristics of this vario-
is calculated. A complete range of field sizes was considered, beginning
gram are:
with the cross-sectional area of analysed core samples (030075 ft2) and
followed in turn by the horizontal cross-sectional area of a mining panel (i) a nugget effect, the amplitude Co of the discontinuity at the origin
(63 000 ft2), the mean horizontal surface of a mine (about 3.4 x 10' ft2) equals about 0.4(O/0 C U ) ~ ;
and, finally, the total surface area of the gold bearing region of the Rand (ii) a transition phenomenon between the origin and a distance of about
(about 3.4 X 10' ft2). 100 f t with a sill value of 1 and a range of about 50 ft;
11. THE TH[EORY OF REGIONALIZED VARIABLES 45
44 MINING GEOSTATISTICS
(iii) beyond 100 f t there is a sharp increase in the variogram values, of the water table) as a function of time t. The measurements are carried
indicating the presence of a trend due to Cu grades decreasing with out daily. The piezometers are placed at intervals of about 500 m, see Fig.
depth. 11.9.
with
CO=0-4, C = O.6(% CU)', a = 50 f t .
This model is shown on Fig. 11.8 as the broken-line curve.
At piezometer no. 3 The water table is very close to the surface and the
water level reacts each time it rains (the amount of water retained by the
ground is proportional to the thickness of the ground through which the
water passes before reaching the water table). The profile from piezo no. 3
is very erratic and is directly related to the daily rainfall. The variogram
I I \
\
150 along this profile is linear with a nugget effect (Brownian motion for small
0 50 100
h tft) lags t ) .
FIG.11.8. Trend effect for large vertical distances. Atpiezometer no. 4 The water table is deeper and the absorption effect of
the ground begins to take effect. The corresponding variogram has a
3. Behaviour near the origin (pluuiometry at Korhogo) smaller nugget effect and a transition phenomenon is becoming evident
(appearance of a sill).
The example given here has the advantage of demonstrating the four
different types of behaviour of the variogram near the origin cited in
At piezometer no. 33 The water table is deeper still and the regularizing
section II.A.3 and Fig. 11.6.
effect of the ground absorption is almost total. There are, however, a
This variography is taken from a prominent unpublished study carried
certain number of discontinuities evident on the profile and these are due
out by J. P. Delhomme (1976) on Korhogo (Ivory Coast) piezomc,atric data. to abrupt refilling of the table. The corresponding variogram has a small
The data z(x, r ) at each position x give the piezo height (height of the top
46 MINING GEOSTATISTICS 11. 'THE THEORY OF REGIONALIZED VARIABLES 47
nugget effect representing the discontinuous refilling of the table followed 4. Coregionalizations a t Exotica
by a parabolic shape which is characteristic of the high continuity of the
R e v z ( t ) . More precisely, the RF Z ( t ) is rnean-square differentiable The experimental direct and cross-semi-variograms for Fe, soluble Cu and
except during the periods of refilling. total Cu, shown on Fig. 11.11, were calculated from the grades defined on
13-m vertical bench heights in the Exotica copper deposit (Chile) and
Atpiezometer no. 18 The water table is even deeper and the effects of the represent the horizontal coregionalization of Fe, Cu S, Cu T grades.
discontinuous refilling are no longer evident. The ]nugget effect has
completely disappeared and the variogram presents an almost perfect
parabolic behaviour near the origin, characteristic of the differentiability of
the profile t ( t ) at any time t.
I I3 - Piezo no. 18
l4 I
- July August September October November December
Confidence intervals for the normal distribution of the n available estimators. The 95% confidence interval of a particular
Amang all the two-parameter distribution functions, thc one most often
true value 2., estimated by z : will then be ' .
used to characterize an error is the normal distribution. The main
justification for using this distribution is that it is the one most often
"*
observed in practice, particularly in mining practice, cf. section II.B.3, Case
This formula will not ensure that the lower bound is always positive,
Studies 1 and 2.
since it beclomes negative when the relative standard deviation is greater
In mining applications, the error distribution functions are generally
symmetric with a slightly more pronounced mode and larger tails than a
normal distribution with the same expectation and variance, cf. Fig. 11.13.
than 5 0 % , u ~ / ~ * > 0 . but
5 , it is sufficient f o r practical applications. The
case where 1 7 E / 5 * > 0.5 corresponds to poor estimations and, in such cases,
i t would be better not: to use confidence limits at all.
I
.&
Thus, in relation to the standard normal distribution, there are more small
errors (in the region of m, = 0 ) and more large errors (in the tails of the
distribution). However, the classic confidence interval [nrE j~~ u E ]contains
approximately 95% of the observed errors. The 95% normal confidence
Formula (I1,21) corresponds to a non-stationary standard deviation of
estimation, since it depends on each estimator z?. It can be shown (cf. G.
'Matheron, 1974, and section III.B.6, Remark 3 and 5, that this formula
1
corresponds to a conditioning of the R F Z ( x ) to the estimator f * , the
interval is then a good criterion for judging the quality of a geostatistical
distribution of Z ( x ) being lognormal.
estimation, but it should always, as far as possible, be verified experimen-
tally through test or control zones, cf. section TI-B-3.
11.8.2. Linear estimators
Consider the problem of estimating the unknown value ZV (e.g., the mean
grade Z v ( x )of a block of size V centred on point x ) from a set of n data
values { Z ( x , ) ,i = 1 to n}.
The estimator Z* will be a function of thc available data:
Z* = f { Z ( x l ) ,. . , Z(xn)},
but not any function of this'data. It must be a function such that
(i) it satisfies the non-bias condition, E{Zv - 2 " ) = 0 ;
(ii) it is reasonably simple, so as to permit the calculation of the esti-
mation variance,
FIG.11.13. Normal standard for the error distribution. --, Gaussian standard;
- - - -, experimental distribution of errors.
For any function f, the calculation of the first-order moment E{Zv - Z * )
Remark Under the non-bias condition ( m E= 0 ) the standard gaussian and the various second-order moments of the expression of c& requires
95% confidence interval can be expressed as [z* *2crE] for the true value z that the n-variables distribution { Z ( x l ) ,. . . , Z(xn)} be known. However,
estimated by z*. It sometimes happens that the value of the estimation
since it is generally not possible to infer this distribution from a unique
standard deviation CTE is such that the lower bound of the 95% confidence
realization o f the R F Z ( x ) , we are limited to the class of linear estimators
interval [r**2uE] is negative, e.R.. with z* = 0.5% Cu and V E = I % , so
(at least in linear geostatistics). It will then always be possible to calculate
that z * - 2 ~ E= - 1 4 % . A p r a ~ t i t - ~%nlution
l to this problem is to define the mean and variance of the error from the variogram or the covariance.
the relative standard deviation a& *, i.e., the ratio of the standard devia- Research into non-linear geostatktics is being carried out and has shown
tion to the experimental mean
that restrictions imposed by linear estimators can be removed, cf. dis-
junctive kriging and transfer functions, G. Matheron, (1975a, 6, c) and
section V1II.B. Such non-linear geostatistics requires a much more
MINING GEOSTATISTICS 11. THE THEORY OF REGIONALIZED VARIABLES 53
The linear estimator z g is the arithmetic mean of the n available data The previous estimatioln variance can then be expressed as
values { z ( x , ) i, = 1 ton}:
, "
The value z g is interpreted as a particular realization of the RV Z ; and Generalizatic n to the conti;nttous case
the unknown error is, thus, a particular realization of the RV ZK--22. Suppose now that the K points xk are located within a volume V centred on
Under the second-order stationary hypothesis, the non-bias is anto- polnt x and the n points x, within another volume v centred on point x'.
matically verified, since As K and n tend towards infinity, the preceding arithmetic means zK and
z : tend towards the mean values in V and v of the point variable z ( y ) ,i.e.,
For the estimation variance, the formulae (XI.23)i and (11.24) for the necessarily be connex (single continuous regions); the domain V to be
discrete case give, for the continuous case of estimating the mean grade of estimated may be, for example, two distinct blocks V = V 1 + Vz and the
V by that of u, informa!tion domain v may consist of several samples v S,, some of =I,
which may be situated within the domain V.
ak = E { [ z v ( x ) - ~ , ( x ' ) ] ~
Remark 3 The variogram 2y(17) can itself be interpreted as the elemen-
tary estimation variance of a variable Z ( x ) by another variable Z ( x + h ) at
a distance h from Z ( x ) :
E { [ z (+~h ) - ~ ( x ) ] * } = 2 ~ ( h ) ,
If the domains V ( x )and v ( x ' ) are three-dimensional, each of the integral cf. definition formula (11.6).
signs of formula (11.25) represents a triple integral and the variance a; is a
sum of sextuple integrals. T o simplify notations, we shall denote by Remark 4 Formulae (11.26) and (11.27) express four essential and intui-
c ( V ,u ) the mean value of the covariance C ( h ) when one extremity of the tive facts which condition all estimations. The quality of the estimation of
distance vector h describes the: domain V ( x ) and the other extremity V by v is a function of the following.
independently describes the domain v ( x l ) . It is then written (i) The geometry of the domain V to be estimated: term -T(V, V )
appearing in formula (11.27). As the semi-variogram y ( h ) is an
increasing function of h , T(V, V ) increases with the size of V.
When the covariance C ( h ) exists, the semi-variogram y ( h ) also exists H e m e , if u and the interdistance (o, V) remain fixed, it is easier to
and these two structural functions are related by formula (11.4): estimate the mean grade of a large block V than the grade of a n
unknown point ( V = x ) . Finally, if the size V is fixed, the term
C ( h ) = C ( 0 ) - y(h). T(V, V), and, consequently, the estimation variance u;, depend
Formulae (11.23) to (11.26) can then be rewritten in terms of semi- also on the geometry of V.
variograms. Since the constant term C ( 0 ) disappears, it is enough to (ii) 'The distances between the domain V to be estimated and the
replace C or ? by - y or -7 and the estimation variance becomes support u of the estimator: term +y(V,v ) appearing in formula
(11.27). As the distance between V and v increases, so does the
value of Y(V,v ) and the value of the estimation variance gi.
(iii) The geometry of the estimating domain u : term -.i;(u, v ) . As the
where T ( V , 0 ) . for instance, represents the mean value of y ( h ) when one dimension of the domain v increases, the value of T ( v , v ) also
extremity of the vector h describes the domain V ( x ) and the other ex- increases and, consequently, cr; will decrease. But, for fixed
tremity independently describes the domain u ( x l ) . volumes V, v and fixed interdistance ( v , V ) ,the estimation variance
It can be shown that this formula (11.27) remains valid even when the will also depend on the configuration of the information v . The
covariance C ( h )does not exist, provided of course that the semi-variogram estimation of the same block V on Fig. 11.14 by the two distinct
is defined (the RF Z ( x ) is then only intrinsic). samples 0 , and v2 will be better than that provided by the two
+
closely spaced samples u; and u ; ; indeed, y ( v ; v i , v ; + v i ) will be
Remark 1 The estimation variance of V by v is sometimes referred to as less than y ( u l + u 2 , V , + ~ 2 )This
. intuitive notion of the importance
the variance of extending the grade of v to V or simply the extension
uariance of v to V and is then denoted by cr:(u, V).
Remark 2 The two formulae (11.26) and (11.27) are completely general
whatever the domains v and V. In particular, the domains need not FIG 11.14. Influence of the information configuration.
56 MINING GEOSTATlSTlCS 11. T H E T H E O R Y O F REGIONALIZED VARIABLES 57
of the sample configuration, which is formalized in geostatistics S o t h e non-bias is ensured if t h e sum of the n weights equals 1:
through the term T(o, v ) , is ignored by the more usual ?stirnation
methods such as weighting by inverse distances o r their squares.
(iv) The structural function, semi-variogram or covariance. T h e quality
of the estimat~onshould depend on the structural characteristics T h e estimation variance can then be calculated from formula (11.27):
(anisotropies, degree of regularity) of the regionalized phenomenon
under study. The semi-variogram y may depend on the direction a
of the vector h , i.e., the semi-variogram will b e a functiion of dis-
tance and direction, y(lh 1, a).In horizontally stratified m~lneraliza-
tions, for example, the variation in grade values will be much more
continuous in an horizontal direction than they will b e vertically, where ?(x,, V) denotes t h e mean value of y(i7) when one extremity of the
and, by taking the anisotropic semi-variogram into account, more distance vector /I is fixed at point x, and the other extremity independently
weight would be given to sample u l in estimating block V in the describes volume V.
same stratum than to sample vz in a different stratum, cf. Fig. 11.15.
This fundamental geological notion of the spatial continuity of a R e m a r k 5 Formula ((11.28) expresses the estimation variance C T ~as a
mineralization as a function of direction is quite often overlooked or linear function of the n weights A,. T h e estimation procedure known as
simply ignored in the application of many estimation methods such kriging (cf. Chapter V)determines the optimal set of weights {A,, i = 1 to
as that of weighting by inverse distances. n } , l.e., the w e ~ g h t sA , which minimize the variance m i subject to the
non-bias condition x , A , = 1. Hence, kriging appears as the best linear
unbiased estimator ( ~ nshort a "BLUE" estimator).
blast-hole samples are taken from it, an average of six being taken from The experimental variance of the 397 errors was 0.273 ('10 C U ) while
~
each block. Experimental observation has shown that the arithmetic mean the predicted theoretical estimation variance calculated from formula
of these six holes, ZV,
can be taken as the true grade of block V. (11.28) was 0,261 ( O h CU)'. Therefore, the estimation variance was predict-
ed with the very acceptable relative precision of 4%: (0.273-
0.761)/0.273 = 0.04.
A n o m a l distribution with zero mean and a variance equal to that of the
experimentally observed errors (erg = 0.273) is also shown on Fig, 11.17. It
tits reasonably well the experimental distribution of errors with the follow-
ing remarks.
(i) The normal distribution underestimates the proportion of both small
Block V t o be est~rnated
Avo~loble ~ n f o r m o t i o n errors (between *O.S0/~Cu) and large errors (greater than
over nlne blocks
1 5 % Cu).
FIG. 11-16. Short-term kriging at Chuquicamata (ii) The standard normal interval (+2aE) correctly estimates the
A test zone of homogeneous mineralization which comprised 397 blocks experimental 95% confidence interval: 96% of all the observed
errors fall within 322.~7~ = k1.04% Cu. I f the predicted theoretical
gave the histogram o f the observed errors (Zv-z$),shown on Fig. 11.17.
variance, c ~ k = 0 . 2 6 1had been used, the result would have made
The arithmetic mean of the 397 errors was +0.01% Cu and, given that
little difference: *2uE = &1.02% CU.
the mean copper grade over the test zone was 2% Cu, it is obvious that the
non-bias condition has been met.
2. Cartography at Mea after J. Deraisme (1976, unpublished)
I- The niclcel silicate deposit of Mea (New Caledonia) was sampled by vertical
bore-holes on a pseudo-regular square grid of 80 m, cf. Fig. 11.18. One of
the problems in the exploration of this deposit was to map the hanging wall
and foot wall of the mineralized bed. Let Z ( x )b e the height of the hanging
wall at the horizontal coordinate point x. A t each point xi where a drill
intersected ihe bed, the height z(x,) of the hanging wall was measured. In
order to map the hanging wall, the unknown value z(xo) of the height at
Hang~ng
wall : x :: xXO :: x ::
"0
' " -, ,.
: X :: X ::
Foot
wqll
d V - 5 . .
FIG. 11.17. Histogram of kriging errors. -, Wistograrn, nt = 0.01, a ; = 0.273; FIG.11.18. Cartography of the hanging wall at Mea. 0,
x, bore-holes; X , x~ node of
- - -, Gaussian distribution, m = 0, ,& = 0.273. the estimated grid. .-.
MINING GEOSTATISTICS 11. T H E T H E O R Y O F REGIONALIZED VARIABLES 61
each node of a regular square grid of 40-m side was estimated. he is sufficiently simple to be carried out systematically, provided that the
estimator used was the BLUE Z*(xo) provided by kriging. support of the elements lo be estimated is the same as that of the data used
Isobath contours of t h e hanging wall were plotted, using the estimated in the estimarion. This is often the case in cartography.
40-m grid pattern. Similarly, the foot-wall depth and the mineralized
thickness (hanging-wall height minus foot-wall depth) were plotted.
T h e kriging estimator z*(xo) used was a linear combination of the nine
bore-hole data closest to xo:
9
z*(xo)= A,z(x,).
)=I
Let z ( y ) be the R e v grade at point y. The mean grade of each unit o ( x , ) real~zationsof RV's Z,(x,)and Z v ( x ) ; the preceding mean square devia-
centred on point x, is tion, s 2 ( s ) ,also appears as a part~cularrealization of a RV s2(x)
defined at
point x, i.e., corresponding to the stope V centred on point x : -.E
a deviation [ r v ( x ) - z , ( x , ) ] . The dispersion of the N grades z , ( x , ) about units o within V. The following notation is used:
their mean value z v ( x ) can be characterized by the mean square deviation:
The dispersion can also be represented by the histogram of the N values Remark It must be stressed that the R V s 2 ( x )and its particular realiza-
z,(x,), i.e., by the frequency of occurence of each value z as shown on Fig. tion s 2 ( x ) depend on the particular position x of stope V ( x ) ,while, under
11.20. In practice, however, at the estimation stage the true grades z,(x,) of the hypothesis of stationarity, D ' ( v / V ) no longer depends on the position
the production units o are not known and neither is the true grade z v ( x ) of x but only on the geometries of v and V and the covariance C ( h ) ,as will
the stope V ; thus, the histogram of Fig. 11.20 is not available. be demonstrated.
For each one of K stopes, identical in shape and size, {V(xk),k = 1 to
K ) , there is an experimental mean square deviation s2(xk)calculated from
(11.29). As iY -+a, the arithmetic mean of these K experimental deviations
tends towards the stationary expectation of s 2 ( x ) , Le., the variance of
dispersion 13'(v/ V ) .
If we are only interested in one particular stope V ( x ) ,then D'(V/V )will
be an estimator of the experimental mean square deviation s 2 ( x ) , just as
the expectation of a RV is an estimator of a particular realization of this
RV.
Probabilistic interpretation
When the point R e v z ( y ) is interpreted as a particular realization of a R F
Z ( y ) , the unit grades tu(xi), as well as the stope grade z v ( x ) , appear as
64 MINING GEOSTATISTICS 11. THE THEORY OF REGIONALIZED VARIABLES 65
T h e variance of dispersion D2(u/ V) of these units u within V is given by It can be shown that the order of the integral sign and the expectation
the definition formula (11.31): symbol can be reversed to give
which is equivalent to
Border
effect
- Now, a limited number n of the N unit grades (n < N ,) mav. b e known and
the mean square d e v i a t ~ o nof these n values can be calculated according to
formula (11.29):
If n is not too small i n relation to N and if the n known values z,(x,) are
FIG. 11.21. Border effect. distributed approx~mately uniformly over the field V , then s,?, can be
66 MINING GEOSTATISTICS 11. THE THEORY OF REGIONALIZED VARIABLES 67
considered as an estimator of the dispersion characteristic s2(x) given by The dispersion variance D 2 ( v / V ) appears as the difference of the two
formula (11.29) and we say that s: is the experimental dispersion variance c
mean values over the two volumes v and V. -,.
Because of the condition v << V, the border effect can be neglected and Thus, thr dispersion variance D'(o/w) of the point variable in a very
the displacen~entof v(y) within V amounts to displacing a point y within large field t' is given, according to formula (II.34), by
V. When the field V(x) consists of an exact number of units v(y), the
condition v < V is no longer required, as the displacement of y within V ( x ) D ~ ( o / ~ lim
)=
V-m V
1V
~ { [ ~ ( y ) - r n ] ' dy.
)
can be limited to the discrete number of centres of units rr, cf. formula
(11.31). If it exists, the a priori variance of the point RF Z ( y )is precisely equal to
the term under the integral sign, i.e.,
68 MINING GEOSTATISTICS 11. THE THEORY O F REGIONALIZED VARIABLES
(cf. relation (II.3)), and, thus, the preceding limit exists and is written I
D ' ( o / ~ ) = V-m
lim D 2 ( 0 / V) = Var {Z(y)}= C(0).
FIG. 11.23. Dispersion of V-block grades in the bench. Block V, 20 m x 20 m x FIG. 11.24. Dispersion of 2 V-block grades in the bench. 7 V block, 40 rn x 20 rn X
13 m; m = 2.16, s 2= 0.346% CuZ. 13 m ; m = 7.1 7'6 Cu, s 2 = 0.769 (96 C U ) ~ ,
ore essentially made up of hematite. Note again the extreme homogeneity (s2= 2.3 (% ~ e ) ' )of the hematite
The extreme homogeneity of this ore is demonstrated by a clearly ore in relation to the variability of the grades of the deeper, poorer ores
unimodal Fe histogram of inverse-lognormal type. If Z ( x ) is the Fe grade ( s 2= 50-1Yo).
FIG. 11.26, Dispersion of Fe grades (global ore), m = 58.8% Fe, s2 = 84.90i02.
3. Grade distribution at Haartebeestfontein after 1). G. Krige (1976) and J. (i) When the dimensions of support u are small compared with those of
M. Rendu (1976) the field of dispersion V (e.g., u representing a core sample and V a
The field of dispersion B is a section (3000 m x 900 m) of the Haartebeest- mining unit or a stope), then (a) the histogram of the grades of
fontein gold mine in South Africa. Each unit u (7.5 m x 7 . 5 m) contains elements of weak concentration (Au, Ag, precious metals, U,Cu,
between zero and four samples (two on average) and the arithmetic mean Ni, impurities in weak concentration such as SiOz, A1203, Fe,. , .)
of the samples, Z,, in each unit u is said to represent th~ecluster of samples are asymmetric with a tail towards the higher grades (direct
inside the unit u. lognormal type); (b) the histogram of the grades of macro-elements
of strong concentration (Fe in deposit, Mn in deposit, phosphate
PzOs,.. .) are asymmetric with a tail towards the lower grades
(inverse lognormal type).
(ii) As the support v increases in size (e.g., v representing a mining unit)
a blending effect is produced, which tends to slowly tone down the
previous asymmetries. Lognormality is not preserved when the
support v becomes too large.
(iii) A cleary multimodal histogram reveals, in general,: the existence of
heterogeneous mineralizations (the presence of several types of ore
or the influence of poorly mineralized areas on the fringes of the
deposit). The interpretation of such a histogram must be made with
due regard to statistics (e.g., fluctuation due to an insufficient
number of samples), geometry (e.g., position of the field of dis-
persion in the deposit) and geology (e.g., mineralogy, faults, direc-
tions of enrichment of the ore).
(iv) The preceding remarks concern the grades of elements. For
Cumulative f r e q u e n c y ( C X )
geometric-type regionalized variables, such as mineralized or over-
burden thickness, or mining type R e v such as percentages of ore
FIG. 11.28. Dispersion of cluster grades at ~aartebeestfontein. recovcly 01 granulometric ratios, the histograms are of a more
symmetrical type, as shown on Fig. II.29(a) and (b).
Figure 11.28 shows the cumulative histogram of the dispersion of the
grades 2, in the field B in normal logarithmic coordinates. Note that for Figure 11.29(a) shows the histogram of the mineralized Ni-laterites
the low grades Z, there is a significant deviation compared with the line thickness of the Prony ore body (New Caledonia). The histogram was
corresponding to the lognormal model. constructed from 176 values measured on the 176 vertical bore-holes of a
The two-parameter lognormal model can be improved by introducing a regular square grid of 100-m side.
third parameter a, so that log (2,
+ a ) , rather than log Z, is assumed to be Figure II.29(b) shows the histogram of the percentage of ore (mesh
distributed normally. With a = 150 g cmlton, a log,normal model can >6 mm) recovered on mining units at Moanda (Gabon) manganese
be fitted remarkably well to the experimental distribution of the grades deposit. This histogram was constructed from the results of mining 1400
(2"
+-a). unit blocks corresponding to an area of about 42 ha.
II.C.4. Type of distribution functions i But the presence of several modes o n a h~stograrndoes not imply that each of these modes
necessarily carresponds to o n e type o f mmeral that can be distinguished at the rniningstage,
It would not be possible within the bounds of this book to present all the For example, a banded hematite quartzite ore surveyed by drdl cores (of a metre or SO in
available examples of density functions of grades (grade histograms) but the length) would show a multimodal histogram o f the type shown on Fig. 11.27 and still be
considered a!; an homogeneous ore at the scale of m ~ n i n g(on blocks of 10 or 20171in
following remarks, drawn from practical experience, can be made. dimension)
76 MINING GEOSTATISTICS 11. THE THEORY OF REGIONALIZED VARIABLES 77
Of course, non-symmetrical histograms of thicknesses can be found just were negative and would have appeared on a global histogram as a strong
as well as symmetrical histograms of grades. The previous remarks are not atom at the origin (zero thickness).
to be taken as laws. As an example, Fig. 11.30 shows a histogram of
exponential type corresponding to the vertical thicknesses of bauxite lenses
within a zone of the Delphi deposit (Greece). Only the positive drill-holes 1I.D. REGULARIZATION
that actually meet a mineralized lens (i.e., with a non-zero mineralized
thickness) were retained for this histogram. Indeed 60% of the drill-holes II.D.l. Regularization a n d estimation variance
Very rarely, in practice, will point data z(x) be available. Most often, the
available data z,(x) will be defined on a certain support o(x) centred on a
point x, v being, for example, a core sample or, more generally, the volume
of a sample. The grade z,(x) of this sample is the mean value of the point
grade z(y) in u(x), i.e.,
Since the point semi-variogram y ( h ) is stationary, the last two terms of formula (II.39), the regularized semi-variogram y,(h) is also of a transition,
the previous expression are equal and, thus, type with a sill value equal to the regularized a priori variance: .
% 2
neglected. T h e mean value over the length 1 of the core sample can then be
written in terms o f a single integral in the d~rectionof the length of the core
as
T<[
rS(
y l ( h ) = $E{[z,(x+ h ) - Z l (x)12}= T(l, lh)- ?(I, 1). (11.47) These integrals can b e evaluated quite simply to give
If u denotes t h e coordinate along the length of the bore-hole, t h e first term [r-1/3, Vral,
T(1, l h ) of the expression (11.47) for the regularized semi-va~iogram is cf. Fig. 11.34.
written For distances lhl 1, the behaviour at the origin of the regularized
semi-varlogram is parabolic. T h e cores of length 1 overlap each other and
the regularizing effect is strong enough to change the behaviour of the model
at the origin from linear to parabolic.
(Recall here that vector h is parallel to the direction of the bore-hole For distances \hi a 1, which, in practice, are the only distances observable,
length,) the linear behaviour of the point model is preserved and t h e regularized
According to whether or not r = IhI is less than I, i.e., whether o r not the model differs from the pomt model by a constant value
two segments 1 and lh overlap (cf. Fig. II.33), we have the following two
expressions.
82 MINING GEOSTATISTICS I,[. THE THEORY OF REGIONALIZED VARIABLES 83
(Note that for the linear point model the approximation of formula (11.42) The relative error involved in this approximation is 8% for r = I and 0.1%
becomes a strict equality.) for r = 51. ,
..
..%
b
V r = lhl G [0,a ] ,
r(h)= 2 a 2 a3'
[ 1 =sill value. 'dr l a = range.
This model reaches a sill value, y(m) = 1, for a distance equal to the range
a, cf. Fig. 111.5.
The expressions for the semi-variogram y l ( h ) regularized by cores are
long and awkward. They can be found along with those for the exponential
model in Ch. Huijbregts (19716).
In practice we can make use of the following charts, given in section
II.E.4:
(i) chart no. 21, which gives the values of the first experimentally
accessible points of yl(h) and also of the sill value yl(m) as functions
FIG. 11.34. The linear model and its regularization of a l l ;
(ii) charts no. 22(a) and 22(b), which give the values of yr(h) for
2. The logarithmic model different values of a l l as a function of Jhlll.
y(h)=logr with r =lh(. Furthermore, for the practical distances l h l a l , the regularized semi-
variogram can be derived from the point model by the approximate
Note that for h = 0, log IhI = --a # 0,SO that this logarithmic model cannot
formula (11.42):
be used for point samples, only models y,(h), regularized on non-point
supports, are defined. y,(h)= y(h)-Y(1, l ) , for l h l a 1,
For distances observed in practice, ( h l 1,~ the semi-variogram the mean value y(1, 1 ) for 1 s a (the length is not greater than the range),
regularized by cores is written as being given by the expression
(r + 112 (r - 1)' 1 l3
YI(~)=- log (r + I) +- 2 1 ~log ( r - 1) T(1, l)=---
2a 20a3'
Vlsa. (11.5 1)
212
(11.49) Given the conditions, j h l a 1 and a 31, which cover almost all practical
r
--l 2 log r -- log I, V r 2 1. cases, this approximation is excellent.
In practice, this exact formula is rather difficult to use and the approxima- 4. The exponential model
tion of formula (11.42) is preferred:
y ( h ) = 1-exp (-rla), V r = lhl.
r 3 This model reaches i t s sill value, j ( m ) = 1, asymptotically. In practice, a
yr(h)-y(h)-T(l,I)=Iog-+- Vr;zl, (11.50)
1 2' practical range a' = 3a is taken, for which
with
y(1, l ) = l o g i-3. cf. Fig. 111.5.
84 MINING GEOSTATISTICS 11. THE THEORY OF REGIONALIZED VARIABLES 85
T h e semi-variogram regularized by cores of length 1 takes the form model. In fact, for the two transition models in current use (the spherical
and the exponential), when Ihl> a and a 2 31, yl(h) can b e approximated
with excellent precision by its sill value y , ( a ) .
Given the conditions, lh 12 1 and a ' = 3 a 2 31, which cover almost all prac-
tical cases, this approximation is excellent.
t
Bench
Regularized range For a transition point model, i.e., for a point model C
with a sill and a (practical o r actual) range a , two point variables Z ( x ) and
+
Z(x h ) are uncorrelated when the distance h separating them is greater
than the range (Ihl> a.)
As a consequence, the two variables regularized by cores of length 1,
Z I ( x ) and Z r ( x + h ) are uncorrelated when the distance between each point
of l(x) and each point of l(x + h ) is greater than the range a , cf. Fig. 11.35;
for which it is necessary and sufficient that jhl> a + 1.
T h e regularized range is, thus, (a + I ) and this is the distance at which the FIG. 11.36. Grading over a constant thickness.
regularized model y l ( h ) reaches (actually o r practically) its sill y l ( 1 ~ ) Note
.
that the previous practical approximation, yl(h) = y ( h ) - y(1, I ) amounts to
If x, and h , represent the coordinates in t h e vertical direction, then the
assigning to the regularized model the same range as that of the point
mean value over the constant thickness 1 of a horizontal bench is given by
1 '
lo
Zd*.,xu) = 7 Z (xu, xu. x W )dxW, (11.54)
Example If Z(x,, x,, x,) is the point grade in a horizontal stratiform Application to point models in current use
formation of quasi-constant thickness I, then ZG(x,, A:,) is the meant grade This regularization by grading will be illustrated for the four point models
of a vertical bore-hole positioned at the point of horizontal coordinates in current use. These point models will be supposed to be isotropic in the
(xu,x , ) on the topographical surface. three-dimensional space. 1
The regularized model is deduced from the point model by subtracting a (ii) The regularized theoretical expression y,(h) (or its approximation
constant term T(1, I ) which depends only on the length 1 of the support of given by formula (11.63)) of this point model y(h) is then calculated
the regularization. and compared with the experimental curve y;(h). The various
paraimeters (ranges, sills, ratios of anisotropies, etc.) of the point
Remark 1 For the regularization by cores of length 1 of transition-type model are then adjusted in such a way as to bring y,(h) in line with
point model with a (practical or actual) range a , the additional condition r: ( h 1.
a >> 31 should also be satisfied before applying the approximation (11.63). (iii) Once the point model y(h) is determined, the theoretical expression
In practice, this additional condition is often met, since the length 1 of the y,*(h) for the regularization over a possible second support u' can be
data support should, generally, be much smaller than the range a of the deduced and checked with the corresponding experimental curve
sought after structure. ~*(h).
Rernark 2 Positive definite-type functiorts for regularized model. As it is an The calc:ulation of the regularized theoretical expressions y,(h)
approximation, the expression (11.63) for the regularized model is not and y,,(h) is simple only when the point model consists o f the
strictly a conditional positive definite function; hence, we are not sure that established models which are in current use, and for which formulae
all variances calculated from this approximate expression of yl(h) will be and graphs are available, e.g., if the point semi-variogram y(h) is a
strictly positive or zero, cf. the condition (11.10) in section II.A.3. spherical model.
Strictly speaking, if the point semi-variogram y(h) is of a certain model -
spherical for instance - the regularized semi-variogram yl(h) is not of the ll.D.5. Case studies
same model. However, for practical applications, it is an acceptable
approximation to consider that the model is the same for the regularized 1. Variograms of grades at Tazadit
-
y,(h) i.e., spherical - with, of course, parameters (regularized range and
The iron (banded hematite quartzite) deposit of Tazadit (Mauritania) was
sill) different from those of the point model y(h). 'This model for yl(h)
would then be a conditional positive definite function and ensures that all sampled by vertical bore-holes. The particular bore-hole considered for
variances calculated from it are non-negative. this example was 177 m (L) long and was cut into core samples of constant
Thus, if the point semi-variogram y(h) is a spherical model with range a length 1 = 3 m.
and sill C, the regularized semi-variogram yl(h) caln be assimilated (for Let Z ( x ) be the point R F representing the iron grade regionalization,
distances I h l a I) to a spherical rnodel with Zl(x) the lZF regularized by cores of length I = 3 m and Zsl(x) the RF
'
regularized by cores of length 41 = 12 m (obtained by grouping the avail-
(i) a regularized sill, CI= C - ? ( I , 1); able core samples in fours).
(ii) a regularized range al equal to (a + I ) if the reg~alarizationis done by Figure Ir.38 gives the two experimental semi-variogtams of the two
cores of length I, and equal to a if the regulariza~tionis a grading over ) circles 0 correspond to the experimental
variables Z I ( X )and Z ~ ! ( X(the
the constant thickness I. points of yl and the crosses x to those of ydl).
These semi-variograms reach a sill value for a range a of approximately
Rernark 3 Decon~~olution (pa.rsing from y, to y). In practice, since the 40 m. The sill C, is approximately 133 (% ~ e ) ' . If a spherical model with
available data are most often defined on a non-point support v , it is often range a = 40 m and sill C is assumed for the theoretical point model y(h),
required to work out a theoretical point model y(h) which is coherent with then the theoretical value of the sill C, is written, according to formula
the observed regularized semi-variogran~y: (h). Deducing a point model (11.5 I), as
y(h) from a regularized model y,(h) amounts to a "deregularization" or a
"deconvolution" of the model y,(h). In practice, this deconvolution is
made through the following identification procedures.
(i) A point model y(h) is defined from an inspection of the regularized with 1= 3 rn, a = 40 m and C = 133. Thus, rhe sill of the point spherical
curve y: (h) experimentally available. model is C = 138 (% ~ e ) ~ . ...-. . -
11. T H E T H E O R Y O F REGIONALIZED VARIABLES 93
92 MINING GEOSTATISTICS
Variance of dispersion within L The experimental variance of dispersion
In Fig. 11.38, the theoretical point spherical model y ( h ) with range of the grades of the 59 cores of length I along the total length L of the
a = 4 0 m and sill C = 138 is shown as a dashed line. The theoretical bore-hole was found tto be s 2 = 120 (% ~ e and ) the
~ mean grade rn =
regularized model y, deduced from the previous point model by the prac- 49.2% Fe.
tical approximation (II.63), The corresponding theoretical dispersion variance D ~ ( I / Lcan) be cal-
( h )y ( - 1) V h 3 1, a>31=12m, culated from formula (11.36) either (i) by using the previous point spherical
model y ( h ) and formula (II.98),
is shown as a first solid line. The regularized model 741 deduced by the
same approximation is shown as a second solid line.
or (ii) by assimilating the core length 1 to a quasi-point support and by
considering a point spherical model with the same range a = 40 m and a sill
C, = 133,
Both these theoretical values are close to the experimental value s" 120.
FIG. 11.38, Regularized semi-variogram3 of Fe grades at Tazadit. Experimental (i) the experimental semi-variogram yT(h) regularized by cores of
points: 0 , 7,; x, y4,, L = 177 m, rn = 49.2O/o Fe, s = 120 (% ~ e ) ' . length 1 and computed from the vertical information (bore-holes);
(ii) the experimental semi-variogram y$(h) graded over the length 21
The theoretical value of the sill C4, is found from formula (11.5 1) to be and computed from the horizontal information (channel samples).
As the experimental points of yF(h) are reasonably well aligned, a
logarithmic mode1 regu1arii:ed by cores of length 1 = 0.5 was fitted accord-
ing to the approximation (11.50):
If the mineralization is isotropic in the three-dimensional space, it should I1I.E. CALCULATION'OF THE MEAN VALUES T
be possible to fit the experimental horizontal semi-variogram y 6 ( h ) to a "..
logarithmic model graded over 21. This graded model is given by the same The main geostatistical operations (estimation and dispersion variances,
approximation as regularization, kriging) make continuous use of mean values 7 of the point
semi-variogram y ( h ) such as
Y ( U , U = I~, =C
nn'' ,)C ~r ~ ( x l - x j ) ,
n discrete points x, which form a regular partition of the domain u. There corresponding covariance will b e systematically overestimated as C(h)=
are, of course, other methods of discrete approximation (the methods of C(O) - Y (h1.
Newton, Gauss, etc., cf. R. W. Hammings, ( 1 9 7 1 ) ) which sometimes pro- Note that the zero effect can only b e produced when the two domains v
vide more precise numerical estimation of the multiple integral j. and u' are not disjoint, i.e., if there is a non-empty intersection o f the two
domains within which points may coincide. It is particularly dangerous
when o n e of the two domains contains the other, e.g., when a mean value of
the type y ( v , v ) is t o be estimated.
For such mean values y ( v , v ) , the zero effect can be avoided either by
slightly shifting each point of the second grid (xj, j = 1 t o n ' ) independently
with respect t o t h e corr~:sponding point of the first grid (x,, i = 1 to n ) , o r by
using a more sophisticated weighting method for the calculation of the
multiple integral T ( v , tl) such as the Cauchy-Gauss method proposed in
remark 5 .
greater than the largest range of the model y(h), then .i;(u, u') = sill value. with
Again, if the dimensions of the two domains u and u', separated by a +m
.,.
.,.,. ""
distance h, are very small with respect to the range of the transition model
y(h), then y(v, v')= y(h). Now, if only the dimensions of v' are small with
K ( u ) = k * I; = [ k(u +x')k(xl) dx'.
respect to this range, then u' can be approximated by its centre of gravity
( n l = 1) and u can be approximated by n discrete points in the usual Formula (11.68) is merely the Cauchy algorithm for reducing a multiple
integral of order 2n to a multiple integral of order n. This algorithm applies
manner. Still again, if the domain u is a parallelepiped very much flattened
in the vertical direction, and if the model y ( h ) is isotropic, then u can be equally well to the calculation of the mean values T ( v , uk) between two
approximated by its median horizontal plane - which is then approximated equal domains v and uh translated from each other by the vector h. The
formula then becomes
by discrete points over two dimensions (n s 6 x 6 points); on the other
hand, if there is a strong vertical variability represented by an anisotropic
model y(h), it is advisable to approximate the vertical dimension of the
parallelepiped u also.
Remark 4 Mean ualues 7 of linear combinations. A s the operation of The Cauchy algorithm for reducing the order of the integral is of interest
taking the mean value f(u, v') is linear in y, more precisely, if y ( h ) = only if the function K ( u ) is known a priori. Now, for rectangular domains
Ck Cky~:(h),then ?(u, 0') = 1,CkYk(v, u'), each of the terms yk(u, u ' ) can be with sides parallel to the system of axes, the corresponding function K ( u ) ,
treated separately. Some of these terms can be estimated numerically, also called the "geometric covariogram of u", can be readily determined.
while others can be determined directly either from charts o r auxiliary (a) I n one dimension, v is a segment of length I and its geometric
functions. covariogram is a function of one variable u :
Remark 5 Cauchy-Gauss method for calculating y(u, u) a n d .i;(u, uh). As
the following developments (rather tedious) are simply exercises in
numerical calculus and do not result in any new geostatistical concepts, the
Consider now the symmetric function of y(h):
reader may go directly to the next section, "Computer subroutines", which
will give, among others, the subroutine corresponding to this Cauchy-
Gauss method.
Then, for aligned segments of identical length 1 translated one from
I Let v be any domain in one, two or three dimensions. The mean value
the other by the distance h, cf. Fig. 11.41, the Cauchy algorithm
T ( u , u) can be written in the form of a weighted mean in which the
(11.69) leads to the following expression for the mean value T(1, lh):
weighting function, (llu)k(x), is proportional to the indicator function
k ( x ) of the domain v:
T(1, lh)= 2 6 I
(1 - u)r(h + b)du. (11.70)
LJ
k ( x ) = ( O if not and u -a k ( x ) d x = l .
By using the function K = k * c, auto-convolution of the indicator
function k, the preceding multipIe integral can be reduced to
This integration can be carried out by choosing a regular centred grid of For a spherical or exponential model y ( h ) and for practical values of the
points within the intervals (0, 1) and considering uniform weights ( l / n if n range, the relative precision is always better than 1%.
points are used). The Gauss numeriaal integration mel.hod provides much
better precision for the same number n of points. For example, four
Computer subroutines
abscissae points (xi, i = 1 to 4) were chosen in the interval (0, 1) with the
corresponding four Gauss weights (Ai, i = 1 to 4). The three integrals of The following subroutines calculate the mean value P ( v , v ' ) between any
formulae (11.70) to (11.72) are then estimated by the: following discrete two rectangular domains v and v', whatever their respective positions in
sums. relation to tlhe system of axes on which the model y ( h ) is defined.
The first subroutine G B A R is completely general and considers a dis-
(a) In one dimension, the domain being a segment of length I :
crete approximation of each of the domains v and v' by a centred regular
4
grid of n points.
?(I, h,)==Z A,r(h +lx,). (11.73)
;= 1 The second subroutine F is more particular but much less time-consum-
(b) In two dimensions, the domain being a rectangle S of sides (I,, 12) ing, and considers the Cauchy-Gauss discrete approximation to calculate
the mean value y(u, v) within a single domain v.
parallel to the coordinate (hl, h2) directions of vector h :
4 4
Y ( S ,Sh)= E C AJ,T(hl+ [IX,,
r=lI=l
k2+1?~,). (11.74) FUNCTION GBAR(- --)
v and v' can be any domain, provided that they are of rectangular form.
(c) In three dimerisians, the domain being a rectangular parallelepiped v The dimensions of the two domains do not have to be the same, e.g., u can
of sides (11, 1 7 , 13) parallel to the coordinate (h,, h 2 , h 3 ) directions of be a three-dimensional parallelepiped (a mining block) while v' is a seg-
vector h : ment (core samples along a bore-hole), cf. Fig. 11.44.
Each of the two domains is approximated by a centred regular grid of n
points, n taking the values given in (11.67). .
When the two domains u and v ' have a non-empty intersection, and
The four abscissae x,, xi, x k of the discrete approximation and the cor-
particularly hen one contains the other, the index I N T E R = 1 allows
responding weights A,. A,. A k have been calculated by A. Marechal (1976,
internal report) and are given in 'Table 11.1.
avoiding the zero effect: each point of the discrete approximation of one
domain is shifted randomly from its corresponding node on the regular
centred grid.
The three-dimensional point semi-variogram y ( h ) is read as a sub-
routine FUNCTION GAM(HX, HY, HZ). The three coordinates
(HX, HY, HZ) of the vector h are relative to the reference system of axes.
(In section III.B.5, an example of FUNCTION GAM corresponding to a
very general structural model y ( h ) is given.)
Relative to the reference system of axes, each domain v or u' is defined
by one apex and one, two or three side vectors according to whether the
domain is one-, two- or three-dimensional, Thus, in the example of Fig.
11.44,
(i) the rectangular parallelepiped v is defined by the coordinates of one
of its apex 01 and the coordinates of the three side vectors C A I ,
CB 1 and CC1.
(ii) the segment o' is defined by the coordinates of one of its extremities
02 and the coordinates of the vector CA2.
co ICC 12=1,r1
CC 1 C C J i = 1 , ? 2
CC 1 O C K 2 = 1 , C 3
IJK= I J K t l
C
C
.................................................
COMPLTES f h E bVERAGE kbLCE CF G b r C A
XZ(IJK,IC)=COi(IC)+( I ~ - O . ~ ) * C A ~ ( I C ) / ~ ~ +
C @€TWEEN T k C C O r A l h S k 1 T H AhY C F I E N T A T l C h k 1 T H
~ ~ J ~ - ~ . ~ ) * C B ~ ( I C J / P ~ + (J *KC ~C z-( C[ .( ~) / ~ 3
1 1 C COhT I h U E
C R E S P E C T TC C C 7 9 0 I Y A T E S SYSTEC..EITt-ER CCtA l h rlGhT
C BE A S E G P E h T t H E C T A N G L E OR P A P b L L E L E P I P E 2 . . C C Y b I h S
C C I G H T BE T F d C I F F E X E h T CIHEhSION..
C
C
C
.... FARAPETERS
CAlLZ) ,C@l(3),CCl(3) kECTCFF D E F I h I F C C C v P I h 1
C VECTORS HAVE CCUPCh O R I G I N LCCATEC
C O Y C N E O F , T H E A P E X E S OF T H E C C C A I h
c C O ~ ( ? ) C ~ O R C I N A T E S C F T H I S F E F E R E ~ C EA F E X
C I D I P l NbHBER O F C I C E N S I O h S C F DOMAI!! 1
C =1 SEGMENT VEClCF CA1
C =Z G E C T A N G L E VECTCFS CAl,CPl
C =3 P A R A L L E L E P I P E D V E C T C F S C b 1 , C E l t C C 1
C U N U S E C V E C T C R S S H O U L C @ E P U T T C C. 2 1 0 CONT I h b E .
C C A Z L ~ ) V C ~ ~ ( ~ ) , C C Z L ~ ) ~ C ~ ~ (C ~E )F I~h II TCI CI hM ZC F C
C CCCb l h 2 C
L
CCMPUTE GEAR
C
t SUBRCLTIhES CPLLEC
C CAM(HXtbYtHZ) SEPI-ibRIOGRPP FIhCTICh
C RANCOW(KJ RANOCM N L L P E R S C E h E G b T C R ( C t i )
106 MINING GEOSTATISTICS 11, THE THEORY OF REGIONALIZED VARIABLES 107
It can be shown that this double integral of the isotropic function y ( r ) can corners of the rectangle and the other extremity independently describes
be reduced to the following simple integral: the entire rectangle A B C D :
a ( L ; I ) = ~ ( A BD)=?
c,
2
1'
0
( ~ - - U ) ~ ( J ( Lu2))du.
~+ (11.79)
H ( L ; I ) = ?(A,A B C D ) .
It can be shown that H ( L ; I ) is also equal to the mean value of y ( h ) when
(11.83)
one extremiliy o f the vector h describes one side of length 1 and the other
Note that the auxiliary function a ( L ; 1) is not symmetric in ( L ; I ) : extremity independently describes the adjacent side of length L:
a ( [ ;L ) = q ( A B , C D ) # T ( A C , BD). H ( L ;1 ) = T ( A C ,A B ) . (11.84)
The two-variable auxiliary function x ( L ; I ) is defined as the mean value The auxiliary function H ( L ; I ) is symmetric in ( L ; I) and represents a
of y ( h ) when one extremity of the vector h describes a side of length 1 and double integral of the isotropic function y(r). It can be shown that H ( L ; 1 )
the other extremity independently describes the entire rectangle A B C D : can be deduced from the auxiliary functions x ( L ;I ) and F ( L ; I ) by simple
and double differentiation, respectively:
x(L;1 ) = j;(AC, A B C D ) = .).(BD, A B C D ) .
1 a 1 a2
The auxiliary function x ( L ; I ) thus represents a triple integral of the H ( L ;I ) = - --i2,y(~; 1 ) s - P ~ ~ L ' F ( L ; I). (11.85)
21 ar 41L a/ a~
isotropic function y(r), and appears as the mean value of the auxiliary
Remarks 1 he two-variable auxiliary functions can be reduced to the
function a ( u ; I)=?(AC, U V ) when the side U V describes the entire corresponding one-variable auxiliary function by contracting one of the
rectangle ABCD, as in Fig. 11.47.The triple integral can then be reduced to sides of the rectangle A B C D to zero, e.g., 1+0, as in Fig. 11.48. The
the'following simple integral: two-variable auxiliary functions then become
1 P L
a ( L ;O ) = j ( A , B ) = y ( L ) but a ( 0 ; 1 ) = ? ( A C , A C ) = F(I);\
x ( L ; 0 ) = ?(A,A B ) = x ( L ) but ~ ( 0I );= T(AC, A C ) = F(1);
Note that the auxiliary function x ( L ; 1 ) is not symmetric in ( L ; I ) : (11.86)
F ( L ; 0) = T ( A B , A B ) = F ( L ) and F ( 0 ; I ) = T ( A C , A C ) = F(1);
x(1;L ) = T ( A B ,A B C D ) f ?(AC, ABCD).
H ( L ; 0 ) = ?(A.A B ) = ,y(L) and H ( 0 ; I ) = ? ( A , AC)= ~ ( 1 ) . j
The two-variable auxiliary function F ( L ; I ) is defined as the mean value
of y ( h ) when both extremities of the vector h describe, independently of
each other, the entire rectangle A B C D :
I
F ( L ; I ) = r ( A B C D ,A B C D ) (11.81) C D
FIG.11.48. Rectangle ABCD compressed (1 = 0).
This auxiliary function is symmetrical in (L; I). It represents a quadruple
integral of the isotropic function y(r). The Cauchy algorithm of formula The auxiliary function F ( L ; I ) corresponds to the highest order (4)multiple
(11.70) can be applied to the auxiliary functions a ( L ; 1 ) and x ( L ; 1 ) to integral of y ( h ) , all other auxiliary functions can be' deduced from F ( L ; 1 )
reduce this quadruple integral to the following simple integrals: by differentiation or by setting a parameter to zero:
1 a2 a
( L ;1
?
L
)
.L
0
( L - u ) o ( u ; 1)du =-
n
L2
1
.t
0
UX(U; 1 ) du. (11.82) a ( ~I );= - -L'F(L;
2a ~ '
I ) = - L ~ ( L ;1);
a~
Note the similarity between these formulae and those of (11.78) cor- 1 a
xIL; I)=- - L ~ F ( L ; 1 ) and ,y(L)=,y(L; 0);
responding to one dimension. In fact, the two-dimensional function F ( L ; I ) 2 ~ aL
is nothing other than a one-dimensional function F ( L ) with the mean value 1
over 1 having been taken beforehand. H ( L ; I)=- -
41L J I ~ L
The two-variable auxiliary function H ( L ; I ) is defined as the mean value
of y ( h ) when one extremity of the vector h is fixed at any one of the F ( L ) = F ( L ; 0).
112 MINING GEOSTATISTICS 11. THE THEORY OF REGIONALIZED VARIABLES
113
In three dimensions, the three-variable auxiliary functions a, X , F and H Mathematical expressions and charts
can be defined for a rectangular parallelepiped ( 1 1 , 1 2 , 13). In practice, these Expressions or charts of these auxiliary functions are given for each of the
three-variable functions are too awkward to use and the three-dimensional four-point isotropic models y ( r ) which are in current use.
case is restricted to the two-variable functions defined on a parallelepiped
P = ABCDEFGH with a square base ( L x 12) as on Fig. 11.49. The follow-
ing auxiliary functions can then be defined: 1. The linear model
(i) the mean value of y(r) between the two square faces ( 1 2 ) of P , y ( h ) = r with r = lh(.
a ( L ;12)= Y ( A C E G ,BDFH); (II.88a) In one dimension.
(ii) the mean value of y(r) between one square face and P,
x ( L ; 12) = T(ACEG, P ) ; (II.88b) XI.)
= L/2 and F ( L ) = L / 3 .
(iii) the mean value of y ( r ) within P , In two dimensions,
F ( L ; 12)= T(P,P); (11.88~)
(iv) the mean value of y(r) between two adjacent faces of I? which is
equal to the mean value between a side 1 and P,
H ( L ; 12) = T ( A C E G , ABEF) = y(AE, P). (II.88d)
25 tan2 0
F ( L ; I ) = l o g l - l o g s i n 0-- log sin 8
12+6
(by considering that ir/3 = 1).
These three terms are also the terms resulting from a limited expansion
of the function log (a + b) with a > b, i.e.,
These exact formulae are rather tedious to use. In pr,actice, it is better to
use their limited expansions as functions of L I I or l/L depending on
whether L is less than or greater than I. For L < I, we have
Thus, the two-dimensional auxiliary function F ( a ; b), for the rectangle
with sides a and b, is approximately equal to the one-dimensional auxiliary
function F ( I ) for the segment 1 = a + b:
F(a;b)-F(l)=logl-3/2 withI=a+b. (11.96)
The segment of length 1 = a + b is said to be the "linear equiualenr" of the
rectangles with sides a and b.
More generally, when v is any domain in one, two o r three dimensions
For L > I, we have and not necessarily rectangular, the linear equivalent of v is the segment of
r length 1 such that
1 l 2 1 1"
a(L; l)=logL+- --- -+...;
1 2 L~ 60 L~ T(v, v) = F(1) = log 1- 3/2. (11.97)
5 . The r e model
Charts no. 10 and 20 give the variance of estimation of a parallelepiped A practical analogy would be a parallelepipedie panel P with surface a x b
P with a square base (L x 1') by its axial drill core I J of length L, cf. Fig. and height c, located within an isotropic horizontal regionalization
11.53: represented by a model graded over the constant height c.
(T
,; =27(IJ, P)-T(P, P)-?(IJ, I J ) = 2 H ( L ; ( 1 / 2 ) 2 ) - ~ (1~2 ;) - F ( I ) .
The cenlred configurations (medians) are obviously better than the I'
F ( a ; a ) = O..S2la and H ( a ; a ) = 0.7'65~. .!
closed configurations (sides). , :
i
This case study shows how geostatistics, by means of the structural .-2
Thus, uk, = 0 . 0 4 7 9 ~for a total sampled length 4a.
characteristic y(h), is able to balance the quality (the median information is
Rectangle evaluated by the two parallel sides AB and CD better located) and the quantity of information (the perimeter has the I . !
largest total sampled length but does not correspond to the smallest esti-
a;, = 27(AB +CD, P)- T(P, P)- T(AB +CD., AB + CD)
-. .
mation variance).
with T(AB -t CD, P) = T(AB, P) = x ( a ; b).
, ,
j
y(ABi-CD, AB+CD)=T(AB, AB+CD) ll.E.4. Charts L J
= B[Y(AB,AB)+ .I;(AB, CD)] All the charts given hereafter correspond to the two following spherical
,
and exponential isotropic models with a unit sill value. All the distances ; "I
= $ [ ~ ( b ) + a ( ab)]
; i..,. (
and, thus, we have considered by the charts are relative distances r/a (divided by the
parameter ,a of the model).
u;, =2,y(a; b)-F(a; 6 ) - i [ ~ ( b ) + a ( a ;b)]. ', !,
Sphericall model: i
,i
The numerical application gives uk, = 0*0778a,for a total sampled length i
2a. ,
~ ," .
'.*
, .,,f
1/
Rectangle evaluated by its two medians IJ and KL :Similar calculations
lead to the following result: ( 1 =sill, Vr 2 a (actual range).
: I
I :,
124 MINING GEOSTATISTICS 11. T H E THECORY' OF REGIONALIZED VARIABLES
125
Chart no.
Spherical Exponential
Object of the chart model model
Grading
yG(L)= a ( L ;1)-o(0; I ) 1 11
Auxiliary functions
x(L;0 2 12
H ( L ;0 3 13
F ( L ;1 ) 4 14
F ( L ; 12) 5 15
a ( L ; 1')-a(0; 1 2 ) 6 16
Estimation variances
d l .& &L, d l 7 17
~ E J
8 18
~6
9 19
4, 10 20
I/o
L/a
CHART NO. 2. Spherical model. Function x ( L ;I).
CHARTNO. 3. Spherical model. Function H ( L ; 1).
128 MINING GEOSTATISTICS 11. T H E THEORY OF REGIONALIZED VARIABLES
129
L /a
CHART NO. 4. Spherical model. Function F ( L ;I ) .
CHARTN O . 5. Spherical model. Function F ( L ; 1').
11. T H E THEORY OF REGIONALIZED VARIABLES 131
I
CHARTNO. 7. Spherical model. Various extension variances.
132 MINING GEOSTATISTICS 11. THE T H E O R Y OF REGIONALIZED VARIABLES 133
L/o
L /a
CHARTN O . 8. Spherical model. Extension variance cr&. CHARTN O . 9. Spherical model. Extension variance u:,.
1/ a
L/a L /a
I /a
CHART N O . 17. Exponential mode!. Various extension variances.
142 MINING GEOSTATISTICS 11. THE THEORY OF REGIONALIZED VARIABLES 143
Representation of the nested structures As a mattcr of fact, each of the main geostatistical operators is linear in
y(semi-varicrgram) o r in C (covariance when it exists), cf. formulae (11.27)
y
As far as the second-order moments of the RF Z ( x ) are concerned, these
for the estimation variance, (11.36) for the dispersion variance, (11.41) tor
nested structures can be conveniently represented as the sum of a number
of variograms (or covariances), each one characterizing the variability at a
regularization. Thus, using the previous linear representation (111.1) for the
point semi-variogram y(h), the regularized semi-variogram on a support v
j
particular scale:
is the sum of the regularization on v of each of the component structures:
I I
The previous representation (111.1) of nested structures is not unique. It is : Y O ( ~ ) = C Oand y(h)=Co+yl(h), Vlh[>ao,
,
currently used in mining geostatiseics because it is convenient in order to fit with y(0) = 0, by definition.
several experimental variograms with different levels (sills) of variability Thus, at the level of observation (e.g., h = 20 m >> ao), the residual micro-
and, above all, because the expression (111.1) is linear in y. structure yo(h) appears on the total variability as an apparent discontinuity
152 MINING GEOSTATISTICS 113. STRUCTURAL ANALYSIS 153
at the origin of amplitude Co, which is called the "nugget consfant", cf. Fig. The corresponding model yo(h) is a transition-type model (spherical, for
111.1. instance) with a range E which is very small compared with the smallest
distance lhj of observation:
Definirion
The term "nugget effect" can be generalized from its specific meaning in
gold-bearing deposits. It is used to characterize the residual influence of all
The corresponding covariance is written according to formula (11.4):
variabilities which have ranges (ao) much smaller than the available dis-
tances of observation (h >>ao). A nugget effect will appear on the vario-
gram (or covariance) as an apparent discontinuity at the origin.
In the course of this book, a pure nugget effect will often be represented
Scale effect It must be stressed that the definition of a nugget effect is by the notation with square brackets: yo(h)= [C,(O)- Co(h)], so as to
closely related to the scale of observation. The same structure y l ( h ) with distinguish it from a structure y,(h) with a greater range.
range a l = 10 m and sill C1 would be quite evident on a sampling grid of 3
to 5 m, but would appear as no more than a nugget effect on a grid of 30 to
50 m, cf. Fig. 111.1. Experimentally, on such a large grid the two Absence of spatial correlation
components Co and C , of the nugget constant would be indistinguishable The corresponding covariance model Co(h) for a pure nugget is, thus, of
unless data are available to put into evidence the range a l . transition type, its value at the origin Co(0) being equal to the sill Co of the
For this reason, it is often advisable when a sampling programme has
semi-variogram model, according to formula (I1.13), cf. Fig. 111.2,
been carried out on a systematic grid (of, for example, sides b = 100 m) to
take a few other samples on a smaller grid (for example, b/10 = 10 m). In
this way, data are available at two different scales of observation, which
makes it possible to study the nested structures and interpret any possible
nugget effect.
The pure nugget effect thus corresponds to a total absence of spatial
correlation between two variables Z ( x ) and Z ( x + h ) , at least for all
Pure nugget effecr
available distances (Il,rl>~), and is equivalent to the well-known
When the observed semi-variogram appears as a single nugget effect phenomenon of "white noise" in physics. In so far as geostatistics is only
(completely flat semi-variogram), it is said to represent a pure nugget concerned with second-(order moments, a pure nugget effect will be inter-
effect, cf. Fig. 111.2 and Case Study 1 in section III.A.5. preted as spatial independence.
In order to establish the standard statistical results for independent
variables, it is sufficient to apply the various geostatistical formulae to a
pure nugget-effect model.
In practice, a pure nugget effect at all scales with a function ~ ( h )
amounting to a single discontinuity at the origin is exceptional.
In fact, if the pure nugget effect at all scales is associated with sta-
tionarity, it represents an incredible degree of homogeneity of the
mineralization. It means, for instance, that no local differentiation is
possible between mining panels and, thus, no selection is possible. There
are no rich zones nor any poor zones (a zone being defined on any scale).
At every point in the deposit the best estimator is the mean grade of the
FIG. 111.2. Pure nugget effect. deposit.
154 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 155
lll.A.3. The n u g g e t effect a n d geostatistical o p e r a ~ t i o n s (iv) u fl V # 0, v and V have a non-zero intersection, the measure of
Let C ( h ) be the covariance, defined on a strict point support, characteriz- which has dimensions much greater than the range ao:
ing the variability of a sequence of nested structures: Meas u fl V Meas v fl V
CO(L,V ) = A vV and TO(u, V) = Co(0)- A vvt '
(III.6d)
In all four caws, we have u, V >> ao, i.e., the range a0 of the pure nugget-
The nugget effect can be distinguished as a transition covariance Co(h) effect mode:l is very small compared with the dimensions of the supports v
with a very small range a. = E : and V; alsoi, A = 1Co(h) dh (recalling that, in the three-dimensional space,
h is a vector and this integral in fact represents a triple integral over the
Co(h)= 0 when lhl> ao. whole space).
It was shown previously that the main geostatistical operators are linear
Proof: By way of example, we shall prove the case u c V.
functions of the covariance C. If Ff: is one of these operators, then, from
(III.3), we have
k
%(C) = %(Co)+ C 8(C1).
i=l
If the distance between every point x E v and the border of V is greater
than the range a o , then, strictly,
In this section we shall be concerned with the effect of these operators on
a pure nugget effect, i.e., with the term 8(Co). In terms of the semi-
variogram, with yo(h) = Co(0)- Co(h), we shall be concerned with the
operation g(yo).
and, thus,
A s yo(h) represents a pure nugget effect, t h e regularized semi-variogram where Coois the nugget constant regularized on the support v and shown
is expressed as o n the regularized semi-variogram y o u ( h ) .
I n the estimation of a block V, the measure u of the data support is often
negligible when compared with the measure V of the block. In such cases,
all terms in A/ V in formulae (111.9) are negligible and w e have
In practice, point data a r e seldom available and the nugget effect will be
revealed by the regularized nugget constant Co, = A I L $when , v is the since the measure V' of' the data support is V ' = nu.
smallest support of information (e.g., piece of core). For all other supports This estimation variance is directly proportional to the a priori variance
V, the regularized nugget constant Co, is glven by the inverse pro- Co, of the grades Z , ( x ) of the samples, and inversely proportional to the
portionality relation: number n of samples.
V
Co, = Couy , with v, V >>no. (111.8) (iii) Dispersiou varianclc of the mean grade of a volume v in a field
V. According to formula (11.361, and taking the expressions (111.6) of yo
(ii) Estimation variance of the mean grade ZV o n the support V (mining into account, this variance is written
block) by the mean grade 2,o n the support v According to formula
(11.27) and taking the expressions (111.6) of yo into account, this variance is
written
LA(,+.- 1 1 2Mea:;"
, if v and V are disjoint,
i f v an" have a
I (111.9)
t Strictly speablng, the standard formula u 2 / n of the statlstlcr of independent variables
corresponds to th; estlmntlon varldnce of the expectation E { Z ( x ) ) of the RF Z ( x )
However, under the hypottlest~ o f \ratlonarlty and ergodmty, the mean grdde Z v =
non-zero intersection, (1/ v ) j , ~ ( x ) d x o f d f~xedd o m a ~ nV, with V > >a,,, tends towards the expectation E { Z ( x ) )
158 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 159
lll.A.4. Case studies These experimental variances are greater than the above nugget constants,
the difference corresponding to the additional variability and nesting effect
1. Pure nugget effect a t Adelai'da of the macr$l.structure observed at large distances (11 > 50 m).
The scheelite (W04Ca) of the Adelaida (Spain) deposit is concentrated in O n the scale of observation, h E [S,50 m], which is also the scale on
veinlets and small nodules which are distributed more or less homo- which production is carried out, this pure nugget effect indicates that
selective mining is not possible in this deposit, unless the size of the
geneously in a network of quartz veins.
Drilling was carried out in a direction perpendiculalr to the plane of the selection unit is of the order of the vein (5 to 40 cm). In fact, the absence of
greatest density of veins. The variable measured was the cumulative spatial correlation makes impossible any local differentiation by estima-
thickness of the quartz veins intersected by a drill core, the support of the tion: at any location in the deposit, the best local estimator of the cumula-
variable being core samples of constant length 1 = 5 nn. Thus, the variable tive quartz thickness is the global mean, m = 5 cm m-' in the direction
measured, Z I ( X )is, the regularization of the point variable "density of studied (approximately vertical).
quartz" over the length of a core sample. The core samples can also be This example is one of the rare examples of deposi'ts in which the main
grouped in pairs to provide the variable ZZ1(x)regularized on core samples variable exhibits no spatial auto-correlation at the scale on which the study
of length 21 = 10 m. is made. There is, however, a macro-structure which appears at greater
Figure 111.3 shows the two regularized semi-variograms y,(h) and y2,(h), distances.
as well as the means and experimental dispersion variances of the data used
in the calculations. Both these semi-variograms arc; flat up to around
h = 50 m (pure nugget effect), after which (h > 50 m) a slow increase 2. Nested structure at Mazaugues, after A. Markchal and J. P. Roullier
indicates the presence of a macro-structure with a range that cannot be (19 70)
determined from the available data.
The bauxite deposit at Mazaugues aval (France) consists of subhorizontal
The respective nugget constants (sills of the flat semi-variograms)
stratiform lenses aligned principally in the North-South direction, cf. Fig.
adopted are 60 and 30 (cm m-I)', which verify the rule (111.8) of inverse
III.4(a). These lenses are characterized by a regular hanging wall and a
proportionality to the support of the regularization.
2 very irregul~ar foot wall which consists of a dolomite karst. The mean
The experimental dispersion variances, s:.=66.8 and szl=
dimensions of the lenses are between 200 m and 300 m in the N-S direc-
3513(cm m-'12, verify well enough the rule of inverse proportionality.
tion and 100 m and 200 m in the E-W direction.
The deposit was sampled by a network of vertical bore-holes with a
regular density; on average, each square 100 m x 100 rn contains one bore-
hole.
Figure II1[.4(b) shows the semi-variograms of the thickness of bauxite in
the two main directions N-S and E-W. These semi-variograms reveal
transition models, the ranges of which correspond to the observed mean
dimensions of the lenses. Thus, the structure of the variability of the
thickness observed on a hectometric scale is that of the lenticular beds.
If a spherical model of range a~~ = 250 m is fitted to the experimental
semi-variogram N-S in Fig. II1.4(b), the nugget constant would be Cor
0.25. On the hectometric scale of this observation, it is not possible to
specify the structures between 0 'and 100 m to any greater degree of
precision. I
transition with a range of approximately 2 0 m can b e noticed, which is Serra (1967a, 6, 1968). Sesra observed and decomposed a n entire series of
characteristic of the karstic structure of the foot wall, cf. Fig. 111.4(a). structures from the petrographic scale (100 p m , t h e regionalization of
Beyond this transition, t h e semi-variogram increases continuously in a way oolites and limonite) LIPt o t h e hypermining scale (5 km, t h e regionalization
remarkably similar to the linear behaviour at the origin of the previous of deposits related to t h e stream currents at t h e time of the deposition of
transition model with range ~ N =S 250 m, which characterizes the macro- the Lorraine basin).
structure of the lenticular beds.
~ ( h ) = ~?y,(h),
i=l I
h?~,(h),
V the weights A,. (111.12)
T o obtain a model with a sill C(O)= C # 1, it is enough t o multiply the Parabolic behaviour at the origin This very regular behaviour at the origin
given expressions for y ( h ) or C ( h ) by the constant value C. is seldom found in mining practice. In t h e absence of nugget effect, it
corresponds to a RF Z ( x ) ,which is mean-square differentiable, and, thus,
to a R e v z ( x ) , which is very continuous in its spatial variability. Thickness
ar behaviour a[ the origin This is the most frequent type of behaviour
untered in mining practice (variogram of grades and accumulations) variograms of contin~zoussedimentary beds sometimes display such a
parabolic behaviour at the origin, but are usually accompanied by a slight
it is most often accompanied by a nugget effect.
nugget effect due to errors of measure.
) Spherical model:
(c) Ganssiatl model:
y(r)= 1 -exp ( - r 2 / a 2 ) . (111.17)
y(r)= T h e sill is reached asyn~ptoticallyand a practical range can be considered
with a ' = a J3, for which y ( a l ) = 0.95 = 1 , cf. Fig. 111.5.
b) Exponential model: A parabolic drift effect (cf. section II.A.3, "Behaviour at infinity" and
section II.A.5, Fig. 11.8) does not stabil~zearound a sill f o r large distances
and, thus, cannot be confused with the parabolic behaviour at the origin of
Note that the spherical model effectively reaches its sill for a finite distance a Gaussian model.
r = a =range, while the exponential model reaches its sill only asymp- However, for small distances (r < 2 a / 3 ) , it makes little difference
totically, cf. Fig. 111.5. However, because of experimental fluctuations of whether a very regular local variation is interpreted as a drift effect or as a
the variogram, no distinction will be m a d e in practice between an effective stationary Gaussian structure.
a n d an asymptotic sill. For the exponential model, the practical range a'
can be used with 2 . Models withou~a sill
a'=3a f o r ~ h i c h y ( a ' ) = l - e - ~ = 0 . 9 5 = 1 .
These models correspond to RF's Z ( x ) with a n unlimited capacity for
T h e difference between the spherical and exponential models is the spatial dispersion, neither their a priori variances nor their covariance can
distance (abscissa) at which their tangents at the origin intersect the sill, cf. be defined The RF's Z ( r ) are only intrinsic, cf, section II.A.5, Case Study
Fig, 111.5: 1.
r = 2 a / 3 , two-third of the range for the spherical model; (a) Models in r s :
r = a = a 1 / 3 ,one third of the practical range for the exponential model.
y(r)=re with 8 ~ ] 0 , 2 [ , (111.18)
Thus, the spherical model reaches its sill faster than t h e exponential model.
the limits 0 and 2 being precluded.
These models have a particular theoretical and pedagogical importance
(they show a whole range of behaviours at the origin when t h e parameter 0
varies and they are easy to integrate).
In practice, only the linear model is currently used, cf. Fig. 111.6:
~ ( r=)wr,
with w the slope at the origin.
Remark 1 For small distances ( h -t O), the linear model can be fitted to
any model that has a lintear behaviour at the origin (e.g., spherical and
FIG.111.5, Models with a sill.
exponential models).
166 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 167
Remark 2 As 0 increases, the behaviour of y ( r ) = re at the origin Krige, 1951; G. Matheron, 1955, 1962; Ph. Formery and G. Matheron,
becomes more regular and corresponds to a RF Z ( . L )of more and more 1963; A . Carlier, 1964). Indeed. up until around 1964-66 this model was
regular spatial variability. practically the only theoretical model used; this is explained by certain
analytical properties which make this model a very convenient one (cf.
section II.E.2, "Linear equ~valents"), and also by the fact that the first
geostatistic~lapplications were carried out on deposits (gold on the Rand,
uranium) for which the variograms of the characteristic variables had no
sill. Since that time, however, geostatistics has been applied to a great
number of other ore bodies for which the variograms show more o r less
clear sills. J. Serra, using a very detailed study of the different scales sf
variability of the mineralization of the ferriferous basin of Lorraine
(France), showed that a logarithmic model is the limit model of a nested
succession of transition models, the ranges of which increase geometrically,
cf. J. Serra (1976b. 1968).
As the nested model has more parameters, it is more flexible and lends
itself more readily to a geological interpretation. Moreover, the particular
FIG. 111.6. Models in re
analytical properties of the logarithmic model are less attractive now that
Experimentally, models in r e for B ~ ] 1 , 2 [are often indistinguishable computers are readily available and it is a relatively simple task to program
from a parabolic drift effect, cf. Fig. 111.6. The choice of interpretation, as a +'
the evaluation of any integral of any nested structure y(h).
drift (non-stationarity) or a stationary model in r e with B close to 2, This explains why the various sill models have increased in use since
depends on whether or not it is desired to make a drift function m ( x ) = around 1966 at the expense of the logarithmic model (except perhaps in
E { Z ( x ) }obvious. South Africa). In our opinion, the use of a logarithmic model or of various
nested spherical models is more or less a question of habit and does not
Remark 3 For 6' 2 2, the function (-re) is no longer ia conditional positive alter the result of geostatistical calculations. What is important is that the
definite function and, in particular, the increase of re at infinity is no longer chosen theoretical model y(h) should be a good fit to the experimental
slower than that of r2, cf. section II.A.3, formula (11.14). This strict limita- semi-variogram within its limits of reliability (jh)< b).
tion 0<1. prohibits, in particular, any blind fitting of a variogram by I n other words, the results of the geostatistical calculations prove to be
least-square polynomials, even though some polynomials of degree higher robust in relation to the choice of the model -provided that the parameters
than two do satisfy the conditional positive definite property, cf. the of this model are correctly estimated (see the following example and, in
expression (111.15) of the spherical model. section III.C.6, the exercise on robustness).
(b) Logarithmic model:
Example Consider the regularization of the logarithmic model y (r) =
y(r) = log r. (111.20)
log r by core samples of length 1 = 1. According to formula (11.50) the
Note that log r -+ -as as h -, 0, so that the logarithmic model cannot be used regularized model can be written as
to describe regionalizations on a strict point support. On the other hand,
this model, once regularized on a non-zero support v , cf. section II.D,
yi(r) = log ( r / l ) + 3/2 = log r + 3/2, V r 3 1 = 1. ,A
formulae (11.49) and (II.59), can be used as the model of a regularized O n Fig. 111.7, a nested model has been fitted to this regularized
semi-variogram y,(h). In particular, the regularization of a point
logarithmic model is a function which is null at the origin: y,(O) = 0. spherical models y, and y2:
. two
logarithmic model. The nested model consists of a nugget effect and .#.
1
Logarithmic model or nested transition models? The logarithmic model r t ( r ) = G f yl(r)+yz(r), V ~ E [ [1311
,
has been studied exhaustively by the initial workers i n geostatistics (D. G. with CO= C1-5; C1= 1.5, a , = 3 ; C2= 2, a 2 = 13.
168 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 169
Within the interval r E [ I , 1311 the relative error involved in fitting the This model has a sill and parabolic behaviour at the origin, cf. Fig. III.S(a):
nested model y r ( r ) to the regularized model y! is less than 3 % , In- y ( r ) = r 2 / 6 when r -+ 0 .
yfl/yl<0.03. In practice, this can be considered as a very satisfbctory fit.
Amplitude of the hole ej$ect Let a be the relative amplitude of the hole
effect, i.e., the minimum value of the covariance divided by the sill value
C(0):
cr = lmin C ( k ) [ / C ( O ) .
For the model given in (III.21), this relative amplitude is a =0.212,
obtained for r,, = 4,4934 = 3 ~ / 2 .
Of course, the, sill value y:' (a)= C0.t C1-k C2 = 5, which is attained for
distances r 3 a z = 13, is a pure artefact of the nested model. Similarly, the
finite character of the a priori variances of the type ~ ~ ( v / o ocalculated
),
using the nested model y:, is a pure artefact. However, when the use of the
FIG. 111.8. Hole-effect model. (a) With sill; (b) without sill,
regularized model yT is strictly limited to the interval r E [ I , 1311, there is
no differencebetween calculations using y! and those made using y:. It can be shown that this value a = 0.212 is the maximum amplitude of
Now, if the logarithmic growth of an experimental semi-va,riogram an isotropic hole effect in the three-dimensional space. If an experimental
continues well beyond the interval [l, 1311, it would be better to use the hole effect \vith an amplitude greater than 0.212 is observed in practice,
regularized logarithmic model y! rather than a nested structure y f consist- then elther the hole effect is not significant and is nothing more than a
ing of too many transition models. fluctuation of the experimental semi-variogram, or the hole effect is direc-
tional, i.e., i t is not present in every direction of the three-dimensional
space.
3. Hole-effect m o d e l s
If we desire to specify a particularly strong directional hole effect, then a
A semi-variogram y ( h ) is said to display a hole effect when its growth is not positive definite hole-effect model in only one dimension must be used,
monotonic. The hole effect can appear on models with or without sills, cf. e.g.9
Fig. III.8(a) and (b).
The following hole-effect model, positive definite in the three-dimen- y ( h ) = 1 -cos h (111.22)
sional space, is given: for one-dimensional h.
In this case, the relative amplitude a is equal to 1 and superior to 0.212.
C ( r ) = sin r/r and y ( r ) = 1-sin r / r . (111.21)
The cosine model C ( h )= cos h is not positive definite in three dimensions
r being expressed in radians and, thus, its use must be restricted to one particular direction.
170 MINING GEOSTATISTICS III. STRUCTURAL ANALYSIS 17 1
Note that the model (111.22) is periodic without any dampening of the structures, the consequence of which is a dampening of the observed hole
oscillations. In practice, however, such a hole-effect rnodel is always asso- effect. For estimation purposes, an experimental hole effect that is open t o
ciated, in a nested manner, with other models with or without sills and the a doubtful interpretation, or is not very marked, can simply be ignored, cf.
considerable oscillations of the cosine will, thus, be dampened, cf. Case section IV.B, Case Study 6.
Study 6 in section 1V.B.
Interpretation A pseudo-periodic component of the regionalization may 111.8.3 Theoretical models of coregionalization
appear on the experimental semi-variograms as a hole effect; for example, Warning Even though the study of coregionalizations is in every way
the stationary sequence of two well-differentiated types of mineralization analogous to that of regionalizations, there is an additional difficulty due to
in an ore body. If this sequence is not isotropic (and, in general, it is not), the number of indices involved. Thus, the reader who is not familiar with
:hen the hole effect will be observed only in certain directions, for example, the index notation of matrices is advised to go directly to the next section,
the vertical direction if the phenomenon is a pseudo-periodic sequence o f III.B.4. A thorough understanding of the study of regionalizations will
horizontal stratification, cf. Fig. HII.g(a). make the study of coregionalizations more understandable on a second
If the grades of Fig. 111.9 are point grades, then the vertical semi- reading of this chapter.
variogram will exhibit a hole effect with an amplitude possibly greater than
0.217 and the abscissae 6 1 and 6 2 of the oscillations can be related to Consider a set of K second-order stationary RF's { Z ~ ( X k) ,= 1 to K}. The
the mean vertical dimensions of the waste and rich ore strata. The object of this section is to establish cross-covariance models Ckkc(h), cf,
horizontal semi-variograms will differ according to whether they are section 1I.A 4.
located within a rich or waste stratum, cf. Fig. II1.9(b).
The posi~ivedefinite condition
z(c1 Just as any function g ( h ) cannot be considered as the covariance or the
Rich variogram of a second-order stationary RF, any set of functions { g k k ( h ) ;k,
k'= 1 to K ] cannot be considered as the matrix of covariances or vario-
grams of a set of coregionalized RF's. The matrix of covariances {Ckk'(h))
Woste must be positive definite so as to ensure that the variances of all finite linear
combinatioris of the basic RF's Z k ( x are
) positive.
In practice, to construct a matrix model of coregionalizations, n e shall
,v( h )
first build a model of the spatial intercorrelation of the RF's Z ~ ( XThe ).
matrix model of this coregionalization will then be, by definition, positive
definite.
The set of K intercorrelated RF's thus generated has a positive definite I n practice, there ir; no need to look for the eigenvalues of the matrix
cross-covariance matrix { C M ( ~ ) ) : [ b k k , ]AS
. this matrix is symmetric, bkkz = bktk, it is positive semidefinite
when its determinant as well as its diagonal minors are strictly positive.
Thus, for a matrix of dimension K,we have the following K conditions:
I b12 . . . 6 1 ~ 1
In the definition expression (111.23) of the R F Z k ( x ) ,some basic RF's,
Y r ( x )and Y,(x), with i # j, can have the same direct covariance, K , ( h ) =
K,(h), while remaining orthogonal, K,,(h)=0. An expression similar to
(111.23) can be written by grouping the RF's which have the same direct In mining practice, colregionalizations more than three variables (K > 3 )
covariances: are seldom considered, mainly for difficulties of statistical inference of the
covariances matrix.
For K = 2, verification of the positive definiteness of the matrix [b;..]
amounts to checking that, V i , the two following inequalities hold true:
with ~f and ~ f having
' the same direct covariance K , ( h ) : Y ! and Y : ' being
orthogonal except when i =j and 1 = 1' simultaneously. The cross-covari- b ; , > 0 ; jbi2(= Ib;, 1 ~ t l ( b i ~ b :(Schwarz's
~) inequality).
ance of Z k ( x )with Z k ( x )is then written as For K = 3, in addition to the two preceding inequalities, one must verify
that the determinant lb;k'l is positive o r zero, Vi.
(ii) for a fixed index i, the coefficient matrix [ b i k , ]is positive semidefinite. cross-covariances:
The cross-covariance Ckk,(h) is then rewritten as yo(h). This model thus corresponds to variables Zk(x), Z&), the vari-
abilities of which are proportionally related to a unique principal cause
Ckk'(hk +A1)= ak,ak,,K,(h'), characterized by the model yo(h). This occurs in mining practice in the
1
Isotropy When the function ?(lhl, a , cp) depends only on the modulus Ihl
Intrinsic coregionalizarion
of the vector h, the phenomenon is said to be "isotropic", The variability of
This particular example of the linear model corresponds to the case in the three-dimensional RF Z(x,, x,, x,), characterized by the semi-vario-
which all the basic RF's Y,(x)of formulae (111.23) and (111.24) have the gram y(lhl), is identical in all directions of the space and, thus,
same direct covariance Ko(h) (or the same direct semi-variogram yo(h)):
(i) in spherical coordinates,
~(lhl,a,cp)=y(lhI), Vaand(~;
The model of intrinsic coregionalization is then
(ii) in reclangular coordinates,
, h w f= r(J(h2 + h i + hk)).
~ ( h , h,,
the matrix of coefficients [bkk,] being positive definite. Anisotropy T h e phenomenon is said,to be "anisotropic" when its vari-
In practical terms, the intrinsic coregionalization model means that all ability is not the same in every direction. The structural function
the direct or cross-covariances (or variograms) can be derived by affinity y(lhl), a , cp) characterizing this spatial variability in mean square thus
(multiplication by the constant bkkt) from a single basic model &(h) or depends on lthe direction parameters a and Q.
176 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS
177
Interpretation The notion of mean variability in the above definition is Anisotropies will be represented by the method of reducing them to the
important. In a massive stockwerk-type ore body, rich, small, local veinlets isotropic case either by a linear transformation of the rectangular coor-
can be found in specific directions, but, on average, these veinlets will occur dinates (h., h., h,) of the vector h in the case of geometric anisotropy, or
over the entire mass in every possible direction of space and the by representing separately each of the directional variabilities concerned in
phenomenon will be isotropic on average. the case of zonal anisorropy.
In practice, a real anisotropy of the structural function y ( h ) corresponds
to the existence of preferential directions at the time of the genesis of the 1. Geometric anisotrop)i
studied phenomenon. These preferential directions are, in general, known
a priori and the variogram quantifies then respective variabihties. For A semi-variogram y(h., h.. h,) or a covariance C(h., hv, h,) has a
example, geometric anisotropy when the anisotropy can be reduced to isotropy by a
mere linear transformation of the coordinates:
(i) the vertical direction in a deposit formed by deltaic deposition;
(ii) the horizontal directions of the deposition currents in an alluvial
deposit; y(h,, h,, h,)= y l ( J ( h f + h f + h f ) )
anisotropic isotropic
(iii) the radial directions around a volcanic pipe in an intrusive deposit.
with
Represen fation
A11 the theoretical models of regionalization presented in section 1II.B.2
are isotropic, i.e., y ( h ) depends only on the modulus r = Ih 1 of the vector h- F I G , 111.10. Geometric anisotropy.
178 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 179
represented by two spherical models with the same sill 1 and different (ii) The graph can be approximated by an ellipse, i.e., by a shape which
ranges a,, and a,,. Thus, since y,,(h) f y,,(h), the phenomenon represent- is a linear transform of a circle. By applying this linear. trans-
ed is anisotropic: the variability depends on the directions a, and a2. formation to the coordinates of vector h, the isotropic case is pro-
In order to make the two curves y,, and y,, in Fig. 111.10 coincide, it is duced (circular graph). The phenomenon is a geometric anisotropy.
sufficient to multiply h~ by the ratio of affinity a,,/a,, or, conversely, to (iii) T h e graph cannot be fitted to a second-degree curve and the second
multiply the distances h2 by the ratio a,,/a,,. Indeed, consider the two type of anisotropy must be considered, i.e,, zonal anisotropy in
spherical models certain directions, ad,for example, on Fig. 111.11.
and
I \
[I 1 Isotropy [II) Geomeir~c onisotropy (1111 Zonal onlaolrooy
Subject to the change of coordinates h i = h,a,,/a,,, the variability in the If, instead of transition structures of range aQ0as in Fig. 111.11, the direc-
two directions crl and a2 is characterized by a single spherical model with a tional semi-variograms are of the linear type, y,(h,)=w,,h;, then the
directional graphs of the inverses of the slopes at the origin w,, will be
range a,,.
considered. An hypothesis o f isotropy, geometric anisotropy or zonal
More generally, if n directional semi-variograms {y,,(h,), i = 1 to n } can
anisotropy will then be adopted according to whether or not this direc-
be represented by n transition models of the same type (spherical, for
tional graph can be fitted to a circle or an ellipse.
example) with the same sill, and the n ranges of which form an elliptical-
shaped directional graph (in two dimensions), or ellipsoidal (in three
dimensions), geometric anisotropy is present. Correction of geometric anisotropy (in two dimensions)
A n example is given on Fig. 111.1 1, which shows four semi-variograms
Consider a geometric anisotropy in two dimensions (e.g., the horizontal
for four horizontal directions al,az,a,,ad.Spherical models with identical
space). The directional graph o f the ranges a, is elliptical, cf. Fig. 111.12.
sills and ranges of a,,, a,,, a,,, a,, have been fitted to these semi-vario-
grams. The directional graph of the ranges, i.e., the variation of the ranges Let ( x u , x u ) be the initial rectangular coordinates of a point, Q the angle
made by the major axis of the ellipse with the coordinate axis Ox,, and
a,, as a function of the d i r e d o n a,,is also shown on Fig. 111.11. There are
A > 1 the ratio of anisotropy of the ellipse. The three following steps will
three possible cases.
-
(i) The graph can be approximated to a circle of radius a , i.e., a,, a,
for all horizontal directions cri, and the phenomenon can thus be
transform this ellipse into a circla and, thus, reduce the anisotropy to
isotropy. ,
V
considered as isotropic and characterized by a spherical model of (i) The first step is to rotate the coordinate axes by the angle Q so that
range a. they become parallel to the main axes of the ellipse. The new
111. STRUCTURAL ANALYSIS 181
180 MINING GEOSTATISTICS
and8
coordinates ( y l , y 2 ) resulting from the rotation can be written in a = cos2 cp I- A sin2 cp,
matrix notation as b = sin2 cp + A cos",
[:I= [R,][~U] with [R,]=
XU
cos cp sin cp
[
-sin cp cos cp
c=(l-A)sincp.coscp.
If h is any vector in the two-dimensional space with initial co-
where [R,] is the matrix of rotation through the angle cp. ordinates A,,, h,), then, to obtain the value of the anisotropic
t xt semi-variogram y ( h )= y(h,, A,), we first calculate the transformed
coordinates ( h:, h l ) from
h,, h,). Each of the components yi(h) of this nested model can be aniso- The model yl(lhl) is fitted to the mean horizontal semi-variogram and
tropic in h, i.e., x ( h ) is a function of the three coordinates (h,, h,, h,) the model y;!(hW)to the difference -y(h,)- y(J(h: + h;)).
rather than of just the modulus Ihl. Moreover, the anisotropy o f yi(h) may
be completely different to that of yi(h). Thus, the structure y l ( h ) may have Remark. T h e model of zonal anisotropy, built from the nest of structures
a geometric anisotropy, while rz(h) is a function of tlhe vertical distance h, having their own anisotropy, is flexible enough to be used as a model for
only: almost any type of experimental anisotropy.
~ 2 ( h ) =~ z ( h ~ ) , Vh,, h,. However, the flexibility of this model must not be misused. When
a specific d~~rectionis differentiated (e.g., the vertical direction h, of
A third structure may be isotropic in three dirnensioms: y3(h)= y3(lhl). the previous example), this direction must correspond to a preferential
The model of zonal anisotropy can thus be defined as a nested structure direction of the mineralization. The known geology should always be
in which each component structure may have its own anisotropy. used as a guide when fitting a mathematical model y(h) to a natural vari-
A n obvious anisotropy of the structural function y(h) will most often ability.
correspond to a genetic anisotropy known beforehand, so that any pref-
erential direction is well known and, thus, can be: differentiated when
modelling the three-dimensional structural function y(h). 111.8.5. General structural model
Now that the different models of anisotropy have been studied, the follow-
ing general expression of a structural model y(h) can b e proposed, y(h)
being made up of the nested sum of N isotropic structures {y,(lhil), i = 1 to
I
N):
directions (h,, h,) with a geometric anisotropy. The previous linear trans-
formation is then written
CAY(4,3) =
0 1 0
and CAZ(4,3)=/: g 61
h2,v= a21hu+a22hv + O . h,,
h,,,=O.h,+O, h,+O. h w = O .
C hST NUYtlFR O F E L E Y E N T A F Y S T R U C T L F E S
The general structural model of formula (111.33) is particularly con- C C(NST) SILLS
C A(NST) R A Y G E S (0. E X P C ~ E ~ I I A L
venient since it is made up of the sum of N isotropic structures x(/h,l) L =O. kLCCE1 E F F E C T
which, in the course of a particular geostatistical operation, will be treated C >Ow SPHERICCL
separately one by one, cf. section III.A.l, "The linearity of geostatistical C CbX(hST,?) 1 LINEAR T R A h S F C F P A T I C h
operators". C CAY(blSTr?) I F A T R I C E S CF
All the case studies and practical exercises given in this book make use of C CAZ(NST,2) I 2 COORDIKATES
this general structural model.
Computer subroutine
T h e subroutine FUNCTION GAM(HX, HY, HZ) gives the values of a
three-dimensional model y(h) of the general type (111.33). The component
structures yi are transition models, either spherical or exponential. A
nugget effect is represented by a transition model with a zero range.
CAx;4,3=
r1 1
0
a,2
0
11
0
0
1 -+for the first coordinate of the first com-
ponent yo
+ for the first coordinate of the third com-
ponent y2
186 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS
Remark 1 Direct and inverse proportional effect. The proportional effect The proportional effect is said to be inverse when the sills of the
is said to be direct if the experimental semi-variogram increases with the experimental semi-vanograms decrease as the experimental mean
corresponding experimental mean, i.e., when the function f[m*(xo)] in increases.
formula (111.35) increases with m*(xo). Such an inverse proportional effect can be expected when the studied
Such a direct proportional effect can be expected when the studied R e v R e v z(x) has an inverse lognormal-type histogram, the mode M being
z(x) has a lognormal-type histogram, the mode M being less than the greater than the expectation m. For example, the grades of Fe-hematite or
expectation m. Examples are the regionalizations of Cu, U, Au grades tricalcite phosphate or, more generally, the grades of mineralizations of
or, more generally, of grades of mineralizations of weak concentration, strong concentration, cf. Fig. III.l6(b).
cf. Fig, III.l6(a).
Remark 2 Proportional effect atld geostatistical operations. As the main
geostatistical operations are linear in y (or in C-covariance), cf. formula
(III.2), a proportional effect present on the structural function y or C will
appear identically in each of these operations.
More precisely, if 8 is anyone of the three main geostatistical operations (i.e.
estimation and dispersio~nvariances and regularization), and if the propor-
tional effect model of (111.35) is considered, then:
and:
8[y(h, ~ o )=fCm
I *(xdI . 8Cr0(h)l. (111.36)
It is enough to work on the stationary model yo(h) and multiply the result
8[yo(k)] by the corresponding proportional effect factor f[m*(xo)].
Consequently, the various relative variances of estimation and dispersion {Z(x)/m*(.ro)}conditioned by the estimator mY(x0)of the mean grade
also appear as stationary and can be cdculated direictly from the stationary
model yo(h), i.e. according to formula (111.36):
Remark 4 Proportional effect and nested structures. When nested struc- (i) When the R F Z ( x )has a Gaussian spatial law and when it has been
tures are present, y = Ciy,, the proportional effect may be different for each conditioned by the best possible linear estimator m*(xo) (i.e., by the
of the component structures, i.e., the function f,[m'k(xo)Jmay be different BLUE kriging) of the mean grade Zv(xo) o n V(xo),then the condi-
for each of the models y,. tioning does not alter the stationary character of the variogram and,
Thus, on Fig. III.l6(c), th: direct proportional effect affects only the thus,
nugget constant Co:
For the linear model and for h > L/2, the relative fluctuation variance a bias becomes obvious. Indeed, if N data are available on the quasi-
) ] ' 2 (i.e., a prohibitive value of 200%).
~ { [ y ( h ) - y a , ( h ) ] ~ } / [ ~ (ishabout stationary field V, the unbiased estimator of this expectation is .
Note that even if the estimation variance vanishes ('very large number N'
1
of pairs), the fluctuation variance allows the choice of any possible m* = -
N
1 z(x,) with E{rn*}=m.
behaviour of the theoretical model y ( h ) for distance lhl> L/2 greater than ,=I
the distance of reliability which is half the dimension of the field L con- A possible estimator of the centred covariance is then written as
sidered.
Note again that this considerable value of the fluctuation variance pro-
hibits the construction of goodness-of-fit tests for the fitting of a theoretical
model y ( h ) to an experimental semi-variogram -y*(h): such tests would and, thus,
almost never invalidate the fit.
Practical rule
Although there is no reason for the RF Z ( x )represen,ting the R e v z ( x ) to
follow a Gaussian spatial law, the preceding expressions for the estimation
and fluctuation variances provide the following orders of magnitude: The estimator C * ( h ) of formula (111.43) then shows a bias which is nothing
more than the mean value of the theoretical covariance C ( h )when the two
/ N ' > 30,SO pairs and / h /< L I Z ]
- (111.42)
extremities of the vector h independently describe the set of the N avail-
able data:
An experimental semi-variogram should only be considered for small
distances (lhl< L/2)in relation to the dimension ( L )of the field on which it
has been computed, and provided that the numbers N' of available data
This bias can be assimilated to the mean value C(V, V) when the N data
pairs are not too small.
are uniformly distributed on the field V. When the covariance model C ( h )
This practical rule holds true for estimators of:
is of transition type, i.e., C ( h ) = 0 when the distance Ihl is greater than a
(i) cross-variograms or non-centred cross-covariances; range a, then the bias C(V, V ) vanishes if the dimensions of the field V are
(ii) nun-centred covariance, the stationary expectation m = E { Z ( x ) } much greater than the assumed range a of the covariance model.
being supposed known, i.e., for the estimator This bias is linked to the simultaneous estimation, from the same set of
data, of the expectation m and the non-centred covariance E { Z ( x +
h ) . Z ( x ) ) . This difficulty does not appear in the estimation of the variogram,
since the expression
E I [ Z (+~h ) - ~ ( x ) l ~ )
Bias of the covariance estimator does not depend on the unknown expectation m.
The preceding estimator [ C * ( h ) + m 2 ]of the non-centred covariance
[ C ( h ) + m 2 ]is without bias, since
II1.C.THE PRACTICE OF STRUCTURAL ANALYSIS
~ { ~ * ( h ) + r n='E} { Z ( X+ h ) . z(x)}
-. ~ ( h ) m+2 . d
knowledge gained from the studied phenomena, and also involves a certain A typical list of questions to be asked when reviewing data is given
amount of "craft" in the choice and use of the various geostatislical tools. hereafter.
While skill can only be gained by experience, structural analysis has certain
general attributes. In particular, it should: Study of the reconnaissance plan
(i) b e adapted to the proposed aim of the study. Structural analysis is What a r e the criteria upon which the plan is based? H a s it been biased in
most often a prelude to an intended estimation and, as such, should favour of a priori geological hypotheses o r financial o r technical con-
b e limited t o the purpose of the estimation. Thus, it is n o use to straints? Is it suited tor the purpose at hand? A reconnaissance plan
specify structures on a petrographic scale if it is intended to estimate designed for a tectonic survey may prove t o b e unsuitable for grade
blocks on a hectometric scale. estimation. The sampling material should also b e suitable, e.g., type of
(ii) take critical account of all qualitative information. T h e experimental drilling equipment, diameter of cores, channel samples o r core samples?
knowledge of a geologist may lead him to give all imporlance to a What type of s a m p l ~ n ggrid is used? e.g., regular grid, pseudo-regular
data file; conversely, and more rarely, an erroneous geological grid with constant sample density, preferential p i d . Have the different
hypothesis may bias the use of such a file. facies encountered and the variables of interest been sampled in a syste-
matic and homogeneous manner?
Before beginning the geostatistical treatment of a data set, i t is advisable T o what extent can we ask for additional sampling?
t o become familiar with both the physical nature of the phenomenon under
study (geology, mining technology) and the available data. 'This pre- Causes of sys!emadc error
liminary phase is essential in any structural analysis, and also for a correct
formulation of the study; very often, it even defines the framework within T h e various geostatistical tools are constructed from the available data. A
which geostatistical methods will be applied. systematic error in these data will b e reflected a t every stage of the
T h e next step is the "variography" phase, i.e., the constructiori, analysis geostatistical analysis. Cieostatistics does not create data, it is a method for
and fitting of the variogram models. T h e purpose of the basic geo~statistical treating information; in other words, geostatistics complies with the in situ
tool - t h e variogram -is to condense the main structural featurest of the reality only as long as the data a r e a n accurate reflection of this reality.
regionalized phenomenon into an operational form. Variography makes as Thus, it is very important, especially at t h e data review stage, to detect all
much use of the previously introduced geostatistical concepts as rt does of possible causes of systematic error.
the gained (qualitative o r not) knowledge of t h e physical nature of the
phenomenon (metallogeny, structural geology, tectonics, mineralogy, Adherence to the samphng plan
mining and treatment of ore, etc.).
Does the actual layout of data correspond to that of the plan? Geometric
errors may have been made, or the drilling platform may have been
displaced for accessibility reasons. Such displacements, even though small,
III.C.l. Review of the data may cause considerable biases in estimation, e.g., if preferentially rich or
Before making any geostatistical calculations, it is advisable t o carry out a poor zones are the ones accessible to the drilling platform,
critical assessment of the available data. Errors at all levels can be cor- H a s the deviation of bore-holes been controlled? Quite often, some
rected and, very often, such an assessment can p r o v ~ d esimple explanations types of samples will tend to follow preferentially rich pockets of soft ore
for results which may seem, at first sight, rather odd. (e.g., bauxite).
What are the causes of breaks in drill cores? M o r e o r less random
accidental causes a r e not so important as definite physical causes, e.g.,
t Most readers have seen feasibility s t u d ~ e sin which excellent s t u d ~ e sof
structural geology or when a boulder of hard waste is encountered.
geomorphology are presented, only to be completely ignored In the estlmat~onphase when
methods such as polygons of influence are used. If estlmatlon 1s d~vorcedfrom g,eology, one Have t h e different facies been sampled in an homogeneous manner?
might r~ghtlywonder about the purpose of the geolog~calstudy Obv~ously,geolog~cal Samplers d o not always take the same care in sampling waste as they d o
information must play a part In the estmation and the varlogram IS used In geostatlstics to ore, but mining may recover part of this waste.
Interpret the structural information to be used In the estlrnatlorl procedures
198 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 199
In space The same ore body may have been sampled in different places A geostatistician should always participate in the definition of the R e v
by very different methods, dragline samples on the surface, drill cores in that he intends to study, i.e., he should at least be consulted at the time that
depth, samples taken along a drive, etc. Topography, rock type and a systematic sampling campaign is being planned. In fact, all potential users
equipment used may vary from place to place and can lead to different of the information to be extracted from the sampling campaign should be
recovery rates and to biases (loss of fines in blasting holes for example). consulted to ensure that the information is relevant to the intended studies.
Such users would include the mining engineer and metallurgist as well as
Has the sampling of the various variables (thicknesses, grades, densities) the geologist. For example, a detailed lithological study in which only the
been homogeneous? In polymetallic deposits, certain trace metals, rich facies are analysed is not suitable for estimation in the frame of
although economically important, are not always as well sampled as other non-selective mining.
metals (e.g., Ag and Au in Pb deposits). Some metal grades are measured In choosing and defining the R e v for the purpose of estimating a
from other variabIes by regression, deconvolution (e.1;. U), construction of deposit, the following rules should be observed.
a metal "balance sheet", etc.
1. Additivity of the variables The R e v must be such that all linear
Compiling tlie data combinations of its values retain the same meaning. If f ( x ) is a grade value,
C,A f(x,) and especially the arithmetic mean ( l / n ) x ,f ( x , ) must have the
How and by whom have the data - from drill log sheet to magnetic tape - same meanmg as a grade.
been compiled? Aberrant grades can sometimes be traced to breaks in the Thus, a volumetric grade (in kg ~ n -for~ ,example) defined on a constant
sequence of data compilation, e.g., changes in groups collecting samples or support is an additive regionalized variable. The same volumetric grade
transferring data from one type of computing device to another. defined on variable supports is no longer an additive R e v and, in fact, the
What type o f error tests are made when the data file is being completed? mean grade of two different supports is not the arithmetic mean of their
Is it possible to retrieve rejected data at a later stage? grades. In such a case, the additive R,eV to be considered should be the
On what basis have the geological maps and cross-sections been drawn? accumulation (or quantity of metal in kg), i.e., the product of the-volu-
how much of the interpretation is ascribable to the geologist or to the metric grad15(kg m-3) and the support volume (m3). However, when study-
draughtsman? ing the spatial variability of this accumulation, it would not be possible to
200 MINING GEOSTATlSTlCS 111. STRUCTURAL ANALYSIS 201
distinguish between the variability due to the grade and that d u e to the biasing the estimation if some facies assumed to be waste were not
variable supports. analysed (in fact, all facies within the bed will be mined and their
If the preceding volumetric grade z (x) is associated with a granulometric grades, however small, should be considered).
fraction r(x) and both analyses are defined on variable supports v(x) (iii) T h e information will b e focused o n the recognition of the horizontal
centred o n point x, then the additive R e v accumulation o r quantity of variabilities since there will be a horizontal selection.
metal is defined as
3. Homogeneity of the tariable within its extension field T h e significance
a ( x ) = r ( x ) . z ( x ) . u(x). and support of a R e v must b e constant, o r vary very little in its field of
Other examples of non-additive variables are permeability, direction of definition. In particular, this homogeneity allows the observed variability
a vector a ( x ) with origin a t point x, indices of all sorts (hardness, colour), between two data z ( x ) and z ( x + h ) to be ascribed to the single regional-
ratios of grades, certain metallurgical recoveries, etc. T h e particular case of ization of z (x).
the non-additive variable Y ( x ) = l o g Z ( x ) , logarithm of a grade, for In a porphyry-copper deposit, f o r instance, grades from the oxidized
zone should not be mixed with those from the sulphide zone. Similarly,
example, is discussed later.
data from two drilling campaigns carried out with different drilling equip-
ment should be considered separately, at least in the first step.
2. Adequateness of the variables chosen for the purpose of the study T h e
It is useful to recall here that the concept of homogeneity and its
chosen ReV's should be such as to encompass the main characteristics of
probabilistic counterpart, the concept of stationarity, are relative to the
t h e problem at hand. They should also lead to operational solutions
scale of observation: the same o r e considered extremely homogeneous
without any superfluous precision or calculation.
during mining and industrial treatment map appear on a petrographic scale
Consider the example of a thin subhorizontal sedimentary bed with
as consisting of a number of very different minerals.
obvious foot and hanging walls.' Rich and poor zones will b e selected by
mining o n a horizontal plane, but no vertical selection is t o b e considered
(the entire hed thickness will b e recovered). Since there is n o vertical Working on logarithms of grades
selection, it ir, of n o use to study the vertical regionalization of the grades, In some cases, the spatial variability of an additive R e v Z ( x ) is s o great
at least for t h e strict purpose of ore reserve estimations. T h e chosen that it is convenient to study it through a transformed variable; for exam-
additive ReV's for this estimation are: ple, the logarithm Y ( x ) =log Z ( x ) . A gold grade in a nugget deposit can
(i) the thickness r(x)of the bed measured along the bore-hole located at vary from 1 to 1000 o r m o r e f r o m one sample to another; it is then more
the point x of the horizontal two-dimensional space; convenient to study the regionalization of the grade logarithm. This new
(ii) the vertical accumulation a (x) = t(x) x z (x), defined as the product R e v is less variable, but is n o longer additive.
of the thickness and the corresponding mean grade z ( x ) sampled o n The structural function y y ( h ) = E { [ Y ( x + h ) - ~ ( x ) ] ~of) the trans-
t h e mineralized section (of length r(x)) of the borehole. formed variable Y(x) is, calculated and a linear estimator Y*(xo)=
1, Y(x,) = 1,log Z(x,) is built. It would then b e erroneous t o think that a
Actually, the thickness variable is used in t h e estimation of the volumes mere exponentiation of Y*(xo) provides a good estimator o f
of recoverable ore, while the accumulation variable is used in the estima-
tion of the quantity of recoverable metal. T h e recovered mean grade within
a zone is given by t h e ratio of the quantity of metal to the recovered First, the non-bias condition may not b e met, as E{Y*(xo)} = E{Y(x)} does
tonnage in this zone. not entail that
In this example, it can b e seen that the following are true.
then, the required estimation variance E { [ Z ( X ~ ) - ~ ~ * " ~is' ] not
~ ) readily
(i) I t is of n o use to consider the three-dimensional regionalization of
the grades. derived from the variance
(ii) It is of n o use to cut the cores into small lengths to represent the EI Y(.xo)- y*(xo)12)
various facies met along a vertical. Moreover, to d o so would risk calculated from the structural function of the transformed variable Y(x).
202 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 203
As a matter of fact, the estimator eY*'"~)is no longer linear with respect The grades of the core samples of constant length 1 are obtained by
to the data Z(x,); therefore, its study requires not only the second-order weighted averages of the grades of the actual core samples. Thus, on the
moment covariance or variogram of Z(x),but also the knowledge of its drill core of Fig. 111.17, the grade 2, of the core sample 1 is made up of the
entire distribution function. When Z ( x ) is lognormal, Y ( x )= log Z ( x ) three actual core sample grades r , , t z , z 3 , weighted by the respec-
being normally distributed, it will be shown that a correct estimator Z*(xo) tive lengths l , , 12, l3 of their intersections with the reconstituted sample
can be derived from the linear combination Y*(xo)= C, A, Y (x,). This I, i.e.,
technique, called "lognormal kriging" will be presented in Chapter VIII,
"Introduction to Non-linear Geostatistics". Until then we shall limit
ourselves to the study of additive regionalized variables.
The previous rules of additivity, adequateness and homogeneity must
always be applied with flexibility. Obviously, recovery rates from drill cores
Smoothing error In such a reconstitution, it is implicitly assumed that the
cannot be perfectly constant. Similarly, a drill core cut into segments of
grade is constant within each actual core sample. This causes a smoothing
very variable length (non-constant support), as unsuitable as it may be,
effect as the. reconstituted grades give a smoothed idea of the actual
represents a considerable investment and it is advisable to make the best of
variability of the phenomenon. This smoothing becomes even more
what is available, e.g., to create an homogeneous set from this hetero-
important as:
geneous information that will allow a geostatistical treatment of the data.
(i) the grades become more variable within each actual core sample
length^, i.e., as the micro-variabilities (nugget effect) become
Reconstiturion of core samples of constant length
stronger;
On a zone considered as homogeneous for the purpose of the study, one of (ii) the reconstituted constant length 1 becomes smaller with respect to
the first requirements of the geostatistical approach is that the available the mean length of the actual core samples;
data be defined on a constarit support. It is recalled from section II.D, (iii) the distance h considered for the semi-variogram becomes smaller.
"Regularization", that the same spatial variability shows up differently In fact, two contiguous supports 1 may be constructed from parts of
according to the size and geometry of the support on which it has been the same actual core samples. This smoothing at small distance h
sampled. may mask a micro-structure, bias the evaluation of a nugget effect
If the cutting and analysis of the core samples on this zone has been and sometimes lead to the adoption of excessively regular behaviour
carried out heterogeneously, then it is advisable to restore the homo- at the origin of the variogram.
geneity, which, in practice, means defining the data on core sections of
The theoretical calculation of this smoothing effect would require the
constant length. Two cases can be distinguished,
knowledge of the distribution of the two variables, length of the actual core
sample and its grade, and quickly becomes tedious. However, the following
1. The actual core samples are not of constant length but have all been practical rules can be stated.
analysed, cf. Fig. 111.1 7 (i) The dimension 1 of the reconstituted constant support must not be
smaller than the mean dimension of the actual supports; if it is, then
the rcamstitution amounts in fact to an artificial creation of
information. Very often, technical considerations largely determine
the choice of the support, i.e., the length I is equal to a submultiple
of the bench height in an opert-pit operation.
i
&
.%
.
1 I 1
I 8 - R e c o n s t ~ t u l i o n of (ii) Use caution when interpreting, the behaviour at the origin of
homogeneous cutting
' -I an experimental variogram derived from reconstituted data; an
1
excessively regular behaviour may simply be an artefact due to
FIG. 111.17. Reconstitution of core samples of constant length. smoothing.
".
" ""
204 MINING GEOSTATISTICS II'I. STRUCTURAL ANALYSIS 205
2. The actual sdmple lengths are not constant and are not all analysed The statistical techniques of histogram and correlation plots are
~ u f f i c i e n t l ~ s i mto
~ l be
e carried out systematically before any variography
This case is more often associated with a preferential analysis (t:.g., of rich is attempted.
facies only) rather than with a random one (e.g., one sample on two The histogram, or the: experimental curve of the frequency of occurence
analysed). of the different values of the random variable, is a remarkably simple tool
Random analysis is more favourable, and the non-analysed samples can t y means of which the homogeneity of the spatial distribution can be
be considered as missing data. The missing datum is sometimes in rich facies verified, extreme or suspect data can be identified and the type of pro-
and sometimes in poor facies and, thus, does not cause any bias in the portional effect to be expected can be predicted (cf, section III.B.6, Fig.
structural analysis. 111.15). In addition, the histogram is an estimator of the stationary dis-
Preferential analysis is much more critical, especially when the non- tribution functlon of the R F under study, as well as a fundamental tool for
analysed facies (generally the poor ones) may actually be recovered by the geostatlstlcal techniques of ore reserves calculation (cf. Chapter VI)
mining. To reconstruct an homogeneous and unbiased data set, it is possi- and conditional simulations of deposits (cf. Chapter VII).
ble to act in one or more of the following ways. Correlndon dlagrams are a means by which the homogeneity of a spatial
coregionallzation can be verified. Minerali~ationsor a set of homogeneous
(i) T o simulate the missing analyses. If the missing analyses correspond
data can somet~mesbe distinguished by apparent groupings in correlation
to intersections with pure waste, a zero grade is given to the inter-
diagrams, cf. the following case study. In addition, the correlation diagrams
section; if the waste is, in fact, poor ore which will be recovered (all
are essential tools for sinlulations to the extent that they reveal inequalities
o r in part) during mining, it is advisable to simulate the missing
of the type
analyses from the knowledge of the characteristics of this poor ore,
deduced, for example, from systematically analysed drill cores. a x grade in metal A + P x grade in metal B s fixed limit.
(ii) T o limit the calculation of the variograms to within the rich analysed Such inequalities often characterize the mineral of interest and it is advis-
facies, thus characterizing the variability of these rich facies only. able to take them into account when simulating the deposit.
This supposes that, during the mining stage, it will be possible to
select perfectly those facies with a negligible amount of contamina-
tion from other facies. Case study
(iii) To calculate an inter-tacies variogram by assigning a global grade The mineralization of Ihe precambrian iron deposits of West Africa
4- 1 o r 0 to the various facies, according to whether they are rich or (Mauritania, Liberia, Guinea) is made up of itabirites concentrated in iron
poor. The drill core is then represented by a series of v,,I l ues of the oxides by departure of the silica and is essentially composed of two types of
"all o r nothing" type with constant support and the sequences of ore:
facies may be characterized by the ranges of the variogram.
(i) type a , rich shallow hematite ore corresponding to a maximum
departure of the silica (67% Fe, and S i 0 2< 1%);
(ii) type P , mixed poor ore (55% Fe and S i 0 2 > 5 % ) located at the
lll.C.3. Statistics of data fringes of the deposits or deep at the contact with the clean itabirites.
Before starting variography (i.e., the study of the spatial auto-correlations), Of course, when mining within each of these two main types of ore,
it is advisable to carry out an elementary statistical analysis of the available several subdivisions will be considered according to the granulometry and
data and to compare the results with the geological hypotheses. the degree of alteration (i.e., presence of goethite).
It is not within the scope of this book to present any of the various These two types a and /3 can be clearly distinguished by
statistics or other techniques of data analysis; it is enough to my that such
techniques can, and in some cases must, complete the geostatistical (i) their statistical distributions (histograms and correlation diagrams of
approach. Among the many works dealing with the statistical analysis of the three main components, Fe, S i 0 2 and A1203of the ore);
geological data, some of the most recent are: G . S. Koch and R. F. Link (ii) the structure of their spatial variability (variograms);
(1970-71), J. C. Davis (1973) and F. P. Agterberg (1974). (iii) their spatlal localization within each deposit,
206 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 207
Figure 111.18 shows the Fe-SiO, correlation diagram derived from the effect); the silica appears as mere non-structured traces in the
core samples (of constant support) available on one of the N m b a range homogeneous hematite ore, whereas the P-type SiOz serni-vario-
deposits (Guinea). This diagram clearly reveals two sets corresponding to gram shows a structure quite similar to that of P-type Fe; the
variability of iron grades is closely linked to that of the silica grades
(since the ore is enriched by departure of the silica).
It is interesting to compare the diagram of Fig. 111.18 with those of Fig.
111.20(a), (b) and (c) corresponding to the Fe-SiO2-A12O3 grades of the
deposit of F'derick (Mauritania). In his study of the F'derick deposit, G.
Matheron (1962) also distinguished two main types of ore, called I and 11,
the characte:ristics of which are very similar to those of the two types of ore
(a and /3) distinguished in the Nimba range deposit.
FIG. 111.21. Data aligned and regularly spaced. (a) Rectilinear drill core;
(b) channel samples along a gallery.
2, Aligned but irregularly spaced data increase to the sill value) has been calculated using at least three or
To construct the experimental semi-variogram in the direction of align- four classes; thus, the tolerance E (r) should not be too large; $ -
ment a , the data are grouped into distance classes. Elvery data pair which is (iii) Ensure that each distance class [r f ~ ( r ) contains
] enough pairs so
separated by a distance [ r * ~ ( r ) ]is used to estimate the value y(r). that the corresponding estimator of the variogram is reliable; thus,
In practice, it is easier to consider a constant tolerance 6 (I) whatever the the tolerance ~ ( r should
) not be too small.
distance r, i.e., E (r) = constant, Vr. Otherwise, e (r) should be smaller for (iv) Detect any risk of bias due to preferential location of data.
smaller distances and larger for greater distances.
This grouping of data pairs into distance classes causes a smoothing of
the experimental semi-variogram y*(r) relative to the underlying
theoretical semi-variogram y(r), cf. Fig. 111.22. If the N' available data This category can be reduced to one of the former two.
pairs separated by a distance r i [ ~r * ~ ( r ) ] are used instead of those (a) By defining approximately rectilinear pathways passing through the
separated by the strict distance r, then it is not y(r) that is being estimated availalble data locations. Each of these approximate alignments is
but rather a linear combination of the y(ri); more precisely, the theoretical then treated separately with a possible grouping into distance
mean value classes, cf. Fig. III,23(a). T h e disadvantage of this method is that all
the available data are not used, and it is difficult to program.
(b) By grouping the data into angle classes followed by distance classes,
cf. Fig. IIIa23(b).T o construct the variogram in the direction a , each
*
The significance of the effect of this smoothing over the interval [r ~ ( r ) ] data value z(xo) is associated with every other value located within
decreases as the tolerance &(r)becomes smaller with respect to the range the arc defined by [ a k d a ] . Within this angle class, the data can be
of the theoretical model y(r) to be estimated. *
grouped into distance classes [r ~ ( r ) ] .
FIG. 111.23. Non-aligned data. (a) pathways; (b) grouping into angle classes.
This grouping into angle classes provides an experimental variogram In the same way, an experimental cross-semi-variogram yZk,(r,a ) given
without distinction between the directional variabilities within the arc by formula (111.47) is associated with
[ai d a ] and, thus, a smoothing effectwill be produced when the theoretical (i) The number of pairs N1(r,a ) of available data;
semi-variogram y(x, a ) varies in the interval [ a k d a ] . (ii) The dimensions and features of the zone V containing the data;
In practice, when the existence of a privileged direction of anisotropy a 0 (iii) The means, fk and f k , , the dispersion variances s: and s:, of the data
is suspected (and such a direction is generally preferentially sampled), then used and the correlation coefficients p,,..
an arc [cuokdcuo] with a small tolerance is defined so as to estimate
accurately the variogram in this direction. Alternatively, when such a
direction is not suspected, it is then enough to consider four directions in Groupirig into mean experimental variograms
the two-dimensional space, each with a tolerance angle of =k7r/4, ,at least as Consider the two followinl: elementary experimental variograms:
a preliminary step. If an anisotropy becomes obvious, the next step would
be to reduce the tolerance of the arcs,
Information associated with each mean variogram In the same way, if the three-dimensional zone V is proven by regular
Each mean variogram obtained by combining elementary experimental sampling of selected horizontal levels, an experimental dispersion variance
variograms is associated with the following information. can be defined over a mean horizontal surface g. .I
Computer subroutines
semi-variogram is calculated by combining the four directi onal semi-
variograms, cf. Table 111.2 and Fig. 111.25. A linear m ~ d e wit1 l 1 no nugget
Practical programs for the construction of experimental variograms are,
effect can be fitted to the mean semi-variogram:
above all, programs for the creation of data files and the classification of
y ( r ) = 4.1 r / a , V r / a E [ O , 3 J2]. data. This section will be limited to the presentation of three basic subrou-
tines which can be adapte~dto any particular data file.
- Direction
Step 1
hT'(l) y(1) N'(2)
Step 2
A (2) N'(3)
Step 3
y1'3)
-- These three subroutines correspond to the three types of data configura-
a1 24 4.1 20 8.4 18 12.1 tion most often encountered in practice in mining applications.
ff 2 22 4.25 18 82 15 10.9 1. One-dimensional configuration of aligned and regularly spaced data,
a3 . 19 5 16 11.9 10 17.3
a 4 18 6.5 14 11.3 8 15.4 cf. Fig. 111. 21; e.g., a rectilinear drill core cut into samples of constant
length I.
2. Two-dimensional configuration of data at the corners of a regular and
r'i -. TABLE111.2, Isotropic mean semi-varlogram
rectangular grid, cf. F J ~111.26;
. e.g., the sampling of a subhorizontal
sedimentary deposit by vertical drills analysed over the entire
-- mineralized thickness, the drills being placed horizontally on a
h regular, rectangular grid.
a a42 2a 2aJ2 3a 3aJ2
- 3. Two-dimensional configuration of data irregularly distributed over a
plane, cf. Fig. 111.27; e.g., the sampling of the previous sedimentary
N'
-Ah)
46
4.1
37
5.7
38
8.3
30
11.6
33
11.6
18
16.3 - deposit by vertical drills placed anywhere on the horizontal plane (but
not preferentially on rich or poor zones).
218 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS
U=U+VRl
\=V+VFl*VRl
C
.................................................
SEMI-VARICGPAC ALChG A L I h E
11=1+1
JH~HIhO(I+KHbXvNC~
C REGULAR G R I C O F POIhTS..HLY t P L E P I S S I h C CATA.. CO 2 1 J a l l p J H
C KaJ- I
C -...PARAPETERS IF(VR(JI.LE.TEST)GG TO 2 1
C VP(hC) C I T b ARRAY hC(K)=hC(K)+l
C NO N L Y B E R CF P C I N T S C F T h E L I A E LRR=kPl-LR( J)
C KHAX P b X I H U P N U P B E R CF C C P P U T b T I C h L A G S G ( K ) = C ( K )+O.S*VRR*VRR
C FbS LENCTt- O F L A G 2 1 CONTIAUE
C NC(KMAX1 NUYBER OF C O U P L E S / L b G 2 C O N T IEIUE
C G(KHAX) SEMI-VPRICCRAC kALLES/LAG
C U A'fERAGE I RESCLTS
C V V A R I A N C E I OF O h T b > T E S T
C N KUHBER I
.... CPTIChS
IS.NE.l RESULTS ARE P R I h T E C EY SLERCL1I h E
...ICC~MChS
IGUT L I N E PRINTER U N I T hlKBEP
TEST I N F E R I C R BCUNCARV C F E X I S T l F G CbTA
PRINT RESULTS IF 1S.NE.I
.................................................
I F YR.LE.TEST M I S S I N G OR E L I H I h A T E C C A T U P
h H I T E ( ICUTv20CO)
~ P I T E ( I O U T ~ ~ O O ~ ) ~ ~ V P N
CC 4 1 K a l g K M A X
C=K* P b S
INITIALIZE 41 h P I T E ( I O ~ T ~ 2 0 0 2 ) K i D r h C v G
4 CONT 1h U E
,-- 220 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS
I
,
hL I NU+lBER OF L I V E S C F C A T A GRIC
hCC - h U M R E H CF C C L U M h S
KMAX M I X I M U C N U M B E R CF C C C P U T b T I C h L A
lo( 4 DIRECT ION IND ICATCP ALOhf L lhES
JD(4) 0 I B E C T JON I N 0 I C P T C R b L O h G C C C l P h
hC(KMAX*4) hUMBEK O F COUPLEC/LAG/GIRECTICh
G(KMAX*4) SCMI-VARIOGPAM VAL1 ES/LAG/C IRECT
U ~vEKAGE 1
r
5 FETURh
EhO C
C
....CPTI
IS.NE.1
ChS
RESULTS APE P P I h T E C
C
2. Dara on a regular rectangular grid on a plane (subroutine GAbdA2)
". C
C
ICUT
TEST
L I N E PRINTEQ L N I T hlPRER
I h F t R I O R BCUNCARY CF E X I S T I h G CATA
, ' 1 2------ N C O columns C I F VP.LF.TEST M I S S I h G CR E L I M I h b T E C C A T U C
C
- .-.
C INDICATOPS I D t J C SHOh TkE R E L C T I V E L l F E b h D
C C O L U M N F C S I T I C h S @F C L n S E S T h E I G H P C L P T C P C I h T
C I,J I N T t E C O N S I C E R E C 0 I R E C T I C N I E h A C ~ L ~ :
C D I R E C T I C V 1 hl-S (//COLUCh:) IC=l JD-C
C 2 hl-E ( //LIhES ) IC=C JC=l
C 3 h'n-SE ( C I b G c h b ~ 11 1c.1 ~0.1
' 0 0 0 0 C 3 NE-SW (DIbGChAL 2) 1 C = - 1 JO.1
NLI C CTI-ER D I R E C T I O h S PAY PE D E F I h E C
0
lines
I FIG.111.26. Data configuration considered by GAMA2.
I
I--
The data for subroutine GAMA2 ( N L I lines and N C O columns) i , input in
the form of a column vector V R ( N L I * N C O ) . The subroutine calc~~lates
the
four experimental semi-variograms corresponding to the four main direc-
tions (the two directions of the grid and the two diagonal directions). The
subroutine also supplies the experimental mean and dispersion variance
corresponding to all the data available on the plane.
C
C
..................................................
SEMI-VARICGRLC I N ThC C I r E h 5 I C h S
C REGULAR G R I C CF POIhTS..MbY b F V E P I S S I h G CATb..
C TI-IS I S A SACPLE SURROUTIYE S k C h I N C C C C F L l A T I C h
C FOR FOUR CI9ECTIUhS.-HOWEVER k E C T O R S I f ChC JC C A Y
C BE D E F I N E C FCR CTHER C I R E C T I C h '
C
C
C
.... PARAPETEFS
VR ( h L I * N C O ) C A T A A R R A Y ( S T C R E C C O L b C h C I 'E I
C
C
CCMPUTE S E P I - V A R l C f Q L M , N E b PCINT
MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 223
STLTISTICS
NEk C I R E C T I C K
CO 30 K C = 1 , 4
J3=JC(KDI
13.1 C ( K C J '
11.1
J 1=J
3. Data irregularly spaced over the plane (subroutine GAMA3)
kEk L A G
SUBROUTlNE G A M A 3 ( V R ~ X ~ Y , h C * K F A X ~ P A S t C P , N C I v
~ALP*DIINCIGV D,UtV,N, IS1
C
C
C
.................................................
S E M I - V A R I O G R L U I N TYC C I M E k S I O h S
C I R R E G U L A R 6 R I D . T H E R E WAY B E C I S S I N G DATb,.
C C A L C U L A T I O N BY C L A S S OF ANGLE b h D DISTbhCE..
C
C -...PARAFETEPS
C VR(ND) D b T A ARRAY
C X(ND1,Y(NDI X AND Y C O O R O l h b T E S CF F C I h T S
C NO NUMBER OF P O I N T S DO 1 0 I K = l , u ~ M C
C KMAX U L X I M U F NUMBER C F C C W P U T A T I C N L A G S NC(IK)=O
C PAS LENGTH OF BASIC LAG D(IK)=O.
C DP W I D T H OF D I S T A N C E CLASS.IF CP.0. 10 G I I K l = O .
C OP=PAS/Z. I S TbKEN IFrVR (ND)-TEST)lZr12*12
C HDI NUMBER O F D I R E C T I O h S 1 2 N=O
C ALP(ND1) ANGLES D E F I N I N G D I R E C T I O N S ( k l T H U=O.
C R E S P E C T TO X A X I S I N D E G R E E S ) vpo.
C DA W I D T H O F A N G L E C L A S S a I F OA=C. GO T O 11
C D A 1 4 5 - DEGREES I S TAKEN 13 N = I
C N C I K M A X * h O I I N U M E E R OF C O L P L E S / L A G / D I F E C T I O N U=VR(LC)
C G(KRAX*NOI) VARIOGRAU VALUES/LAG/DIRECTION V=VR(hD)*VR(NC)
C D(KMAX*NDI AVERAGE D I S T A N C E / L A G / O I P E C T I O t 4 11 C O N T I N U E
C U bVERAGE I
C V V A R I A N C E I OF D A T A > T E S T COMPUTE SEMI-VARIOGPbC,NEW POINT
C N NUMBER
C
C 0-..OPTIONS
C I S - H E W 1 R E S U L T S ARE PR1HTED:ChE PAGE CF P R I h T C U T
C PER E A C h 5 D I R E C T I C N S
C
c
C
....c o n n o ~ s
IOUT L I N E P R I N T E R L N I T hLMBER
C TEST I N F E R I C R BCUNCARY OF E X I S T I h G D A T A
C I F V R - L E O T E S T M I S S I N G OR E L I M I N A T E D C A T U U
C
NEW L A G
C ..a .CAPACITY:
C
c
C
..................................................
10 D I R E C T I O H S : S A N ~ 1 O ) t C A N ~ 1 C 1 DO 2 1 J ~ 1 1 , N c l
IF(VR(J).LE.lESTdGO
DX=X(JJ-X(1)
TO 2 1
DIMENSION V R ( l ) i X ( l ) r Y ( l ) OY=YIJ)-YII)
DIMENSION ALP(1) ,CAN(lC),SbN(lO) H=SQRT(DX*OX+DY*DY
DIMENSION G ( 1 l i D l l l v hC(11 IF(H.LT.1.E-03)GO TO 25
COMHCk I C U T * T E S T K x I N T IH/PAS+0.5) *1
DATA I A / ' '/
226 MINING GEOSTAT1STIC:S
111. STRUCTURAL ANALYSIS 227
4 1 CONTIhUE
c
2 0 0 0 FORMAT(1H ,'**OOU@LY D E F I h E C PCINT,*EATUH'i'
174,' X=',F9.4r1 Y='1F9.4v1 O A T L P ' i l 4 1 ' X=',
CO 2 2 K D = l , N O I 2 F 9 * 4 r 1 Y=',F9-41 .h
*.*I,
must be accounted for, either by the geology of the phenomenon o r as a experimental variograrns calculated from the two sets of data, rather than
structural artefact due to measurement (preferential d a t a locations, biased t o compare a single vertical drill core with a single horizontal channel
analyses, etc.). sampllng profile.
Again, through geological studies the various scales of variability of the Second, if the two experimental variograms a r e defined on different
phenomenon can b e qualitatively evaluated, e.g., the petrographic scale of supports (channel sample and core sample in the preceding example) it is
the transitions from o n e element to another, t h e lithological scale of the advisable to determine whether the difference in support is sufficient to
various facies, the tectonic scale of the transitions delimited by the main explain the observed difference between the variograms.
faults, etc. Through the concept of nested structures, ~ a ~ r i o g r a p h y Third, ~f the means of the data used to calculate the variograms are
quantifies the respective part played by each of these scales of variability different, then i t is advisable to look for a possible proportional effect; for
and evaluates their importance in relation t o the various problems of example, by plotting and comparing the relative semi-variograms y*/m*'
estimation, optimization of production, quality control, etc. For example, if a dlrect proportional effect is suspected.
what is the residual macroscopic effect of the micro-variabilities o n the Fourth, the preliminary review of the data may immediately reveal the
grade fluctuations of a production size unit? T h e geostatistical answer to causes of an observed difference between two variograms: a single aberrant
this problem is provided by a study of the effect of regularization of these data value, \{,hen squared, is enough to completely change the behaviour of
micro-variabilities when changing from the data support (core sample of the variogram; the two sets of data may not have been collected with the
several kilograms) t o t h e support of a production size unit (several hundred same equipment, etc.
tonne block) cf. section IV.F, Case-Study 15 - the case study of the grade Variography does not create data, it simply presents it in the synthetic
fluctuations in a nickel deposit. and quantified form of the variogram. A significant heterogeneity dis-
Each of the case studies given in this book, and particularly those in played by a varlogram usually has an obvious physical explanation; when
Chapter IV, demonstrates the complementary nature of the quantitative this is not the case, an (artefact of measurement o r calculation should be
approach provided by variography and the more qualitative geological suspected. If only a slight difference is observed between two variograms
approach. a n d there is no obvious physical explanation, then it is best t o ignore it - by
Because of this essential interaction between the geostatistical approach assumlng isotropy, for llnstance - rather than developing a complicated
and a qualitative experimental knowledge of the studied phenomenon, it is model for anisotropy that may turn out to be illusory. T h e rule in such a
not possible to define a universal model o r plan for the analysis of vario- case is to adopt the simplest model.
grams which would apply to all types of deposits and to all methods of
proving these deposits. A t the most, several general remarks can be made
Pure nugget effect or experimentnl fluct~lations?
on the various data artefacts which may bias the structural picture given by
the variograms. Let y,(h) be a point tralnsition model with a range a. which is very small
when compared w t h the dimznsions of the data support: a o e v ,
In practice, when u > ao, t h e regularization on t h e support v covers the
Causes of heterogeneity between variograms entire variability yo so that it no longer appears o n the regularized semi-
Before ascribing the observed difference between two variograms to a variogram except for a nugget constant, cf, formula (11.7):
physical heterogeneity of the phenomenon under study, it should be you(h:)= Co, =constant, Vjh 1 > v ,
ensured that this difference is not d u e to unrepresentative variograms or,
more generally, to a n artefact of measurement o r calculation. Causes of with the rule of inverse proportionality to the support: CoU = (vf/v)CoU8
heterogeneity d u e to measurement are numerous. When the variability consists only of micro-structures of the type yo(h),
First, each of the two variograms may be calculated from a n insufficient i.e., structures with ranges much smaller than the supports of the data, the
number of pairs o r a neighbourhood which is too small (cf, practical rule experimental semi-variograms will appear flat or, more exactly, fluctuating
(111.42)) t o b e representative. For example, if the available data consist of slightly around their sill values, as in Fig. III.28(a).
vertical drill cores a n d horizontal channel samples, and it is intended to Before assuming the hypothesis of a pure nugget effect, it should be
detect anisotropy, then it is advisable, if possible, to compare the two mean ensured that nugget-like behaviours (~.e.,without apparent increase) of the
230 MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS 23 1
experimental variograms are not due to a s m o o t h ~ n geffect in the cal- reflects the various experimental features which have been found (experi-
culation of the variograms (e.g., due to reconstruction of core samples of mental vario,grams, anisotropies, a priori geological knowledge, etc.). This
constant length, cf. formula (111.45)) or a fluctuation effect (e.g., too few representation must also be operational, i.e., simple and adapted to the
data used for the construction of the variogram). Contrary to an apparent purpose of the study. In particular, certain minor features revealed by
nugget effect due to experimental fluctuations, a pure and stationary structural analysis may be omitted from the adopted representation.
nugget effect only shows slight fluctuations around the sill value, cf. Fig. Similarly, when two geological o r mineralogical causes contribute in the
III.28(a) and (b). A good way to verify a pure nugget effect is to test the same way to the observed variability at some scale, it is useless to compli-
rule of inverse proportionality of the sills to the supports, cf. in section cate the model by distinguishing between these two causes.
III.A.4, the case study of the Adela'ida mine.
Support and regularization
The dimensions of the support v of the data are often very small with
respect to the distances involved in the intended estimations, and also with
respect to the dimensions of the units V to be estimated. In such cases, the
support v of the data can be approximated to a quasi-point support (o =0)
and the experimental regularized semi-variogram y: ( h ) can be treated as
if it were an estimator of the point model y(h).
T h e techniques of regularization (passing from y to y,) or, conversely, of
deconvolution (passing from y , to y ) are required only when comparing
variograrns defined on different supports (y, and y,, with u 71 tl'), cf. in
section II.D.4, remark 3.
FIG.111.28. Nugget-like variograms. (a) Actual pure nugget effect; (b) apparent These techniques should not be used to deduce a point model which is
nugget effect. illusively detailed, e.g., a sophisticated point model with a large number of
parameters. 'To be rigorous, it is not possible to reach a greater degree of
It should be stressed that under the hypothesis of stationarity a pure precision than that of the smallest support v of the data without introduc-
nugget effect corresponds to an hypothesis of extreme homogeneity of the ing supplemcntary and unverifiable hypotheses. It is always better to avoid
underlying mineralization, cf, section III.A.2. In rnininl; practice, an actual such hypotheses and stick to the available data; for example, by assuming a
pure nugget effect is extremely rare; on the other hand, insufficient or quasi-point support.
inadequate data will quite often result in experimental variograms having
the appearance of a pure nugget effect. Sill and dispersion varinrlce
The choice of a theoretical model with or without sill is always made by
lll.C.6. Fitting a model examining the experimental variograms, taking account of the fact that
these variograms are subject to significant fluctuations at large distances.
As was the case for the analysis of the variograms, there is no unique In the same way, a sill is chosen by fitting a line through the experimental
technique or method for fitting a theoretical model to a regionalization fluctuations. T h e experimental dispersion variance
identified by a set of experimental v a r i o g r a m s . - ~ tthe most, the foliowing
remarks can b e made.
Operational representation can sometimes be used in fitting the sili, provided that it is in fact a good
estimator o f this sill (i.e., of the a priori variance D ' ( v / c ~ )when
, v is the
The representation of a natural variability by a point or regula.rized vario- support o n which the semi-variogram is defined), cf, section II.C.1' and
gram can never be unique. In practice, it is enough that this representation section III.C.4 (the note about the experimental dispersion variance s2
MINING GEOSTATISTICS Iill. STRUCTURAL ANALYSIS 233
particular to that point; this error varies according to the separation dir-
associated with each mean variogram). In particular, it should be stressed tance h, the zone o r data set from which the variogram is calculated, etc.
that the existence of a sill is not related to the existence of an experimental The best method for weighting each estimated point y*(h) with its
dispersion variance s2; the experimental variance of N data can always be "coeficient of reliability" comes from a critical review of the data and from
calculated even when a sill does not exist ( t ~ corresponding
e RF being only practical experience.
intrinsic).
Among other things, physical knowledge of the phenomenon (geology,
for example), common sense involved in approximations and adapting
Behnuiour at the origin available information to the ultimate purpose of the study are all as
essential to a structural analysis as the calculated values of the experi-
Most variables encountered in mining applications yield a semi-variogram mental variograms.
with a linear behaviour at the origin and a nugget effect. cf Fig 111[.?9.The
nugget constant Co is obtained by extrapolating the average linear
behaviour of the first experimental points to the ordinate axis. results with respect lo rhe chosen type of
Robusrness of the geosta~~istica/
model
Consider the two transition-type models in current use which have a linear
behaviour at the origin, i.c., the exponential and the spherical models.:
y(r) = 1- exp (-r/a);
and
Automatic fitting
As far as mining geostatistics is concerned in this book, any attempt at the parameters of the two spherical component structures C, y, and Cty2
hlind
-- automatic fitting of parameters to experimental variograms - such as being
least squares methods - should be avoided. ci = 0.45,
It should be recalled that each estimator point ~ ' ( hof) an experimental
semi-variogram is subject to an error of estimation and fluctuation which 1s
sills
1
CZ= 0 ~ 5 3 ,
ranges{ a l = 1.2
a2 = 3.6.
MINING GEOSTATISTICS 111. STRUCTURAL ANALYSIS
For these two models, exponential y ( r ) and its fit y*(r), the values of the Note that !he relative deviation between two results derived from the
following geostatistical operations in the one, two and three-din~ensional two models is o f the order of 2% for the dispersion variances and of several
spaces are given in Table 111.3: per thousand for the estimation variances. The excellent concordance of
the results is to be expected since the fitting of y * ( r ) to y ( r ) on Fig. 111.30 is
(i) The dispersion variance of a point within a segment, square o r cube remarkable, at least within the working interval r E [O, 31.
of side I = 3a = 3, cf. formulae (11.98) and (11.99) and Charts no. 4, 14 and
5, 15;
(ii) The estimation variance of this segment and this square by a central
point, and of the cube by an axial core of length 1, cf. Charts no. 7, 17 and
10, 20.
Dispersion Estimation
variance var~ance
R' R2 R~ RL R~ R3
Exponential y 0.544 0.73 0.82 0.42 0.57 0.18
Fit y* 0.535 0.72 0.81 0.419 0.571 0.179
Relative deviation 0.017 0.014 0.012 0.002 0.002 0,006 .
IV. CASE STUDIES OF STRUCTURAL ANALYSIS
237
had an approximately constant length L = 3 5 0 ft, thus providing a set of
some 7000 samples analysed for oil expressed in p e r cent saturated weight.
A s there was little variation in the data support, it was not deemed
necessary to reorganize the data into strictly constant lengths, However, an
effective grouping into distance classes was made by approximating the
data to a point grade z ( x ) centred o n the mid-point of the corresponding
core sample, the class interval being equal to the mean length, 1 = 2 ft, of a
core sample. It will be recalled from section III.C.4 that the grouping into
distance classes causes a smoothing effect particularly significant for very
Case Studies of Structural small distances, and, for thls reason, the first point of the experimental
variogram, I.e., y*(2), has been rejected.
Analysis T h e mean vertical semi-variogram obtained from the weighted average
of all the elernentary vertical semi-variograms for each drill core is shown
on Fig. I V . l . T h e following information is relevant to this variogram.
(i) T h e first experimental point y Z ( 4 )was estimated with 3363 pairs of
SUMMARY data values, and all points up to h = 1 5 0 f t were estimated with at
least 2000 pairs each. Toward the limit of reliability (h = L / 2 -
The 15 case studies given in this chapter were selected to cover the main structural 175 ft, one half o f the mean length L of the drill cores), the number
models of varlograms in current use as well as the most crltical problems vvhlch may o f pairs was still greater than 700. Thus, the reliability of this mean
be encountered In the course of an actual structural analysls
The first five examples of section 1V.A show how to fit the various variogram vertical semi-variogram can be considered as excellent.
models currently used and Introduced in sectlon III.B.2.
Section 1V.B gives a typical example of model fittmg to a regionalization whlch
shows a hole effect and a proportional effect.
Section 1V.C gives two case studles of coregionahzatlon and theu fit to the l~near
model introduced in section III.B.3.
The two examples in sectlon 1V.D present the two models of anisotropy intro-
duced in section 1II.B 4, i.e., the geometric and the zonal anisotrop~es.
While the presentation of each of the previous ten case studles is focused on a
limited number of theoretical concepts, the three case studles of sectlon IV E glve
three examples of structural analyses in three d~mens~ons, where most of the
previous concepts are involved s~multaneously.
The two examples of sectlon 1V.F are part~cularlyimportant in mmng practice,
as they show how the two geostatistical tools - estimation and dispersion variances
-can be used in a simple but operational way to control grades durlng production
A A A This value of D'(o/L) is very close to the value 18.24 calculated assuming
D~(I/L)=---=-
1 L l a support of length 1 = 2 f t . Note that the quasipoint support approximation
greatly simplifies the calculations (see the remark on support and
and, from formula (11.36), the variance contributed by the structure y1 is regularization in section III.C.6).
given by
Remark 2 Exponential models, like spherical models, also have a h e a r
behaviour at the origin and a sill. However, if such a model is adopted for
yl(h), the practical range would be three times the abscissa (h = 24 ft) of
where y ; ( h )is the underlying point model of the mode! y , ( h ) regularized intersection of the tangent at the origin with the sill, cf. section IIIiB.2,
I V . CASE STUDIES: OF STRUCTURAL ANALYSIS 241
3'- MINING GEOSTATISTICS
distance u p to r ='60 m , equal to the limit of reliability, a n exponential
model with a nugget effect can be fitted remarkably well:
.*-*-
E-W
L
5 000
I I I
90' l SOo 270' 360'
Long~tudeEast
-
is shown on Fig. IV.4. By contrast, the E-W profile of the mean values of experimental semi-variograms of Fig. IV.3. As these semi-variograms were
the geopotential within vertical bands of lo0, shown on Fig. I V S , appears calculated from the quadratic deviations of the data z(x,), they are, in fact,
stationary. estimates of the moment $ E { [ z ( x + h ) - 2 ( x ) 1 2 } which, according to the
ikm) above model, is
.50 5 2 54 56 5,8
Pole 90
The directional experimental semi-variograms showed no clear anisot- t(x) of the support, i.e., a ( x ) = t(x).z(x), and is expressed in per cent
ropy over small distances (h < 300 m), and so they were grouped together Al2O3-metres.
to provide a mean isotropic semi-variogram, as shown on Fig. 11v'.6. Using the data on the 50.-m grid, the semi-variograms for the N-S and
A model without sill was fitted to this semi-variogram: y(r)==ere, with E-W directions were calculated for distances IP in multiples of 5 0 m .
C = 0.7, and 0 = 1 - 4 < 2 , cf. formula (111.18).
FIG.IV.7. Survey of Mehengui deposit. (a) Vertical section; (b) horizontal section.
is not possible to determine visually which is the best fit, at least within the Calculation of the dispersion variance
interval of reliability of the experimental semi-variogram (r < 500 m).
The theoretical calculations using both models lead to completely The experimental dispersion variance of the accumulations a ( x ) within the
similar results, cf. the following evaluation of the theoretical disuersion deposit is s 2 ( 0 / 0 ) = 15.3 ( O h m)'. This variance was calculated from the
variance. data on the regular square 50-m grid (uniform distribution within D) and
can thus be considered as an estimator of the theoretical dispersion
D ~ ( O / D ) within the deposit approximated by a rectangle D =
112.5 m x 6710 m. The support of the horizontally regionalized variable a ( x )
is the horizontal section of a core and is assumed to be a point.
The experimental value s 2 ( 0 / ~ ) =15.3 will now be compared with the
two theoretical values derived from the two previous models.
From the wneral formula (II.36), the theoretical dispersion variance is
written as
D ~ ( o / D ) =jj(D/D)-y(O)= T(D, D).
For the logarithmic model, the mean value r(D, D )in the rectangle D is
given by formula (II.97), which considers the linear equivalent (1 = 1125 +
670 = 1791 rn) of the rectangle D ;
For the nested sill model, y (r) = Co-t.C l y l(r)+ C2yz(r), the mean values
F1(D, D ) an~dT2(D, D ) of the spherical models y l and y2 are found from
chart no. 4 in section II.E.4:
y(D, D ) = C o + C 1~ 0 . 9 9 9 + C ~ x 0 * 9 1 15.5.
8=
The latter theoretical value of 15.5 agrees with the experimental value
15.3. The theoretical value 16.6 derived from the logarithmic model seems
a little large, due to the fact that this model assumes a continuous increase
of the theoretical semi-variogram for large distances r E [500, 1125 m]
which are beyond the limit o f reliability of the experimental semi-vario-
gram. In this case, a slight preference should be given to the second model.
FIG. XV.8. Horizontal regionalization of accumulations (Mehengui). (a)
Logarithmic scale.- , Experimental variogram; -. -, two nested spherical
models; -----, logarithmic model. (b) Arithmetic sca,le. - -, Two nested
spherical models, y(h)= Co+Clyl+C,y,, Co= 3.2, Ci == 7.3, C z =5 . 5 , a , = 50, IV.E3. HOLE EFFECT AND PROPORTIONAL EFFECT
0 ~ ~ 3 5 -----
0 ; , logarithmic model y ( r ) = 2.88 log r - 0.69.
Case S t u d y 6. Regionalization of uranium grades
This study concerns the regionalization of uranium grades along a vertical
This case study shows that the important thing in fitting variograms is not cross-section of a sedimentary uranium deposit in Niger (Central Africa).
the type of models, but ratherthe agreement of the model with the various Twenty-five vertical drill-holes were available, each of which was cut into
experimental curves available, cf. section III.C.6 (the paragraph on core samples of 20 cm length and chem'ically analysed for U.T h e mineral-
robustness). Note the importance of the graph scale in plotting the vario- ized length L,, is defined for each drillrhole i. The experimental mean m? of
grams. uranium grades and the corresponding dispersion variance s f have-been
248 MINING GEOSTATISTICS IV. CASE STUDIES OF STRUCTURAL ANALYSIS
239
calculated? along each of these lengths L,. This experimental varrance sf is
an estimator of the dispersion variance D ' ( I / L , ) of a core sample 1 in the
vertical segment of length L,.
Figures IV.9 and IV.10 show the experimental vertical semi-variograms
for four holes S4, S14, Slo and Slh.The semi-variograrns are clearkj different
f r o m each other both in shape and variance: semi-variograms S4 a n d SIo
show a marked hole effect between 1 m and 3 m ; the increase of S I 4is very
rapid while that of Sl6 stabilizes quickly. A proportional effect is also
evident: the variability increases with the experimental mean rn:.
FIG. IV.9. Semi-variogram S4 and its fit. -, Experimental; - -- --, model. (m)
FIG. IV.10. Semi-variograms S,,, Slo, SI6.
T h e experimental mean semi-variograrn obtained by grouping all 25
elementary semi-variograms is shown on Fig. IV.11. This mean serni-
variogram corresponds to a mean mmeralized length t= r?j L, = 1 3 rn z,
and is calculated from more than 1500 samples with a mean grade rn* =
2.2% U. T h e hole effect is n o longer evident. Two nested s p h e r ~ c a models
l
with a nugget effect have been fitted to it, i.e.,
with Co= 3.5 (Yo u)', as nugget constant, and C l y l and C2y2being two
spherical models with sills C1= 3.7, C2= 1.1 (X0 u)*and ranges a = 1 rn
and a 2 = 3 m.
-I_
T h e two ranges correspond roughly to the abscissae of the maximum and
minimum of the hole effect observed on the semi-variograms S4 and Slo.
T h e two ranges can b e interpreted as the mean thickness of t h e rich seams
( a l = 1 m) and the mean distance between two such seams ( a 2=. 3 m) which * 4o I --2L----I------
3 4
h (m)
+For proprietary reasons, all the mean grades quoted in t h ~ example
s have been rnultipl~edby
a factor. FIG,IV.11. Mean experimental semi-variogram and its fit.
250 MINING GEOSTATISTICS IV. CASE STUDIES OF STRUCTURAL ANALYSIS 25 1
can also be seen on the vertical profile of the grades of drill-hole S4 on Fig. variance s j is calculated with a larger number of samples (five per metre) and
IV.12. can be considered as an estimator of the a priori variance ~ ~ ( l /ofa the )
Fitting a proportional effect grades of core samples of support 1 = 20 cm. A scatter diagram of experi-
mental variance s f is shown on Fig. IV.13. A parabola has been fitte? to the
The preceding nested spherical model is only representative of the mean scatter diagram:
vertical structure over the zone sampled by the 25 drill-holes with mean
grade m * = 2.2% U. In order to represent local values on a zone with a
mean grade m? f 2.2, a proportional effect should be added to the model.
with
- y l ( h ) = A [ l - e x p (-phlA).cos p h ] ,
Bh, Vh E [0, a ] ,
y2(h)= {Bat Vh 2 a,
U3O8grade on cores of variable length, or by radiometric measurements. by the squares of their respective mean. These three relative semi-vario-
The purpose of the study is to provide a local and global estimation of the grams are shown on Fig, IV. 16, the coordinate axis being graduated now in
deposit through "cokriging", i.e., considering simultaneously both types of arithmetic scale. A remarkable isotropy between the three directions
information available, the chemical analyses and the radiometric studied is observed and a last grouping provides the mean isotropic relative
measurements. For this, it is necessary to study the U grade-radiometry semi-variogram of chemical grades, yi(h)/m,XZ, cf. Fig. IV.17.
coregionalization.
E-W
Vertical
N-S
mean rn? and the experimental dispersion variance s ? calculated on each The radliometric measures reconstituted on the same support @ores of .-II
chemically analysed drill core: a parabolic relation (sf, m * ) is observed. constant length 1 = 1 m) provide experimental semi-variograms with struc-
Thus, a direct proportional effect in m:, cf. formula (111.37) has been e s similar to those of the grade semi-variograms. A direct
tural f e a t ~ ~ r quite
considered, and the three semi-variograms of Fig. SV.15 have been divided proportional effect in m i has been considered and the various directional
256 MINING GEOSTATISTICS IV. CASE STlJDIES OF STRUCTURAL ANALYSIS 257
semi-variograms have been grouped into a unique mean isotropic Mining is carried out using a bore and pillar method in a n E-W direction;
relative semi-variogram of radiometric measurements, y g (h ) / m g2, cf. pillars are 19 ft wide and stopes a r e 33 ft wide, cf. Fig. IV.19. O n each
Fig. IV.17. mining level, both faces of each pillar were sampled continuously by 10-ft
This same figure, Fig. IV.17, shows the mean cross semi-variogram, chip samples, providing a set o f data o n a regular grid measuring 1 0 ft in
y,*_R/m,*.m;, isotropic and relative to the product of the two experimen- the E-W direction and alternately 33 f t and 19 ft in t h e N-S direction.
tal means of the grades ( m : ) and radiometries ( m g ) which have been used Each chip sample is anzlysed for Pb, Z n and Ag.
t o calculate it.
O n Fig. IV.17 it should be noted that the cross-variability y,*-R/m,*.ntg
is always greater than the direct vx!abilities y,*/m,*' and y$/nig2, which
would entail that the experimental correlation coefficient p * ( h ) between
the two increment variables [Z,(x + h)-Z,(X)] and [ Z R ( X+ h ) - - Z R ( X ) ]is
superior t o 1. Indeed, the theoretical correlation coefficient is written
Thus, the three experimental curves of Fig, IV.17 are not consistent. O n e
possible explanation is that the experimental cross-semi-variogram y:-R
has not been calculated f r o m t h e same data as the direct semi-variograms
(in practice it is difficult to d o otherwise, unless we consider only the few
points where the two measurements - grade and radiometry - exist simul-
taneously). Inconsistent and excessively small valucc of the means m l and FIG. IV.18. Cross-sections showing the three types of ore. (a) Vertical;
m g can then explain the excessive variability of :he experimental relative (b) horizontal.
cross-semi-variogram yz-R/m t .m g as observed on Fig. IV. 17.
Fitting a model
The simultaneous presence of the same transition structure with range
a l = 6 0 ft o n all direct and cross-semi-variograms suggests that the linear
model of coregionalization (111.27) would provide a good fit. This model is
given by
Y ~ (Lh )= 2 b;~,y,(h),
I Vh horizontal E [ O , 150 ft],
The indices (k, k' = 1 to 3) represent, respectively, the grade in lead, zinc
and silver. The index (i = 0 and 1) represents the two basic structures; a
transition model y o ( h ) with a very small range or nugget effect, and a
transition madel y l ( h ) with range a l = 60 ft, respectively.
As the nugget effect is not present on the cross-semi-variograms, the
coefficients b t k , , Vk $ k' are all zero.
A spherical model with a unit sill and a range a, = 60 ft has already been
fitted to the basic structure y,(h), and Table IV.1 gives the two matrices of
coefficients [ b i k ] for i = 0 (nugget effect) and for i = 1 (ylstructure). The
corresponding fits obtained using this model are shown by the dotted lines
FIG. IV.20. Direct semi-variograms. (a) Pb grades regionalization; (b) Zn grades on Figs IV.20 and IV.21.
regionalization; (c) Ag grades regionalization.
TABLEIV. 1. Coeficient matrices of the linear model
Pb Zn Ag '
(ii) There are obvious sills which are well approximated by the experi-
mental dispersion variances sZ and which, in all1 cases, are reached at
distances of about (hi= a, = 6 0 ft.
Pb 11 0
[Zn
Ag 0 09 1.1 ]
5.!:['
14.5 5
15 3.81
3.8 1.8
1
. *"I*
(iii) There are obvious nugget effects.
260 MINING GEOSTATISTICS IV. CASE STUDIES OF STRUCTURAL ANALYSIS 261
Remark 1 T h e coefficients hip verify the constraints of t h e linear model, Remark 3 O n e possible: way of interpreting t h e chosen linear model of
i.e., both matrices of Table IV.l are positive definite, cf. conditions coregionalization is to consider the three RF's Zk(x) as i s s u ~ n gfrom two
principal causes of variability represented by the basic RF's YOk(x) and
Yl(x), cf, the genesis of the linear model in formula (111.23):
(i) Nugget constants strictly equal to those found for the horizontal
FIG. IV.21. Cross-semi-variograms. (a) Pb-Zn coregionallzation; (b) Ag-Pb
coreglonalization; (c) Ag-Zn coreg~onallzation. structures: the micro-variabilities YOk(x)would then b e isotropic in
the three-dimensional space.
Remark 2 Note t h e quality of t h e fit achieved for small distances (h < (ii) A common transition-type structure with a range a ; = 30 ft
range a l = 60 ft). This coregionalization model is thus ideally suited to the representing the mean vertical dimension of the micro-basins of
geostatistical procedure of co-estimation of o n e grade (e.g., Ag) by the deposition, cf. Fig. IV.22. A proportional effect is added to this
three data sets (Pb, Z n and Ag), cf, section V.A.5, Case Study 2. geometric anisotropy o t ratio a ; / a l = 0.5, d u e to the fact that the
This co-estimation will be carried out within an horizontal neighbourhood three types of gangue a r e not encountered in the same proportions in
of approximately 60 ft around the unit to b e estimated. each level.
262 MINING GEOSTATISTICS I\/. CASE STUDIES OF STRUCTURAL ANALYSIS 263
Case Study 9. Geometric anisotropy, after J. Serra (1967a) There is a marked hole effect on the E-W semi-variograms, with a maxi-
mum and a minimum variability occurring at around lhl=200m and
As presented here, the study concerns the regionalization of iron grades on Ihl= 350 m, respectively. This hole effect is also present on the N-S semi-
a hectometric scale (from 100 m to 2 km) in the ferriferous basin of variograms with the same mean sill of variability (around 8.3 ( % ~ e ) ' )but
Lorraine (France). T h e data come from two mines, Hayange and Bois with the maximum and minimum variabilities occurring at around Ih[=
d'Avril which, although geographically distinct, mine the same sedimen- 700 m and lh/ = 1000 m, respectively.
tary iron seam known as "red seam".
This red seam is subhorizontal, very regular, and has a fairly constant
thickness of about 3 m. The horizontal regionalization of the iron grades of
this seam has been sampled along the length of mining galleries by
regularly-spaced small drills (90, 100 and 120 m accoirding to direction and
mine). Data is thus available from both mines on a regular grid jn the E-W
and N-S directions o f the galleries. These galleries are very long, from 2 to
3 km, and, consequently, the limit of reliability of thr: experimental vario-
grams extends up to 1 km.
The E-W and N-S experimental semi-variograms for Hayange and Bois h(m)
d'Avril are shown o n Figs IV.23 and IV.24, respectively. The N-S semi- FIG.IV.2J. Horizontal regionalization at Bois d'Avril (red seam).
variograms for both mines are compared on Fig.Iv.25.
d'Avril, and, thus, the experimental semi-variograrn from Sancy will not Data auaii'able
reveal the hole effect quite so clearly. However; the beginning o f the hole
The data come from fans of drills set in an E-W vertical plane approxi-
effect is quite obvious on the NW-SE semi-variogram which shows the
mately perpendicular to the N-S direction of advance of the main drives,
most rapid variability.
cf. Fig. IV,28(b). T h e cuttings from the drills were analysed in l.S;m and
The first step in establishing a model of the geometric anisotropy at
3-rn lengths according to the zone. Because of this configuration, no data is
Sancy was to approximate the sill of the various semi-variograms by the
available in the N-S direction.
experimental dispersion variance, s 2= 11.5 (% ~ e of ) all~ the data con-
sidered in the calculations. Next, a line representing the average slope of
the first few experimental points of each semi-variogram was extrapolated T y p e l ore Type 3 ore
to this sill, providing an abscissa which was transferred to the directional
graph on Fig. IV.27(c) (e.g., 260 rn for the NW--SE dlirection). An ellipse of
geometric anisotropy can be fitted to this directional graph and shows that,
at Sancy and in the grey seam, the sedimentary units seem to be elongated
in the NE-SW direction and appreciably longer than those of the other
two mines, all facts which are corroborated by various lithological
observations.
"
15
Ihl ( m )
FIG. IV.29. Regionalization in the type 2 ore. - , Experimental; - - - -, FIG IV.30. Regionalization in the type 1 ore. Support, 1 = 1 . 5 m.
model; support, l = 3 m.
regularization o f support 1 iiS a nugget constant Col; C l y l ( r )is an isotropic
The combined semi-variograms for directions 1+ 4 and 2 + 3 are spherical model with sill value C I and range a , ; C Z Y ~ ( ~isDa) spherical
shown on Fig. IV.30. They correspond to the central type 1 ore for which model with sill C2 and range a>, depending only on the distances hD in
the data support is 1 . 5 m. In this case, the main variability direct~on(1 + 4 ) direction D.
tends towards directions D (30" W), perpendicular to the mean dlrecticrn of The values of the various parameters of this three-dimensional point
stratification in type 1 ore. model for each of the three ore types are given on Fig. IV.32.
The semi-variograms for directions 2 and 4 and the combined semi-
variogram for directions 4 + 2 + 1 are shown on Fig. IV.31. They cor-
F~ttingthe parameters
respond to the lower type 3 ore for which the data support is 1.5 m. In this
case, no significant directional anisotropy can be distinguished, which The parameters of the point model are derived from the experimental
confirms the absence of stratification in the type 3 ore (dolomitic sand semi-variograms regularized on the support I by the standard procedure of
filling the karstic cavities). deconvolution and ~dentification,cf. section II.D.4.
270 MINING GEOSTATISTI12S IV. CASE STUDIES OF STRUCTURAL ANALYSIS 27 1
and
4 I 2 I-J
0 15 301
Ihl (m)
Finally, the regularized model y l ( h D )in the direction D of zonality is
FIG.IV.31. Regionalization in the type 3 ore. Support, 1 = 1.5 m.
! = 3 r n i C o= I ( % z " ) ~ O W / =I 5 m , ~~'4.7(%*")2
a1 = 21 m , CI= 4 ( % ~ n ) ' o1=12m , C l = 12(%~n)~
a2=36rn,C 2 = O ( % z n ) ' a2=24m ,C2= 2 0 ( % ~ n ) ~
Zans E Zone W
TYPO2 Type 1
'
Q
Zone W
Type 3
...:,., ."
L ' 1 5 rn , Co = 3 , 3 ( % ~ n ) ~
01=9m i C, = 1 5 ( ' % , ~ n ) ~
C2 = 0 ( " ~ > ~ n 1 2
C Z I Y ~ J ( ~being
D ) a spherical model with sill Czl and range a21. This (ii) drills are.coilared in each drive and developed in fans. T h e directions
regularized model yl(hD) is shown on Fig. IV.29 and is a good fit to the a r e variable but the cores h a w been analysed in constant 10-ft
experinlental semi-variogram for direction 1. lengths.
T h e channcl sample and core sample supports can be mixed with good
Verification T h e expressions for the theoretical semi-vario,grams
approximation and statistical studies have shown that there is n o significant
regularized o n the double support 21 = 6 m, y2l(h), were calculated from
bias between the two data sets.
t h e previous point model (formula (IV.1)). These theoretical curves were
then compared with the corresponding experimental semi-variograms
(obtained from the data of support 21) on Fig. IV.33, for the type 2 ore.
T h e verification is positive.
increases with the mean m * of the data from which the semi-variogram was nugget cclnstants are observed. On the N-S curve, there is a change of
calculated. slope around 10 m and the curve stabilizes around 100 m, which suggests
All the data of the channel samples from six levels were combined to the fitting of two nested structures with 10-m and 100-m ranges, respec-
provide the two mean semi-variograms for the N-S and E-W directions tively.
These experimental points (C:, m * ) are shown graphically on Fig. IV.38; vector p of the ellipsoid and is given in terms of spherical coordinates (p, a,
the parabola CO= 0.18 [ m - 0 . 5 1 is~ a remarkably good fit. as
To take account of the radial drift in the E-W directions, it is necessary
to add an additional anisotropic term to the preceding isotropic model
(formula (IV.2)). The deviations from the isotropic model of the E-W
FIG.IV.39. E-W horizontal deviations from the isotropic model. 0 ,Core samples;
x , channel samples.
T o sum up, the three-dimensional model adopted to represent the T h e three types of samples ( D D H , R D H and B H ) were studied
regionalization of t h e C u grades of 10-ft samples is: separately. T h e vertical regionalization can only b e studied through the
D D H , which are the only sampled long enough. O n the horizontal plane
(by grading over the constant 1.3-m thickness of a horizontal bench), these
where nz > 0.596 Cu is t h e mean grade of the zone considered. DDH provide data over large horizontal distances ( h a 1 0 0 m.) T h e RDH
y , and y2 are spherical models with ranges a l = 9 m and a2 = 1 0 5 rn,the provide a comparable horizontal information, i.e., graded over 13 m, for
three-dimensional vector h being defined by its spherical coordinates horizontal distances (h 2 5 0 m). T h e BH provide horizontal information
(r, cr, p). This model is valid for r = 1/11 < 150 m and for horizontal E-W graded over 14 m (14 = 13), but at smaller distances (h 2 8 m),
distances less than 6 0 m.
This model proves to be an excellent fit to a11 the available experimental
figures, cf. the dashed curves o n Figs. IV.35 to IV.37. Horizorlral variograrns graded over 13 m
DDN samples. The horizontal semi-variograms of t h e Cu T and Cu S
' C a s e S t u d y 12. Structural a n a l y s i s of Exotica Cu d e p o s i t , after Ch. were calculated bench by bench over 1 4 benches and in three directions
Huijbregts a n d R. S e g o v i a (1973) (the two directions of the grid and o n e diagonal direction). T h e number of
T h e mineralization of the Exotica copper deposit in Chile is generally data pairs used in the calculations were too few for the bench variograms to
supposed to be of detritic sedimentary origin, consisting of material coming be significant. However, a strong proportional effect can b e seen on the
from the neighbouring massive Chuquicamata porphyry-copper deposit. nugget effect: the nugget constant increases with the mean grade of the
T h e deposit consists of a large stratiform slab with an average plunge of data from the bench considered.
10". T h e total surface area is approximately 1 km x 2 km, with a thickness F o r each of the three directions considered, t h e 1 4 bench semi-vario-
varying between 50 and 1 0 0 m. grams were combined into a mean directional semi-variogram, cf. Fig.
IV.41. A s the three directionall semi-variograms proved to be isotropic,
they were combined into a single mean isotropic semi-variogram, cf. Fig.
Available data
IV.41. For both Cu T and C u S, the mean isotropic semi-variogram shows
T h e results from three types of sampling campaigns are available. a macro-structure with a range a2= 200 m and a very strong nugget effect
engulfing all the structure with ranges less than the horizontal scale of
(i) The entire deposit was sampled by vertical diamond drilling ( D D H )
observation of the DDH (100 m).
on a pseudo-regular 1 0 0 m x 100 m horizontal grid. T h e diamond
drill-holes were cut into constant lengths 1 = 13 m and analysed for
copper. T h e 13-m core sample length corresponds to the height of
mining benches.
(ii) The primary production zones were sampled by rotary drilling
( R D H ) on a regular 5 0 m X 5 0 m horizontal grid. These holes were
also cut into 13-m lengths.
(iii) The production benches were sampled by the same rotary drills used
for blast-holes. T h e drill cuts coming from these blast-holes ( B H )
were approximately 1 4 m long. The blast-holes are located on an
irregular horizontal grid with constant density and a mean 8-rn
distance between two B H . A t the time this study was carried out, the
ore production had just begun and the information from the BH was
very limited. -50 Zlo abo 340 460 4;o
Ihl (m)
All samples were analysed for total copper (Cu T ) and soluble copper FIG. IV.41. Horizontal regionahzation of DDH samples (graded over 13 m).
(Cu S). --- - - - - - -, Directional; -, Mean isotropic.
282 MINING GEOSTATIST1C:S IV. CASE STUDIES OF STRUCTURAL ANALYSIS 283
-
The two other groups of samples R D H and BII -were used to evaluate {Co(0)- CO(r)} represents the nugget effect which appears on the grading
these small-range structures. Care was taken to choose benches and zones over the 1= 13-m support as the nugget constant Car, cf. formula (111.7).
for which the C U T and Cu S grades were similar to the corresponding Clyl(r) amd C2yZ(r) are two spherical models with respective sills C1 and
DDH grades (l~lO/oCu T and 0.8Ol0 Cu S). This precaution cancelled the Cz and ranges a l = 45 m a 2= 200 m.
influence of the proportional effect. The theoretical model of the semi-variogram graded over the constant
13-m thickness is obtained from the point model by applying the practical
RDH and BH samples. The mean isotropic semi-variograms obtained rules of section II.D.4, and is given by
from the RDH samples (on a regular 50-m grid) and the BH samples
yG(h) = 1701+ Cllyl(r)+ CZ1y2(r), V horizontal vector with modulus r E
(grouped into classes because of the irregular grid) are shown in Fig. VI.42. [13,400 m], (IV.4)
There is a remarkable agreement between the stiructures revealed by the
three types of samples. The long-distance structures revealed on the RDH where the graded sills are written as
and DDH semi-variograms are almost exact extensions of the small-dis-
tance (h <50 m) structures revealed by the BH samples. Conversely, the
Cll = C1[1- TI([, l)] and C2, = C2[1- .i;2(1,l ) ] .
B H semi-variogram explains part of the strong nugget effect observed on The graded parameters Col, Clr, Czr are fitted to the experimental
the RDH and D D H semi-variograms, graded semi-variograms. By way of example, the fitted parameters for total
copper are
Cur= 0.22, C l l = 0.10, CZ1
= 0.64 ('lo cu T)'
and, thus, the parameters for the point model (IV.3) are
C 1 = = C I 1 x 1 . 1 6 = 0 . 1 2 and C z = C 2 1 ~ l . 0 3 = 0 . 6 6 ( 9 / CUT)^.
o
The graded model (IV.4) thus considered is shown on Fig. IV.42 and is a
good fit to the experimental curves.
This three-dimensional model is t h e s u m of a n isotropic component and vertical semi-variogram and the theoretical 13-m graded semi-variogram,
a n anisotropic, specifically vertical component C3y3(hw). For every cf. Fig. IV.43.
horizontal vector h (such that h, =O), this three-dimensional model A spherical model, C31y3(hH,),with range a31 = 6 5 m and sill Gl= 0.61
reduces to the preceding horizontal point model (IV.3). (% C u T ) ~ without
, nugget effect, was fitted to this experimental difference
[yT (hw)- y ~ ( h , , ) ] .T h e underlying point model C3y3(hw)is, thus, spherical
without nugget effect, with range
--
Exper~mental
seml-variogrom . /
/
/
/
The deposit was evaluated on a pseudo-regular, square (80 mx 80 m) tonnage of nickel metal, Q A ;
grid of vertical drill-holes. These holes were used to determine the hanging mean nickel grade, g, = QA/ TA.
and foot walls of the silicate seam. Kriging was used to map the isobaths af The tlhree additive regionalized variables regularized on the 1-m core
the hanging wall and foot wall as well as the isopacks of the silicate sample support that were used in the study, are
thickness, cf. section II.B.3, Case Study 2. the accumulation of waste, zw(x,)= I, x d l ;
the accumulation of ore, zT(xL)= (1- I,) x a, ;
the accumulation of metal, zo(xl) = (1 - I,) x a, x g,;
where I, is the length of pure waste measured on each core sample, dl
is the density of this pure waste, a, is the density of the altered ore (of
- S ~ l ~ c a l e ore
Wostt
d
length 1-1,) and g, is the corresponding Ni grades of this altered ore, cf.
Fig. IV.45.
The preceding tonnages are obtained by multiplying these three
regionalized variables by the volun~ein cubic metres.
T h e following regionalizations were studied, using the data available
from the vertical drill cores cut into 1 = 1-m lengths:
(i) the vertical regionalizations through the semi-variograms y,(h)
regularized on 1, cf. Fig. IV.46(a);
FIG.IV.44. Vertical cross-section of the Mea deposit. (ii) the horizontal regionalizations through the semi-variograms y ~ ( h )
graded over the constant thickness 31 = 3 m of a horizontal mining
The drill cores were cut into constant 1-m lengths within the silicate bench, cf. Fig. IV.46(b).
seam. The pure waste was removed from the softer, altered material with a
geologist's pick and the lengths of the two separated materials were
measured and analysed for density and grades (of dry ore) cf. Fig. IV.45.
Length, I-/,
Grades,N,, CO
C o and lmpurlt~es Bench
h e ~ g h l . 31
Length, I,
Dens~ty,a',
FIG. IV.45. Ore and waste materials separated on a core sample. FIG IV.46. Available structural information. (a) Regularization over I ; (b) grading
over 31.
Besides mapping of the silicate seam, the study consisted of a local
estimation on panels and a global estimation 011 the deposit of the four Vertical regionalization L
quantities The 1-rn and 3-m regularized vertical semi-variograms, y l ( h ) and y31(h)
tonnage of pure waste, Tw; for each of the above thiee variables, were calculated from the 167 vertical
tonnage of altered ore, TA; drill-holes and are shown on Fig. IV.47(a), (b), (c).
MINING GEOSTATISTICS IV. CASE STUDIES O F STRUCTURAL ANALYSIS 289
-;;I
formula (111.45).
(a) Support / Using these variables graded over 31, the graded experimental semi-
05- _____ _-- ---- variograms were calculated, bench by bench, and for various horizontal
(a) Support l
(b) Support 31
for the point support; (r 3 8 0 m). T h e nested structure of spherical models (IV.7) is preferred in
practice, essentially for reasons of homogeneity with the models chosen for
the other two variables used, accumulations of waste and ore.
\kI
1
P/
t
Tolol tonnage,
Day j Tfue grode, T,
r/i Esttrnaled grode, T,*
FIG. IV.49 Fitting a logarithmic model to the regionalization of metal accumula-
tion. FIG IV.50. Production planned from several blocks.
294 MINING GEOSTAT1STIC:S IV. CASE STUDIES O F STRUCTURAL ANALYSIS 295
tonnages p,, are calculated so that the predicted mean grade TT of the total If the exact location of the tonnage p,, in block B, were known, the
production P,of the day j satisfies the sales or millling constraints. Thus, for extension variance of the grade t,, of this tonnage p,, to the grade 8,. of
a day j, the following relations hold: block B,, i.e., the term E{(t,,-el)'}, could be calculated. In practice, it is
total tonnage produced, P, = C,=, p,,;
N sufficient to consider the mean value of this term E{(t,,-8,12} when the
tonnage p,, is taken anywhere within block B,, i.e., the dispersion variance
predicted mean grade, Ti*= xipiie*/pi; of an average production unit p, in block B,:
true mean grade actually recovered, T, =. C,p,, . r,,/P,.
The object of the study is to control the difference between the predicted
grade Ti* and the true grade actually recovered. This is dnne by means of the
cf, formula (11.33).
estimation variance:
If the block B, is mined in k days, then the average production unit pi
can be expressed in terms of tonnage as
Suppose that the N blocks B, a r e within the same stationary regional-
ization characterized by thegrade expectation m. The various grades o:, e,,
tii thus have the same expectation m and the non-bias for the prediction of
the daily grade is verified: As for tlhe term E{(0,- o : ) ~ ) ,it is, by definition, the estimation variance
a:, of the mean grade of block B,. If this block has been estimated by
kriging, then a:, is simply the kriging variance of the block given by the
EIT, - T: ) = C p,,E{r,,- e:}/P, = 0. kriging system of equations.
1
Finally, the estimation variance of the mean grade T, of the total pro-
The estimation variance thus becomes
duction P, of day j is written
~ - , -o?)}/P;.
&, = X I P I.,p,,, . ~ { ( t , eT)(t,
I 1,
corresponding production unit p,, is not known (e.g., extraction by block approximately 8125 tons, corresponding to a mean support p of dimen-
caving of p,, tonnes from block B , ) , o r because t h e necessary data is not sions 60 m x 60 m X 0.75 m.
available. These two cases will b e studied later in the numerical appli- T h e mean dispersion variance D 2 ( p / ~is) calculated from the three-
cations. dimensional semi-variogram model of the grades. This results in
I n such cases, it is obvious that the grade control has an absolute limit D ~ ( ~ /=B0.026,
) i.e., a mean irelative standard deviation of dispersion
imposed by the dispersion of the production grade units in the blocks, and
obtained by equating the estimation variances ui,t o zero in formula (lV.8)
(i.e., assuming that the true grades 8, of blocks B, are perfectly known). T h e estimation variance of the daily production grades is then given by
T h e minimum value of the estimation variance a:, of the daily grades is,
formula (IV.8):
thus,
approximated to this grid unit, i.e., to a panel B of dimensions 6 0 m x between the two auxiliary functions F(p,) and F ( B , ) defined on the two
60 m X j m, cf. Fig. IV.51. A preliminary estimation of the mean grade 8, rectangles p, and B,, cf. section II.E.2, formula (II.82), e.g.,
of each panel B, is obtained by taking the mean grade 8: of its central
drill-hole (classical but not optimal estimation procedure by polygons of D ~ ( ~ I / B
T(B1,
~ ) P~)-y^(pi,
= P I ) =F(60; 60)-F(20; 5).
influence), the estimation variance cr: is, thus, given by chart no. 7, and the ) , that as the supports p,
(For the calculations of the values l I 2 ( ~ , / B ,note
relative estimation variance is and B, are considerably larger than the support of the 5 m-long core
samples, the influence of the nugget effect disappears and the relative
a ? , / ~=?C'o/rn2
~ + (C'/m2)x 0.074 = 0.056, V the panel i. spherical model with range a = 300 rn and sill c/rn2= 0.075 is sufficient.)
Immediate calculation and chart reading provide the following values:
Thus, from formula (IV.8), the estimation variance of the daily grade
production is
1
a',, ===[900' ( 0 . 1 1 9 + 0 . 6 8 6 ) + m ' (0.0033 i-0.724)
i i 6300~
2700 1 d a y "I + 2700 1 day-'
FIG. IV.51. Daily production model. i.e., a standard deviation a ~=,0.30% Ni.
Thus, adopting the Gaussian standard, the mean grade T, actually re-
I t is desired to produce 6300 t of ore per day with a grade average of 2% covered each day will, in 95% of the cases, fluctuate in the interval
Ni from three shovels located in workings with different mean grades, cf. Fig. [ 2 k0.61 YO Ni around the predicted value T: = 2 9'0 Ni.
IV.5 1. The mediocre result of this grade control is due to the rather hasty
The following situation is typical of any given day j. estimation of the panel grade 8, by the grade of the central drill core
(i) Shovel no.1, which has a small capacity (900 t dayb') is located in (polygon of influence method). Using the same data (drills on a regular
workings B I with a rich mean grade estimated to be 0: = 3.5Oh Ni. 6 0 m x 6 0 m grid), simple kriging would, without any additional cost,
This shovel moves a unit of size p , = 20 m x 5 m x 5 rn each day. reduce the estimation variance v:,by half, and daily grades could then be
(ii) Shovels no. 2 and 3 have thrice the capacity (2700 t day-') of shovel controlled with a variance of cr;, = 0.051 1 (% ~ i ) ' ,i.e., a standard devia-
no.1. Shovel no.2 is located in a workings of average grade (8?= tion of I T E , = 0323% Ni, which represents a noted improvement.
2?b Ni), while shovel no. 3 is in a poor grade site (8T = 1.5% Ni). The lower limit of the variance r r k , corresponding to thc perfect know-
ledge of the mean grades 8, of the working sites is given by formula (IV. 10):
Each shovel moves a unit of size p2 = p 3 = 60 rn x 5 m x 5 m per day.
The predicted mean grade of the 6300 t day-' so produced is min {a:,) = 0.0120, = 0.1 1% Ni.
i.e., min {q)
*
The interval of fluctuation is, thus,J', E [ 2 0.221 % Ni.
Ni. In practice, if the grade control is deemed insufficient in the interval
T, E [2 *:0.46j0/o Ni provided by krigin'g and the data on a 6 0 m x 60 m grid,
The dispersion variances D2@,/B,)of the grades of units p, in the it is advisable to establish some sort of balance between the cost of
working sites B, are given by the standard procjedure of the diference reorganizing the dally production plan and the cost of additional sampling.
300 MINING GEOSTATISTICS IV. CASE STUDIES OF STRUCTURAL ANALYSIS 30 1
C a s e S t u d y 15. Dispersion of p r o d u c t i o n g r a d e s Similarly, for the first method, the term T(s, s) is given by
Consider again the New Caledonian type of nickel silicate deposit con-
sidered in t h e latter numerical applications of Case Study 14.
I n the present example, only t h e two large capacity shovels are retained, T h e calculation of the term .F(s,s ) for the second method is a little more
providing a weekly production of 5400 x 6 = 32 400 t, which can b e involved. A s s is the union of t h e two rectangles s t , s2, it follows that
achieved in either of the two following manners, cf. Fig. IV.52.
1. T h e two shovels can advance in line o n the same front. T h e weekly
production unit s can be approximated by a rectangle with horizontal .j;(sl, s l ) is given immediately as
dimensions 360 m x 5 m. T h e height of the bench is 5 m , and the
average density of the dry o r e is 1.8.
2. T h e two shovels can advance in parallel on two fronts, getting T h e term ?isl, sz) depends o n the distance 2x between the two working
progressively farther away f r o m each other. T h e weekly production sites sl and s?, cf. Fig. IV.52(b). T h e production year consists of 50 weeks
unit s thus consists o f two rectangles sl and sz with equal horizontal (i = 1 to 50) and, during week i, the distance between the two sites is
dimensions 180 m x 5 m. 5(2i- 1 ) metres. Approximating each working site by its median line of
In the first case, the annual production (1.62 million tonnes over 50 180 m length, the term T(sl,5 2 ) for week i becomes
weeks) can b e approximated by a rectangle S with horizontal dimensions
720 m x 125 m ; in the second case, it will be approximated by a rectangle S
with horizontal dimensions 500 m x 180 m.
Taking t h e weekly production as the homogenization unit, it is desired to
calculate the dispersion variance D 2 ( s / s ) of the weekly mean grades
actually obtained in one production year, using both types of 'jhovel
advancement.
T h e horizontal regionalization of the Ni grades graded over the constant
thickness (5 m) of the benches is stationary with expectation m = 2% Ni
a n d characterized by an isotropic spherical model with a nugget effect,?
range a = 300 m and sill C = 0.3 (% ~ i ) '
T h e dispersion variance is given by t h e standard forrrula:
f o r method 1,and
for method 2
V.A. THEORY OF KRlGlNG T h e experimental data to be used consist of a set of discrete grade values
The problem {Z,, a = 1 to n } . These grades a r e either defined on point o r quasi-point
supports, or else they arc- the mean grades Zu0(x,) defined on the supports
T h e problem of local estimation is to find the best estimator of the mean v, centred o n the points x,; the n supports may differ from each other.
value of a regionalized variable over a limited domain, the dimensions of Note that under the hypothesis of stationarity the expectation of each of
which a r e small compared to the dimensions of the quasi-stationary these data is m : E{Z,} = m, Vcr.
(homogeneous) zones of the deposit, e.g., the mean grade of a block T h e linear estimator Z f f considered is a linear combination of the n data
located well within a zone of homogeneous mineralization. Local estima- values Zg = Z:=, A,Z,. T h e n weights A, are calculated to ensure that the
tion differs from global estimation in that global estimation considers estimator is unbiased and that t h e estimation variance is minimal (the
distances larger than the limits of quasi-stationarity and, thus, sometimes estimator ZE is then said to be optimal).
engulfs various heterogeneous mineralizations.
T h e available information used for local estimation within a quasi-
stationary zone is generally made up of a set of data (e.g., n core grades) I Noon -bias condiiion
and structural information (e.g., the variogram model characterizing the T o obtain a zero error [ Z V-.Z;?,] xn expectation (i.e., an average over a
spatial variabihty in the studied zone). great number of similar estimations) it is enough to impose the condition
Kriging is a local estimation technique which provides the best linear I:=,A, = 1, and, thus,
unbiased estimator (abbreviated to BLUE) of the unknown characteristic
studied. This limitation to the class of linear estimators is quite natural,
since it means that only the second-order moment of the R F (i.e., the
covariance o r variogram) is required, and, in general, it is possible in
practice to infer the moment, cf, section II.B.2. 1 which entails E { Z v - 2;:) = 0.
Of course, when the available structural information includes more than
the simple knowledge of the second-order moment, then non-linear esti-
mators - more precise than simple kriging - can be defined, cf. Chapter
1
i Minimum estimation variat~ce
VIII, "Introduction to Non-linear Geostatistics". T h e present chapter will
b e limited to linear estimators only. 1
1
T h e estimation variance E { [ Z V-z:]'} can be expanded as follows:
(The standard notation C(V,u,) designates the mean value of the covari- Matrix form
ance function C ( h )when the two extremities of the vector it independently
describe the domains V and v,, respectively.) Both kriging systems can be expressed in matrix form as
The estimation variance can, thus, be expressed as a quadratic form in [Kl . [A 1 = [M21 07.4)
A,, Ag, and is to be minimized subject to the non-bias condition La A, = 1.
The optimal weights are obtained from standard Lagrangian techniques by from which, for system (V.1),
setting each of the n partial derivatives a[E{[Zv-Z~]2]-2p C pAp]Pah, [A] = [K]-' . [M2],
to zero. This procedure provides a system of (n + 1) linear equations in
(n i- I) unknowns (the n weights A, and the Lagrange parameter p ) which and, [A]' being the transposed matrix of [A],
is called the "kriging system ". (T: = E(v, V)- [A]'. [M2],
where the matrix of unknowns [A] and the second member [M2] can be
written as two column matrices:
This kriging system can also be expressed? in terms of the semi-vario- The first member [K], o r "kriging matrix" is written as:
gram function y(h), particularly when the RF Z ( x )is intrinsic only and the
covariance function C(h) is not defined:
t In practice, even when t h e covariance is not defined, the system (V.1) written in terms of
covariance is preferred f o r reasons of programming efficiency. For this purpose, it 1s enough
to define the "pseudo-covariance" C ( h )such that y ( h ) = A - C ( h ) , the constant A being
any posi~ivevalue greater than the greatest mean value used in the krlglng system (V.3). Seven important practical remarks
The non-bias c o n d ~ t i o n1, A, = 1 allows rhe elimmation of this constant A from the system
(V.31, which then becomes a (V.l)-type system written in terms o f the pseudo-covariance Remark 1 The existence and uniqueness of the solution. T h e kriging
C@). system (V.1) has a unique solution if and only if the covariance matrix
3W XlISIZiG GEOSTATISTICS V T H E ESTIMATION OF IN SlTU RESOURCES 309
flits. c,)] tl i z . - ~ i > positive definite and, thus, necessarily has a strictly Remark 5 T h e kriging matrix [K] depends only o n the relative
pclijtive -
d~~---:-. -
-J.....L.J. For this purpose, it is enough that t h e point c'ovari- geometries of the data supports u,, v~ and not at all o n the support V of
the domain to be estimated.
"z' m d e ! C : k ) used is positive definite and that no data s q p o r t
CO m i n c i d s cozpletely with adother one, In fact, v, = v a t entails Thus, two identical data configurations would provide the same kriging
that C(C,: ~ I U , . .c g ) , V p . and the determinant ( c ( v , , u p ) / is, thus, matrix [K] and i t is then enough to take the inverse matrix [K]-' only once.
ZFrO. The two solution column matrices [ A ] and [A'] are then obtained by taking
conditix i?: the existence and the uniqueness of the solution o f the the products of the single inverse matrix [KJ-' and the respective second-
k r l ~ n gsysteir, :hc; entails that the kriging variance (V.2) is non-ne@,ative, member matrices.
cf. seaion II.A.2. rslation (11.8). [A] = [ K ] - ' ( ~ 2 1 and [ A ' ] = [K]-' . [M'L'].
Remark 2 i(,:!glng. which is an unbiased estimator, is also an exact In practice. t h ~ samounts to using a program for the resolution of linear
interpolator. i.:. if the support V to be estimated coincides with any 'of the systems uith two o r more seclond members (such as subroutine GELG in
W P o r t s c, o! ~k,tt ~ \ a i l a b l edata, then the kriging system provides: the IBM package).
I f , in addition to identical data configurations, i.e. [K] = [K'],the two
panels to be estimated have t h e same geometry, then [MZ] = [M2'] and the
result~ngkriging w e ~ g h t sare identical: [ A ] = [A']. It is then enough to solve
the kriging system once, which results in a considerable saving of comput-
In cartograph>. it is said that "the kriged surface passes through the ing time. For this reason. when evaluating the same stationary o r quasi-
experimental points". s o t every estimation procedure has this property, stationary phenomenon, it is advisable to ensure, as far as possible, that the
especially procedures using least square polynomials. data configurations are systematic and regular. In this way, a single kriging
plan that can be repeated identically over all or part of the deposit, can be
Remurk 3 The expressions of the systems and the kriging variances using used.
the notations and y are completely general,
Remark 6 T h e kriging system and the kriging variance take into account
(i) w h a l e ~ e rthe supparts v,, up of the data and the support V to be the four essential and intuitive facts, previously mentioned in section
estimated, some data supports may partially overlap, u, n up f: 0, II.B.2, remark 4, u.hich condition every estimation. These a r e as follows.
but for u # 9, it is imperative that v,+ ug - some data supports may
be included in the volume V to be estimated, t., c V ; (i) T h e geometry of the domain V to be estimated, expressed in the
(ii) wharecer [he underlying structure characterized by the model C ( h )or term y ( V , 1') in the expression of the kriging variance cr:.
?'(A), the structure may b e isotropic o r anisotropic, nested o r not. (ii) T h e distances between V and the supports v, of the information,
expressed by the terms .v(v,, V ) of the matrix [M2].
J h m u k 4 T h e kriging system and the kriging variance depend only on (iii) T h e geometry of the data configuration as expressed by the terms
the Structural model C ( h ) o r ? ( h ) and on the relative geometries o f the .i;(u,, u p ) of the kriging matrix [K]. T h e accuracy of an estimation
various supports t.,, u,, V,but not on the particular values o f the data .To. depends not only on the number of data but also on their configura-
Conscqucntly, once the data configuration is known and before drilling is tion in relatron with t h e main features o f the regionalization as
undertaken, the kriging system can be solved and the correspolnding characterized by the structural function y ( h ) in the various terms
minimurn estimation variance can b e forecast. Thus, t h e kriging variance ?(um 0,).
can bc used to balance the cost o f a new drilling campaign with its forecast (IV)T h e main structural features of the variability of the phenomenon
utility (new data would decrease the estimation variance, thus providing under study as characterized by the semi-variogram model y ( h ) .
11:LrIOw~rconfidence intervals) cf. 0 . Bernuy and A. Journel (1977). Consider t h e estimation in two dimensions of the panel V by the sym-
metric configuration of the four data A, B, C, D as shown on Fig. V.1. The
underlying mineralization shows a preferential direction u o f continuity
310 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 311
which appears on the anisotropic semi-variogram y(h,, h,) as a slower if all supports v,, up, V are disjoint, and
variability in the direction u. The kriging system thus gives a greater weight
to the data B and D, although they are at the same distance from V as A
and C.
If all supports v,, vp, V are not disjoint, then each of the terms of the
kriging system must be calculated.
A case often found in practice is that of a configuration of data disjoint
among themselves, some of which (v,.) being located within the support V
to be estimated:
va,c v, v a v n v p= 0 ,
but V a , a',p,
v,nv=O v,~v,=D,
A -
Dlrectton of c o n t l n u ~ t y
= -+
v C1(ua*,V),
FIG. V.1. Influence of the variability struclture on kriging.
and
The mean values t ( v a ,u p ) appearing in the kriging system are written,
according to formulae (III.6), as
if vp c LI,,
C(uu, u p ) = cl(ua, up)+
~ ~ / v ifQ va and up are disjoint,
I
where A is a constant given by A = Co(h)dh. (It is recalled - cf. section In mininlg applications, the support V is very often much larger than the
-
III.A.3, formula (111.7) that this constant A appears in the nugget data supports u,, u,,, and the term A/ V is negligible when compared with
constant value CO,= A / v of a semi-variogram regularized over the the terms A / v a or A / v a . . When all the data supports u,, up are disjoint,
support v,) system (V.5) is obtained.
The kriging system (V.l) can then be written as If the semi-variogram y(h) = {Co(0)-Co(h)}-t y l ( h ) is to be used instead
of the covariance C ( h ) , it is enough to substitute ( - y l ) for CI in the
preceding systems OJ.5) and (V.6).
Note in system (V.5) that the influence of a nugget effect with nugget
constant Co, = A / u affects only the elements of the main diagonal of the
kriging matrix as an additional term A / v a .
312 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 313
The influence of a plire nugget effect panels' V within the neighbourhood a r e equal to the unique mean
value I,, Z J n of the tt data available in this neighbourhood. The
If the macro-structure C , ( h ) disappears, leaving only the pure nugget
effect Co(lz),the kriging system (V.5) becomes
nugget effect is said "to smooth the estimation ".
A,A/v, - p = 0, Va, V.A.2. Taking a c c o u n t of a drift
xAq=l In all the kriging equations pr~esentedheretofore, it is assumed that the RF
4
Z ( x ) is stationary o r at least quasi-stationary over a given estimation
and, thus, neighbourhood.
In practice, it may be known that a drift (non-stationary expectation
E { Z ( x ) ) )exists within certain zones of the deposit and that there is not
enough data available in these zones to allow quasi-stationary estimations
(by considerably reducing thle neighbourhoods of estimation). It then
becomes necessary to take the drift into account when estimating these
when all the supports c,, c4. 1' are disjoint, and zones by kriging. For example, consider the mineralized seam on Fig. V.2,
which thins out towards one end. W e want to estimate the mean thickness
(or thc o r e tonnage) of a panel V located on the fringe of the seam, by
drill-holes all located further along the seam.
T h e rl weights A, are all equal and the kriging variance c r i = Co,/n is Available
simply the variance of estimation given by the statistics of spatially uncor- d r i l l - holes
related variables. T h e nugget constant Cot is the a priori variance of the FIG. V.2. Drift of a seam thickness.
support v data, according to formula (111.1 1):
C,, = A/v = D ~ ( V / W=) Var {Z,,(x)}. Kriging under a stationary hypothesis, like any other "stationary" esti-
mation procedure such as inverse-squared distances weighting, will inevit-
Thus, for a pure nugget effect, we have the following. ably overestimate the real mean thickness of the panel V,For this reason, it
(i) All data having the same support v (disjoint from V) have the same is essential to take the drift of the thickness into account in the direction of
weight whatever their distances from the volume V to be estimated. thinning of the seam.
T h e close data no longer screen the influence of the more distant T h e following method of "unbiased kriging of order k" (also called
data a n d the nugget effect is said "to remove the screening effect". "universal kriging") provides an unbiased linear estimator which takes the
(ii) Within a given neighbourhood of quasi-stationarity, it is n o longer drift into account, provided that both the form of the drift E{Z(x))=m ( x )
possible to make different local estimations: all estimators Z z of all and the covariance o r variogram of the non-stationary RF Z ( x )are known.
314 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 315
1. The form of the drift neighbourhood D(xo) in the form of a polynomial of the first or second
Let Z(x)be the non-stationary point RF in the studied zone. By definition, degree (the first terms of the Taylor expansion of the function m ( x ) , i.e.,
the drift is the non-stationary expectation of this R:F: m(x)=at+a2x (V.lOa)
E { Z ( x ) }= m ( x ) . for a so-called "linear" drift,
The non-stationary RF Z ( x ) can then be expressed as the sum of a drift
term and a residual term Y ( x )= Z ( x ) - m ( x ) .This residual term Y ( x )may
or may not be stationary but has a zero expectation: for a so-called "quadratic" drift.
It should be recalled that x represents a point in the working space. If
Z ( x > = m ( x ) + Y ( x ) with E { Y ( x ) ) = = O , Vx. P.8) this space is one-dimensional (e.g., time with coordinate t ) , then a quadra-
The drift m ( x ) expresses the regular and continuous variation of the RF tic drift would take the form
Z(x) at the scale of observation of the experimental data. The RF Y ( x )
accounts for the erratic fluctuations around the drift, cf. Fig. V.3.
If the working space is two-dimensional (e.g., horizontal regionalization of
vertical thickness with the two rectangular coordinates u, v ) , then a
quadratic drift would take the form
Of course, this drift may only be evident in some directions. For exam-
ple, the phenomenon may be stationary in the horizontal plane with a drift
in the vertical direction, in which case the previous quadratic drift would
reduce to
to estimate y ( x , y ) directly from the initial experimental data z(x,) This D ( x o ) is a neighbourhood centred on xo and includes t h e panel V and
would entail the simultaneous estimation of the drift m ( x ) and the semi- the supports of all the data used in the estimation. Within D(x& the
variogram y(x, ),) from a single realization z ( x ) of the RF Z ( x ) , which is covariance C(1r) o r the semi-variogram y ( h ) are known.
not rigorously possible.+ A linear combination of Ihe n data is then formed to provide the
estimator:
In mining applications, experience has shown that, in most instances, it is n
possible to take either of the following approximations for y(x, y): Z: = c AaZu.
a=l
(i) the quasi-stationary semi-variogram y ( h ) calculated o n neighbour-
ing zones with a possible correction for a proportional effect; Non-bias conditiorl T h e non-bias condition can be written as
(ii) more simply, the behaviour at the origin (in general, linear, with or
without nugget effect) of the neighbouring quasi-stationary semi-
variogram, y(x, y ) = y ( h ) - wlhl, where h = ( x - y ) is the kector
joining the two points x and y. where
parameters P I ) . This system is called the "unbiased kriging system of order imply that cl = 0 , V l = 1 to k. For example, consider a data set in which all
k" or the "universal kriging system". the data are aligned perpendicularly to the direction u, as on Fig. V.4. It is
obviously impossible to take a drift in the direction u into account, using '
just this set of data, even if this drift is only linear: rn ( u )= al + a2u.
Remark 1 T h e kriging system (V.l) in the stationary case can be seen as Remark .3 Matrix form.The first member matrix [ K k ] of the system
an unbiased kriging of order k = 1, the stationary expectation being (V.12) of unbiased kriging of order k can be obtained by adding (k - 1)
reduced to an unknown constant a , : rows and (k - 1 ) columns to the first member matrix [K] of the system (V.1)
~g under the stationary hypothesis, i.e.,
r n ( x ) = a ~ f l ( x ) =a1 with f l ( x ) =1, Vx.
Note that the mean value of the constant Eunctron f l ( x ) = 1 over any
volume v is equal to 1: b t = 1 , Vv.
Remark 2 Existence and uniqueness of the solution. As was the case under
the stationary hypothesis, cf. section V.A.l, remark 1, the covariance 1 ...
matrix [C(v,, ua)] is assumed to be strictly positive definite. It can then be ... bt"
shown that th'e system (V.12) of unbiased kriging of order k has a unique
solution if and only if the k functions h ( x ) are linearly independent on the
set of the n data, i.e., if the n following linear relations ... b '0, ,O .:. 0-,
k (n -t- 1) columns ' (k - 1) column
C cb;.
I- 1
= 0, V a = 1 to n , (V. 15 ) Note that the matrix [Kk]is still symmetric.
320 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 321
The column matrices of the unknown [A] and of the second members ZD= ( 1 1 0 ) J,, Z ( X )dx directly over the entire zone D. A linear combina-
[M2k]are obtained likewise: tion of the available data can be formed to provide the estimator 25 =
C, A,ZQ and the standard kriging systems, (V.12) or (V.1), to be con-
sidered can be used by replacling the index V by D (depending on whether
or not a drift is to be considered).
In mining applications, the global mean of an entire deposit D would
very rarely be kriged directly for the two following reasons.
(i) It is not usually possible to assume stationarity or a single drift of
known form over the entire deposit D , but only over limited neigh-
bourhoods (local quasi-stationarity).
(ii) Even if stationarity could be verified over the entire deposit D , there
are usually too many data in D to construct a kriging matrix and then
The system (V.12) for unbiased kriging of order k can then be expressed to solve the kriging system. Moreover, the construction of such a
in matrix notation as kriging matrix would imply that the structural function, C ( h ) or
y(h), is known for distances Ih/ of the order of the dimensions of the
deposit D and, as was seen in section III.B.7, rule (III.42), the limit
of reliability of an experimental semi-variogram is a distance L/2,
half the dimensions of the field D.
Thus, to estimate the global mean value? over D, it is usually advisable
to combine the estimates of the component volumes of D. If these local
[A]' being the row-matrix transpose of the column matrix [A]. estimations are done by kriging, then the term "combination of kriged
estimates" is used.
V.A.3. Combining kriged estimates
Very often, in practice, we will be required to combine the kriged estimates Theorem of combining kriged estimates
of local units to form a "global" estimator of a zone or even of the entire
deposit. Assuming an hypothesis of stationarity for the entire zone, the Consider again the point RF Z ( x ) , which is assumed to be either second-
theorem for combining kriged estimates shows that the direct kriging OI[ the order stationary over the extent of the zone or deposit D, or to have a drift
required global mean yields the same results as combining the local kriged with a known form over the extension D.
estimates. In practice, the second procedure is preferred for its greater The kriging systems (V.1) and (V.12) have the remarkable property that
flexibility, and because it does not require the preceding stationarity over only the second-member matrices [M2] or [M2k]depend on the support V
the entire zone. to be estimated. The solutioln of the kriging system, [ A ] = [K]-' [M2], is .
thus linear in V.
More precisely, let AQ(x)be the solution of the kriging of the point value
Kriging rhe global mean Z ( x ) at point x from the data configuration {Z,, a = 1 to n). The mean
Let Z ( x ) be a point RF supposed to be either stationary of second order
over the entire zone or deposit D, or to have a drift of known form on this
extension D. The covariance or the variogram of Z ( x )is also supposed to t Remark The global mean ZD = i:l/D) ID Z ( x ) dx, which has a precise physical meaning,
be known. should not be confused with the stationary mathematical expectation m = E { Z ( x ) ]of the
probabilistic model. ZD is an estimator of the expectation m only when the dimensions of
Instead of estimating the mean value Zv, of each of the panels If, wlhich the deposit D are very large with Iespect to the actual range of the stationary variogram or
make up the zone D , it is possible to estimate the global mean value covariance representing the model (ergodic property).
322 MINING GEOSTATIST1C.S V THE ESTIMATION OF IN SITU RESOURCES 323
value ZD = ( l / D )JD Z ( x )dx on the support D is then kriged from the inside the polygonal contour which best approximate$ the form of the zone
same configuration and the solution A,(D) is written as to be estimated, cf. Fig. V.5.
.,
Variance elf the global esrimaror
The kriging of the global mean on D can be expressed as a linear Kriging variances cannot be combined as simply as the kriging estimators.
combination of the point krigings. Let Z$ = (1/D) 1, V Z $ , be the combination of the N kriged values Z'b,
This relation also applies to the Lagrange parameters p, which are and let CT;.V,= E{[Zvl-Z$,ll?) be the N elementary kriging variances of
simply the (n +l)th terms of the solution matrix [A]: the N units V,.The global estimation variance of ZD by the combination
22 can be written as
possible if the semi-variogram y ( h ) is known for all distances lh] between T h e estim'ator Z*v,, is a linear combination of all t h e available data values
the data in D. of all the K variables in coregionalization:
I n mining applications, such a calculation is very awkward and, in addi- "k
tion, the semi-variogram y ( h ) is usually not defined f o r t h e large distances a,= C x
K
~ k = l harZak.
k = o (V.20)
lhl approaching the dimensions of the deposit D. T h e solution is then to
calculate the global estimation variance (T&, with the aid of approximation T h e expectation of the (error involved is
principles which \vill be given in section V.C. 1.
V.A.4. Cokriging
Sometimes, one variable may not have been sampled sufficiently to provide
T h e non-bias, E{Zv,, -i!$,,}= 0, is then expressed by the K conditionst
estimates of acceptable precision. T h e precision of this estimatior~may
then be improved by considering the spatial correlations between this 1A 1 and CA,,=O, Vk#ko.
ak
variable and other better-sampled variables. For example, the silver grade Oko
of a panel can be estimated by lead and zinc grades in addition to those of Note that the condition Cake
A,,, = 1 cannot hold t r u e if nb= 0. In other
silver. words, cokriging require:^ a t least one experimental value of the main
From a theoretical point of view, cokriging is no different from kriging; variable Zk,.
difficulties arise essentially from notations, since an extra index k is T h e minimization of the variance a?v,,= E{[Zvk, - ~ * v , , ] ~ }under the K
required to differentiate between the variables. non-bias constraints results in a system of (Ck nk - t K )linear equations (the
Consider a regionalization characterized by a set o f K spatially intercor- unknown being the Cknk weights A,, and t h e K Lagrange parameters p k )
related random functions { Z k ( x ) , k = 1 'to K}. T h e first- and second-order called the "cokriging system": -,
moments of these RF, assuming stationarity, are
E { Z k z ( x+ h ) . Z k ( x ) } - m k . . mk = C k , k ( h ) , cross-covariance;
"r
cf. section II.A.4. .
T h e estimation of the mean value of the point RF Z k o ( x )over the
support Vko(x,) is required, i.e.,
1 Aok
Pr=1
=O, V k +ko.
iIt may happen that one of the secondary variables Z t t , with k' # ko, has the same meaning
T h e available data {Z,,, a k = 1 to nk, for each k = 1 to K} are defined on and the same expectation ( r n k . - mk,) as Zro,e.g., ZI, and Zkf are two lead grades, but one is
obtamed from core analyses and the other one from cutting analyses. T h e two variables Zb
t h e supports {t',,}, e.g., and Z,, may have the same expectation, but different structures of spatial variability
(generally, their nugget efiect would be d~fierentdue to different measurement errors). The
K prevlous non-bias conditions are then reduced to
~ ~ , , ~ + L h , , =and l xA,,=O, Vk#ko,kl.
-1 ' 0 k
326 MINING GEOSTATISTICS V T H E ESTIMATION OF IN SITU RESOURCES 327
(the notation ck,k(vpk.,v m k )is the standard representation of the mean V.A.5. Ca'se studies
value of the cross-covariance Ck,k(h) when one extremity of the vector h
describes the support upk.and the other extremity describes independently The following case studies are simple enough to be carried out by hand,
the support u,,). without the use of a computer. They are intended as a comparison of the
kriging method and the more usual weighting methods, and to demonstrate
the influence of the structural function, y ( h ) or C ( h ) .
Remark 1 Unlike the kriging system, the cokriging system can be written
simply in terms of the cross-semi-variogran~ yk,,k(h) only if C k ' k ( h ) =
C k ~ ( h 3 which
, k = C k ' k (0) -- Ck-k( h ) , i.e., if the cross-
entails that y ~ (11) Case Study 1. Estimation on a square
covariances are symmetric in (h, - h ) , cf. section II.A.4. The cokriging
Consider a two-dimensional regionalization characterized by a point R F
system written in terms of cross-semi-variograms is then simply derived Z(u, v ) which, as a first approximation, is supposed to be intrinsic with an
from system (V.21) by replacing Ck,kby (-7k.k). isotropic stationary semi-variogram: y ( h ) = y ( r ) with r = Ihl, For example,
Z ( u , v ) may be the vertical thickness of a sedimentary seam.
Remork 2 The existerrce and uniqueness of the solution. It can be shown The mean value Zv is to be estimated over a square panel ABCD of side
that the cokriging system (V.21) provides a unique solution if the covari- 1 from the non-symmetric configuration of the four data of support v,
ance matrix [ck,k(upk.,v,,)] is strictly positive definite, for which it is located at ,SI (central sample) and S3, 04,Q 5(peripheral samples) as shown
enough that a positive point coregionalization model [ C k s k ( h )be ] adopted, on Fig. V.6.
cf. section I I I . B . 3 , and that no data value is totally redundant with respect
to another, i.e., that vak+ vpk.,V a k # Pk,.
However, as explained above, the nonvbias condition is satisfied only if
the set of data values of the main variable Zk0is not empty, i.e,, it is
necessary that nk, Z 0.
FIG. V.7. Weights used for the estimation of the square panel.
330 MINING GEOSTATISTIClj V THE ESTIMATION O F IN SITU RESOURCES 33 1
Remark 1 In the case of a pure nugget effect, the conditions of formula not give too much weight to data preferentially placed with respect
(V.7) are met: CT; = Co.14 = 0.25 and the weights are proportional to the to the panel to be estimated. O n the other hand, the smaller the
supports, A 1 = A S =Az/2. There is no advantage linked to the privileged nugget effect, the better are procedures (such as ID2) which favour
location of the central datum S1. well-located data (by screening the influence of furthest data).
(iv) In all cases, an unbiased weighted average-type estimator such as ID
Remark 2 As the nugget effect decreases and, thus, the spatial cor- or I D 2 fares much better than the POLY estimator. Had the
relation increases, the influence of the central information becomes pre- configuration of data been more clustered, e.g, if data 04 was located
ponderant (Al increases from 0.25 to 0 6 3 ) . Note that for a # 0 the weight nearby data Os, the advantage of kriging over ID and ID2 would
of the datum S2 (of support 2 u ) is always greater than twice the weight of have been much more decisive.
Sf (of support v ) ; this is due to the fact that S2 is nearer to the right-hand
side of panel V, which is less sampled. Nan-stalionary case Let us consider now a non-stationary regionalization
characterized by a RF Z(u, v ) with an anisotropic h e a r drift which is
Other estimation procedures The standard weighting methods (Poly, ID,
evident only in the direction o f the coordinate u, i.e.,
ID2) do not take into account any of the characteristics peculiar to the
regionalization under study and, consequently, provide the same estimator
whatever the structural function y(h) may be, cf. Table V.I.
The polygon of influence method gives all the: weight to the central
datum:Al=l andA2=A3=0. For example, Z ( u , v ) may be the vertical thickness of a subhorizontal
Both the inverse and inverse-squared distance weighting methods (ID sedirnen~.aryseam which thins out towards one end in the direction u,
and ID2) give the same weight to the three peripheral data: A 3 = A 2 / 2 . The or i t may be the depth of the continental shelf, the direction u being
central datum S1 is given greater weight by ID2 ( A l = 0,727) than by ID perpendicular to the coastline, cf. Fig. V.8.
( A , = 0.484). The mean distance from S1 to the panel has been taken as a
quarter of the diagonal, i.e., lJ2/4.
In the stationary case, these standard methods ensure the unbiasedness
of the estimation, as they satisfy the unique non-bias condition A, = 1,
but they do not, by themselves, provide the estimation variances cr;. For
this purpose, it is necessary to characterize in some form the spatial
variability of the phenomenon under study. The geostatistical method is to
use the structural function y(h) with which the estimation variance of any /" estimated
unbiased linear estimator can be calculated from formula (11.28).
The following remarks can be made on the results o f Table V.I.
FIG. V.8. Linear drift in the direction u.
(i) In all cases, kriging provides the best unbiased linear estimator
(BLUE).
(ii) Depending on the nugget effect (i.e., the degree of spatial cor-
relation), one or the other of the three standard estimators is close As previously, the mean value Zv is to be estimated over the square
enough to the kriging optimum, but only a structural analysis, i.e., a panel V from the non-symmetric configuration of the four point data
geostatistical approach, can tell which one is the closest. located at S1, 04,05,S3,cf. Fig. VL6.
(iii) Generally, the greater the nugget effect (thus, the greater the screen For reasons of symmetry, and sin:e the drift does not occur in the
removing effect), the better are procedures (such as ID) which do direction u, the two data O4 and o5will receive the same weight and they
332 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 333
can thus be grouped into the set S2= { 0 4U 0 5 } o f support 2v. A n d , as Review of other weighting tncthods In the absence of a structural analysis
previously, a linear estimator with three weights can be defined: providing evidence of the linear directional drift m(u), the three usual
weighting methods (Poly, 113, I D 2 ) would provide the same weights as in
the stationary case, i.e., the weights given in Table V.l and reproduced in
Table V.2. Consequently, t h e I D and I D 2 estimators are biased, since they
d o not fulfil the second non-bias condition (A3=0). A s for the Poly-
T h e drift m ( u ) = a l + a 2 u = alfl(u)+a2f2(u) consists of the two known
estimator, it remains unbiasl-d, since it gives a zero weight to the datum S3
functions, f l ( u ) = 1, Vu, and f 2 ( u ) = u. T h e mean values of these functions
(but also - and wrongly - to the information S2).
over the supports S, and V can thus b e written as
If, as in the case of t h e unbiased kriging of order 2, the form of the drift
m ( u ) = a , + a 2 u is s ~ , p p o s e dknown, this drift can then be estimated by
bk- = b: = 1, b:a = u, (coordinate u, of the support S,), least squzres (abbreviated to LS). This estimated drift obviously passes
through the ~ a l u eZ ( S 3 ) and also through the mean of the three values
Z ( 0 1 ) , Z ( S 1 ) and Z ( 0 5 ) , i.e.,
Remark 1 When there is a pure nugget effect (total absence of spatial Remark 3 In the presence of the directional drift m(u), the ID and ID2
correlation of the true residuals Z ( x ) - m(x)) the re:sults obtained from the
estimators are biased. This bias, E { Z v - Z $ } , depends on the unknown
least squares method are the same as the optimal results given by kriging. It
drift m(u) and is itself unknown, The existence of this unknown bias thus
can be shown that this property remains true whatever data configuration is
prohibits the definition and calculation of the estimation variance CT: =
used (cf. G.Matheron, 1971, p. 162).
Var {ZV- 2 6 ) .
TABLEV.2. Estimation of a square panel in he presence of a drift m ( u ) Addendum. The tensorial invariance property of the kriging systems
-
- (This subsection may be omitted in a first reading.)
Nugget
effect Kriging LS Poly ID Obviously, neither the kriging system (V.24) nor the corresponding kriging vari-
ID2
- ance u: should depend on the arbitrary c h o ~ c eof the origin of the coordinates, i.e.,
A1=1/3 113 of the para~meterc. In particular, this system must be equivalent to the following
A l = 1 0.484 0.727
A 2 = 213 A z = 213
system (V.25) corresponding to the choice c = 0:
0 0.344 0.182
Pure A3=0 A3=0 0 0.172 0.091
d = 113 , vf = 1 / 3 1 o: non-defined
A = 0.52
A 2 = 0.48 idem idem
Partial A,= 0 idem idem
V' = 0.34 a: = 0.39 0.'734
A 1 = 0.66 In fact, w t h the change of variables p', = ~ ~ + p p~i c= p, 2 , the two systems
A 2 = 0.34 . idem idem (V.24) and (V.25) are equivalent. Thus, X b =A,, VP, since the solution of the
Absent A,= 0 idem kriging system is unique.
idem This invariance property of the kriging system with respect to changing the origin
a; = 0.25 v: = 0.45 0.468
- of coordinates is related to the polynomial character of the component functions
Unbiased Unbiased Unbiased Biased Biased f , ( x ) of the drift m ( x ) and is the reason why only functions of t h ~ stype are used, in
- practice, to characterize the form of the drlft.
The equation corresponding to a = 3 can be eliminated from system (V.24) as it
only serves to define the parameter p 2 , which has no purpose when c = 0, The
system (V.:!4) then reduces to (V.26) which, in the stationary case, is the same as
kriging panel V by the two data S , and S 2 :
and
FIG.V.9. Weights used for estimation of the square panel. Case Study 2. Cokriging on a square
Consider the coregionalization of the two stationary RF's Zl(x)and & ( x )
Remark 2 As the nugget effect becomes smaller, the LS method deviates defined on a horizontal plane and characteri~edby the matrix of isotropic
more and more from the optimum provided by kriging. In the case of cross-semi-variograms [?kk'(r)], k, k' = 1 to 2,
absence of nugget effect, the deviation amounts to 180% in terms of The mean value 2, of the main variable Z l ( x )is to be estimated over a
estimation variance, 0.45/0.25 = 1.8. square panel V of dimensions 1 x I, as shown on Fig. V.lO. The available
MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 337
data consist of the central point value Z1(So)of the main variable and the T h e two non-bias conditions have already been taken into account by the
five point values Zz(So), Zz(A), Zz(B), Zz(C), Z2(D) of the secondary form of the estimator (V.28)itself, and the cokriging system (V.21) reduces
variable located at the centre and the four corners of the square This to the following three linear equations in three unknowns ( A and the two
corresponds to the case of a variable Z 1 , undersampled with respizct to Lagrange parameters, ~1 arid p2):
another variable (Zz),and the problem is to evaluate the gain in precision
provided by cokriging ( Z l v is estimated from the two data sets on Z 1 and
Z 2 )over kriging (Zl is estimated from the single data set on Z 1 ) .
and
Kriging amounts to estimating the panel V from its central value Z1(So).
T h e kriging variance is, thus, given by the general formula (11.27):
uk = E { [ z i v-Zi(so)12}= 2 r i l ( s o 3 v ) - ? ~ i ( V ,v ) - T ~ i ( S oSo)
,
(V.27)
= 2H11(1/2,[/2)-Fii(&0, By simplifying, the preceding system gives t h e following solution:
where H l l and Fll are the auxiliary functions corresponding to the point
model y l l ( r )which characterizes the regionalization of Z l ( x ) .
F o r cokriging, reasons of symmetry (the coregionalization matrix being
isotropic) allow the four peripheral samples in Z 2 to b e grouped together
into o n e single set SR with a grade of u: being given by formula (V.27).T h e gain in precision of cokriging over
kriging is thus equal to
By putting A = A 3 , A being the weight of &(SR), the cokriging estima.tor is Numerical application 1 Consider first t h e case in which all the semi-
t o b e written as variograms consist of a linear isotropic model with nugget effect:
Z:v = Zi(So)-A&(So)+ ~ & ( S R ) . ('V.28) ykkr(r) , > E (very small range)
= C k k ' + ~ ~ k k ' r'dr
338 MINING CEOSTATISTICS V THE ESTIMATION O F I N SITU RESOURCES 339
The auxiliary functions H and F of the linear model are given by Numerical application 2 Consider now the silver-lead coregionalization
formulae (11.91) and, for 1 >> E , give model adopted for the Broken Hill deposit (Australia) in section I-V.C,
Case Study 8. The silver grade is considered here as the main variable
and Zl(x), which is undersampled with respect to the lead grade Zz(x),this
being often the case in silver-lead deposits.
For calculation purposes, yll(r) will be fixed so that a$ = 1 and the gain
TABLEV.3. Injluence of the cross-structure on cokriging
g =AN will thus be relative: g = g, = AN/U;. For example, the parameters
Cll= 0.5 and a l l 1 = 2.05 will be taken, which entails u i = 1, according to CIZ ~ 1 2 1 gr
formula (V.27).
Similarly, yz2(r) will be fixed so that the variance D is small, which means 0 0 Increments of Zland Z2independent
that the gain, g = N ~ / D , will be large. For example, the model yz2(r)= 0.38 0.05
0.84 0.25
j y l l ( r ) will be taken with the parameters Cz2= O * l and a 2 2 1 =O.4l. 0.28 0.03 y , , is a pure nugget efiect
According to formula (V.29), D thus becomes 0.66 0.15 Intrinsic coregionalization
1.12 0.45
0.63 0.14 y 1 2is a pure nugget effect
1.01 0.36
1.47 0.77 y,, has its maximum value
With the two direct fixed models y l l and y 2 ~ the, cross-model y12(r) will
be made to vary, taking care that the coregiona1i:ration matrix [ - ~ k k ' ( r ) ]
remains conditionally positive. For this purpose, it is enough that the two The adopted semi-variograms y k k (r) (with a quasi-point support) consist
parameter matrices [Ckk']and [ w k k . ] be positive definite, cf. section III.B.3, of an isotropic spherical model y l with nugget effect and range a = 6 0 ft, cf.
conditions (111.26). It is thus enough that the two following determinants Table (IV.l):
be positive or zero: for silver grades, y l l ( r ) = 1.1 + 108yl(r)(oz Agt-')';
for lead grades, yz2(r)= 11+ 39yl(r)(O/0 ~ b ) ~ ;
for silver-lead y12(r)= 5 yl(r)(oz Agt-I)(% Pb).
The auxiliary functions H1and F1for the spherical model yl are given by
charts no. 3 and 4 in section II.E.4. The following results for 1 = 100 ft. are
Table V.3 gives nine pairs (C12, m121)which satisfy these conditions, and thus obtained: for the kriging variance,
gives for each one of them the value A of the weight assigned to the
secondary sample Z ~ ( S R ) as
, well as the relative gain g, provided by c+; = 2.38(oz Ag t-'12;
cokriging. for cokriging, according to formulae (V.29),
Remark 1 When the increments [Z1(x + h)-Zl(x)] and [Z2(&+r)- N = 4.l25(oz A g t-I)(% Pb); D = 62.5(% ~ b ) ~ ;
Z2(x)J of the two variables are independent, the: cross-semi-variogram hence, A == N / D = 0.066(oz. Ag t-')(% ~ b ) - ' , which corresponds to a rela-
y12(r) is zero, and the secondary data on ZZ(x) adds nothing to the tive gain g, = AN/*: = 0.114= 11%
estimation of the main variable Zl(x). Thus, although the silver grade is quite undersampled with respect to the
lead grade, the gain of cokriging oyer kriging is negligible for the
Remark 2 When the cross-variability (i.e., the parameters CI2and wi2)
configuration studied here. This is due to both the fact that the only
increases, the gain of cokriging also increases. The maximum 77% gain is available data for silver, Z1(So) is well located, and that the cross-vari-
obtained with the maximum cross-variability. ability yl:! is not high.
MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 341
T o sum up this case study o n cokriging, the following advice can b e and
given: before undertaking a coregionalization study, first proceed to the
calculation of the possible gain of the planned cokriging, considering a
simple data configuration.
It should be recalled, from section V.A.4,remark 2, that the set S, of the
Case Study 3. Cokriging with systematic errors unbiased data Z l ( x o , ) must not b e empty. This restriction is quite natural,
since the unknown bias o f the data set Sz cannot be corrected unless there
In a deposit mined by open-pit, two types of data a r e available o n t h e grade are actually some unbiased data in S 1 . Under the condition S1# 0, the
of interest. influence of the unknown bias o f the data set S2is made nil by the relation
(i) A n unbiased data set derived from drill cores. T h e analyses of the
cores are considered as the true grades Z l ( x ) of these cores.
(ii) A biased (e.g., because of selective loss of fines being blown away)
and imprecise data set derived from blast-hole cuttings. T h e analyses and there is no need to evaluate this bias E{F(x)) when cokriging.
of these cuttings are considered as realizations o f the RF Z2(x), However, when this bias is known, i.e., when an unbiased cutting data
which differs from the true grade Z l ( x ) by a random error E ( x ) , i.e., set Sz can be defined, the two non-bias conditions of the preceding system
(V.30) reduce to the unique standard non-bias condition:
Z~(X)=~~(X)+E(X).
In general, the poor quality of the cuttings data set is balanced by the
large number of blast-holes on any mining bench.
T o estimate the true mean grade Z l v of a block with support V, t h e
But there is still a need for cokriging, since the two spatial structures y l ,
following estimator is used:
and y22 of the two typ:s of data m a p be different.
ZTv = C A a l Z i ( x ~ , > +C Aa2Z2(x&
OISSI azeS2
Numerical application T h e three RF's Z l ( x ) , Z2(x), E(X) have the
S1 and S2 being the two sets of core data and cuttings data, respectively. following characteristics.
T h e supports of the core lengths and cuttings are negligible with respect to First, they are all stationary with unknown expectations:
the support V of the block.
A s the matrix of coregionalization of the two variables Z1 and 22, is
[-~~.'(h), k , k t = 1, 21, the cokriging system (V.21) is written in terms of
semi-variograms:
equal to the unknown bias between the two data sets.
Second, the structure of t h e true grades Z l ( x ) is characterized by the
quasi-point model y l l ( h ) derived from drill-core data.
Third, as a first approximation, the error ~ ( x can ) b e assumed to be
spatially uncorrelated (the loss of fines not being regionalized). Thus, the
semi-variogram o f the error r: (x) reduces to a pure nugget effect:
y 2 ( h ) = i ~ { [ ~+
( xh)-,e(x)12} = Var { ~ ( x )=} C,, Qlhl> ao,
a. being a very small range compared to the distance Ihl of the experimen-
tal observation.
Fourth, the error ~ ( x is) sulpposed to be independent of the true grade
Z l ( x ) (this is not always true, e.g., in banded hematite quartzite deposits,
342 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 343
the quantity of fines is inversely proportional to the iron grade ZI). Thus, In spite of their bias, taking the cuttings data into account provides a
relative gain of more than 50% in terms of estimation variance. -e
~~~(h)~~~Eil~~(x+h)-~~(x)l~~alx+h)-~;!(x)l~=~l~(h) > ,
as
Zz(xf ~ ) - Z ~ ( X ) = [ Z ~f (hX) - Z ~ ( X ) ] + [ E ( X+ h )- E ( x ) ] ,
and the increments of E(X) are independent of the increments of ZI. V.B. THE APPLICATION OF KRlGlNG
Similarly,
V.B.1. A comparison of kriging with other weighting m e t h o d s
Y~~(~)=~E{[z~(x+~)-z~(x)I~)=YI~(~)+Y~(~)
If we are limited to the class of linear estimators (the estimator 2" =
=yl~(h)+C,. C:=l h,Z, is a linear combination of the n data values) and if the structural
Thus, the following regionalization model is obtained: analysis is sufficient to infer correctly the underlying structural function
(covariance or variogram), then, by definition, kriging provides the best
unbiased estimator (in the sense of minimum estimation variance). This
Indeed, in practice, the .two direct semi-variograms y l l and y 2 ~are property is irrefutable and is not shared by the more classical linear
calculated from the corresponding experimental data and they usually estimation methods such as polygons of influence (Poly), inverse distances
differ only by the nugget effect: (ID), inverse:-square distances (ID2) and least-square polynomials (LS), to
cite only the methods more commonly used in mining applications.
The two criticisms most often levelled at kriging are:
The cokriging configuration of Case Study 2, above, cf. Fig. V.10, will be (i) the gain in terms of estimation variance provided by kriging is
taken as an example, i.e., the estimation of the mean grade Zlv over a negligible in practice when compared with more common estimation
square block V o f side 1 = 20 m by the two data sets .Z,(So) with weight 1 procedures;
for the central core analysis, Z@O)with weight -A and Z2(SR)with weight (ii) the application of kriging, as well as geostatistics in general, is
+A for the cutting analyses at the centre Soand the corners SRof the block. difficult and costly in terms of both sampling and computer time.
The quasi-point semi-variogram yll(h) is the model fitted to the
experimental semi-variogram of the iron grades of core length at Tazadit The following responses can be made to these criticisms.
For the direct estimation of Zlv by the only central core grade Zl(So),
First, as kriging is optimal, small deviations in the choice of weights
around their optimal values will only result in a second-order increase in
the estimation variance. This means that if, for any given data configura-
the variance rrk, given by (V.27), is tion, the weights provided by any of the usual estimation methods differ
little from th~ekriging weights, then the gain due to kriging can be neglec-
ted, and the more usual method will be preferred for this particular data
Cokn'ging, Le., taking the cutting data into account., gives the following configuration because of the simplicity of its application. However, it still
results from (V.29): must be known which method most closely approximates kriging, and this
can only be determined by a structural analysis to calculate the variogram
N = 0.243C = 32.3 and D = 1.25 i-0,451 C. to be used in the kriging solution. When a strong screen effect is present,
\By taking the variance of the measurement errors C. equal to the first
exphimental value y11(3), i.e., C,= 15, D becomes 78t7, and the weight
Poly will usually be adopted, whereas I D will usually be better when the
screen effect is weak. A modified version of ID2 may prove better when
assigned to the corner cuttings data &(SR) is A = N / D = 0.41. The gain of anisotropies are present, and LS should usually be preferred in cases of
cokriging over kriging is thus equal to non-stationarity, cf. Case Study 1, section V.A.5. Before a structural
analysis is made, it can nqver be determined which of these methods could,
in practice, replace kriging; i t depends not only on the variogram (i.e., on
Le., a fairly good relative gain: g, = g / ~ & = 0.54. the structure of variability), but also on the data configuration. Since, in
V THE ESTIMATION OF IN SITU RESOURCES
344 MINING GEOSTATISTICS 345
The various steps in the practical application of the kriging of each pane]
mining applications, the configuration of the data used to estimate a panel
V are as follows.
may vary considerably throughout the deposit, kriging remains the surest
solution (and not necessarily the costliest). (i) Select the data Z, to be used in the estimation of Z,,
Second, the understanding of geostatistics and the application of a struc- (ii) Calculate the mean covariances C(v,, UP)between all the selected\
tural analysis followed by kriging is, without doubt, more difficult than the data (these are the elements of the first member matrix [K] of the k.
arbitrary choice of one of the usual weighting methods. kriging system). \.
Is geostatistics really costly in terms of sampling? Geostatistics quantifies (iii) Calculate the mean covariances c(u,, V) between each of the
the quality of the available data by means of the kriging variancie, and this selected data and the panel V to be estimated (these are the ele-
allows cases in which the amount of data is dangerously insufficient to be ments of the second member matrix [M2]).
identified. T o reject the determination of variability structure by using a (iv) Choose an algorithm for the solution of the linear system.
method which does not have an associated estimation variancei amounts
T o carry out a 'crude" kriging without any consideration for computer
to ignoring completely the problems involved in estimation.
cost would consist of taking all available data without distinction within a
Geostatistical studies can also result in the conclusion that there are too
neighbourhood of quasi-stationarity centred on each panel V, calculating
many data for a given objective. Used in time, geostatistics can result in
exactly the various nican values C and then applying a precise algorithm
considerable savings in sampling, cf. A. Journel (19736). which would give an exact solution of the linear system (such as SIMQ-SSP
Is the computer cost of kriging really prohibitive? Even though
IBM). Such an application would yield very precise results, but would also be
computer costs are always negligible when compared to the cost of obtam-
very costly in terms of computer time.
ing a few core samples, the answer to this question is yes, if no precaution is
taken when writing and solving the kriging system. Now, if all1 the sym- T o establish a practical alternative, each one of the preceding steps must
metries and approximations are taken into account by establishing a "kri- be critically examined to provide an approximate procedure which will
ging plan", then both the dimension of the kriging system and its solving minimize computer time while yielding results that are within acceptable
cost can be reduced. limits of approximation.
The theoretical optimum of kriging is of little practical use if the appli- The five essential points to be used in reducing the kriging cost are:
cation of kriging is not competitive in terms of computer cost.
1. the reduction of the dimension of the kriging system;
2. the reduction of the number of kriging systems;
V.B.2. Elaborating a kriging plan 3. the rapid calculation of the various mean values C (or 7);
4. the preparation of a data file adapted to the kriging plan;
Consider the estimation of the mean grade Z v of a panel of support V 5. the choice of a time-competitive algorithm for the resolution of the
from the rt data Z, defined on supports v,. The kriging system ( V . l ) is
kriging system.
written as
(1A ~ C ( ~ (up)-
, , p = C(U,, V), V a = 1 to n,
Thus, o n Fig. V.14, the estimation neighbourhood of the square panel V systems can be solved for an average data configuration; i t can then be
is taken as the union of nine squares of the same size (the square 1 ' ' itself
determined whether t h e gain in computer time justifies t h e loss in precision.
plus the eight squares which constitute the first aureole around V ) . T h e However, grouping t h e d a t a into sets does not always involve a loss in
available data are classed into nine sets Sk according to their locations precision, particularly w h e n the grouping is based upon a structural sym-
within o n e o r the other of the nine squares. Let {Zk,k = 1 to 9 ) b e the mean metry.
grades of these nine sets. T h e estimator considered is, thus, a linear
combination of the nine mean grades Zk,i.e., Structural symmetry All data that would receive strictly equal weights
from kriging can be grouped together. This grouping reduces the dimen-
sion of the kriging system but does not alter the kriged value o r the kriging
variance.
Such data must b e "structurally" symmetric with respect to the support
V to be estimated. This symmetry will be identical t o geometrical sym-
metry only if the structural function y ( h ) is isotropic. For example,
consider the estimation o f the square panel V o n Fig. V.15 by the
geometrically symmetric configuration of the five quasi-point grades
Z ( x l ) ., . . ,Z(xs).If the structural function y(h,, h p ) is isotropic, then the
four exterior data will have exactly the same weights a n d they can, thus, be
grouped into one single set called the "first aureole" with a mean grade
2, = [ Z ( X , ) + Z ( X ~ ) + Z ( X ~ ] I + X ~ ) ] / ~ .linear estimator is
T hZe (two-weight
then Z : =AlZ(xl)+AzZ,.
u
FIG. V.15. Structural a n d geometrical symmetries.
2. Reducing the number of kriging systems on Fig. V 16, the centred configuration A is preferable to the displaced
configuration B.
In section V.A.l, remark 5, it has already been noted that if the data
For the size of the support to be estimated, the practical rule is to define
configuration is repeated from one panel to another, the first member
a support equal to the grid size or some simple submultiple of it, cf. Fig.
matrix [K]is the same for each panel and it is necessary to take its inverse
V.17.
only once. Moreover, if the panels to be estimated are equal and have the
same geometry, the second member matrix [M2] is also the same for each
panel and, thus; there is a unique kriging system which needs to be solved
only once. The system of solution weights, [ A ] = [K]-'. [M2] is unique and
applies to all the panels to be estimated, cf. the case: study of the kriging of
Donoso copper mine in section V.B.5, Fig. V.27.
Systematic sampling on a regular grid is, in this sense. particularly FIG. V.17. Relative size panel -grid of information.
favourable for precise and rapid kriging In addition, such regular sampling
plans facilitate the structural analysis which can be carried out precisely It may sometimes happen that units have to be defined on variable
and without any risk of bias,due to preferential positioning of the data. supports om the borders of a deposit or a lease. The kriging of these units
In practice, it is, thus, very convenient to have a data configuration that is would then be considered separately, especially as specific problems (non-
repeated identically from one panel to another. This may be achieved by: stationarity, preferential location of data sets) often arise which would
(i) insisting on a systematic and regular sampling pattern when the require particular forms of kriging.
sampling campaign is being prepared; I t often happens that certain points on a
Krigirzg of gaps in the data grid
(ii) defining the panels to be estimated on a constant support, at least
within the central homogeneous zones of the deposit; data grid have, £or various reasons, not been sampled (difficulty of access,
(iii) considering a unique "mean" data configuration ht. using the tech- accidental interruption of drilling, not enough time, etc.). Before consider-
niques of random kriging or kriging of gaps in the data grid, cf. the ing these gaps as missing data, it should be verified (cf. section III.C.l,
following paragraphs. "Review of the data") that, in fact, they are not zero values nor pref-
erentially located in some way. Thus, consider a regular grid with non-
preferentially located gaps. There are as many different kriging systems as
Definition of the support fo be estimaud A constant estimation support is
very frequent in mining applications as it very often corresponds to there are different locations of the gaps, cf. Fig. V.18. The presence of gaps
in a regular data configuration thus increases the number of kriging systems
constant mining or selection units.
to be solved.
The geometry of the estimation support V should be simple (rectangular
in two dimensions or parallelepipedic in three dimensions) so as to make In order to return to a unique kriging system corresponding to a
complete data configuration, each gap can be replaced by its kriged esti-
the calculation o f the terms T(u,, V) as easy as possible. The division of the
deposit into panels to be estimated should be in geometric harmony with mate. It has been shown, cf. A Journel (1977, French version), that the
the regular data configuration. Thus, for the two-dimensional case shown
Gops
FIG.V.16. Relative locations of panel and information. FIG. V.18. Regular data configuration with gaps.
MINING GEOSTATISTICS V THE ESTIMATION OF IN SlTU RESOURCES
353
kriging using a particular and simple data configuration. I n practice,
panel kriged estimator so obtained is identical to the direct kriged estima-
random kriging is used during the first stages o f the estimation of a deposit
tor, using the configuration with the gaps. Of course, the kriging variance
when no great precision is required f r o m local estimations. During this
given by the completed configuration will underestimate that obtained
phase, local estimations a r e used m o r e as a n i n t e r m e d i a t e step before
using the configuration with the gaps. I n practice, prior kriging of data gaps
proceeding to a global estimation by combining all t h e local kriged esti-
will b e o f interest only if there is a small number of such gaps. T h e n , the mates; the simplified procedure o I random kriging is t h e n entirely justified,
kriging o f these gaps can often be replaced by simply taking the arithmetic A further justification occurs when there is a nugget effect in t h e structural
mean of the neighbouring existing data, cf, the Prony case study in se:ction model y(ii): the nugget effect removes t h e screen effect, i.e., it tends to give
V.B.4. the same influence to all data regardless of their distances from the pane]
V, In such a case, the precise location o f t h e d a t u m Z ( r , )in p, is not
R a n d o m kriging Random kriging results from a principle of random data important. When there is a pure nugget effect, all d a t a with the same
positioning withln determined domains.
support will have the same weight; this is the limiting case o f random
For example. let Z(x,)b e o n e of the data considered in the estimatl~onof
kriging in which there is only o n e domain P, with r a n d o m locations, and p
a panel V. T h e location of Z ( x , ) relative to t h e panel may vary from one
is the entire deposit o r the neighbourhood o f quasi-stationarity around the
panel V to another V', cf. Fig. V.19, but the d a t u m Z ( x , ) remains inside a
panel V to be estimated.
domain p, of fixed location relative to the panel to be estimated. It is
assumed that the location x, in PI is uniformly random. Similarly, P, is the
The diagonal terms in the random kriging system R a n d o m kriging
domain of the uniform random location of t h e d a t u m Z ( x , ) .
amounts to replacing t h e rectangular terms T ( x , ,x,), i ij, by ?(pi, Pi)and
the terms )'(x.. V ) of the second m e m b e r matrix by J(Bi. V). However, this
approximation does not affect t h e diagonal terms P(x,, xi) which only
depend on the support u, of each d a t u m Z ( x , ) .
If all the data Z ( x , ) ,Z(x,),. . . , have t h e s a m e support u, then
4. Preparation of a file for a kriging plan 5. Choositg a n algorithm ,for solving kriging systems
A considerable amount of the cost of kriging is spent in defining the A n important element of kriging costs is the time involved in solving
estimation neighbourhood and searching for data within these neighbour- the sj~stemof kriging equations. For a three-dimensional deposit sampled
hoods. This cost can very often be cut down by establishing a data file by drill-holes in all directions on an irregular grid, it is common to have to
adapted t o the particular kriging plan. T h e set of data would be classified estimate a number of blocks of the order of 10 000, each of which requires
and organized into a file which would permit rapid access to thc data a particular kriging system which may consist of five to 12 equations. Thus,
required for solving each kr~gingsystem with a minimum of intermediate it is essential to make use of a s o l u t ~ o nalgorithm that is both adapted to the
calculations. kriging plan and is particularly competitive in machine time.
Consider the example on Fig. V.14. In the two-dimensional space, the T h e kriging (stationary o r otherwise) system of linear equations has two
plan for kriging the square panel V consists of grouping all the data ion the important properties that should be taken into account.
nine grid squares {L;, k = 1 to 9}, and considering the linear comb~nation (i) T h e first member matrix of the system is symmetric.
of the nine mean grades on the nine domains V k , i.e., (ii) if all data supports v, are compact, which implies n o grouping of the
data, and if the structural model does not present any hole effect, t h e n
first equations of the system written in terms of the covariance* are
such that:
In order to build a data file adapted to this kriging plan, the space is divided
into a certain number of grid squares, each square being indexed by a pair
of integer coordinates (iuk,jvk). For each domain Vk the following
information is recorded in t h e data file. T h e symmetry property of the first member matrix [K] suggests the use
of the square roots method (the matrix [K] is expressed in the form of a
(i) T h e number nk of data (with support supposed constant and quasi- product of two triangular ]matrices, o n e the transpose of the other: [K]=
point) within Vk. [TI'. [TI), o r the Cholesky method ([K] = [R] [Dl . [R]', where [Dl is a
(ii) T h e coordinates (u,,, v , , ) of the location x,, of each of these data diagonal matrix). Both these methods can be found in B. Demidovitch and
(ik = 1 to n ~ )This
. information allows the calculation of the mean I. Maron (1973, pp. 287-294). These two methods were tried and rejected
values 7 of the first member kriging matrix [K]: at the Centre d e GCostatistique of Fontainebleau, because of the machine
time involved. Two other methods are used at Fontalinebleau: the standard
pivotal method and the method of successive approximations when the n
number of unknowns is greater than 20 (which is not often the case in llai/= max
I 1]aI,l< 1
i=l
mining applications).
[A '"1 .
= [ B ]+ [ A ] [ A "'1,
[A'"] =[B]+[A].[A("],
The times given in this table are only relative (in IRIS-80 seconds), the
same FORTRAN programs translated by another computer would result
in different absolute times.
Note that, for the most cbmmon dimensions (n S 15), RELMS is three
times faster than the standard GELG, and, for large dimensions, ( I : 40)
REITER is more than 10 times faster than GELG. Moreover, the per-
formance of RELMS and REITER can be improved if they are revvrikten FIG.V.21. Sorting of the data in a network of blocks.
taking account of the particularities of the machine language of the tThe Centre de Giostatistique of Fontainebleau gives out (under certain conditions) its own
computer used. Mining Geostat~sticspackage, cf. (3.Huijbregts (19756).
362 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 363
distinct drill-holes (e.g., two pieces of core for the block of Fig. V.21). The The CLAS version given here corresponds to a direct word coding on file
characteristics of these pieces of core are put into memory. (CII French computer). This type of coding can easily be adapted to the
If required, the program CLAS provides a listing of this classification particular coding required.
I
once completed.
G
1
MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES
IAD=hL!P(hOBL)
1FIIAC.GT.OIGC TO 11
ICL= I C L * 1
hUM(hC8L ) = I L L
11 I F ( I k C . E C . O ) G C TC 2
c
C SORT I h G 6NC WR I T I N G SACPLE/SAHPLE
C
hPB(NCeL)=NPB(NOBL)+l
IF(NPE!(NCBL)wCT.HlHB)GC 7C 13
1 2 hCUT=KCUT+l
C
C k R I T E ON D I R E C T ACCESS F I L E IE 47 6CCRESS
C I D A I N MORES
SORTING D R I L L - k C L E / D R I L L - k C L E
END OF D R I L L - - h C L E
IDA=IRE*(NUM(hOBLI-1 ) + S * ( h F e ( N O B L j - 1 )
WRITE D I S C I E , I D A ~ N U S ~ X T C V ~ T O ~ Z TD,~~FIYTF~
l Z T F t HLITV
ZZ=ZOE-C~(~)*(L-~J
URITE(IOUT~~CC~)LVZZ,TEST
00 4 7 I=lrNY
0 0 46 J=l,NX
IJL=I+NY*(J-l)+NX*NY*[L-11
IFINP~(IJL).EC.O~GO Ta 4 ~
YY=YO@+OE(2)*(NY-I)
N R I T E t I O U T ~ Z C 0 6 I) v Y Y
IAD=hLHI I J L I
XX=XOE+OEt l ) * ( J - 1 )
WRITE(I@UT~20C7)J~XX,NFB(IJL)~IbC
23 N O B l = h O B L NO=NPE(IJLI
hPB( K C B L I = N P @ ( N O B L I + l IDA=IRE*IIAC-1)
I F [ N P B ( N O B L ) - G T q N l H B )GC T C 1 3 0 0 4 5 K=l,NG
READ C I S C I E t I O A , N U S ~ X O ~ L C , Z D , X F , ~ F ~ Z F ~ H L , T
I N I T I b L I Z E NEY I N T E R S E C T I C h Y R I T E ( IOUT,20081NUS,XD,YC,ZO,XF,YFtZf*HL~T
IDA= ICA+S
NCT= 1 4 5 CONT IhUE
NTUT=NTUT+l 4 6 CONT I h U E
47 CONT I h U E
C
JNO-0 2 0 0 0 FORMAT l l h l , ' S C R T I h G CN F l L E O F C R I L L - h C L E ',
XTD=XC l.'INFORMATIONS ' / 1 H , 5Xv1BLCCK GRID : C R I G I h ' t
YTD=YC
ZTC=ZC
2'XO= ',FlO.3*' YO= ' r F 1 0 - 3 , ' ,
LO= ' r F 1 0 * 3 / l H
3 1 7 X t ' NUMBER / O X = ' , I 4 , ' /tj=',I4,' /CZ='rI4
ZTF=XF 4/1H v17Xi' SIZE DBX='rFlC*3i' DBt='rF10-3,
YTF=Y F 51' D B Z = ' , F 1 0 . 3 1
ZTF= ZF ZOO1 F O R M A T ( 1 k i 8 N U M B E R OF D R I L L - H O L E S = ' , I 4 * 1 0 X ,
TV=HC*T l 1 n A X I M U M NUHBER OF I H F O R P A T I O N S T A K E h I N A ' r
HLtHC 2 ' 8 L O C K = ' r I21
! CONTIhUE 2 0 0 2 F O R M P T ( 1 k ,'**BLOCK , I s r I Y C = ' , 13,' I Z C = f l
XO=X F 113,' IZC=',I3r4Xi'NUPBER CF OATA t ' t 13r'ABObE 1 ,
YD=Y F 2 ' A L L O k E O WAXIMCH' )
ZO=Z F 2003 FORMAT(lh0,' NUHBER OF B L C C K S C O h T b l A I h G D A T A '
31 CONTIKUE 19' = ' r I 5 / l h , ' T O T A L NUHBER OF S A P P L E S ~ ' r 1 6 ,
IF( 1NC.EC.O.bhC. JND*EQ.O j1 2 ' N U H B E R OF S A M P L E S T A K E h ~ ' ~ 1 6 )
4 CONTIhUE , 2 0 0 4 FORMATilCO,' h U M B E R OF B L C C K S C O h T P I h I N G D A T A '
DO 4 1 l J L ~ l i N 0 1.1 = * r 1 5 1 1 H , ' T O T A L NUMBER OF S A P P L E S = ' r I b r
2 ' NUMBER OF S A M P L E S T A K E h = ' r I 6 r ' l h THE ' 9
316,' INTERSECTIONS WRITTEh')
URITE D I S C I E l , I A O r ( h U F ( l j v I = l , k B J ~
2005 F O R M ~ T ( / / ~ H O I * * * L E V E L '* I5r8*** ELEbATIOh ',
YRITE OlSC I E 2 , I A O ~ f N P B I I ) ~ I ~ l , N B J ~
I F ( I N C * E C - O I G Q T O 42
1 ' 2 ~' r F 1 0 - 3 r ' ** L I S T OF L A L U E S PER B L O C K * * '
2 / 1 H r E I 1 ~ 5 r ' ~ ~ I S S I NV AG L L E Z ' I
Y R I T E ( I O U T t 2 0 0 3 ) I C L t NECTnkCUT
2 0 0 6 F o R M L T ~ ~ H o ~ ~ ~ L I N EI S , * Y=',FlC.3)
GO T C 43 2 0 0 7 F O R H A T ( 1 H ,'CaLUWH 'rI5r' Xa'*FlC.3,' hUHBERV
42 Y R I f E ( I 0 U T 1 2 0 0 4 1 I C L * N E C T v H C U T , H T U T
1,' O F O A T A = ' r I 2 r S X r t F I L E A D D R E S S 5 '916)
C
43 I F ( I S . E P . l j G 0 TO 5
2008 F O R M L T ( 1 H * ' D R I L L - H O L E ' 9 18,' START A h 0 END '
LI'COOROINATES 's6(2XtFB*3),5X,'LENGlk ='rF10.3
C P R I h T R E S U L T S I F IS.hE.1
2 / 1 H r 'VALUE ='rFlO.3 J
c 2010 F O R M A T ( / / / l H O I
kRITElICUTt2010)
C
CO 4 7 L = l i N Z
5 IRETURN
IEHD
368 MINING GEOSTATISTICS V TIHE ESTIMATION OF IN SITU RESOURCES 369 i1
Program KR 13D for the definition of the nugget effect, the structure of this subroutine is
similar to that of FUNCTICIN GAM given in section III.B.5.
This program is designed for the kriging estimation of parallelepipedic
units from three-dimensional drill-hole data, previously sorted by the
program CLAS.
C
C
C
....**.***..........~........................*...
KRI-30 K R I G I N G PROGPCC
The unit v k to be estimated is a submultiple of the block B of CLAS c
network. There are NSB units to be estimated in each block B. The C K R I G I N G O F A T h R E E - O l M E N > I C N d L SYSTEM C F Bl,OCKS
location of these units v k are deduced from the location of the block A
3 by a C D E F I N E D B Y O R I G I N 4NC G R I D F R C ? O R I L L - I - C L E S M I T H
C ANY ORIE~NTATIOH..CATA H A V E B E E h I N I T I A L L I SORTED
system of integer coordinates similar to that used in CLAS (the unit no. 1 C I N BLOCK SYSTEM 131 S U B R O U T I N E CLAS*.EACH BLOCK
within B is located at the superior left corner of the upper level of B, cf. C M I G H T BE S U B O I V I O E O I N A SYSTEC OF SUB-BLCCKS M I T H
Fig. V.22). C SAME K R I G I N G NEIGhBOURHOOO.*
The same estimation neighbourhood is considered for all the units v k C
C ..*.PARAMETERS
comprising block B. The form of this neighbourhood is parallelepipedic, C : S O R T I N G O F CATA ( I H BLOCK SYSTEM)
centred on block B, and consisting of an integer number of blocks 6,.The C X O B t Y O B v Z O B O R I G I N OF B L O C K S t S T E M : CEFTRE OF
data within the estimation neighbourhood are grouped into NL sets. Each C BLOCK L O C A T E D DOWN L E F T ON TOP LEIlEL
of these sets consists of the data which are within a certain numb~trNP C CBXIOBYICBZ S I Z E O F B L O C K S ALONG A X I S CX*OY,OZ
(L= 1 to NL) of blocks B, comprising the neighbourhood. A parlicular C NX*NY,NZ BLOCK SYSTEM I S ARRLNGED AS I h <CLAS>
C H I M 0 SEE SUBROUTINE CLAS
kriging weight will correspond to each of these NL sets of data. All the C TESTIIS SEE COMCON CND P R l h 7 C P T I C N
units or, of the same block B are estimated from the same data configura- C ~SBINSYINSZ SUB--BLOCK S Y S l E V b I T H I N EbCH BLCCK
tion; their kriging systems differ only by their second member matrices. C ::KRIGING P L A h
C NL NUMBER O F DATA GROUPINGS I K R I G I N G L E I G H T S )
The mean value C ( L ~L,z )calculations, between two sets L1 and L Z z , are C IJV(3) NUMBER O F L I N E S V C O L L # ~ S I L E V E L S CF
carried out taking into account the precise locations of each datum (sample C P A R A L . L E L E P I P E D I C U P I G I NG NE16HBCURHC00
or piece of core) constituting L1 and L Z .For the calculation of the mean C NPINL) NUMBER CIF B L O C K S C O P F E S P O N D I N t 10 E A C H
values C(L,,vk), the unit vk is approximated by a regular set of discrete C C4TC GROUPING
C NOB(MAX-NP) L O C A T I O N O F T H E f E BLOCKS k 1 T H I N
points. C NEI6HBOURHOOO.BLCCKS ARE NLMBERED
The algorithm used for solving the system of linear equations is RELMS, C ACCORDING TO L I H E r C O C U M h * L E V E L
cf, following section. C P O S I T I O N I N NEIEI-BOURHOOC
The total volume to be estimated is contained in the network of blocks B C :STRUCTURAL FCFAHETERS
C NSTvCO hUHbER O F S T R U C T U R E S l h U G G E T E F F E C T
defined by the program CLAS, and is const~tutedby an integer number of C C(NST) SILLS
blocks B, (those indicated by the index IEXP = 1). All the units v k , com- C AAtNST) RANGES:<O. EXPCNENl~IdL*>O. SPHERICAL
prising a block B to be estimated, are estimated. C C L X ( N S T e 3 ) * C A Y ( h l S T v 3 ) * C b Z I N 5 7 * 3 ) SEE <CCVA>
C N B X * N B Y t N B Z S U B C I V I S I O N OF SUB-BLCCUS FOR
Subroutines used KRIGMS + RELMS, for the writing and solving of the C
C
COVARIANCE CALCLLAT I C N
:CIRECT-ACCESS F I L E S U P 1 7 hUFBERS
kriging system. When kriging systems with large dimensions (more than 20 C IE: SORTED O A T l l P R O V I D E O BY < C L A S > )
equations) are to be considered, RELMS (pivot method) should be C IEX: ARRAY I E X P L O C A T I O N O f KHIGEC BLCCKS
replaced by a modified version of REITER. C IEXP(NX*NY*NZJ = l UR1GEO BLOCK = C NO K R I G I N G
CBARS, for the calculation of the various mean values c. The structure C I N SAME ORCER 4 s BLCCK SYSTEM
C I E l < ARRAY HUM F R O V I O E O e Y < C L A S >
of this subroutine is similar to that of FUNCTION GBAR given in Sixtion C IE2: ARRAY N P B P R O V I O E O eY i C L A S >
1I.E.1. C IE2: ARRAY N P B P R O V I O E O B Y C L A S
VOIS, for the data search within the estimation neighbourhood of each C 10s OUTPUT F I L E OF RESULTS.ONLY B L C C F S SUCH
block B. C T H A T I E X P r l A R E W R I T T E H I N SEQUENCE.MITHIN
C EACH BLOCK SUB-BLOCKS AFE NUCBEPEC ACCORDING
FUNCTION COVA, for the calculation of the covariance function C ( h ) C TO LIhE*CCLUMN*LEVEL
defined as a nested sum of basic models (exponential or spherical). Except c
370 MINING GEOSTATIS'TICS V T H E ESTIMATION OF IN SITU RESOURCES 37 1
C
C
... HEHORY R E C U I R E M E N T S
c81CBlrCB2 hSB=NSX*NSY*hcZ
C
C IP
C NIN
C XAvYAaZA h8T*NIHB
C TA h'L
C TKISK HX*NY*N S B
C NUNr N P B NX*NY*I J V 1 3 1
C R,RRIX (NL+lJ*NSB
C A 1 N L + 1 ) * (NL+Z ) / Z
c
C .--.COHHChS
C IOUl L I N E PRINTER U N I T hlM0Ei1
C TEST I N F E R I C R BOUNDARY C F E X I S T I b G D A T A
C I N C A S E CF F L O C K W I T H NC I N F C F H A T I C N
C VALUES A P E T H E N =TEST,
C /VOlSI/ PARAPElERS FCR k C I S
C /STRUC/ P A R A F E T E R S FOR CCVA
C
C
C
.. oCPT1 OhS
IS-EO.1 S T A T I S T I C S OF KRIGEC BLOCKS LRE PRINTED
R E A D I N P U T PARAMETERS AKC F R l N l T I T L E
C JS.NE.1 A L S O R E S U L T S B L O C K / B L O C K ARE P P I h T E D
C
C .-.oSUBROUTIhES CALLED
c KRIGHStRELMS K R I G I N G SYSTEC*:OLUTION
C VOIS NEIGHBOUPHOCO OF E S T I H b T E C ELCCK
C CBARS AVERIGE COVARIAhCE
c cove C O V A R IANCE F U ~ ICNC ~
C
C o...COMMEhTS:
C BLOCKS ARE NUMBERED AS I h S L E F O U T I N E CLAS:BLOCK
C h U n B E R I + N Y * ( J - ~ I + N Y * N X * ( L - ~ I I S CN L I N E 11
C C O L U M N J v L E V E L L..LINE 1 CCPRESPONOS L L k A b S TC
C P A X I H U M Y C O O R O I N A f E v C O L U F h 1 TO C R I G I N , L E V E L 1 C E F I N E NEIGHBCURHOOO AFD C b T A G P O L F I F G
C T C TOP L E V E L * . KRIGING PLAh:hElGHBCUPhCCC I S
C D E F I N E D AS b P A R A L L E L E P I P E C I C E N V E L O P F E M A D E U P
C C F BLOCKS.. I N F O R F I T I O N S B E L C h G I N G t C P L C C K S C F
C TI-E E S T I C A T I O N N E I G h B O U R H C C C C P E G R G U F E E I N N L
C SUB-SETS EACH O F WHICH W I L L I-bVE A P A P T I C L L A R
C K R I G I N G WEIGHT..CHOICE OF T h E J E S t B - S E T S S H C C L D
C BE CAOE A C C C R D I N G T C L O C A T I C R I S Y H P E T P ~ , ETC.. TO
C C E F I N E A P A R T I C U L A R S U B - S E T L THE N U M E E F S O F T H E
C BLOCKS D E L I N E A T I N G I T A R E G I b E N A S A L I S T OF h P ( L )
C VALUES (ARRAY NCB).-THESE BLCCKS ARE hUPBERED READ BLOCK NUCBERS O E F I N I h G GROUPING L
C A C C O R D I N G T O T H E I R R E L A T I V E L I N E (1. T C I J V ( 1 ) 1
C P O S I T I O N I 1 ( L I N E L C O R R E S P C h C S A L M 4 Y S 7G Y
C C A X I H U M J t C O L U H N ( 1 T O I J V ( 2 1 1 A N D L E V E L ( 1 TO
L I J V t 3 ) ) P O S I T I O N S 12 A N 0 I 3 8) I l + I J V ( l ) * ( 1 2 - I ) +
C IJVf2)*IJV(ll*(f3-I)..SUCk hCPBERS bRE CChVERTED
C I N T O R E L b T I V E P O S I T I O N S U S E t EY S U B R O L T I h E V O I S
C ( S E E ARRLY IPI..
372 MINING GEOSTATISTICS I V THE ESTIMATION OF IN SITU RESOURCES 373
N2=3*(hl-11 XF(L,GT.hBZ)GC TO 1 2
LB=(hCB( 1 ) - l ) / N L C + l J2=-J2
JB=(NCB( 1 ) - N L C * ( L B - 1 ) - l ) / h L I + l YOtYO+JZ*DSY (2-*NBY 1
IB=NO@(I)-NLI*(JB-11-NLC* (18-1) DO 1 0 J a l v N B X
IP(N2+l)=IB-IJV( 1)/2-1 DO 1 0 I = l , N B Y
IP(N2+2)=JB-IJV(2)/2-1 I l = I l + l
IP(Ni+3)=LB-IJV(3)/2-1 XDB( I1 ) = X O + ( J - l ) * D S X / N B X
1 0 0 CONTlhUE YDB( I l ) = Y O + ( I - l ) * D S Y / N B Y
C 10 ZDB( f l ) = Z O
C READ A h 0 P R I N T S T R U C T U P A L F A R A P E T E F S ZO=ZO-DSZ/NBZ
C 11 CONTIhUE
12 CONTIhUE
13 COHTIhUE
14 CONT IhUE
C
c COMPUTE V A R I b k C E OF A F O I H l b I T H I N SLB-BLOCK
C
HIN( 11-1
N I N I 2 )=NCB+l
C I N P U T CUTPUT C I R E C T ACCES! FILES i CALL CBARS(XCEtYCBr2DBt l r C r N I N r X C @ , Y C B ~ Z D B t
l h D 8 , NSBvCB)
PEAOlINP,1006)IE~I E l t I E i t l E ' t l O C
NXYZ( 1)=hX C K R I G I h G E L C C K PER BLCCK
NXYZ ( 2 ) =NY c
NXYZ(?I=hZ h BL= 0
XYZl l)=XCB UK-0,
I1
X Y Z [ 2 I = Y CB VK=O.
X Y Z I ? )=ZCB HK=O
C B ( 1)=CBX IAD=O
CB(Z)=DBY hO=NX*hY
IF(IS.NE.l)UIRITE( IOUT,iOCC)
L
C
C@13)=CBZ
S U B C I V I C E SUB-RLCCKS I N FCIhTS
I DO 1 L B = l , K Z
ZE=ZOE-CEZ*(lL@-1)
e I F ( IS.hE.1lWP I T E ( I O U T t i O C 5 )LB,ZE
NBK=O
LO=MAXO( l v L E - - I J V ( 2 1 / 2 )
Ll=HIhO(hZ,Ll?+IJV(3J/2)
i C READ I E X F 9 N b P t h P e
I C ON R E S P E C T I V E U N I T 5 I E X . I E l t l E 2
1 C PDCRESSES S T A R T 4 1 0
IADl=hX*hY* (LE-11
i
1 READ C I S C I E l i * I A O l c ( I E X P ( I J ) t I J = l , h C )
1 IAD2=hX*hY*(LO-13
hl=NX*NY*Ll-I bCZ+l
I READ C I S C I E Z I I A D ~ ~ ( K U C I(J I r I J = l , h l )
READ C I S C I E 3 r l A C 2 r ( N P B ( J J I r I J = l t h ' l I
C COMPUTE L I N E B Y L I h E
CO 2 I B = l , N Y
'i CO 3 J B = l t N X
I
3
4
i
IJB=I@+NY*(JB-1)
I F ( I E X P ( IJB),EQ,O)GO
hEK=HEK+l
hBL=NEL+l
XB=XOE?+CEX*( J B - 1 )
TC 3
-
37 W R I T E ( I O U T r 2 0 1 2 ) I B p J B
381 0 0 3 9 K - l r N S B
J=K+NSB* IHBK-1) ,,,,
TKLJ)*TEST
39 S K I J ) = T E S ~
SORT CUT I N F O R C A T I C N W l T k l h N E I G H e C U P h C O D 3 CONTINUE
2 CONTINUE
NZ=NBK*NSB
WRITE RESULTS O h F I L E
GO T O 34
SOLVE K R k G I L E SYSTEC
3 1 CALL K R I G ~ S ( N ~ ~ X A ~ Y A ~ Z A ~ N I ~ ~ X D E , Y C @ ~ ~ C B ~ N D B , 1000 FCRHAT(6F8.3r313rF8.2r 11 1
lNSBrRrRR*A,Xr IERpCB11 1001 FORMAT(312)
I F ( I E R e N E o 0 I G G T O 36 1002 FORHbTt4I2)
SCLUTION 1003 FGRMAT(4012)
CO 3 3 K = l , N S B 1004 FORMbT(I2rF10.4)
J=K+NSB* 1 N B K - 1 ) 1005 F O R M A T (F 1 0 . 4 ~ F 8 . 3 ~ 9 F 4 - 2 )
TTK= 0. 1006 FORHLTI512I
IK=( N b + l l * K C
SSK=CE(lJ-XL I K J 2 0 0 0 FORMAT( l H l * 4 [ / ) *4eX, I * * * * K F l c I h G OF e L O C K S * * * * '
00 3 2 I x l r N A ].//'*DRILL-HOLE D A T A S O R T I N G G R I D ( I h BLOCK S ' r
IK=I+(kA+lJ*(K-l) ZtYSTEP)*'//' SIZE : O r = ' i F 8 . 3 9 ' CY= 'tF8.3
TTK=TTK+X( I K l * T A ( I ) 3,' D t s ',F8.3/1H 9 ' O R I G I N OF BLOCKS ! X O a ' r
3 2 SSKzSSK-Xf I K l * R R ( I K )
TK( J 1r T T K
4tF11.3r' YO- ' r F 1 1 . 3 9 ' '
201 9 F l l . - ? / I H 9
5' G R I D OF ' r I 3 9 ' L I N E S . ' r 1 3 r 1 C O L L t k S AND ' r
3 3 SK(J1zSSK 613'' LEVELS ' / I *KRIGEO BLCCKS* S I Z E : OSX. ' r
34 0 0 3 5 K = l r k S B IpF8.3r' OSYx ' r F 8 - 3 9 ' OSZx ' r F B e 3 r ' hUMBER ' t
JrK+NSf!*(NBK-1) 8 * O F SUB-BLOCKS= 13)
UK=UK+TK ( J J 2 0 0 1 F O R W A T ( 1 h r 1 K R 1 G I k G PLAN: ' r I 2 r ' h E I G H T S ' / l H r
VK=VK+TK(J ) * T K ( J I 2"YEIGHT:CORRESPONCING BLUCKS W I T H I b hEIGHBOUR'r
3 5 NK=NK+1 3'HOOD (NUMBERS)')
IF(IS.EQ.1IGO TO 3 2002 FORHIT(1h ,2XrI2r2X,':'~S$q40( 1 ) r ~ I i )I
C P R I N T O U T I F IS,NE.,l 2 0 0 3 F O R W L T I I H O r t S T R U C T U R A L P A P A M E T E R S OF T H E ' r
YRXTE(IOUf*20C91 10~JB,XBv~'IerNA '
I'EST I M A T E O V A R I A B L E ' / l h p bUMBER O F I T R L C T U R E S '
00 4 0 K t l r N S B 29' ',I1r1 NUGGET E F F E C l r r r F 1 2 . 4 / ' SILL 'r
JxK+NSE*(NBK-1, 3' RANGE I CAX I CAY
40 Y R I T E ( I O U T ~ ~ O ~ O ) K ~ T K ~ J J ~ S K ( J ) 4'1 CbZ' 1
GO TO 3 2 0 0 4 F O R M A T ( 1 H rFl0.4rlX~FB.3rl~rQI'I'rF4~2))
C ERRGREO S Y S T E C 2 0 0 6 FORHAT(lHOr4(/lr48Xrr**KR161NG R E S U L T S BLOCK ' t
36 W R I T E ( I O U T ~ ~ O O B I I B I J B '
1 PER ELOCK** / / 1 '
GO T C 38 2007FORWATllhO~/rlHr' **LEkEL I Z"c
C hO I N F O R H A T I C h 1F8*3//)
376 MINING GEOSTATISTICS JT V THE ESTIMATION OF IN SITU RESOURCES 377
SUBRCUTINE K R I G H S t N A r X A r l E r L A ~ h I N t X C E , Y C e ~ Z C B
I~NDR,~SBIR,RRVA~X,IER~C~J
C
C
C
..................................................
COMPUTE K R I E I N G M A T R I X A b C S C L V E SLSTEM k N E0
C C
C ....PARAMETERS C C A L L SYSTEP S C L W T I C N R C U l l h E
C NA hUMBER OF E F F E C T I V E DATA E F O U P I h G S C
C X A v Y A v Z A C O C R O I N C T E S OF O A T 4 b I T H l h NEICHBOURHCIDD CALL R E L M S I X ~ ~ ~ R I C ~ N S M ~ I E P )
C h I N L O C b T I O N O F CATA GROWPlhG I N XAVLAIZA RETURN
C NSB NUPBER OF SUB-BLOCKS END
C R ( I N A + l )*NS@) RIGHT-hAND S I T E MATRIX
C LI ( N A t l )*(NP+Z)/ZJ COEFFICIEhT MATRIX
C X I ( N A + l )*NSB SOLUT I O N
C KTILT ERROR I N D I C A T C R I F KTILT.NE.0
L C ....................... ..
. *.*
. .o.
, ..
. ..
. ..
. ...
c
c
.........-....................................... C ASSURED TC BE STORED COLUHhYISE..
C
C .... PARAMETERS
OIMEbSIOh X A ( l ) r Y A ( l ) r Z A ( I ) , X O e ( l ) ~ Y c B ( l ) r C X(W*N) F I N A L SOLUTION
l Z C e ( l l r N I N ( 1) C AlM*(M+l)/Z) UPPER T R I A N E U L P F P A R T OF
DIMENSIOh C B ( 1 ) C COEFF I C I E N T M A l P I X DESTROkEC B Y
DIMEhSIOh R ( l l r R R ( l ) r A ( l ) r t ( l ' C TRIANGULATION
hEO=NP+l C BIM*N) RIGHT-HAND S I D E M A T R I h D E S T R O I E C
hN=( N E P + l I * N E C / Z C M NUMBER OF E O U A T I O h S
c C N NUHBER OF RIGHT-HAND S I D E LECTCRS
C COMPUTE C O E F F I C I E N T C A T R I X C K T I L T R E S U L T I N G ERROR PARAPETER
C C K T I L T = O NO ERROR
378 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 379
C KT ILT-1 hO R E S U L T HC I R = A ( II ) * P 1V
C K T I L T = K ZERO P I V O T Y 4 S FOUND 17 LP=LP+l
C E L I C I N A T I O N STEP K IJ=II-KW1
C DO 4 J * l t M 1
C .ee.COMHEhTS: I J = IJ + J
C S O L U T I O N I S DONE BY MEANS CF G A U S S - E l I P I h A T I O N LL=LL+J
C W I T H P I V C T I N G I N R A I N DIAGOhPLo.HEhCE NC S E A R C H 4 A(IJ)=A(IJI-R*4(Ll)
C OF M A X I M U M ELEMENT ..EOUATlOblS A R E K E P T I N ORDER 0 0 5 LLB=KrNMrC
C PNO R I G H T - H A N C S I D E F A T R I X I S F I R S T KEPI..SYSTEW IN=LLBtLP
C 1s THEN S O L V E D BY S I M P L Y h O R K I N G @dCKhPPDS.. 9 El( I N ) = @ (I N 1 - R * B ( L L B )
c K T I L T ~ K I ~ C I C A T E S HIGHEST R A N F M I N O R C E T ~ G ~ KI I 45 CONTIhUE
C I S ZERO..IN GENERAL A F T E R T R I L N G U L A T I C N T H E M A I N 7 CONT Ih U E
C D I A G O N A L E L E M E N T OF L I N E I H b S T H E V A L U E RIA[ IJ)
C P ( I * ( I + l J / Z J = O E T ( G ( I ) ) / C E T ( G ( 1 - 1 1 1. IF(R-TOLl9i8tS . .
C RELMS S H C U L C BE USED ONLY FCR S Y S T E M S W I T H NON- B KTILT=M
C ........*....*
,. *......*.*..........
C Z E R O H I G b E S T RANK M I N O R S * .
.... *...*.*. C
C
PETURN
END T R I A h G U L b T I O N
C
C START @ACK SOLUTION
C
IF(A(~I.EO.O. 9 PIV=le/R
0 0 19 I a 1 , N 0 0 1 0 LLB=M*NMrM
1 9 X ( I l = @ ( I] / A ( 1 10 X ( L L B ) = B t L L B I * P I V
K T I L T =O I=M
RETURN . KK=I J
20 K T I L T = l 00 13 II*l*Ml
RETURN kK=KK-t
C PIV=I./A(KK)
1-1-1
00 1 2 L L B ~ I I N W , M
IHrCLB
RIB( IN1
IJ = K K
DO 11 J ~ I v M l
I J = I J+J
IN-lNt1
C START TRIANGULATION 11 R = R - A t I J j * X ( I N 1
C 12 X ( L L B ) = R * P I V
00 7 K ~ l r M l 1 3 CONTINUE
KK=KK+K C
AKyA ( K K 1 C END S C L U T I O N
IF(AK-TOL13r2~3 C
2 KT1LT.K RETURN
PETURh c
3 P I V = l./AK C ERROR R E T U R N
lI=KK C
LP=O 1 4 KTILT*-1
K H l r K- 1 RETURN
00 6 f=KtMl END
LLaII
11x1 It1
MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 381
CIUENSION N P ( l ) , J L ( l ) v I P ( l I
DIMENSION N X Y Z ( 1 ) r N P B ( 1)r h l P l l )
DIUEhSION Xbfl)vYA(l)~ZA(lIiNIN(1)vTL~11
C
AVERAGE COVPRIPNCE B E T k E E h S E T S I PNC J COUHOH/VOI SI/KIHBIXB vYBvZBvIJVL
C
)rl=NIh(J) NX=NXYZ(l)
CZ=NIh(J+I)-1 NY=NXVZ( 2 )
CO 1 I 2 = N l v h 2 NZ=NXYZ(3)
DO 2 J2=F'l,R2 NA-0
C B ( ~ ) = C B ( ~ ) + C O V A ( X A1( 2 ) , L A ( I Z ) v Z A ( I 2 ) v X A [ J ~ ? ) , 11-0
1YA(J2)vZb( J2) J= 0
NC=NC+l N I N ( 1)=1
2 C O N T Ih U E N=O
1 CONTIhUE C
CB(l)=CB(l)/PAXO(lsNC) C NEW G R O U P I N G C O M S I D E R E C
PETURh.
DO 1 I L = l , N L
AVERAGE C O V A R I A N C E B E T M E E N S E T I b N 0 S L B - B L O C K S NB=NP(ILI
IA=NA+l
3 DO 4 K 2 = 1 i N S B PA=O
NC-0 TA(Ib)=O.
00 5 1 2 - k l t h 2 C h E Y B L O C K CF l t - I S G R O U F I h G
00 5 J l = t v N O B GC 2 C + l , N B
J2=Jl+hOB*(KZ-1) N=N+1
CB(KZ)=CB(KZJ*COVA(XA( 1 2 ) , t A t I 2 ) r N1=3*(h-1)
1XDBt J21iYDB( J2)r ZDBl J211 IO=JL(l)+IP(Nl+I!)
NC=NCtl JO=JL ( Z ) t I P I N l + i ? )
5 C O N T Ih U E LO=JL(?)+IP(Hlt3I
CB(KZ)=C@(K2)/CAXO(l(NCI LB=IP(hl+3)+IJVL
4 CONTIhUE IF(IO*(NY+l-IG),bLE.OJGC It 2
END IF(JO*(NX+l-JCI,vLE*O)GC 1C 2
382 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES
t
C
C ................................................
O B T P I N b P P R C X I M I T E S O L U l 1 CN
FUNCTICN COVI f X l r Y l r Z l * 2 i l \ i r t i l
C OF S I H U L T A N E O U S L I N E A R E C U b T I C b S AX=R..LCPER
C
C
c
...*.........................................,.*.
COVARlAhCE F L h C T I G h
C TRIANGULAR PART CF C O E F F I C I E h T MATRIX' I S ASSUPED
C T C B E STORED CCCUPNNISE..
C
c DEFINED A S T ~ sur E O F E X P O ~ E ~ T I A LC R SPHERICAL
C -...PARAHETEPS
C STRUCTURES W h I C h ARE I S O T R C P I C AFTER L I h E A P
C X t U l VECTOR O F F I N A L S O L L T I C h S
C T R A k S F O R M b T I O N OF T H E I N I T I A L C O O R l D I h A f E ! CF C A(H*(M+1)/2) UPFER T R I A N G U L P P P A R T OF
C TbE VECTOR ( X 1 - X Z t Y 1 - Y 2 t 2 1 - 2 2 l . . C COEFFICIENT WAlEIX (LNWCCIFIEO)
c C RIM) RIGHT-HANC S I C E VECTOR
C .ee.COHMONS ISTRUC/ C E P S TOLERAHCE CN S O L U T I O h ( G I V E N b S F R E C I S I O N
C N S T N U n e E R O F ELEMENTARY S l P L C T U R E S
C OF ( K R I G I N G I V A R I A N C E )
C CO N U G G E T E F F E C T C I E R DIVERGENCE INDICATCR
C CINST) SILLS C IER=O GGCD S O L U T I C N
C AAiNST) RANGES < O w E X P O N E H T I b L C IER=-1 DIVERGING SYSIEF
C >O. S P H E R I C A L C
C C C X ( N S T ~ ~ ) ~ C A Y ( N S T ~ ~ ) V C A Z IL Nl h~E TA ~P ~ )
C
C
C
...... T R A N S F O R F A T I C N O F I h I l I A L CCCEC l h A T E S
**....m.....*............*....r.......,....
C .g-.COMMENTS:
C T H E M E T H O D IS B Y S U C C E S S I V E b P P R C X l C b T lCNS
+As it is REITER works only for simple kriging systems (without the condition E m A, = 1).
384 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 385
C
C
I N W H I C H T h E F I N A L S C L U T I C N I S IMPROVED (REF.
CEM I D O V I T C H / H A R C N E L E M E N T S C E C A L C U L h U I E R I Q L E
> ... L
C T E S T C F b C C U R 4 C Y AND C C N I E F G E N C E
C ED. D E P O S C O U 1 9 7 3 1 . . T H I S kE1l-OD I-AS B E E h c
C A C A P T E D TO T H E S O L U T T C N O F A W R I G I h G $YZTEM..THE L=O
C S C L U T I O N @ E I N G C F L Y b P P R O X I P b T E Tt-E I I E R A T I V E hITER=hITER+IT
C P R O C E S S I S S T O P P E D WHEN T H E K F I G I N G V b R I A N C E CNMU=O.
C I S O B T A I h E O W I T H A GOOD P R E C I ' I O N (PAPPCETER DOMU=O.
C EPS E.G.:A GOOD L E V E L O F P F E C I S I O N I 5 C B T A I N E D CSIG-0.
C WITH EPS=l.E-04 OR 1 . E - 0 6 FCP A K R I G I h G VARIAhCE DO 6 I = l t M
C ABOUT l.)..YAXIHLP h U M B E R O F I l E R A T I C h S I S 2C. L=L+I
C V=X(I I/A(L)-TI11
C *..CAPACITY: CNCU=CNHU+V*Y
.
.
C
c
.
C
.
...
.
5 0 ECUATICNS
I.
... ....... ............. .....*....*..
( S E E ARRAYS T P h D D l )
0 0.
CDHU=CCt'U+V*CTlI)
6 CSIG=CSIG+V*V*A( L )
ANU=CtPU/DhPU
DIHENSIOh T(5O)t CT(5C) C C I V E R C I h G SYSTEM
C IF(A@S(AhU).LT.l.Ol)GO TC 5
D I M E h S I C N A ( l ) r R ( i ) c X ( 1) DS=DSIG/(AW*(hNU-l)*EPS)
C C CIVEN A C C l R b C l ACtiIEbEC
C I N I T l b L SOLUT I C N IF(DS.LT.l.)GO TO 7
C I F ~ N I T E R ~ G E ~ Z O I TGOO 9
CO 1 I = i , H IT12
T(IJ=C. GO T O 2
1 X(I)=R(I) c
C C FINAL SOLUTICh
C NEW I T E R P T I O h C
c 7 ANU=l./(LNU-1.)
IER=O L=O
NITER=O CO 8 I x l t M
IT = 4 L=L+ I
2 LO 5 K = l ? I T V=X(I)/A(L)
L =0 8 X(I)=V+(V-T( I ) I*ANU
C RETURN
C PRECEDIhG S C L l T I O N 9 IER=-I
C RETURN
0 0 3 I=l,M END
L=L+ I
V=X(I)/A(L)
C T ( II = V - T ( I )
T(Il=V V.B.4. Typical kriging plans
3 X(I)=P(I)
C
Numerous complete geostat~sticalstudies of actual deposits, from struc-
C CCHFUTE N E k SCLUT I C N tural analysis to kriging, can be found in the bibliography. In this section,
some of these studies have been selected because they are typical examples
of kriging plans and their presentation will be limited to this kriging aspect.
Three types of deposit can roughly be distinguished: stratiform deposits
proven by vertical drill-holes; massive deposits (e.g. porphyry-copper)
proven by drilling on an irregular grid and in all directions; and vein-type
deposits proven by drilIing or by drifts with channel samples.
This rough classification, based on geostatistical practice, does not cor-
respond to any precise metallogenic limits. From a geostatistical point of
386 MINING GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 3 87
view, a thick vein (subhorizontal or not) can be treated as a stratiform The horizontal dimensions of this small unit panel are 10 m x 10 m.
deposit in the plane of a vein. A porphyritic mass may also be regarded as a Various nickel cut-off grades were applied to the small panel prafile, cf.
stratiform deposit on each horizontal bench. Conversely, for a deep sedi- Fig, V.23(b): g,= 0, 0.8, 1, 1 , 2 % Ni. A thickness r ( x ) and two accumula-
mentary deposit sampled by drives and percussion-drill fans it may be tions in Ni and in MgO impurity correspond to each cut-off grade g,. A n
advisable to use kriging plans analogous to those used for three-dimen- accumulation a ( x ) = t(x).z(x) is defined as the product of the thickness
sional porphyritic masses, etc, t ( x ) at point x and the corresponding mean grade r ( x ) . These variables,
called service variables, are regionalized in the horizontal two-dimensional
1. Stratiform deposits space, These horizontal regionalizations are sampled on a systematic
In geostatistical applications, the techniques of stratiform deposit are rele- regular grid of 100 m x 100 m and by two 20-m in-fill drilling crosses to
vant when the structure of the mineralization and the layout of the avail- detect small-scale variability structures. The fitted variogram models are
able data, in the three-dimensional space, favour one dimension with isotropic and consist of the nested sum of a nugget effect and two spherical
respect to the other two, The most characteristic example is a sub- models.
horizontal sedimentary deposit sampled by vertical drilling: the vertical The data on the regular 100 m x 100 m grid were used to estimate by
direction is the best known, and it is often the direction of preferential kriging the service variables at each node of a 10 m x 10 m grid, cf. Fig.
variability (due to surface leaching by water for example). V.24. This kriging is not intended to provide estimates of such small
The study of these deposits can very often be reduced to a two-dimen- 10 m x 110 m unit panels, since the relative estimation standard deviation
sional problem, by considering, for instance, the hiorizontal regionaliza- would very between 25% and 95% depending on the variable and the
tions of variables accumulated over the total vertical thickness, location of the unit. The object of this kriging is to condition a simulation
of the deposit, cf, section VIII.B.1, or to provide estimates of larger panels
of various shapes by combining the kriged estimates of the 10 m x 10 m
Prony nickeliferous laterites (New Caledonia), after Ch. Huijbregrs and A. unit panels.
Journe1(1972), A. Journel(19736).
The deposit was sampled by vertical drills on a regular grid. The drill cores The kriging plan The zone that was studied covered 176 squares of the
were cut into constant 1-m lengths and analysed for Ni grade and MgO 100 m X 100 m grid, i.e., 17 600 small 10 m X 10 m units. It is, of course,
impurity. From the experimental profile of the Ni gra.des of each drill-hole out of the question to solve 17 600 kriging systems for each of the 1 2
Si, a new, more regular and representative grade file of a small panel studied service variables (three variables for each of the four cut-off
centred on the drill Si was defined, using an unbias~edkriging of order 3 grades). First, note that provided the regular data grid is complete, the data
(with a quadratic vertical drift), cf. Fig. V,23 and Fig. 11.1 in section II.A.1. configuration is the same for all 100-m grid squares. This limits the number
of kriging systems to be solved to 100 for each service variable; there are
100 unit panels within each 100-m grid square, to which correspond 100
kriging systems with a unique first member matrix [K] and 100 second
member matrices [M2] (v), v = 1 to 100. Furthermore, because of the two
symmetries with respect to the medians and diagonals of the 100-rn grid
square, the 100 second member matrices [M2](v) can be reduced to 21, cf.
Fig, V.24. Thus, for each service variable, the problem of solving 17 600
kriging !systems is reduced to that of solving a single system with 2 1
different second member matrices. In such a case, kriging proves no more
costly than any of the more traditional estimation methods.
There are several non-preferential gaps in the 100-m data grid due to
nine accidental breaks in the drilling, These gaps were filled in, before
(0) (b)
kriging was undertaken,'by simply taking the arithmetic mean of the eight
FIG. V,23, Prony deposit (vertical cross-section). data values surrounding each of them on the 100-m grid, cf. kriging of gaps
388 MINING GEOSTATISTICS V T H E ESTIMATION O F IN SITU RESOURCES 389
in section V.B.2. All 1 0 0 unit panels within the 100-m grid square were - fraction and for the tricalcic phosphate grade z ( x ) of this fraction. T h e two
estimated considering the 16 data {Z,, a = 1 to 16) o f the first and Second studied regionalized variables are the percentage p ( x ) and the accumula-
aureola, cf. Fig. V.24. T h e kriging estimator is thus a linear combination of tion a ( x ) = p(x ).;(.I). Thesc variables arc defined on the constant sup-
these 1 6 data, port, 0.5 m x horizontal section o f the pit, and are regionalized in the
three-dimensional space. In addition to the 50-m regular square grid,
data were also available o n Iwo diagonal 10-m in-fill grids to detect
micro-structures.
for each of the 1 0 0 unit panels ( v = 1 to 100). T h e dimension of the T h e three-dimensional variograms that were fitted consist of three nest-
stationary kriging system is, thus, 1 7 (16 weights A, plus the Lagrange ed structures:
parameter p,).
weights, a l , crz, a3, for the vertical data of pit S.5 which runs down on one normal for the mean characteristics of slices P of 25 m X 25 m x 0.5 m =
side of the panel slice P, 312-5 m3 to be much less dispersed than the characteristics of data defined
Of course, if any of the data corresponding to the weights Pi or a, are on pi1 sections of 0.5 x @-pit 2 0 . 4 m3, and, indeed, production data did
missing, the dimension of the kriging system will be reduced accordingly. verify this remark.
The eight weights resulting from kriging the accumulation variable for
the particular configuration considered are also given on Fig. V.26. Note
that the value & I =0.11 is significantly different from the value 1 which
50 crn
elongation o r flattening. In general, the mineralization is diffused a~ndthe Figure V.27 shows the weights of each core as a function of the distance
scale of the heterogeneities is small when compared to the scale of mining to t h e median horizontal plane of the block to be estimated, Note the
Estimation usually consists of evaluating a mean grade or a mean recovery following:
rather than determining the exact geometry of each ore lode. The: most
characteristic examples of this type of deposit a r e porphyry-type brecciated (i) the rapid decrease of the, curve S 1 ;
masses (copper o r molybdenite). Such masses may be proven on any 'sort of (ii) the screen effect cores S2, S4 o n S3, which means that enlarging
grids with drills in all directions. the estimation neighbou~rhoodwould serve no purpose.
(iii) the relative importance of the weight assigned to the global mean
gradc, A, = 0 . 1 4 3 , which is d u e to the presence of the relative
Los Bronces copper deposit (Chile), after J. Dantay and M. Durocher nugget effect C o / C I= 0.28.
(1 9 70)
T h e first campaign consisted of vertical drilling and horizontal drilling from
drives. All drill cores were cut into constant 2-rn lengths and analysed for
copper.
T h e structural analysis, given in section IV.A, Case Study 2 , resullted in
an isotropic structural model consisting of an exponential model with a
nugget effect:
with r 2 = I h 1 2 = h ~ + h ~ , + h ~ .
A second campaign was then carried out with vertical drilling, on a
regular 25 m x 25 m grid determined by the previous practical range 3 a ,
25 m . T h e object of this second campaign was to estimate the upper part of
-
the deposit (Donoso zone) which was to b e mined by open-pit. T h e mining
units were cubic 12 rn x 12 m x 1'2 m blocks, the hor~zontal dimiension
being approximately equal to one quarter of the 25 m x 25 m square grid,
cf. Fig. V.27. A single data configuration was used for the kriging ol' these
blocks. This configuration consisted of the following.
(i) T h e estimated global mean grade of the Donoso zone.
(ii) the 4 x 24 of core lengths of 2-rn support, located on the four
drill-holes S i , 5 2 , S3, S4 in the estimation neighbourhood of each
block. Taking symmetry and structural isotropy into account, the
dimension o f the single kriging system can be'reduced to 3 8 and
consists of:
o n e weight for the global mean grade; FIG. V.27. Donoso's kriging plan. Global mean grade weight = 0.143.
12 weights for each of the three cores S1, S2, S3 (because of
symmetry, core S4 receives the same weights as S2);
o n e Lagrange parameter required for the non-bias condition,
37 Chuquicamafa copper deposit (Cltile), after I. Ugarte (1972)
2 A,=l.
cr = 1 This deposit is mined by open-pit and estimated by two types of sampling.
3 94 MINING GEOSTATIST1C:S V THE ESTIMATION OF IN SITU RESOURCES 395
(i) Drilling in all directions. Drill cores are cut and analysed on 1 ~ 5 - ~ Thus the terms ?(a,P ) between two blocks a and /3 for which data is
lengths. These samples are used essentially for long term estimation, available, each with six blast-holes, are approximated by the mean value
(ii) Vertical blast-holes analysed over the thickness of 13-m benches.
These samples are used for short-term estimation of mining units.
The regularized variograms calculated from the 1-5-rn core lengths were
compared to the horizontal variograms graded over the constant 13-m
bench thickness calculated from the blast-hole data. The three-dimen.
sional regionalization of the copper grades, with the 1.5-m support assim-
ilated to a point, is characterized by a model of anisotropic, nested
variograms, specifying two horizontal directions and the vertical direction.
The presence of a proportional effect means that neighbourhood of quasi-
stationarity must be considered.
Block a B l o c k ,8
i = It o 6 j:I t o 6
Short-term estimation of a mining unit (20 rn x 20 m x 13 m block) by
blast-hole data on previously mined neighbouring blocks is illustrated in
Fig, V.28, Note that the blast-hole data available do not completely sur-
round the block V to be estimated. On the Western front, with mining FIG. V.29. Discrete approximation of the blocks (a) blocks for which data is
known; (b) block V to be estimated.
advancing from East to West, all the blast-hole data is located on the
Eastern side of block V, and the reverse is true for blocks on the Eastern
front, If the object is to estimate the mean grade of the entire block V, then the
terms 7(a, V) are calculated using a discrete approximation of V consist-
ing of 30 blast-holes on a regular centred grid within V, cf. Fig, V,29(b).
1::
On the other hand, if the object is to estimate the arithmetic mean of the
six blast-holes within V, then the terms T(a, V) are calculated using a
discrete approximation of V consisting of six blast-holes on a regular
6 m x 10 m grid positioned at random within V:
30 k-I .Y(X,-X~),
I/ Avwlable lnformot~on
over nine blocks, a a I t o 9 1
Block I/ to be estimated ?(a,V )= - y(x,, V) with ?(xi, V) =
6 1-1
FIG. V.28. Short-term kriging plan at Chuquicarnata.
Using the blast-hole data file from the benches already mined, the
For kriging purposes, the blast-hole data are grouped together inside experimental arithmetic mean 2%of the blast-holes within V can be
each of the nine neighbouring blocks already minled, cf. Fig. V.28. Since calculated and compared with the kriging estimator 2: corresponding to
the number and exact arrangement of the blast-holes within each of these the discrete approximation (k = 1 to 6). Note that the same experimental
blocks are variable, it is assumed when calculating the mean values f of the value Z:t can be compared .to two kriging estimators Z t w and Z*KE
kriging system that each block contains, on average, six blast-holes on a depending on whether the block V is considered as coming from the
regular 6 m x 10 rn grid, the grid being positioned in a uniform random Western front (the data a are then located East of the block) or from the
manner within the block, cf. Fig. V.29(a). Eastern front (the data a are located West of the block). The comparisons
1
*L
i
396 MINING GEOSTATISTICS
- V THE ESTIMATION OF IN SITU RESOURCES 397
(Z&, Z $ ) and (Z&, Z $ ) a r e given by the histograms of errors I(z*,- which depends on whether the block is considered as coming from the
, ( z $ - z $ on
Z ~ V )and ~ )Fig, V.30. Western front o r from the East~ernfront. This bias can b e explained by a
If the results corresponding t o the East and West data configurations are local East-West drift that cannot b e corrected because of the anisotropic
grouped together, the following remarkable verifications can b e made: data configuration, More precisely, as shown on Fig. V.31, t h e Chuqui-
(i) the non-bias of kriging, the mean error being 0 . 0 0 % C u ; camata deposit is bounded by a large fault oriented N-S, and t h e variability
(ii) the representativity of t h e kriging variance, ok =0.117, which is of the grades is related to a network of small faults with the same main
very close to the experimental variance of 0.118. direction. O n average, over the whole deposit, this E-W variability appears
on the mean stationary variogram as a marked geometric anisotropy in the
horizontal plane. Locally, at the scale at which the comparison (Z;, Z $ )is
made, this E-W variability may ,appear as a local drift of the copper grades
% W front
E front
(non-stationarity). In fact, as all the data are located o n only o n e side of
block V, any separation of these data by a mineralized fault will result in a
locally biased estimate; if all the: data are located on the other side of the
block, this bias will be reversed.
Small faults
T h e solution consists in taking this local E-W drift into account in the
short-term kriging system, which means introducing more non-bias condi-
FIG. V.30. Histograms of errors for the stationary case in which the kriging system tions than the single condition,
is of dimension 10. 9
For the Western front, mean error = -0.11% Cu; experimental estimation
variance=O-125 (% C U ) ~number
; of values=410. For the Eastern front, mean C= 1 A a = 1 ,
oc
error= +@11% Cu; experimental estimation variance = 0.1 11 (% C U ) ~number; of
values = 400. present for the stationary case. T h e variogram model y remains
Grouping the two fronts, mean error = 0.00% Cu; number of values == 810; unchanged, and a linear drift is considered which, in the three-dimensional
experimental estimation variance = 0.118 (% C U ) ~kriging
; variance 0 i = 0.117 space, amounts to introducing three additional non-bias conditions:
(% CU)*; mean grade of the considered zone = 2% Cu.
of the centre of gravity of the block n for which data is known and the case: what is gained in correcting the bias is lost in the kriging
block V to be estimated, cf, relation (V.10) in section V.A.2. variance.
Using the same data file coming from the same already-mined benches
considered above, the comparison of the kriging estimates Z&, 2:. and An important aspect of this short-term estimation is the simple manner
the experimental values Z*V yields (cf. Fig. V.32): in which the local.drift'was taken into account. It was enough to introduce
several additional non-bias conditions in the stationary kriging system
(i) an almost null bias, whether the East or West data configuration is without modifying the structural model. It is advisable to use such a
used; procedure when the data is anisotropically positioned and if a local drift is
(ii) an excellent agreement between the kriging variance & = 0.261 suspected, e.g,, at the edges of a deposit.
and the experimental estimation variance. Note that the value 0.261
is greater than the value 0.1 17 previously obtained for the stationary Long-term estimation of blocks B (60 m x 60 m X 13 m) from the core
samples only, is illustrated in Fig. V.33. The estimation neighbourhood
consists of a set of 3 x 9 = 27 blocks of the same size, centred on the block
=Gzi&-- W front
E front
3 to be estimated, i.e., nine blocks on each of the three horizontal benches.
Each intersection of the drill cores with each of the three blocks of the
vertical column containing B is considered separately; on Fig. V.33, these
intersections correspond to the five weights 1, 2, 3, 5, 6. Each intersection
FIG. V.32. Histograms of errors for the non-stationary case in which the knging
system is of dimension 13.
For the Western front, mean error=-0.02% Cu, experimental estimation
variance- 0.247 (% C U ) ~number
, of values= 397. For the Eastern front, mean
error = +@01% CU;experimental estimation variance = 0.273 (% C U ) ~number
; of
values = 397.
Grouping the two fronts, mean error= 0*00% Cu; number of values = 794;
experimental estimation variance = 0.261 (% Cu)', kriging variance, r r 2 - 0.261
(% Cu)'; mean grade of the considered zone = 2% Cu. FIG.V.33. Long-term kriging plan at Chuquicamata.
400 M I N I N G GEOSTATISTICS V THE ESTIMATION OF IN SITU RESOURCES 40 1
of the drill core with the aureole of 2 7 - 3 = 2 4 remaining blocks is also '- within this square), a random kriging plan was adopted for this estimation
considered separately; on Fig. V.33 these interesections correspond 1.0 the of 1 0 0 m x 100 m X 15 m blocks. T h e estimation neighbourhood consisted
three weights 4, 7, 8. A length I, corresponds to each of these inter- of 27 blocks: the block B to be estimated and the 2 6 blocks of the first
, sections, identified by its centre of gravity x, and a mean grade Z.The aureole surrounding it, cf. Fig. 11.34. These 2 7 blocks were divided into six
stationary kriging system can be written in the standard form: domains {D,, a = 1 t o 61 within which it was assumed that t h e data were
randomly located within each 1130 rn X 1 0 0 m square.
Dl is the domain consisting of the block B to b e estimated; D2 and Dj
are the blocks located above and below B, respectively; D4 is made up of
the eight blocks of the first aureole surrounding B on t h e same level; D 5
and D6 consist of the eight blocks located, respectively, above and below
the eight blocks of D4.
where there are n intersections 1, (n = 8 in t h e example of Fig. V.33). T h e data inside each domain D, consist of n, drill intersections, all
F o r t h e calculation of the terms ?(la, Is), a + P , the lengths I, are vertical, 15 m long and of mean grade Z,.In t h e general case, the six
assimilated to their centres of gravity: ?(la, lP)= y(x, -xs), V(Y # P. domains D, will all contain some data (n, >O, V a ) , and t h e kriging
Similarly, ?(la, B ) = .i;(x,, B ) where the block B is represented by a set of estimator will be the linear combination of the six mean grades 2,:
discrete points. O n the other hand, the length 1, must be taken into 6
account when calculating the diagonal terms 2; = 1 A,Z,.
,=I
As the three intersections S1, S2, S3 come frolm the same vertical drill the second member matrix. Thus, for example, -F;(Dl, B ) = T(B, B ) and
intersecting the block B, cf. Fig. V.35, then .i;(D1, D2) = .i;(Sl, S2), and 9 R
similarly for the other two rectangular terms T(D1, D3) and 7(D2, D3).
This relation holds, strictly, only when there is a single vertical drill inter-
secting each block B. For this reason, only one intersection per block was In the case of a complete configuration (nu 3 0 , V a = 1 to 6, and
retained in the classifying program of the data (nl, n2, n3 s 1); in any case, na = n 5 ==n6 = 8), the following results are obtained:
only rarely did more than one drill intersect a block.
- ,
rlU.
"
- (iii) the set of channel samples taken along the drifts which run along the
base of block B.
This provides a total of 21 weights for the 21 data sets {G,, ac = 1 to 211,
, and a stationary kriging syste~nof dimension 22. As the results of this
kriging were to be used for production planning, it was essential that they
be as accurate as possible. For this reason, the number and exact location
of each intersection S, constituting the cluster G, were taken into account
when calculating the terms Y(G,, Go) rather than assuming random loca-
tions of the clusters G, in their respective domains:
Similarly,
explains why geostatistical applications to vein deposits are not as In the plane of the vein, the three levels are about 600 ft apart, which is a
developed as those for stratiform or massive deposits. much greater distance than the largest range a 2 = 100 ft. Thus, local esti-
The first problem is to determine when, froin a geostatistical point of mations of ore resources can be made only within 100 m of workings for
view, a deposit should be considered as vein-type. A brecciated deposit which samples are available, i.e., the zone of influence of the levei's'. This
with multiple interlaced veins can be regarded as a massive deposit. The
estimation of the mean width of a single vein estimated by drifts can
sometimes be reduced, in the plane of the vein, to a two-dimensional
problem of the stratiform type, cf. the following case study of Eagle vein. If
the vein is thick enough, intra-vein estimations C~UI be made: the geometric
problem vanishes and kriging plans used for massive-type deposits can be
applied,
Data for this study consist of vein width and copper assay values taken at
approximately 10-ft intervals along three main horizontal development
drifts located in the plane of the vein. The two regionalized variables used
in the study were the vein width and the corresponding copper accumula-
tion.
Nested structures, y ( h ) = y ~ ( h ) +
y2(h), were fitted to the variograms of
these two regionalizations. The two structures are spherical with respective FIG.V.39. Local estimation of Eagle.
ranges for the width of a l= 20 ft and az = 100 ft. These two ranges can be
interpreted as pseudoperiods of the successive bulges in the vein, as shown The three vertical bands of influence are split into panels of vertical
by the thickness profile obtained on one of the development levels cf. Fig. dimension 100 f t x 2 0 0 ft. For the estimation by kriging of the mean
V,38. horizontal thickness and the corresponding mean copper accumulation of
MINING GEOSTATISTICS V THE EST,MATION OF IN SITU RESOURCES 409
each panel, the information o n the central level is grouped into three data kriging estimates). Similar results were obtained for the estimation of the
sets, S1, S2, S3, cf. Fig. V.40. S1 is the data set within t h e pancl to be quantity of metal and the mean copper grade.
estimated; S2 and S, are t h e next data sets taken along the central level.
Uchucchacua argentiferous vein (Peru) after E. Tulcanaza (1972)
The data for this study consist of samples taken along drifts and raises. T h e
samples were taken at 1-m intervals perpendicularly to the direction of the
workings o n an approximately constant support. T h e mineralization
essentially consists of argentiferous pyrite with a little galena, and the
problem was to determine the intra-vein variability of the silver grade. The
regionalization of this grade is characterized by a marked zonal anisotropy:
the horizontal variability as determined o n the drifts is greater than the
FIG.V.40. Data configuration for the estimation of a 100 f t X 200 f t panel. vertical variability as determined o n the raises. This is explained by the
vertical direction of the main faults. T h e chosen structural model is aniso-
A s the mining company planned to carry out additional sampling flor the tropic and consists of the nested1 sum of two horizontal structures (with 4-m
local estimation of all the mining panels, several kriging plans were and 70-m ranges) and a vertical structure (with a vertical 20-m range).
established to predict the improvement to be gained by different additional T h e problem was to estimate the m e a n silver grades of blocks of dimen-
sampling plans, cf. Fig. V.41. It was shown that, with a n additional level sions 5 0 m x 5 0 m or 40 m x 2 0 m in t h e plane of the vein (the constant
between each of the existing ones, the relative standard deviation q / m , of thickness of the vein is fairly well known). T h e kriging configuration simply
the global estimation of the ore-tonnage, would b e reduced from 4.ti0/0 to consists of two data sets, cf. Fig. V.42: (i) for the 5 0 m x 5 0 m block
2 . 3 % , the relative standard deviation uK/m, of the local kriging of the sampled by drifts and raises, t h e two mean grades Zd and Z, of the drifts
ore-tonnage of 1 0 0 ft x 200 ft unit panels, would b e reduced from 17% to and the raises were used; (ii) for the 4 0 m x 20 m block sampled by a single
15%. This latter decrease is not especially significant; however, all the central drift, the two mean grades Zd,and Zd2of the samples in the section
units in t h e deposit could b e differentiated by kriging, while, with the three of the drift inside the block and the two symmetric sections outside the
existing levels, only 4096 of t h e panels were differentiated (had different block were used.
For t h e calculation of the 7 terms of the kriging matrix, t h e supports of
the channel samples were assimilated t o line segments; thus, for the third
configuration, cf. Fig. V.42, the support of the datum Z d , is the line dl of
4 0 m length. Thus, T(d1, d l ) = F(4O) and y ( d l , B)=,y(20; 40), where F
and x are the standard auxiliary functions.
This assimilation of the support of t h e discrete set of channel samples to
a continuous line amounts to assuming that t h e mean grade of the line
FIG. V.41. Kriging plans with additional data. (a) Two additional drifts; (b) one
additional vertical raise. FIG. V.42. Kriging configuration at Uchacchacua.
410 MINING GEOSTATISTICS V T H E ESTIMATION O F IN SITU RESOURCES 41 1
segment is perfectly estimated by the mean grade of tlhe channel samples unique as well as its mean grade z ~ T.h e probabilization? of the global
within it. It is then necessary to add a correction terln to the preceding error ( z D- - z $ ) is obtained, no longer by considering various particular
kriging variance. This correction term corresponds to the estimation of the deposits D,, but by considering various locations of the data configuration
line segments by their component channel samples; thus, for the third as a whole over the unique and fixed deposit D: a particular global
kriging configuration, 2; =A&, + (1 -A )Zdlr the three segments of 120 m estimate z & corresponds to each particular position i of the origin of this
total length are estimated by 120 channel samples spaced at 1-m intervals, data confi,guration, cf. Fig. V.54, the example of two surface estimates
and the variance correction is a,2/120, where crz is the extension variance S * ( x o ) andl S * ( x b ) linked to two particular locations xo and xb of the data
of a central point to its segment of influence of 1 m length, cf. section grid origin. Thus, the global estimation error ( z D - z $ ) has been prob-
V.C,2, the principle of independence of elementary errors. For this third abilized, and the global estimation variance is defined as E { [ Z D- z:]').
configuration, a: = 11.8 and ( ~ , 2 / 1 2 0 = 0 . 3 ,which gives an estimation However, the distribution of the error ( z D- z $ ) will always remain
variance E{[ZB-z;]') of the mean grade Ze of block B equal to 11.8 + inaccessible since, in practical applications, only one particular location io
0.3 = 12.1. of the data configuration over D is available. Hence,.all confidence inter-
vals (e.g., the +2a Gaussian standard) attached to the global estimation
variance will have but a mere conventional meaning.
V.C. VARIANCE OF GLOBAL ESTIMATION Choice of the structural model -y(h) or C(h), used in the calculation of the
global estilnation variance
V.C.1. Definition of the global estimation variance Consider [he "local variogram" on D, i.e., the variogram 2 y ( ~ , ( hthat
)
A global estimation differs from a local estimation in two important would be available if the domain or deposit D were perfectly known, cf.
aspects. section III.B.7, formula (111.40):
(i) The distances involved in global estimation are usually greater than
the limits of quasi-stationarity defined by the structural analysis, and
sometimes include several heterogeneous mineralizations.
(ii) The globaI estimation does not have the repetitive character of the In practice, this local variogram is estimated by the experimental vario-
local estimation. As the deposit is unique, there is only one global gram, cf. formula (III.39),
estimation error and it is, thus, difficult to probabilize it.
The first aspect means that a global estimator should be built by combin-
ing various local or partial estimators, as shown in section V.A.3. I n where N' is the number of pairs of data (distant of vector 11) available over
general, kriging techniques can be used to provide local kriged estimates the whole domain D.
within each zone or neighbourhood of quasi-stationarity; the global esti- As this experimental variogram 2-y%(h),and, thus, its fitted model
mator is then defined by the weighted sum of these: local estimators, 2 y ~ ( h are
) defined over the whole deposit D, they may differ from the
according to formula (V. 18). quasi-stationary models x ( h ) used in the local estimation of each partic-
The problem of the definitian and calculation of a norm of t'le quality of ular zone or block V , o f the deposit D. Often, in mining practice, these
a global estimation remains. In lacal estimation, the estimation configura-
tion for a particular block Vi can be considered as repeating itself identic-
t Nore The global estimation error probabilization as well as the conceptual basis OE the
ally for a large number of blocks V within a zone of homogeneous global estirnation variance calculation is related to an approach different from the approach
mineralization. In such a case, the probabilistic interpretation of the local of regionalized variables considered in this book. This approach is called "transitive
-
error ( Z V , 2 5 , ) as a particular realization of the random error (Zv,- Z $ , ) theory", cf. G . Matheron (1965, p. 17) and (1971, p. 9), and results in calculation algorithms
and solutions remarkably similar to those provided by the regionalized variables theory. In
has a physical meaning, and the local estimation varianlze E { [ Z v-z$,,)'}
, the present text, this transitive theory approach is briefly outlined in section V.C.4, which is
of a block V was then defined. In global estimation, the deposit D is devoted to the surface or volume estimation problem.
."
f-
'-
4 12 MINING GEOSTATISTICS h V THE ESTIMATION OF IN SITU RESOURCES 413
quasi-stationary local models y , ( h ) are linked by a proportional effect, cf. " T h e difficulty lies in the calculation of the N ( N - 1 ) error covariances
section III.B.6: the local model y,*(h) corresponding to the estimator E{[Zv,- Z$,][Zv, - Z $ , ] } .Such a calculation quickly becomes tedious
m* = zg of the mean grade of I) should then be taken as the global rnode] when N is great. T h e practical solution consists in expressing the global
to b e used in the calculation o f the global estimation variance. ,, estimation error [ Z D- Z g ] as the s u m of n (with, in general, n G N )
I
with o:s=2H(i/?; l/2)-F(1; 0,according to the formula for o:, cor- In this case, each of the elementary errors [ Z ,- Z ( x , ) ] has its own
responding to Fig. 11.50 in Section II.E.3. estimation variance r;,, which depends o n the particular position x , of the
Note that this sum of elementary independent errors, Z,, - Z ( x , ) , is drill-hole In the grid square s,. Over a large number N of such grid squares,
related to the approximation x will take all possible locations inside the grid square s and the elementary
estima'tlon variances a;,, will have the following mean value:
{T\
N>50 gives an excellent approximation. extension E { [ Z ,-z(x)]'} of a point datum Z ( x ) to the mean value 2, is,
by definition, the dispersion variance D ~ O / ofS )this point datum in s. T h e
0 0 X X X X global estimation variance is then written as
O O d X X X X X
0 X X X X X X X
O X X X X X X
$
416 MINING GEOSTATISTICS I
V THE ESTIMATION O F IN SITU RESOURCES 417
s*t
T h e dispersion variance D"O/S) is expressed very simply by means of the i.e., to the assimilation of the global estimator Z$, weighted sum of local
auxiliary function F: kriging estimators, to the polygclns of influence estimator. When N is large,
D 2 ( 0 / s ) = y(s, s)= F(1; I), there is very little d~fferencebetween these two global estimators. Never-
theless, at the local level, the kriging estimators should always be preferred
if s is a square of side 1. ' to the polygons of influence estimators (because the local kriging variances
The principle of uniform random location of the data within the grid unit are, by definition, minimum).
can be applied in a similar manner to data configurations in one, two or
three dimensions. In three dimens~onsfor example, for a parallelepipedic
grid unit v containing, on average, o n e datum o f support v'<< u and located
in a uniform random manner within v , t h e global estimation variance of a
large volume D cons~stingof N grid units v can b e obtained by direct
combination:
1 1
E{[ZD-z;]'}=-D2(v'/v)=-[y(o, v)-y(v', v')]. (V.39)
AT N
Irregular grid Consider the same sedimentary deposit, this time with the
N pos~tivedata on an irregular grid, cf. Fig. V 45 (the N surfaces of
influence s, are unequal). Note that under no circumstances can a pref- FIG. V.45. Direct combining of elementary errors from an irregular grid.
erential grid, with all its associated risks of bias, be assimilated to an
irregular grid, cf. section III.C.l.
T h e global estimator ZT is the weighted sum of a certain number of Other configurations for directly combining elementary errors
local estimators obtained, for instance, by local kriging; note that these I n one drmension, let L be a line sampled by N points regularly spaced at
local estimators may not be defined on the surface of influence s, of each intervals I, cf. Fig. V.46(a). This line may represent, for example, a drift of
positive hole. By expressing the global estimation error [ZS-z;'] as the length L with a negligible transverse section which has been sampled by
sum of the N independent elementary errors [Z,, -Z(x,)], of extending the vertical channel samples spaced at regular intervals I. T h e line L thus
thicknesses Z ( x , ) to their respective surface of influence s,, the follolwing consists of N segments 1 each having a central, centred point datum. T h e
formula is obtained: global estimation error [ZL -Z,t ] of the mean characteristic on L is consid-
ered as the sum of the N eleme,ntary errors of extension of the central data
Z ( x ) to their segments of influence I. A s a first approximation, these errors
can b e considered independent, resulting in the lollowing expression
where S = ( l / N ) 1, s,, and u:,, is the variance of extension of a central similar to (V.37):
point datum to its surface of influence. This variance is expressed by means
of the two auxiliary functions H and F,cf. section II.E.3, if the surface s, is
a rectangle. If s, is not rectangular, it can, for instance, be assimilatedl to a
square of the same surface area s, = I ? ,and, thus, with
a:,, = H ( W ; 1,/2)- F(1,; I,). a:, = E{[Zl - z ( x ) ] ~ } = ~ X ( N ~-) F(1).
Note that t h e previous expression (V.40) of the global estimation vari-
ance is related to the approximation The same line L = N l may be sampled by ( N S 1) points in a closed
arrangement, each segment 1, having two extreme data Z ( x , ) and Z(xi+l),
as shown o n Fig. V.46(b). Two contiguous segments I, and li have o n e
datum in common and, as a consequence, the two corresponding local
418 MINING GEOSTATISTI(:S V T H E ESTIMATION O F IN SITU RESOURCES 419
estimation errors can n o longer be considered as independent. It can be the spacmg ( / / a> @ 5 , the number N must be great (N > 3 0 ) to provide a good
shown that, for N r 6 , a closed arrangement of (N+ 1 ) points can be enough approximation. .".~.
assimilated, with a n excellent approximation, to a centred arrangement of
N points, and formula (V.41) can then be applied. This practical rule can
also be applied for closed arrangements in two o r three dimensions (with
N 3 36).
the elementary error of extension of the mean grade of a drlft L, to its can then be obtalncd by directly combining the elementary estimation
rectangle of influence s, = L, x I, is calculated using the auxiliary funci ions variances ui,,,
:
a n d F, o r read directly from charts no. 8 and 18, cf, section II.E.3. The
global estimation variance over the surface
N
s= C s, where, if the median section s, is assimilated to a square of the same area,
,=I 1: = St,
can then b e obtained by directly combining the elementary estimation c;", = 2,y(hl/2; l ? ) - ~ ( h ,1;; ) - ~ ( 1 , ; 1,).
variances a 2~ , ,T.h e result is a formula similar to (V.40):
Note In the last two examples, corresponding to Fig. V.48 and V.49, the
elementary extensions of the mean grade of a median drift o r a median
level to its rectangle o r parallelepiped of influence were considered. But
where both the median drift and the median level have been estimated with a
certain estimation error. Thus, a term corresponding to this estimation
error must be added to :he estimation variances given by formulae (V.43)
In three dimensions, consider the estimation of the mean grade over a and (V.44); this results in combining the various terms of the global
volume V divided into N rectangular parallelepipeds v, = s, x h,, each of estimation error.
which being informed by its median section of surface s,, cf. Fig. V.4'9. The
variance a;,, of the elementary error of extension of the mean grade of the
median level s, to its parallelepiped of influence v, = s, x h , is calculated 2. Coinbining rlle t e r m of a gilobal estimation error
using the auxiliary functions x a n d F, o r read directly from charts no. 9 and T h e data available for the estimation of a three-dimensional phenomenon
19, cf. section II.E.3. T h e global estimation variance over the volum,e is very o f t e n concentrated along alignments (e.g., pieces of core along
drill-holes, sampling along drifts) and planes (e.g., planes of vertical
sections along profilcs of vertical drill-holes o r horizontal levels of the
drifts). Thus, the data available for the estimation of the mass shown on
Fig. V.50 are vertical channel samples taken along horizontal drifts on each
of the three horizontal levels.
T h e procedure f o r estimating the mean grade of this mass can be consid-
ered as consisting of three steps corresponding to three terms of the global
estimation variance.
(i) The estimation o f the drifts by the channel samples. The estimation Slice tcprm T h e final term is the error of estimation of t h e three slices OF
variance corresponding to this step is called the "line term", T . the mass by their median levels supposed to b e perfectly known. Each slice
(ii) The estimation of the levels by the drlfts which are supposed to be may, for example, be assimilated to a rectangular parallelepiped estimated
perfectly known. The estimation variance corresponding to this step by its median level Sk, and the slice term is then obtained from the"direct
is called the "section term", Ts. combining formula (V.44):
(iii) The estimation of the slices of the mass by their median levels,
which are supposed to be perfectly known. T h e estimation variance
corresponding to this last step is called the "slice term", Tv.
As a first approximation, the three estimation errors corresponding to where
these three steps can be considered as independent, and the global estima-
tion variance of the mass then consists of the sum of the three preceding
terms:
and &", is the extension variance of the median rectangle of surface Sk to
its parallelepiped of influence u k .
The following example o f the calculation o f these three ternis is based on
Fig. V.50.Another example is given in the study of Ity Mount Flotouo gold Remarks T h e terms combining to give formula (V.45), (+;=
deposit in section V.C.5. Ti + Ts + Tv, express the intuitive notion that the estimation would not be
improved by increasing t h e density of channel samples in the drifts (Ti
Line term If each drift is assimilated to a line L, along which channel decreases) without also increasing the number of drifts o n each level (Ts
samples are taken at regular intervals, then the estimation varlance of this decreases) and the number of levels within the mass (Tv decreases). T h e
line can be expressed using the direct combining formula (V.41): line term is often negligible with respect to the section ( T s )and slice ( T v )
terms. If the costs of channel sampling and drift development are known,
the two-term formula, Tl + Ts, can be used to make optimum use of the
sampling budget on any particular level.
where t i=N,x 1 and c& = 2X(1/2)- F(1).
The estimation of all the drifts (of total length L = 1,k, = ,Vl) by the set
of N = S I N , channel samples, regularly spaced at intervals 1 along these
drifts, is thus characterized by the line term
V.C.3. Estimation of a q u o t i e n t o r a p r o d u c t
Warning Ttie product and quotient of two random functions underline
T h e quantity to be estimated is often in the form of a quotient o r a product difficult problems of existence and convergence. Thus, the quotient of two
of two quantities estimated separately. Consider, for instance, a sub- normally-distributed variables is a variable with a Cauchy distribution
horizontal sedimentary deposit of surface S and variable vertical thickness which has no moments (no expected value, no variance); the expectation of
t(x) as shown on Fig. V.52. F o r each vertical drill at point x, two additive a limited expansion is not always t h e limit of the sum of the expectations of
variables are defined: mineralized thickness t(x) and accumulation a ( x ) = the terms, etc. T h e following demonstrations are more symbolic than
t ( x ) . z(x), where z ( x ) is the grade measured along [(s).This grade 2 (x) is rigorous, which should b e sufficient for the practitioner of mining estima-
not an additive variable because it is defined on a variable support t(x), and tion. All the following results can b e established rigorously for the product
s o the regionalization of t(x) cannot be studied directly, cf. section III.C.2. and quotient o f two lognorma'lly distributed variables.
T h e mean grade Zs of the deposit is estimated by the quotient of the two
global estimators of mean accumulation and mean thickness:
Estimation variance of the product P = A B
A$
zs= 1 / S S s a ( d d x estimated by Z $ =-. T h e true global values P, A , B a r e considered to be known, and their
11sJS t(x) d x Tf estimators P*, A*, B* are interpreted as random variables:
Similarly, if the mineralized surface S is estimated by S*, the estimator
of the global v o l u n ~ eof o r e is in the form of the product of the two global
estimators S* and T z : where & A , E B , c p are the random estimation errors.
T h e estimations of the factors A and B are assumed to b e unbiased, i.e.,
V = S x Ts =
Ist(x) d x estimated by V* = S*x T $ . E{E~ = E{cB}
}
E { E ~ are
} , known.
2
= 0 , and their estimation variances, a: = E{E:} and cr8 =
Thus,
and
In practice, the true values A and B in formula (V.46) are replaced by If the two variables a ( x ) and b(x) are intrinsically coregionalized, i.e., if
their unbiased estimators A* and B*. T h e calculation of p , ~ will be the two structural functions y a ( h ) and y b ( h ) are proportional, cf. section
detailed later. III.B.3, formula (III.30), and if the two configurations for the estimation of
their mean values A and B are identical (similar in practice), then it can be
Estimation variance of the quotient Q = A / B shown that the coefficient pas is equal to the standard correlation
coefficient pab between the two variables a ( x ) and b(x).
Using the preceding notations, the estimation of the quotient can be If the two variables a ( x ) and b(x) are not intrinsically coregionalized,
expressed as then the covariance o f the two errors E { [ A- A z ] [ B- B * ] } must be cal-
culated, i.e., the coefficient p , ~ is calculated from the cross-structural
function
If the relative error E ~ / isBsmall with respect to 1 - and this is generally Ya~(h)=~~{~a(~+h)-a(+~)][b(x+h)-b(x)]}.
the case for global estimation - the following l i m i t e d ' e ~ ~ a n s i oisn obtained: W h e n the two configurations for the estimation of the two factors A and
B are. identical (or very similar), a "reduced difference" variable can be
defined, d ( x ) = a (x)/A - b(x)/B. Its structural function can then be
and determined:
similarly,
E { [ A- A * ] [ B- B * ] }= % ( r o b ) .
T h e relative estimation variance of the quotient A / B can then be written
as
Note first that if the two estimations of the two factors A and B are
independent, the covariance E { [ A-A*][B - B * ] )of the two errors is null,
C
T h e relative variance U & / Q ~is obtained directly by applying t h e operator (ii) econoAic cut-off grade, e g., in d e p o s ~ t ssuch as porphyry-copper,
8 to the structure yd of the "reduced difference" variable. all zones wlth grades greater than 0.4% C u will be considered as
In the case of the estimation of the product AB, the "reduced sumw mineralized zones;
variable, s ( x ) = a ( x ) / A+ b ( x ) / B would b e defined, its semi-variogram
L
(iii) with the aid of geological interpretations made on planes o r cross-
y,(h) would b e calculated, and the relative estimation variance of the sections.
product could then, similarly, b e written as
When this estimation of the geometry of the deposit is directly related to
geological hypotheses (metallogenesis, tectonics, etc.), the geometric error
is o f a deterministic type not favourable to its quantification (this does not
mean that the error will be large o r that such geological methods of
estimation should be disregarded). In such a case, it is advisable to examine
these geological hypotheses and possibly consider different ones, There
will be a different global estimator of the ore tonnage for each inter-
V.C.4. Global e s t i m a t i o n o n a d o m a i n of u n k n o w n g e o m e t r y pretation of the geology which returns a certain random character to the
est~mation.Thus, even i f the variance of the geometric error cannot be
U p to this point, the geometric error has not been considered, i.e , the evaluated, at least its limits can be.
error d u e to t h e ignorance of the exact limits of t h e mineralization The However, the geometric error may intrinsically have a purely o r partially
volume of the area to be estimated was assimilated to the sum of N positive random character which can be modelled.
blocks o r zones of influence v,, i.e., the estimator of the total mineradized
volume V was taken as the sum (i) From the experimental knowledge of similar deposits (e.g., random
location in space of mineralized lenses of random volume and
geometry). T h e morphology of the deposit can b e simulated several
times. T h e available data configuration can be applied to these
various simulated deposits, and an estimation variance can, thus, be
of the volumes of these blocks. In doing so, an error ( V - V*) was incurred determined. It should b e stressed that such a variance depends
and must b e evaluated, for example, through a variance of the geometric heavily o n the model adopted for the morphology of the deposit.
error. This geometric error ( V - V * ) not only affects the volume and the (ii) There are, however, some cases, frequently encountered in mining
o r e tonnage, but also the estimations of metal tonnage and mean grade of applications, which allow an approach entirely free from any a priori
t h e deposit D. Thus, it is equally important to combine the influence of the hypothesis o r model. This is the "transitive theory" approach, cf. G.
geometric e r r o r with that of the different quality errors, e.g., the quality Matheron (1971, p. 9), the main aspects of which are outlined in the
error d u e to t h e regipnalization of the grade variables in the total volume following subsection.
v.
The transitive approach Let k ( x ) b e the indicator variable of the mineral-
ized volume V to be estimated. This indicator is a n all-or-nothing variable
1. Variance of the geometric error defined as
T h e determination of the limits of the mineralization is a constant problem 1 if thepoint x is in V,
in global estimation of a deposit. It sometimes happens that some of these O if not, xiZ V. (V.52)
limits a r e known a priori, e.g., the limit of a lease, a fixed depth for drilling
and estimation, clearly defined vein walls, etc. O t h e r limits require esl.ima- The mineralized volume V is then the triple integral, in the three-
tion by o n e of t h e following methods: dimensional space, of this indicator variable:
In the transitive approach, the structural function "geometric c o ~ a r i o g r a m * ~ For each new position xb of the grid, there is a corresponding estimator
of the volume V is defined as S*(xb I and, thus, a new realization [ S - S K ( x & )of] the estimation error of
the surface S, cf. Fig. V.54. O n e possible way of defining the quality of the
~ ( h ) = kj ( x ) k ( x + h ) d r , (V.54) estimator S* is to take the mean value of the quadratic error [ s - ~ ~ ( x o ) ] "
when the origin xo o f the grid takes all possible locations inside the
where h is the standard notation for a vectoi in the three-dimensional elementary grid surface a l a z . This mean quadratic error is expressed as
space for the estimation o f a volume V, o r in the two-dimensional space for
the estimation of a surface S.
The real unknown volume (or surface) to b e estimated is simply the
value of the geometric covariogram at the origin: V = K(0). Note the similarity of this expression for the mean quadratic error to that
In the two-dimensional case, when the surface S is estimated by a
regular grid (a,, az), the origin of which is located at the point xo of
-
of the estimation variance E{[S s * ] ~ Using
} . this similarity, even though
no probabilistic interpretation is involved in this "transitive" definition, the
coordinates (xo,, xo.), the estimator S* of the surface area can be written, notation u: for the estimation variance can b e used, i.e., in one-dirnen-
cf. Fig, V.53, as sional notation:
+m +w
S*(xo)=S*(xou,xov)= ala2 1 1 k(x0,, i-p, . a , , XO, + p , . n2),
pu=-m p,=-m
p, and pu being two integers defining the various nodes of the regular grid. This mean quadratic error, called the "variance of the geometric error", is
If n ( x 0 ) is the number of positive drills, the estimator S*(xo) is then written a funciion of the available data grid; here, a = ( a , , a d .
as S * ( . ~ O = )n l a 2 . n (xo).Note that this estimator depends on the choice o f If S*(xo) is replaced in the integral o f (V.55) by its one-dimensional
the origin within the elementary grid surface a , a 2 , cf. t ig. V.53. expression
+9
S*(xo)= a C k(xo+pn),
P 3 -a
Using this property, the first term in the limited expansion of the S, the more jagged (i.e., the longer) the perimeter, the more difficult is the
expression r r i ( a ) can be evaluated when a tends towards zero. This first surface estimation.
term is sufficient to give the order of magnitude of the geometric error When a portion of the perimeter is perfectly known (e.g., the limit of a
variance. lease), this portion must not be considered in the calculation of the values
For the two-d~mensionalproblem of the estimation of a surface from a 2 N , and 2 N Z , and this will reduce the relative variance cr;/s2by approxi-
regular, rectangular grid ( a l ,a > ) ,it can b e shown (cf. G. Matheron (1965, mately the ratio of known perimeter to total perimeter.
p. log)), that the relative estimation variance can be expanded up lo the
first order as
,Posltlv
.
where vi = li Si,
I
hole
The estimation variance o f the mean grade Zs = ASITSby the quotient V.C.5. Case s t u d y
Zg== A ~ * / T : *of the previous estimators defined o n the known surface
S* is obtained from formula (V.48): Globai estimation of fry Mt Flotouo deposit, after Ph. Fornlery (1963)
This case study is well known to geostatisticians and has been used in
courses a t the Centre d e GCostatistique of Fontainebleau since 1968. It
concerns the Ity M t Flotouo deposit in the Ivory Coast, w h ~ c hconsists of a
The quantity of metal in the deposit is the product Q = A s .S of the gold-bearing crust of laterite and clay. T h e thickness of this crust varies
mean accumulation AS and the true mineralized surface S. T h e global from 3 m t o 5 m with a slight overburden. T h e deposit was sampled by
estimation variance of Q by Q* = A $ . S* is obtained from formula vertical pits o n a regular 20 m x 30 m grid, cf. Fig. V.58. Channel samples
(V.49): were taken down each pit along the entire mineralized thickness.
V d = T s x S x d by V * d = T $ . S " . d is X X X X X
-
X X X
I
X
-
X
X x
-
X X
X
-
x
where the density d of the ore is supposed to be constant and known. If this
density, d(.r), is regionalized and sampled along the drill-holes, it should : I 2 3 4 5 6
then be studied by means of a new accumulation variable b ( s ) =
a (x) . d(.r). FIG. V.58. Data configuration on Ity deposit
approach by means of f o r n ~ u l a( V . 5 7 ) is used to characterize the surface Direct cornbitling of eletnenlarp errors. As the estimator A* is simply the
error: arithmetic mean of the n = 5i3 positive accumulations, the global estima-
(i) the estimated mineralized surface is S* = n b l b 2= 58 x 20 x 30 = tion errors [ A S S -A*] is made u p of the 58 elementary extension errors of
34 800 m *;
a central accumulation datum, to its grid of influence ( b l , bz). Thus, from
(ii) the polygonal contour of S * , including the NW gap, consists of formula (V.37), the global eslimation variance of the mean accumulation
2N1 = 44 and 2N2 = 20 grid elements ( b l , b2); o n t h e known field S* is
(iii) the order o f magnitude of the relative estimation variance o f the
surface is then
with
Table V.5 gives the variances u&, as well as the variances weighted by and D ~ ( o / sis) the dispersion variance of the point accumulation in S, and
the square L:, of their length of influence, for each of the six lines L,. T h e is estjirnated by the a priori experimental variance of the 58 'positive
section term is expressed by formula (V.43): accurnulation data, i.e., ~ 2 ( 0 / ~ ) = 1 T~ h0e . border tern1 is* then
~ { d d i )=~0.0014,
}
Taking t h e surface error into account, t h e re1ative"istimation variance
of the mean accumulation is finally given by formula (V.61):
where S* = b 2 .Ci Li,
i.e., .
where &S2= 0.0014 is the relative estimation variance of the surface, t Recall that, since the regionalized variable is the relative accumulation, the calculated
varlauces are also relative.
VI SELECTION AND ESTIMATION OF RESERVES 445
practice, amounts to saying that the sampling of G by the set of drill cores (ii) an experimental dispersion variance s 2 ( c / ~ ) ;
is unbiased - no particular zone of G is preferentially sampled. (iii) a certain shape, asymmetric for example, which could b e fitted to a
lognormal distribution. hi.
Represenrativity of an experimental histogram
Consider the grade z o o n the abscissa: the hatched area represenis the
It is very important to ensure that the experimental histogram of the grades proportion of core samples with grades z , 3 zo.
of the available core samples is truly representative of the histogram (called Supplose now that the true grades z,(x) of all the mining units of support
"local over G" from the terminology of section 1II.D 7) that would result u are known; t h e histogram of dispersion of the grades r , ( x ) can then b e
from the set of core samples obtained by dividing the entire deposit G into constructed as o n Fig. VI. 1. This histogram has
drill cores. Thus, the experimentally available core samples should be
uniformly distributed in the volume G ; this will be approximately so if the (i) a mean equal to the mean of the core ;ample grades;
drilling grid is regular over G. But if the drilling grid is irregular, with, for (ii) a n experimental dispersion variance s-(c/G), (Note that s ' ( v / G ) <
example, significant gaps o r clusters of drills, it is advisable to assign an s ' ( c / ~ ) since, in general, the support o is much larger than the
adequate weight to each grade value when constructing the histogram. This support c.);
will prevent a non-representative mode appearing in the rich-grade area of (iii) a certain shape, symmetric for example.
the histogram if the drills tend to be concentrated in rich zones. T h e Consider the cut-off grade zo; the dotted area represents t h e proportion
weighting of the grade values will, of course, influence the mean and of panels v with true grades z, 2 ZO.
variance of the histogram. In particular, the mean of the weighted histo-
gram will no longer be the arithmetic mean of the data set available over G,
but it will be closer to the true mean grade over G (correction of the bias
due to the preferential concentration of drills). It shouId be ensured that
the new variance agrees with the theoretical variance of dispersion
D ~ ( c / Gof) the grades of support c in G ; if the mineralization is globally
homogeneous over G, then D ~ ( c / Gcan ) b e calculated from the structural
function (covariance o r variogram), cf, section II.C.2. Among other possible
declustering weights, o n e may consider to take as weighting factor for each
of the data values z, either
(i) its volume of influence in G, in which case the mean of the histogram
is considered as the polygonal influence estimator (unbiased but not
0 rn* I0 5 % Cu
optimal) of the mean grade over G ; or
(ii) the weight attributed to it in the global kriging of G , in which case FIG.VI.1. Influence of support on recovery.
the mean of the histogram is considered as the kriged mean grade
of G. Just as the semi-variograms y, and y, regularized o n core-sample
lengths c and panels v , respectively, are different, so are the two histograms
Selection of units of support u and the two hatched and dotted areas of these histograms, T h e r e may b e a
Consider the histogram of Fig. VI.1, which is supposed to be representative non-neglig~bleproportion of core samples c with grade z, 3 z o = 5% Cu,
of the dispersion in G of the grades z, of core samples of support c. This but there are no panels v with a mean grade z , 3 5% Cu. F o r high grades,
histogram has z o > m,the hatched area of the experimental histogram of core grades may
seriously overestimate the possible mined recovery in both tonnage and
(i) an experimental mean m*,which for the moment is supposed to be mean grade and, correspondingly, underestimate the tonnage of metal left
equal to the true mean m of the deposit; this amounts to neglecting. in the waste, i.e., underestimate the area corresponding to the panels with
the global estimation error of the deposit; true grades z, < zo.
44 8 MINING GEOSTATISTICS VI SELECTION AND ESTIMATION OF RESERVES 449
In mining, the actual selection is made on panels v and not o n core '" the grades Z, of the mining units of support v (cf, section II.C.2), either
samples c, and for the estimation of reserves, it is essential to take [he support formally by the general formula (II.36), D'(v/G)= y ( G , G ) - y ( v , v ) , o r
v of the selec~iorlwrit rrlro account. Otherwise there is a r ~ s ko f bias w ~ t h expcrimentally by means of Krige's additivity relation (II.37), D'(U/G)=
serious economic consequences. D2(c/~)--D'(C/V). T h e two dispe~sionvariances of the core grades in G
This remark is not so trivial as it may sound. For example, consider a . and v can be obtained e~perintentally.
sedimentary deposit sampled by vertical drills analysed over the total T h e mean over G o v h e grades of t h e units v can b e estimated, for
mineralized thickness. If the mean mineralized thickness is f, the example, by the mean m * of the grades of available core samples. But the
support c of the data is a core of section 4 and mean length as shownr first two moments, mean 07 and dispersion variance D'(v/G), are not
sufficient to determine the distribution o f the grades Z,.It is also necessary
to know the type of function.
(i) In some cases, this distribution may b e known a priori; for example,
as a result of mining, o r from blast-holes o n benches, or from zones
that have already been mined, cf. section II.C.3, the case study of the
block grade distribution at Chuquicamata.
(ii) In other cases, an hypothesis of permanence of law can be postu-
lated, and the shape of the histogram of experimental data Z, can
be adopted for the distribution of Z,. Thus, if a lognormal dis-
tribution can be fitted to the histogram of the Z,, the 2, are
supposed to be lognormally distributed with the variance D'(v/G).
T h e problems involved in determining one distribution (2,)from
another (Z,) defined on a different support by means of
FIG.VI.2. Estimation using polygons of influence permanence-of-law hypotheses will b e treated later, in section V1.B.
If the polygonal influence method is used, then the mean grade of the
central drill core S, is attributed to each polygon v,. Now, if th: cut-off
VI.A.2, influence of t h e level of information
grade io is applied to the estimates obtained by this method, the estirnated
recovered tonnage will correspond exactly t o the hatched area of Fig. V I . l , In the preceding section, VI.A..l, we were concerned with the estimation of
with all the associated risks of bias. T h e results of thc case study given in the histogram of the true mean grades Z,, defined on the support v on
section VI.A.4 are particularly good illustrations of this point. which selection will actually be made. But is it really useful to know this
Another typical example concerns the method of estimation by inverse histogram of true grades?
squared distances and, more generally, all estimation methods which do In fact, except in rare cases, these true grades z, are unknown a t t h e time
not take account of the particular geometry of the panel to be estimated. of selection, and selection is made on t h e estimates z:. This means that the
For a given data configuration, such a method would give the same esti- actual recovery will consist not of the units with true grades z, 2 zo, but of
mated value regardless of the estimation support v. Accordingly, if !;elec- the units with estimated grades zz 2 t o corresponding to the dotted area in
tion is m a d e o n these estimated values, the result will not depend on the Fig. VI.3.
size v of t h e selection unit. Now, it is obvious that mining a deposlt with a This dotted area, corresponding to t h e units actually recovered (i.e.,
small hammer (very selective mining with a very small support c ) will yield those for which iz 3 zo), differs from t h e hatched area corresponding to
results very different from block-caving (the mining unit v being very perfect selection m a d e o n the true grades (z, a z o ) . This difference
large), see P. Switzer and H. Parker (1975). becomes more important as the deviation between the estimators 2: and
Within any homogeneously mineralized zone G (at t h e scale of mining), the true values i, becomes grleater, i.e., as the estimation variance E{[Z,-
geostatistics can be used to calculate the variance of dispersion D'(v/G) of Z;l2} becomes larger. Now kriging minimizes this estimation variance.
450 MINING GEOSTATISTICS VI SELECTION AND ESTIMATION OF RESERVES 45 1
Influence of technological conditions be related by o n e o r the other of the following two relations, called
In the preceding selection, it was supposed tha~tit was possible to recover "smoothing relations".
all units v with estimated grade z: greater than the cut-off grade zo, If the dimensions of the field G are large a n d if the available informatior.
regardless of the position of v in the field G. In practice, this is not always is sufficient to provide a global estimation variance u $ c which is negligible
possible; the actual recovery of a unit also depends on its accessibility and with respect to t h e mean local estimation variance&, of a unit u, then
the proposed mining method. In an open-pit operation, for example, a
block can be mined only if all the blocks inside a cone with its vertex o n this
block have already been mined; in underground mining, a new shaft will b e (the notations & and 9 are defined in the following paragraphs).
sunk only if sufficient global reserves are accessible from it. In fact, tech- If the kriging procedures that provide the estimators ZZ, can be assim-
nological considerations define a first rough selection, from which a second ilated to a complete kriging procedure, i.e., to a unique kriging procedure
and more precise selection (by cut-off grade, for instance) can be made. using the same data configuration comprising all the data available in G,
then
7
D'(v/G) = D? ( V / G ) + ~ ~ , , = - (D?
T ~(u/G)+&,,
~ (VI.2)
system ((VI.1 o r (L7.2)), there is no restriction that u& 2 2 p , , SO that it is It is then 'poss~ble to evaluate the increase in the recovery of the
7
not necessary that & r 2 p . However, the smoothing effect . D ~ * ( v / G ) r resources of a deposit by improving local estimation: a decrease in a K , will
D'(v/G) has always been observed in all data configurations encountered result in the histogram o f the z;, (used by the actual selection) becoming
in mining applications and for the standard variogram o r covariance closer to the histogram of the true values z,.
models. 2. Very often, at the time of the study of the evaluation of reserves, the
T h e kriged dispersion is smoothed with respect to the true dispersion ultimate data (e.g., blast-holes in an open pit) from which the estimators
and the amount of smoothing increases as the mean -2 u ~ , o fthe local kriging z:* used in selection will be constructed, are not yet available. This means
variance increases, i.e., as the estimators ZE,(x) of Z,(x) become worse. that the only histogram available at this stage is that of the estimators zg,,
This smoothing efiect means that the histogram of the kriged values constructed from the data available at the time of the study, rather than the
z;,(x) will display more values around the mean (close to m * ) but less histogram of the z:* on which the actual selection will b e made. This first
extreme values than the histogram (unknown in practice) of the true grades set of data 1s often very inferior to the ultimate set of data available for
Z,(x), cf. Fig. VI.3. This may be awkward when the cut-off grade is high, selection, e,g.. the study o f evaluation must be performed from drill-hole
z o > m*. core data, while actual selection is based on a dense grid of blast-hole data,
Note that the smoothing effect expressed in r e l a t ~ o n s( V I . l ) and (VI.2) is plus important visual information from the mining front. As a
characterist~cof the krig~ngestimator. T h e linear estimator obtained from consequence, the two estimation variances u? = E { [ Z V - ~ z , ] Z }and
the polygonal influence method, for instance, has a tendency to over- E{[Z, -z:* 12},as well as the two histograms, may differ considerably.
evaluate the real dispersion D 2 ( u / G ) , cf. Fig. VI.9 in section VI.A.4 Note that from the fact that a? i: E { [ Z ,-Z:* I?), the dispersion vari-
ance D2**(v/G) of the ultimate estimators Z:* will generally be greater
than the d ~ s p e r s ~ ovarlance
n D$* ( v / G ) of the kriging estimator ZE, cor-
Practical estintation of the recovery f responding to the primary information.
T h e tonnage recoverable f r o m future selection (z:* zo) corresponds to
A t t h e time of selection, the true values z, are generally unknown, and the the dotted area o n Fig. VI.4.
actual selection is made on the estimators z,* (preferably on the ksiging
estimator z;,). It is, thus, imperative to evaluate the recoverable torinage
from t h e histogram of the estimators 29, i.e., by the dotted area (zz 2 zo)
o n Fig. VI.3.
1. If, at the time of the evaluation of the reserves, the ultimate estima-
tors z:, upon which selection will be based, are known, then the histogram
of the z: is experimentally available and the recoverable tonnage (2: 2 zo)
can b e estimated without any problem. In the next section, VI.A.3, rt will
b e shown that t h e krjging of t h e selected units (those for which z;, 2 20)
provides a fair estimator of the mean grade of this recoverable tonnage.
From the histogram of the kriging estimators z;,, i t is possible to obtain
a n estimator of the histogram of the true values z, by the following means.
FIG. VI.4. Distr~but~ons
of estlrnated values corresponding to two levels of
(i) Use of a permanence-of-law hypothesis to extend the type of law inforrnat~on.
observed o n the distribution of the z;, to the distribution of the 2,.
(ii) U s e of a correction of variance: the variance of the distribution of In practice, at the time of the study, both the histogram of the core
t h e z, is D 2 ( u / G ) . This variance can be deduced from ~ P ( v l ( 7 by ) grades z, and that of t h e kriged grades z z , are available, these two
o n e of the smoothing relations (VI.1) o r (VI.2), o r calculated from histograms having respective variances D 2 ( c / G ) and D$*(u/G). T h e
the general formula (11.36). problem is to obtain an estimator of the histogram of the ultimate estima-
i.Formulae for deriving recovered tonnage and average grade from a known distribution are
tors z z * (to b e used for selection) from these two experimentally available
given in the addendum VI.B.3. histograms. T o d o this, it is necessarj to act as follows.
451 MINING GEOSTATISXICS VI SELECTION AND ESTIMATION OF RESERVES 455
(i) T o evaluate the level of ultimate information which will be used for be written as a sum of the N true grades of the units 0:
section V.A.l, remark 4. However, the prior calculation of crz*is The dispersion variance of the true grades of the un~tsv in V can be written
often approximate, since it must take into account future informa- according to formulae (11.31) and (11.35).
tion which may be difficult to quantify, e.g., the experience of the
shovel operator o r inforrnat~onabout the stope face. 1
(ii) Using one o f the smoothing relations (V11.1) o r (VI.2), to deduce the
~ ' ( 0v)=
1 ElN 5 [ Z ~ ( ~ , ) - Z , ( X=) ~C(u.
' ] v)-C(V, V )
dispersion variance D'**(v/G) of the ultimate estimators 2;".
Thus, for example, if the following approximate smoothing relation Consider now the N unb~asedestimators ZZ; ( x , ) of the grades of the units v The
is used: global mean grade est~matedover V is then
D'(v/G)= D ' * * ( v / c ) + ~ ~ * ,
the variance D ' * * ( v / G ) is written as In a similar way, we can define the dispersion variance of the estimated grades of
the units u in V :
(iii) To deduce an estimator of the histogram of the Z:* from one of the
two experimentally available histograms by a variance correction
and assuming permanence of law.
If the core grade histogram is considered, the variance correction is a T h e care of kriged estimators If the estimators used above are kr~gedestimators,
smoothing, D " * ( U / G ) < D ~ ( C / GIt) .is then advisable to verify that the then Z: (x,)=Z:,(x,) 1s a 11near combination of the n, data {Z,(.r,), a = 1 to n , } in
the nelghbourhood of estimation, i e., Z:,(x,)=~,h,Z,(x,), and the krlging
same permanence-of-law hypothesis will also reproduce the histogram of system (V I ) IS wrltten as
the 2: by correcting the variance D"c/G)+ D? ( O I G ) , with
D?{v/G)< D ~ * * ( u / G )D <~ ( c / G ) .
If the histogram o f the kriged grades 2:" is considered as the starting
point, correcting the variance then amounts to increasing the variability:
L Z ' * * ( U / G ) ~ D $ * ( V since,
/ G ) , in general,zVz
2:.
T h e case study o f section VI.B.4 gives an examlpie of the debelopment of
such permanence-of-law procedures. and
Note that, sometimes the sought after hlstograrn of the ultimate selection
estimator Z:* can be obtained directly from data coming from zones or
benches already mined.
Addendum. Demanstrarion o f the smoothing relations Considering that C ( h ) is a centred covariance, and that the stationary expec-
tation is m = E { Z z , ( x ) }= E{ZZv(x)},the followmg relatmn can be expressed:
(This demonstration does not provide any new insight into the problems of sclec-
tion and can be omitted on a first reading.)
Let V be a stationary field made up of the union of N units, all with the same
support u, centred on the points {x,,i = 1 to N].The global mean grade over V can
456 MINING GEOSTATISTICS VI S E L E C T I O N AND E S T I M A T I O N O F RESERVES 457
Now, by putting
VI.A.3. Bivariate distributiolq of t h e e s t i m a t e d a n d true v a l u e s
In the previous section, it was shown that the actual selection is carried out
on the estimators z: and not o n the unknown true values z,. T h e recover-
and noting that c(v,, u , ) = C(c, v), V i (all units v, have the same support v ) , the able tonnage ( I : 3 z o )must thus b e estimated from t h e distribution of the
following relation is obtained for the dispersion variance of the kriged grades:
- estimator 2:.But once this selection is made, the mill receives the true
D ? ( v / V ) = C(C,~ ) + r ) / i -u:, -E{[Z:\,(X)-VI]~}. values, not the estimates; more precisely, it receives the true values z,
By substituting D 2 ( v / V ) +C ( V , V ) f o r C ( U ,u), this relation becomes conditiotted by the fact that their estimators z: are greater than the cut-off
20. The set of recovered values is denoted by {z,/z: a z o } .
By interpre-t~ngt h e true and estimated grades z, and z:, as realizations
T h c term between the brackets is of the order of magnitude of the global of the random functions Z , ( x ) and Z : ( x ) , the bivariate distribution of the
estlmatlon carlance over the field V, and c a n generally be disregarded when two variables 2,and Z : car! b e conveniently represented by the scatter
compared t o mean local estirnatlon varlance o:, T h e first a p p r o x m a t e smoothing
diagram (or correlation cloud) o n Fig. VI.5. T h e true grades appear along
relatlon (VI 1) 1s then obtamed.
- the ordinate axis and the estimated grades along the abscissa+axis. The
D 2 ( t / V)- D? (v/ V ) + u : , - 2 ~ distribution of the true grades, with variance D ' ( v / G ) , can thus be seen as
The case of complere krrging T h e term "complete kriging" is used when the N t h e projection of this diagram o n to the ordinate axis; similarly, the
krigings Z i , ( x , ) have been carried out using the same data configuration compris- distribution of the estimated grades, with variance D 2 * ( v / ' ~ )can, be seen
ing all: the available data in V. From section V.A.3, "Theorem of combining as the projection of the diagram o n t o t h e abscissa axis.
kriged estimates", it can then be shown that the s u m Z$ = ( l / h r ) ~ Z ~ , ( xcor-
,] A regression curve f ( z ) , giving the mean of t h e true grades of the units
responds exactly to the direct kriging of Z v ( x ) from the previous data confi,gura- f o r which the estimated grades a r e equal to a fixed value 2: = z, is also
tion. For E { [ Z i v ( x ) - m ] ' } w e then obtain an expression similar to (VI.3): shown o n the same figure. In probabilistic terms, f ( z ) is the "conditional
expectation" and is denoted by f ( z ) = E { Z , / Z : = z } . Note that the
recoverable tonnage for a cut-off zo, { Z v / Z c ~ z o }corresponds
, to the
where dotted area on Fig. VI.5.
T h e mean local kriging variance 6;" is always much greater than the global
kriging variance u i v , and, thus, the previous relation reduces to
? I n practice, the lnformatlon used for the kr~glngof a unlt v IS ilrn~tedto a nelghbourhood
surrounding v However, and prov~dedthere IS no strong nugget effect (s~ncethe nugget
effect removes the screen cf sectlon V A 1, remark 7), thls nearby ~nforrnat~on screens the
~nfluenceof more d~stantdata and, thus, the k r ~ g ~ nofg u can be assumed to be complete and
taking all the ava~lablemformatlon over V Into account FIG.VI.5. Bivariate distribution of t h e true a n d estimated grades, Z, and 2:.
VI SELECTION AND ESTIMATION OF RESERVES 459
Conditional non -bias area differs in two ways f r o m the dotted area {Z,/Z: a zoj corresponding
The true grades {Z,/Z: = zo} of the selected units with estimated grades to the tonnage which can be recovered in practice from the information
Z? = Z O are distributed along the vertical chord with abscissa 20,and their z$.
mean conditional expectation is f(zo)= E{Z,/Z$ = zo}. In practice, this (i) A certain number o f units estimated as being too poor w h i n they
true mean grade is unknown and 1s estimated by the cut-off value 20. It is are, in fact, above the cut-off z o ; these units are wrongly rejected and
very important that this estimation zo b e as close as possible to the true correspond to the hatched area only.
value ~ ( z o )T. h e ideal case would be f(zo)= zo, V z o , i.e., an estimator Z $ (ii) A certain number of units estimated above the cut-off grade when
such that its conditional expectation curve E { Z ; / Z $ = I } is identical to the they are, in fact, too poor; these units are wrongly accepted and
first bisector. T h e conditional non-bias relation can then be defined as correspond to the dotted area only.
The. problem is to minimize these two areas corresponding to wrongly
which can also be written as estimated units by reducing the length of the vertical chords of the scatter
diagram (z,, z : ) , i.e., t o minimize the dispersion of the selected true grades
around their mean f(z)=E{Z,/Z: = i } , which amounts to minimizing the
Note that this conditional non-bias condition is more severe than the global condi~ionalvariances E { [ Z ,- f ( z ) ] ' / ~ : = z } for all values z.
The desired estimator Z : is, thus, the one which minimizes both
non-bias condition E{Zv- 2 : 1 = 0. Note also that the conditional non-
bias (VI.5) entails the unbiasedness of the estimation of the true mean the accuracy criterion E{[f ( ~ :) - 2: 12} and the dispersion criterion
grade of the recoverable tonnage by the mean of the estimated grades of EQZ, - f ( Z)I2\.
Consider the variance of estimation of Z, by 2: :
the selected unit, i.e.,
Deposit G
FIG. VI.6. The simulated data basis.
variance &* = 1 . 3 5 % ~ .The kriging system thus provides an excellent The global quantity of metal actually recovered, Q(ro)=
prediction of the error variance. T(z0).E{Z,/Z: 3 zo}, is proportional to N(r0).E{Zu/Z13 zo}. The
If each of these 384 blocks had been estimated by the polygons of difference between the two quantities of metal recovered, corresponding to
influence method (Poly), by assigning the grade of the central sample to the two types of selection estimator considered, is a direct measure of the
each unit v , Le., A 1 = 1 and A:! = A 3 = 0, the experimental estimation vari- economic benefit to be obtained from a geostatistical approach to selection
ance would have been = 6 . 9 2 ~ Using
~. geostatistics, this variance problems, i.e., an approach which takes into account the two fundamental
would have been predicted, according to formula (II.28), as = 7.31%'. influences of the selection support and the level of information.
The results of a hypothetical selection based on the true grades Z,, of the
TABLE VI.1. Compared results of the kriging and 384 selection units (i.e., perfect knowledge of the deposit) can also be
polygons of influence methods o/'estimatlon given. This ideal selection will obviously lead to the best results.
For the three selection hypotheses (based, respectively, on true grade,
KI ~g Poly kriged grade and grade estimated by polygons of influence) Table V1.7
gives, for the three cut-off grades zo = 4, 9 and 14%:
AI 0.204 1
A? 0.458 0 1. the number N ( z o )of recovered blocks;
A3 0.338 0 2. the true mean grade of the recovered tonnage, E{Z,/Z: 3 20);
m i , theoretical 1.323 7.307
u:, experimental 1.348 6.919 3. the estimated mean grade of the recovered tonnage, E { Z $ / Z : 2 ro};
4. the conditional bias, E { Z , - z : / Z : zo};
5. the experimental estimation variance of the grades of the recovered
Estimation of recoverable reserves blocks, E { [ Z ,- 2: 12/2:3 ZO};
Suppose, now, that at the time o f mining, the only information available on 6. the global quantity of metal actually recovered, N(ZQ).E{Z,/Z; 3
which to carry out selection is the u X u grid of samples. Recovery consists 20);
of all units ui which have an estimated grade I:, greater than the cut-off 7. the estimation of this quantity of metal, N(z0). E { Z : / Z : 2 zo}.
grade r ~ .
T A B L EV1.2. Comparative reslrlts of selecriot~based on vario~csestitnators
The tonnage of recovered ore T(zo), i.e., the number of units N(zo)
recovered, is determined from the histograms VI.9(a) and (b) of the 384
-
ro (1) (2) (3) (4) (5) (6) (7)
estimated values, Zg,, by kriging or Z? by polygons of influence. Note
that histogram V1.9(b) of the estimators Zt is that of the 384 grades of the True grade 4 370 10.08 0 0 3730
central samples; thus, using the polygons of influence estimator ZT selection, 9 203 12.97 0 0 2633
amounts to ignoring the influence of the selection support v , cf. section z" 14 67 16319 0 0 1085
VI.A.1. The histogram of the 384 true grades of the blocks u, is given on
Fig. VI.g(c).
Once the selection is made, Z;, 2 z o or 2::2 r,, the conditional non-
bias condition must be verified, cf, section VI.A.3, formula (VI.S), i.e., the
true mean grade E{Z,/Z: 3 ZU} and the estimated mean grade E { Z : / Z : 2
ro}of the recovered tonnage must agree. For each of the two estimators
considered, Fig. VI.8(a) and (b) give the two curves E{Zu/Z:s z o } and
EU'IZ1 ~ Z Oas} functions of the cut-off value z,. Note that kriging Note lhat the selection made on kriged estimates results in a better re-
ensures conditional non-bias, whereas selection carried out on polygonal covery from the deposit, in terms of the global quantity of metal actually
influence estimators is subject to a serious bias (systematic overestimation of recovered. Also note that the standard polygons of influence estimator
the actually recovered grades); this bias increases with the selectivity (i.e., as alwayc, overestimates the effectively recoverable quantity of metal, which is
zo increases). particularly dangerous during a feasibility study. Selection based on the
-
L-
464 MINING GEOSTATISTICS VI SELECTION AND ESTIMATION O F RESERVES 465
standard estimators of inverse distances (ID) or inverse squared distances The smoothing relation dwhto kriging
(ID2) would have yielded better results than selection based on polygo~nsof
influence estimators (Poly), but these results are always strictly inferior to The smoothing relation (VI.2) due to kriging can be verified from the
those obtained from selection based on a kriged estimator. experimental dispersion variances of Fig. VI.9(a) and (c):
D~(~/G':)=D~*(u/G')+~~,
(the global estimation variance sic* is disregarded). Indeed, the experi-
rncntal values show
FIG. VI.9. Histograms of true grades and of their estimators. (a) Kriging: m * =
9.72, u2*= 14.71. (b) Polygons cd influence: nt* = 9.70, u2*= 23.13. (c) True
grades: m* = 9.82, u2= 16.14.
FIG. VI.8. Conditional bias of the estimators. (a) Kriging. -, E{Z,IZ:X, The bivariate distribution of true and estimated grades
zO);----, E { Z ~ , / Z~~r, , } (b)
. Polygons of influence. -, E{Z,/Z: 23 2 0 ) ;
----, E { Z z / Z : 3 z,,}. It was shown on Fig. VI.5 of section VI.A.3 that the conditional non-bias
E{Z,/Z: = z } = z, Vz, is verified when the conditional expectation curve
f ( z ) = E{Z,/Z: = z } is identical to the first bisector. The best linear esti-
If, instead of a regular grid, the 384 blocks were estimated from data on mator of this regression curve f(z) is the regression line Dl of the true
a pseudo-regular grid with clusters of data and gaps (which is often the case values Z, given the estimated values Z:. It is thus advisable to verify the
in mining applications) the results obtained from kriging would be still even proximity of D l to the first bisector on the two scatter diagrams (z,, 2:)
better than those obtained from ID, ID2, and Poly. corresponding to the two types of estimator used.
MINING GEOSTATISTICS
VI SELECTION A N D ESTIMATION OF RESERVES 367
It was seen in the preceding section that the estimation of the recoverable
tonnage: requires the inference of the histogram of the estimators Z:*
which are ultimately used in the future selection. For example, this
inference is made from the histogram of the grades 2, of the available core
sampler,, or from the histogram of the estimators 2: available at the time
of the study.
More generally, we have the experimental histogram of a stationary
FIG.VI.lO. Scatter diagrams, true grade/estimator. (a) Kriging; (b) polygons of variable: 2,from which it is required to deduce the distribution of a variable
influence. Y that has the same expectation, but a different, known variance: E { Y }=
E { Z ) =m and Var {Y) # Var { Z ) .This problem as posed has no solution,
since the type of the distribution of Y is undetermined. However, the
i
i
468 MINING GEOSTATISTICS VI SELECTION AND ESTIMATION OF RESERVES 469
w
problem can be solved if we assume that the distributions of Y and Z have m e preceding transform function QZ can then be written as
the same "shape". The following section will show how this permanience of
shape can be interpreted in practice.
z = cpZ([ J ) = m . e-"2/2 . e"; U. (VI. 10)
A
The hypothesist of permanence of lognormality amounts to considering
VI.B.l. Simple hypotheses of permanence that the variable Y follows a lognormal distribution with the same expec-
tation m = E { Y }= E { Z ) and a variance a: # a;. The Gaussian transform
1. Permanence of normality function (py which expresses Y as a function of the standard normal
variable U is thus written as
Let Z be a Gaussian stationary variable (in practice, a variable the: histo-
gram of which can be approximated by a normal distribution), with expec-
Y = p y ( l J ) = ~n, e-"'$'2 . e"b ", (VI.1I )
tation m and variance a$. Let U be a standard normal variable with zero The logarithmic parameter is
(T'~ deduced from the known arithmetic
expectation and unit variance. These two variables are related by the variance a$ by the relation
standard formula Z = m + U Z . U.
The hypothesist of permanence of normality means that Y is considered
a:.= Var {Y)= m2(e"';- 1).
to follow a normal distribution with expectation m = E{ Y }= E { Z Jand a It is thus possible$ to construct the required distribution of Y from the
variance a: P a;. Y can be expressed in terms of the standard normal law standard normal distribution 01; U. The distribution function is then
as
F y ( y ) = P r o b { Y < y } = P r o b { U < u } = G(u), (VI.12)
where: }I = cpu(u) or u = Q? ( Y ) ,
The distribution of Y can then be readily determined from tables of
standard normal distribution. and G ( u )is the standard normal distribution function.
4. A f i n e correction of variance
Instead of the preceding indirect correction, it is also possible to make a
direct variance correction &-tu?y by means of the very simple affine
relation
Y-rn 2-m
has same distribution as -
'JY 'Jz
FIG. V1.11. Indirect correction o f varlance through a lognormal model.
(a) Lognormal fit; (b) correction of variance. This s~mplevariance correction on the standard variable (2- m)/uz
results in a variable Y with the required mean m and variance &.
(i) Using the simple two-parameter distribution which fits best the Knowing the distribution of Z (or its experimental histogram), it is a
experimental histogram of Z, make the variance correction cr; -+ a$, simple matter to deduce the distribution of Y:
cf. the two lognormal distributions on Fig. VI. 1 l(b).
(ii) From these two two-parameter distributions, read the recoveries F,(y)=Prob { Y < y ) = P r o b { Z < z ) = F z ( r ) , (VI. 14)
corresponding to the cut-off grade zo. The quotient of these two
recoveries gives an estimate p of the correction coefficient, to be where
applied to the recoveries deduced from the experimental histogram
of Z to obtain the recoveries correspondirlg to the histogram of Y.
Tnus, if the recovery read on Fig, VI.ll(b) from the lognormal dis- This affine variance correction results in a distribution for Y, the shape of
tribution (m, c&) is 20% (hatched area) and if the recovery read from the which 1s very similar to that of 2 , i.e., all secondary modes, significant or
lognormal distribution (m, m$) is 25% (dotted area), the correction not, of the 2-histogram are reproduced on the Y-histogram, cf. section
coefficient is 25/20= 1.25. The hatched area on Fig. VI.ll(a) gives the VI.B.4, Fig. VI.16. However, if u y / u Zis greater than 1, there is a non-
recovery read from the experimental histogram of 2 say, for example, zero probability of obtaining negative values of Y, which can be awkward if
18%. The estimator of the recovery of Y for this cut-off grade z, IS, then, Y represents a grade. If they are few, these negative values can be cor-
Prob* { Y a z c l } =1 8 x p = 2 2 5 % . rected by hand (set equal to zero, for example) on the Y-histogram.
In practice, this affinevariance correction produces results very close to
? ? h ~ sindirect correction procedure retains all the secondary modes observed on the hlsto- those provided by the indirect correction by permanence of a two-
gram of Z As a c o ~ ~ e q u e n cand
e , ln minlng appllcat~ons,this procedure w~llbe used if these parameter distribution. Either one of these two procedures can be used
sewndary modes are signlficant and therefore should be reproduced on the histogram of Y. when the variance correction is small, in practice when /(~$-&1/&<
For thls, the varlance correctior~a:-.& must not be s~gnificant;in practice, the relat~ve
differences I&-u$l/& must be less than 30%. In fact, ~f this varlancc correction 30°/o, and when a great deal of confidence can be placed in the experimen-
represents a signlficant smoottrlng (&cc cr:), this strong smoothing should mask the tal Z-hi,rtogram. Note that the indirect correction, as opposed to the affine
secondary modes, which then should not appear on the Y-hlsrogram. correction, guarantees that no value of Y will be negative.
MINING GEOSTATISTICS VI SELECTION AND ESTIMATION OF RESERVES 473
VI.B.2. Generalized permanence The order n is such that the distribution of the variable cp,(U)correctly fits
As the simple two-parameter distributions are often insufficient to take the the experimental histogram of the data z. The method of fitting the ( n + 1)
various characteristics of an experimental histogram into account, a more coefficients ICr, of the expansion of cpz from the experimental Z-histogram
I
general solution would be to use n-parameter distributions or, more pre- is given in the following section, VI.B.3. It can be shown that
cisely, order n expansions with Hermite polynomials. n $ ; *
Let Z be a stationary variable with any distribution, represented by an 40= nz, 2 7= cr,,
irl I .
(VI. 16)
experimental histogram with mean m and variance ui.We can then define
the Gaussian transform function cpz which expresses the variable Z in where tn and a; are the mean and variance of the experimental Z -
terms of the standard normal variable U, i.e., Z = cp~(U). This transform histogram.
function is fitted to the experimental Z-histogram, and the variance & Applying the hypothesis of permanence to this Hermite expansion '
appears as one of the parameters defining the function cp,. consists of considering that the Hermite expansion of the transform
One possible hypothesis of generalized permanence of law consists in function cpy of the variable Y can be written as
considering that the Gaussian transform function cpy of the variable Y is
identical to p, with the variance parameter u: replaced by cr$. The (VI. 17)
application of generalized permanence then consists in:
(i) determining the transform function cp, of a variable Z from a where
representative experimental histogram;
(ii) determining how this function cp, depends on the variance cr: so as
to make the variance correction a$ + cr$. Thus, cp, and cpy have the same Hermite expansion, except for the vari-
ance correction coefficients c,. The variance correction is written ci= a' ( a
Fitting cp, by a Hermite expansion to the power i ) which entails
In general, the distribution of Z is given in the form of an experimental
histogram which is often too complex to be correctly represented by a
simple two- or three-parameter model. A way around this problem is to fit
The justification for adopting the form ci = a ' for the variance correction
an expansion up to an order n to this histogram; more precisely, to its
is given later, in remark 1.
Gaussian transform function cp,. The order n increases with the complexity
Since the n parameters {$,, i = 1 to n} are known, the function f ( a ) can
of the histogram. It can be shown that the Gaussian transform function cpz
be plotted as on Fig. V1.12. The known value cr; then determines the
of a variable Z which follows any distribution with a finite variance can be
parameter a of the variance correction. Formula (VI.17) gives the expres-
expressed in the form
sion of the Hermite expansior~of py, i.e.,
(VI. 19)
r=O
where the H,(u) are the Hermite polynomials. The definition and main Remark 1 If Z is a normally distributed variable with expectation m and
properties of these polynomials will be presented in the following section, variance & the Hermite expansion of its Gaussian transform function is
VI.B.3. reduced to two terms:
In practice, the Hermite expansion is carried out up to an order rr, i.e.,
Thus, formula (VI.18) adopted for the variance correction assures strict
permanence of normality and, above all, of lognormality, which is a
frequently observed experimental fact. By extension, this variance cor-
rection formula is used for all variables 2 , whatever their distribution.
Remark 2 The variance correction by a H e m i t e model must always be
made in the direction of smoothing, i.e., &-to$, where u$<u$. This
decrease of the variance corresponds to a convolution, and, in general, it is
much easier to convolute a curve than to deconvolute it. The variance
FIG. VI.12. Determination of the variance correction factor n, correction procedure, with c, = a', is particularly unstable for an increase in
the variance (u$ > &) corresponding to a correction factor a > 1, and this
instability increases with the order n of the Hermite expansion. This is
If Z is a lognormal variable with expectation m and variance u$, its illustrated by the example on Fig. VI.20.
Gaussian transform function can be written, according to (V1.10), as The decrease in the variance, u?y<u$, corresponds to a < 1, which
entails a smoothing of the various secondary modes observed on the
2-histogram; the Y-histogram tends, at the limit (when a$<< &), to the
where u: is the variance of the normal variable log, Z. I t can be shown that normal distribution (m, a;).
the Hermite expansion of this function rp, can be written as As a consequence, if the Y-histogram must faithfully reproduce the
shape of the 2-histogram with all its secondary modes, an affine variance
correction procedure should be used, cf. section VI.B.1, formula (V'I.13).
The problem consists of determining whether or not the Y-histogram
thus, (with the lesser variance u$< us) should smooth the secondary modes
observed on the experimental 2-histogram. In mining applications, Z
usually represents the grade of a core sample, for instance, and Y the grade
of a panel on a considerably larger support, which would involve a
smoothing of the secondary modes, and the Hermite procedure would thus
be preferable. However, this is not an absolute rule, cf. the case study in
section VI.B.4. The choice between methods will be made on the basis of
The variance correction factor is given by formula (VI. 18): physical information about the studied phenomenon.
(This section can be omitted at first reading. For mathematical theory, see
i.e., u =o&/uL the ratio o f the standsrdlduviations of the viiri;iblci log. Y
P. Beckmann (1973); for geostatistical applications see G . Matheron
and log, Z.
(1975c, d) and A. Markchal (1975a).)
L
476 MINING GEOSTATISTICS VI SELECTION AND ESTIMATION O F RESERVES 477
A Hermite model of the Gaussian transform function c p will~ be fitted to Consider the condition
the experimental histogram of the variable Z and, by permanence, this
model will be extended to the variable Y. The experimental Z-histogram
must thus represent the distribution of 2,i.e., the unbiasedness of this
histogram should be verified beforehand, cf. section VI.A.1, the remark on This condition is satisfied for any variable Z with finite variance, which is
the representativeness of an experimental histogram, and section III[.C.l, always the case in practice since this variance is deduced from the experi-
"Critical review of data". mental histogram of 2.The following property then results: any function
p(x), such that
Let g(x) be the density of the standard normal distribution; the cumullative
distribution is then written as
the equations aI/a~,b,= 0, Vi, providing the minimum of I, give the relations
An orthonormal basis can then be defined by considering the standardized (VI.23) which define the coefficients &.
Hermite polynomials v,(x) = Hi(x)/Ji!, which are such that Now, let qt be the Gaussian transform function of 2 ;the moments of Z
are written as
which entails
478 MINING GEOSTATISTICS VI SELECTION AND ESTIMATION OF RESERVES 479
There are, thus, two relations relating the coeficients of the Hermite within each class interval (zk,z ~ + ~This ) . way, a one-to-one correspon-
expansion of ipz to the mean and variance of 2: dence z e u can be defined graphically, i.e., not only the transform z =
cpz(u), but also its inverse u = cp,' ( 2 ) .
If there are N data values zk,one solution consists in sorting these N
.
values by ascending order: z 1 s r2G . . S z ~The . experimental cumulative
frequencies are approximated by: F s ( z k )= k / ( N + 1). The graphical trans-
form, cf. Fig. VI.13, consists in identifying the two cumulative distributions,
In practice, the Hermite expansion of (4, is stopped when Ifn=, @ : / i ! is i.e. G ( u k )= F : ( Z ~ ) Hence
. the ordered series of gaussian values uk is
close enough to the required variance cr: The contributions $(/i! of the given by:
terms of higher degree (i > n) tend rapidly toward zero.
p, being expanded up to the order n, it must be verified graphically that
the distribution of the variable
An excellent approximation for the inverse function G-'(r) is given in
Abramovitz, M. and Stegun, I. (19651, p. 933.
Let To be the total in situ tonnage, corresponding to a zero cut-off. The written, cf. Aitchison, J. and Brown, J. (1957) "The Lognormal dis-
following expressions are classical for tribution", Cambridge U. Press:
(i) the recovered tonnage:
T(ZO)=TO.[~-FZ(ZO)I=TO.
Q ( Z ~ ) = T Jz0
~ . ziz(z)dz,
lo j,
FIG. VI.15. Kriged thicknesse:s histc)gram (100 m x 100m support, 200-m sam-
pling grid). Experimental values for m = 11.83 and DF = 24.1 are shown by the
histogram. -, Herm~tianpermanence, n = 15; - - - -, Gaussian fit.
(i) A normal distribution with the same mean and variance. There is a
slight probability that this normal distributio~lwill result in negative
thicknesses.
(ii) An Hermite expansion, deduced by variance correction D2(0/G)-+
D$*(v/G) from the n = 15 order Herrnite expansion fitted to the
histogram of thicknesses measured along the drill cores (cf. Fig.
VI.14). The correction factor a used in this variance correction is
FIG. VI.14, Thickness histogram (core section support). Experimental values for given by formula (V1.18) and a = 0.7064. Note that this histogram,
m = 12.09, D2(O/G)=48.98 are shown by the histogram. -, Hermite fit, deduced by Hermitian permanence, tends at the limit towards the
n = 15; ----, Hermite fit, n = 30.
normal distribution with the same mean 11.83 and variance 24.1.
Indeed, the decrease in the variance is particularly significanr here,
the difference between D2(0/G) = 48.98 and DZ ( v / G )= 24.1
Kriging of mining panels being more than 50%. Note that if an Hermite model were fitted
The experimental data z(x,) on the 200-rn sampling grid were used to directly to the experimental histogram on Fig. VI.15, the secondary
estimate (by kriging) the mean mineralized thickness z,(x) of 2214 panels mode around 8 m would show up more clearly, cf. Fig. V'I.20.
v with horizontal dimensions 100mX 100 m. 'The mean of these 2214
kriged values Z$,(X) is m = 11.83m, the dispersion variance is Selech'on study
~ ? ( v / ~ ) = 2 4 m2,
. 1 and the histogram of these kriged values is shown on The 100 m x 100 m panel u is considered as the selection unit on which
Fig. V1.15. future production will be carried out. The selection variable is the mean
Note that the secondary modes for the large thicknesses are no longer minerrllized thickness.
present, but the secondary mode around 8 m remains. The following If the only information available at the time of study is the data from the
models were fitted to this experimental histogram, cf. Fig. VI.15. 200-rr~sampling grid, the cut-off will be made on the preceding kriged
484 MINING GEOSTATISTICS VI SELECTION AND ESTIMATION OF RESERVES 485
values z $ , , ( x ) , i.e., on the experimental histogram on Fig. VI.15, the results The distribution function (cumulative histogram) of Z$: is then given by
of which are shown in column 1 of Table VI.3. From this table, for the relation (VI. 14):
example, a cut-off thickness 2, = 5 m provides a rccovery of 94.5% of the Prob {Z$': < y } = Prob {Zg, < z }
panels, i.e., 94.5% of the kriged values z$,(x) are greater than the cut-off
zo=5m. with
In fact, much more data are available at the time actual selectl~onis z =O.86(y - l l . 8 3 ) + 11.83.
madc; this future information is assimilated to a regular, square 25-m grid, The histogram of the future values z$g thus estimated is shown on Fig.
i.e., 16 samples per panel. The actual future selection is thus assumed to be V1.16. The corresponding recoveries for different cut-off thicknesses zo are
made on the ultimate estimators zg:(x) deduced by kriging from the data given in columns i,5 and 6 of Table VI.3.
on the 25-m grid. Therefore, the problem is to deduce an estimator of the
histogram of these 2214 ultimate selection values 2;: from the histogram
available at the time of study, i.e., the histogram of the kriged values t;,.
T o d o this, the procedure outlined in section VI.A.2 is followed.
(i) and (ii). The kriging variance a;:* corresponding to the ult~~mate
25-m grid is calculated, and then the dispersion varlance of the ult~~mate
kriged values is calculated by means of the smoothing relation (741.1).
This results in a value D;** (v/G)= 32.8 m'.
(iii) Using an hypothesis of permanence of lab, an estimator of the
future histogram of the 2;: is deduced, either from the sample thick-
n e s s e s histogram by decreasing the variance D2(0/G)-+ D;** (tl/G),
o r from the histogram of the now available kriged values by
increasing the variance D? (v/G)-+ D;** (v/G). Applying the cut-off
thickness zo to this estimated histogram of the zg; results in the required
prediction of actually recoverable reserves. F ~ GVI.16.
. Estimated histogranls of the future kriged thicknesses (100 m x 100m
Two procedures based on two different permanence-of-law hypotheses support, 25-m sampling grid). --, Hermitian permanence; ----, affine cor-
were considered, namely: rection. m = 11.83, Di** = 32-8.
(i) the afine variance correction, cf. section VI.B.l, point no. 4, the
correction by increasing the variance .DF+ D;** is made directly
on the zz,-histogram, according to formula (VI.14);
(ii) the correction by decreasing the variance D 2 ( 0 / ~ ) +DF* is made TABLEV1.3. Esrimations of recoveries
on the Hermite expansion fitted to the sample thicknesses histogram.
z:, ZkX
1. Afine variance correction Let Zg, be the variable kriged from the o 1 (2) (3) -
f4) (5) (6) (7) (8) (9f
200-m grid, with mean nz= 11-83 and variance D? =24.1. The future
OlOcl 100 11.83 99.399.9711.91 100 100 11.83
kriged variable Zz: with the same mean and a variance D?* = 3 2 . 8 is 5 94.5 94.7 12.4 88.6 96.7 12.9 88.3 97.1 13.0
then considered as having a distribution identical to that of the variable 7 82.5 92.3 13.2 77.3 91.0 13.9 79.2 92.5 13.8
11 56.6 72 7 15.2 55.6 74.3 15.8 53.4 72.6 16.1
15 22 5 35.4 tn.6 25.9 42.0 19.2 27.8 44.7 19.0
20 6.0 12.1 23.8 7.3 15.4 24.9 8.9 17.2 22.9
-
cf. relation (VI.13), i.e.,
Z$, Selection made uslng the data on the 200-m grld.
Zg:, Future selection to be made uslng the data on the 25-m grrd.
486 MINING GEOSTATISTICS VI SELECTION AND ESTIMATION OF RESERVES 487
For zo= 5 m, the estimators are: Remark 1 The practical values of the cut-off thickness are between
z~ = 3 n1 and t o= 7 m; thus, the results of Table VI.3 appear to be 'robust
88.6Oh for percentage of recovered panels;
with respect to the adopted permanence hypothesis. In the absence of any
96.7% for percentage of recovered volume;
test data which could be used to choose between the affine vafiance
12.9 m for mean thickness of the recovered panels.
correction and the correction on the Hermite model, the more pessimistic
2. Variance correction on the Hermite models After fitting an n = 15 results ljhould be retained; thus, for zo = 5 m, 96.7% of recovered volume
order Hermite expansion to the experimental histogram of the sample with a rnean thickness of 12.9 m.
thicknesses (cf. Fig. VI.14), variance correction was used to deduce the This robustness should not be observed on the histograms (relative
15th-order Hermite expansion of the variable 2:;$. In this case, the vari- frequency curve) on Fig. VI.16, but on the corresponding cumulative
ance is decreased: D ~ ( oG) / = 48.98 -+ DP* = 32.8. The correction factor histograms on Fig. VI.17; this robustness is even more evident on the
a is found from formula (VI.18) to be a = 0.823. The resulting estimated curves on Fig. VI.18 and VI.19, which give the cun~ulativepercentage of
histogram of the Zg: is shown on Fig. VI.16. The corresponding re- volume recovery and the mean thickness of the recovered volume, respec-
coveries are given in columns 7, 8 and 9 of Table VI.3. tively. By eliminating the cut-off thickness zo from Figs V I .18 and VI.19, a
For a cut-off thickness zo = 5 m the estimator of the percentage of curve giving the recovered volume as a function of the corresponding mean
recovered panels is 88,3%, which is very close to the value 88.6% resulting thickness can be obtained.
from the affine variance correction.
If f(z)is the probability density of the future values ZL:, according to the
classical formulae recalled in the addendum to section VI.B.3,
0 d = f d di,
if(i)
0
FIG. VI.17. Cumulative percentage of panel recovery. -, Hrrrnitian
expressed in per cent. . permanence; -- - -, nttine correction.
The total in situ volume (for a zero cut-off) is
Tne efect of the adopted hypothesis of permanence on the two histo-
2214 x 100 x 100 x 11,83 = 261.9 x lo6 m3; grams on Fig. VI.16 is twofold:
(iii) the mean thickness of the recovered panels, cf. columns 3 , 6 and 9 is: (i) the affine correction reproduces the secondary modes of the
experimental histogram of the kriged values zg, (compare Fig.
+m VI.15 and V1.16); 4
P o= z f ( z )d . ~ / ~ + * / (dz,
~) (ii) the secondary modes of the experimental histogram of the sample
20
values (cf. Fig. VI.14) are smoothed by decreasing the variance of
expressed in meters. the Hermite model.
488 MINING GEOSTATISTICS VI SELECTION AND ESTIMATION OF RESERVES 489
However, this smoothing D2(0/G) = 48.98 + D;** = 32.8 is not enc~ughto The affine correction does not guarantee a zero probability of obtaining
eliminate the mode at the origin (zero thickness corresponding to rises of negative thicknesses 2:: ; thus, from Table VI.3, 0.7°h of the 2:: values
the bedrock). For the greater decrease of the variance, D'(o/G)+ D? = are estimated to be negative. However, in practice, these 0.7% aberrant
24.1, the smoothing is sufficient to almost entirely mask this mode at the values do not afIect the estimation of the recoveries in the practical range
origin, cf. Fig. VI.15. zo E [3 to 71 m.
SUMMARY
A regionalized variable z(x) is interpreted as one realization of a certain random
funct~onZ(x).This RF - more prec~selythis class of RF's - is characterized by a
distribution function and a covariance or variogram model. The idea of simulations
consistcj in drawing other realizations z,(x) from this class of RF's, and to retain
only those realizations z,,(x) which meet the experimental values at the data
locatio~nsx,, i.e., z,,(x,)= z(x,), Vx, E data set. The realizations z,,(x) are called
"conditional simulations" of the regionalized phenomenon z(x).
The theory of conditional simulations is outlined in section VILA together with
the original "turning bands" method, which provides three-dimensional simulated
realizations from one-dimensional realizations simulated on lines rotating in space.
The practical application of conditional s~mulat~ons is given in section VII.B,
through two-case stud~es.A bas~cFORTRAN program is given, which should
allow each user to wrlte his own simulation package. Then several typical uses of
simulated deposits in the field of mine project and plannmg are presented.
z * ( x ) is, on average, closer to the real curve .ro(x),the simulation curve experimental data points the simulated values and the experimental data
zTc(x) is a better reproduction of the fluctuations of the real curve. The values must be the same: ,, .
estimation curve is preferable to locate and estimate reserves, while the
simulation curve is preferred for studying the dispersion of the charac- zSc(xa)
= z0(xa), V X ,E data set I. ...
.,
teristics of these reserves, remembering that in practice the real curve is Consider the true value zO(x)and its kriged value z t K ( x )deduced from
known only at the experimental data points x,. the avadable data {zo(x,),x, E I ) . These two values differ by an mknown
error:
ZO(X)= z$K ( x ) +[ z o ( x ) - z ~ K ( x ) ] WII. 1 )
or, in terms of random functions,
= zk ( x )-t [ z o ( x ) - Z k (XI]..
ZO(X> fyII.2)
Now, a characteristic property of kriging is that the kriging error
[Z&)-- ZCK(x)] is orthogonal to the kriged values, i.e.,
(The proof of this result is given in the addendum at the end of this
section).
To olbtain the desired conditional simulation, it is thus enough to replace
the unknown kriging error [ Z O ( x ) - Z & ( X ) ]in formula pII.2) by an
FIG. VII.1. Real, simulated and estimated profiles. -, Reality; -, isomorphic and independent kriging error [Z,(x)- 2:~ (x)]. More pre-
conditional simulation; - - - -, Kriging; 0,
(conditioningdata. cisely, we consider a R F Z,(x) which is isomorphic to, but independent of
Zo(x).Given a realization z,(x) of this IZF Z,(x), the kriging procedure,
The right-hand side of the profiles o n Fig. VII.1 illustrates the fact that a when applied to the same data configuration {z,(x,), x, E I } will result in a
simulation cannot be used to replace sampling, which is always necessary kriging error [ . ? , ( x ) - z ~ ~ ( x )isomorphic
] to the true error [zo(x)-z&(x)]
for a good local estimation of the deposit. In any case, the better the and independent of z & ( x ) . The required conditional simulation is then
deposit is known, the better the structure of variability can be modeiizd written as
and the denser the grid of conditional data will be, all of which will result in
a simulation closer to, and more representative of, reality.
or, in terms of random functions,
VI1.A. THE THEORY OF CONDITIONAL SIMULATIONS zSc(x)= Z$K( x ) + [ Z ( x ) - z T (x)I*
~
VII.A.l. The principle of conditioning The requirements of a conditional simulation are satisfied since we have
the following.
Consider the regionalization of a variable zo(x),which may be, for exam-
ple, the grade at point x. This regionalized variable is interpreted as a (i) 'I'he RF t,,(x) has the same expectation as &(x). This follows from
realization of a stationary RF Zo(x), with expectation m and centred I he unbiasedness of the kriging estimator:
covariance C ( h )or variogram 2 y ( h ) .The problem is to build a realization E { z g K( x ) } = E{Zo(x)) and E { z $ ~
(x))= E{Zs(x)) .
of a R F Z,,(x) isomorphic to Zo(x), i.e., a RF which has the same expec-
tation and the same second-order moment, C ( h ) or y(h). Moreover, the which entails
realization z,,(x) must be conditional to the experimental data, i.e., at the
496 MINING GEOSTATISTICS VII SIMU'LATION OF DEPOSITS 497
(ii) The RF Z,,(x) is isomorphic to Zo(x).This follows from the facit that In the case of a kriging system without the unbiasedness conditions - the
the simulated kriging error [ ~ , ( x ) - z ;( x~) ] is isomorphic to the stationary expectation of the F W Z,,(x) being known (for example, zero) - the
kr~gingsystem ( V . l ) reduces to, cl'. section VIII.A.l, system (VIII.3),
true error [&(x)-ZgK(x)] and independent of Z & ( X ) . In the
addendum, it will be shown that this isomorphism applies rigorously
to the increments of the RF's Zsc(x)and Zo(x)only. In other words,
the variograms of these two RF's are identical, but not neces!;arily
their covariances. This does not pose any problem since, in geo-
statistics, the variogram structural function is always preferred to
the covariance.
The covariance between the kriging error and the kriged values then becomes
(iii) The simulated realization r,,(x) is conditional to the experimental
data, This follows from the fact that, at the experimental data
points, the kriged values are equal to the real values, i.e.,
and the kriging error R(x) is thus orthogonal to the kriged values Z;tK(y).
In the usual case of a kriging system with a non-bias condition, the ordinary kriging
which entails zsC(x,)
= ZO(X,). system ( V . l ) is written as
Remurk 1 The kriging weights {A,, Vcv E I}are the same for Z & ( x ) and
Z:K(x),since the two RF's Zo(x)and Z,(x) are isomorphic and the: two
kriging configurations are identical.
From expression (VII.3) for conditional simulation, this variogram is exactly that illustrated on Fig. VII.2. The thickness of this slice is the equidistance
of the R F Z,,(x), since the error [Z,(x)-Z:,(x)] is isomorphic to R ( x ) and
independent of Z$;E(X). Thus, the conditional simulation ZJx) ensures the between the simulated values on the line Dl. ? .
reproduction of the variogram of the RF Z,(x), but not necessarily its covariance. We then consider N lines D l , D2, . . . , DN corresponding to the direc-
tions of ?he unit vectors k , , k 2 , .. . , kN uniformly distributed over thevnit
sphere. O n each line D, a realization y(xD,) of a RF Y(xD,) is generated
Vll.A.2. Non-conditional simulation in t h r e e dimensions isomorpnic to Y (x,,), the N RF's { Y(xD,),i = 1 to N ) being independent.
The first step in carrying out the conditional simulation z,,(x) given in Using the method outlined above, a three-dimensional realization z,(x)=
formula (VII.3) is to simulate a realization z,(x) of a RF Z,(x) isomorphic y(xD,), Vx E R ~ is, made to correspond to each one-dimensional realiza-
to Zo(x), i.e., with the same mean and covariance C(h).The simulation tion y ( x ~ , ) .
z,(x) is non-conditional, Le., it does not necessarily take on the experi-
mental values zo(x,) at the data location x,,.
It will be recalled that the notations z,(x) and zo(x) represent regional-
ized variables spread in the three-dimensional space, x E R ~ .
Many methods are available for simulating one-dimensional realizations
of a stationary stochastic process with a given covariance. However, when
these procedures are extended to the three-dimensional space, they are
often inextricable or prohibitive in terms of computer time. The originality
of the method known as the "turning bands" method, originated by G.
Matheron, derives from reducing all three-dimensional (or more generally,
n-dimensional) simulations to several independent one-dimensional simu-
lations along lines which are then rotated in the three-dimensional space
R' (or Rn).This turning band method provides multidimensional simula-
tions for reasonable computer costs, equivalent, in fact, to the cost of
one-dimensional simulations.
\u
which, when N + a,tends toward the following isotropic covariance: cf. formula (V.9). One such simulation of a functional drift will be presen-
ted in section VII.B.l.
'I
C(r) = -
2 71. 1 / 2 unit sphere
C"'((h, k)) d k, (VII.5) However, if the drift m(x) is evident only in a restricted section of the
zone, to be simulated, a sufficient representation of this local drift may be
where t h , k>is the projection of the vector h onto the axis k. obtained by conditioning the simulation to true or fictitious data in this
, r = (hI =lJ(h:+ h t + h i ) is the modulus of the vector h. section, cf. the left-hand side of the simulated profile on Fig. VII.1. More
The preceding integral can be written in terms of spherical coordinates, precisely, a large number of conditioning data zo(xa) are placed in the
cf. Fig. VII.2: section in which the drift is evident; these data may be real (x, E I ) or, if
the real data are insufficient, fictitious. These conditioning data will force
C(r) =l
2%-
j2vd~j:2
0
c(li(lr cos sin 9 d q = f jorC(])(S)ds. the initially stationary simulation to follow the drift m(x) within the cpn-
sidered section. The conditiorring confers robustness to the simulation with
In practice, the three-dimensional covariance C(r) is fixed and the respect to the local unmodelled characteristics of reality, and for this
one-dimensional covariance c")(s) to be simulated on each of the A1 lines reason, it is always advisable to use conditional simulations z,,(x) rather
is thus given by the derivative than nonconditional ones z,(x). The robustness and hence the represen-
tativity of the conditional simulation z,,(x) will improve as the number of
real conditioning data {zo(x,), xu E I ) increases: simulation is not a substi-
tute for a good survey of the ireal deposit.
From Bochner's theorem, it can be shown that the function C(')(s) given by
formula (VII.6), is a positive definite function in one dimension and can, Remark 2 Simulatiort of anisotropic regionalization. The three-dimen-
thus, be used as a covariance function. sional covariance C(r) of formulae (V1I.S) and (VII.6) is supposed to be
It is thus always possible to determine the one-dimensional covariance isotropic, whereas, in practice, the three-dimensional RF's to be simulated
~ " ' ( s ) to be imposed on the R F Y(u,) simulated on the lines 13,. By often have various anisotropies. However, as was seen in section III.B.5,
turning these lines and their corresponding bands in R ~ the , required any anisotropic covariance C(h,, h,, h,) can be modelled as a nested sum
three-dimensional simulation z,(u, u, w ) , with the specified isotropic of isotropic models in spaces of dimension n s 3 . It is, thus, sufficient to
covariance C(r) is obtained. simulate independently the realizations corresponding to each of the
component models and then to sum up these realizations at each point.
Remark1 Specified expectation. In general, the expectation of the three- Take, for instance, an anisotropic covariance of the type
dimensional RF Z(x), a realization of which is to be simulated, is nolt zero:
E { Z ( x ) }= nt # 0. To obtain the desired expectation, it is enough to add a
constant, equal to the specified mean m, to each of the previously simu-
lated values z,(x), which constitutes a realization of the zero-expectation where r = lhi; KO,K , , K2, K3 are positive constants; and Co, C,, C2, C3 are
RF Z,(x). This additive constant does not alter the second-order moments isotropic covariances defined in three, one, two and three dimensions,
(covariance or variogram) of the simulated RF Z,(x). respectively. These covariances are, for example, spherical models with
Note that this additive procedure can also be used for the simulation of a ranges ao, a l , a2, a 3 and unit sills.
non-stationary R F interpreted as the sum of a drift and a stationary In this case, the R F Z ( x ) can be interpreted as the sum of four inde-
residual, Z ( x ) = m(x)+ R(x), cf. relation (V.8) in section V.A.2. The sta- pendent RF's:
tionary residual R(x), with zero expectation and known covariance, is
simulated by means of the turning bands method. At each point of this
simulation, a simulation m,(x) of the drift m(x) is then added. The sl~mula-
tion m,(x) of the drift may be based, for example, on the expansion of m(x)
as a series of known functions: where To(x)characterizes an isotropic nugget effect, i.e., a micro-structure
with covariance Co(r) and a range a. very small with respect to the
502 MINING GEOSTATISTICS j03
VII SIMULATION OF DEPOSITS
distances involved experimentally or in the simulation. All the random
variables {To(x),for each x } are spatially uncclrrelated with zero expec- (i) select from the simulation file the realization, spherical or exponen-
tation and a variance E{Ti(x))=KO.The simulation of a realization of tial, that has a range closest to the specified value;
T&) can, thus, be carried out very simply by drawing numbers at random (ii) multiply all values of this realization by the constant JK.
..
from a uniform distribution with a variance of KO;there is no need in this
case to use the turning bands simulation method. Remark 3 The icosahedron approximation. Formula (VII-5) of the turn-
T,(x) characterizes a structure of one-dimensional variability with a ing bands method shows that a three-dimensional isotropic covariance is
covariance KIC,(h,). This structure may correspond to a vertical vari- only attained as a limit for an infinite number N of lines D,. In practice, of
ability of grades (for example, surface weathering phenomenon) that is course, I his number N will have to be finite. For example, we could take a
particular to this direction. Tl(x)varies only in the vertical direction, i.e., large number ( N around 100) of lines distributed randomly over the unit
on any given horizontal plane all the realizations t , ( x ) are equal, and can, sphere. An equally good but much less costly approximation can be
thus, be simulated directly on a single vertical line. Alternatively, this obtained by taking the N = 15 lines joining the mid-points of the opposite
simulated line may be drawn from a previous three-dimensional simulation edges of a regular icosahedron, which is the regular polyhedron with the
(simulation file on tape or disc) with the isotropic covariance C,(r).All the maximum number of faces (20).
The sum z,= (C: z , ) /J 15, corresponds to the covariance
simulated values along this line are then assigned to their respective
horizontal planes, providing a three-dimensional realization with covari-
ance K1Cl(h,).
Tz(x) characterizes a structure of two-dimensronal, isotropic variability
with covariance ~ , ~ ~ ( J ( h : + h t )This
) . structure may correspond to a
large-scale variability of grades specific to the horizontal directions (e.g., which is the icosahedron approximation of the limit covariance C(r) given
stratifications); the range a 2 will thus generally be much larger than a l and in formula (VII.5). The small difference between C*(h) and C(r)can be
a3. Tz(x) varies only in horizontal directions, i.e., on any given vertical line, theoretically corrected using the geometric properties of the regular
all the realizations t 2 ( x ) will be equal, and T2(x)can be simulated directly icosahedron. This refinement is unnecessary in mining applications.
on a single horizontal plane, or the plane can be drawn from a file of The geometric properties of the regular icosahedron are detailed in A.
three-dimensional simulation with isotropic covariance K2C2(r). Journel (1974b). For practical applications, it is sufficient to know that the
T&) characterizes a structure of three-dimensional, isotropic variability 15 lines defined on the icosahedron can be. divided into five groups of
with covariance K3 C3(r). This structure may correspond to a small dis- three, each group constituting a trirectangular trihedron. Each group of
tance (small range a3), average, isotropic variability within each bed or three lines can be obtained from the others by five successive rotations.
mineralized lens. T3(x) varies in all directions of the space R~ and its More precisely, let (Ou, Ov, Ow) be any reference rectangular axes, and
simuIation requires the turning bands method. (Oul, Ov,, Ow,) the trirectangular trihedron with the three base vectors
A simulation file - o r a simulation data base - can be created by storing defined by the following direction coordinates:
an tape or disc different realizations of three-dimensional RF's having
standard types of isotropic covariances with variable parameters, e.g., a
spherical or exponential covariance with unit sill and variable ranges. As an
experimental anisotropic covariance can be represented as the sum of
isotropic covariances, it is, thus, possible to build a simulation with the
desired anisotropic covariance by selecting appropriate realizations from
the simulation file.
wheie k is either one of the roots of the equation k'i- k - 1 = 0 , i.e.,
The sill K of a transition-type covariance KC'(h) appears simply as a
multiplicative factor. Thus, to obtain a simulation corresponding to a k=0.618 or k=-1,618. Note that the number ( l + J 5 ) / 2 = 1 . 6 1 8 is the
"golden number" well known in geometry and architecture.
particular spherical or exponential model (with specified range a and sill
This trihedron (Ou,, Ovl, O w l ) is rotated four times, with the same
K), it is enough to
rotation matrix [9]thus defining five trirectangular trihedrons, the aies of
5 04 MINING GEOSTATISTICS VII SIMULATION O F DEPOSITS 505
which are the preceding 15 lines that form a regular partition of the The method of niouing averages
three-dimensional space
Consider a stationary random measure T(r) dr with a Dirac covariance
measure, e.g., the differential forms of a Poisson process or of a Brownian
motion. The discrete version of this process T(r) dr is simply a succession of
independent random variables T, with the same distribution function.
Each random measure T(r) dr is then regularized by the weighting
function f(u), and this value defines the following one-dimensional RF
The matrix of rotation is Y(u):
+w
Y ( U ) = ~f ( u + r ) T ( r ) d r = T * f (VII.10)
-m
with
if r # r',
E{T(dr)T(drt)} =
{ t 2 d r ifr=rl.
Indeed, the two random measures T(dr) and T(drl) are orthogonal if r Z r'.
Vll.A.3. Non-conditional simulation in o n e dimension The preceding double integral then reduces to
The turning bands method uses the simulation along lines of the one-
dimensional W Y(u), with covariance c(')(s), to provide realizations of a
three-dimensional RE with the required covariance C(r). which, except for a multiplicative factor which is the density of variance a2,
By means of the theorem of harmonic analysis of a second-orcler sta- is simply the convolution product f * f of formula (VII.9).
tionary process, one-dimensional realizations y ( ~ )with any specified
covariance c(')(s), can be simulated, cf. A. Journel (1974b). However, this Discrere approxintation I n practice, the random measure T(dr) is
method requires the Fourier analysis of the function C")(S), wh~chcan approximated by a series of random variables located at discrete intervals
sometimes be dificult; in addition, there are problems with the con- along the line, as on Fig. VII.3. These RV's T, are independent of each
vergence of the results. other and have the same distribution with zero expectation and a specified
In mining applications, the usual three-dimensional models C(r),e.g., variance E { T ~ )= a2.
spherical o r exponential, are such that the corresponding one-dimensional Realizations t, of these RV's T, are assigned to points at regular interval
covariances c")(s) can be expressed in the form of a convolution product b on the line D, as on Fig. VII.3. The values of ti could be, for example,
of a function f(u) and its transpose f(u)= f(-u): independent realizations of the random variable with uniform distribution
function as provided by a random generator such as IBM's RANDU. Let
. . +
. . . , t,, . . . , t,+k be the realizations at points i - k, . . , i, . . , i k.
The next step is to take the moving average y, weighted by the function
This means that the one-dimensional simulations can be carried out as f(u) and defined at each point i as
moving averages with weighting function f(u), which is a much simpler +m
This formula is practical only when the function f(kb) decreases rapidly
The spherical model is expressed by formula (111.15):
towards zero as lkj increases, so that the discrete sum does not have to be
taken over the range k = - co to f a.
Thte constant K = C(0) = Var {Z(x))is the a priori variance of the three-
dimensional RF Z(x) and is also the sill of the corresponding semi-
variogram y(h) = C(0)- C(h).
The corresponding one-dimensional covariance function is given by
formula (VII.6):
FIG. VII.3. Moving average on the line.
- --
As the RV's 7 ; + k and T+i are independent if k # I, and have a variance
E{T'} = ( r 2 , the covariance reduces to This result can be demonstrated by direct iritegration. This odd weighting
funcrion f(u) = - f(-u) differsfrom zero only on the interval [-a/& C aiZ].
In practice, an odd integer number (ZR + 1) of elementary values 1,+1f is
used for each moving average y, in the discrete formula (VII.1 I), i.e.,
This discrete sum cil)(s) is simply the discrete approximation of the
integral in (VII.9), which defines the specified one-dimensional covariance
~ " ' ( s )This
. discrete approximation C'P(s)can be calculated for each
weighting function f ( u ) and for each value b of the discrete interval. The in which k progresses by integer values from -R to +R. The central value
bias d"(s)-Ci1'(s) caused by the discrete approximation can then be ti corresponding to k = 0, is assigned to the same point as the value y,, cf.
corrected. In practice, a multiplicative correction is enough, and this can be Fig. V11.3. The interval b of separation of the random variables 7; on line
achieved by manipulating the parameter 02= E{T2}to ensure the equality D is, thus, equal to b = a/2R. Program S I M a , given in section VII.B.2,
of the variances: c(')(o)= c ~ ) (An o )example
. of the correction of this uses a number greater than or equal to 20. In practice, this number can be
bias is given in the case study of the simulation of the Prony deposit, cf. reduced (R < 20), particularly when the range a is one of the ranges at of a
section VII.B.2. nested model C(h) = 2, C,(h), in which case the object is to reproduce
globally the nested sum C(h), rather than precisely to reproduce each of its
Application to the usual covariance models All that remains is to show component structures C,(h), cf. the Prony case study in section P?.VII.B.1. '
that the usual three-dimensional models can be expressed, by means of
formula (VIIS), in terms of one-dimensional covari,itices of the form T'he exponential model is expressed, according to formula (III.16), as
P ( s ) =f * f C(r)=Keb"', Vr30, with A = l / a . - i
508 MINING GEOSTATISTICS VII SIMULATION OF DEPOSITS 509
The corresponding semi-variogram y(r)= C(0)- C(r) reaches its sill realizations z,(x) of a RF Z,(x) with a Gaussian (normal) multivariate
y ( W ) = C(0) asymptotically, cf. Fig. 111.5. A practical range a ' = 3 a is uljed, distribution (because of the central limit theorem). Assuming the hypo-
for which y ( a l )= 0-95K, or C ( a f )= 0 - 0 5 K thesis of stationarity, upon 'which the method is based, the univariate
The corresponding one-dimensional covariance is given by forrrlula distribution of the RF Zs(x) is Gaussian and independent of x. Thus, the
(VII.6) as histogram of the simulawd values {z,(x), x E deposit) is also Gaussian.
In practice, however, the lhistogram of the experimental data {zo(x,),
x, E I)and, more generally, the univariate distribution of the stationary RF
This covariance can be expressed as a convolution product, C"'(s)= Zo(x) is not restricted to the Gaussian type. In particular, the real values
f * f: with the following weighting function: {z0(x), x E deposit} are often non-negative, as in the case of grades or
thicknesses, whereas the simulation {z,(x), x E deposit}, having a Gaussian
distribution, may have a significant proportion of negative values. Thus, it
is particularly important to ensure that the simulated values z,(x) have the
The weighting function f(u) approaches zero asymptotically as u tends same histogram as the experimental histogram of the available data zO(x,).
to infinity. In practice, the fufiction is limited to the interval [O,4a] and, on Of course, it must be assumed that this experimental histogram is
this interval, the random measure T(dr) is approximated by ( 8 R - t 1 ) representative of the distribution of the variable zo(x) in the deposit, cf., in
discrete random variables T+k.The separation interval b is thus equal to section V I A 1, the remark on the representativeness of a histogram.
b = 4a/8R = a/2R, with R 2 20. This high density discrete approximalion Let &(x) be a stationary random variable with a distribution represen-
is required to reproduce the asymptotic tendency of the covariance c")(s) ted by the experimental histogram of the available data {zo(x,)e I}. By
towards its sill. means of a Gaussian transform function, this variable Zo(x) can be
The Gaussian covariance is expressed by formula (111.17) as converted to a standard normal variable U(x), i.e.,
Zo=qzo~(U)and u=cpg(~o). ,
cf. Fig. 111.5. the corresponding one-dimensional covariance is Instead of simulating the grades zo(x), the transformed grades u(x) are
simulated. This simulation {u,(x), X E deposit}, obtained by the turning
bands method, has a Gaussian histogram and the transform c p ~ results
, in a
simulation {z,(x)= cpa(u,(x)~]} with the required histogram, i.e., a histo-
which can be expressed as a convolution product, c")(s)= f * with the gram similar to that of the experimental: data zo(x,). The covariance
following weighting function: imposed on the simulation u,(x) must, of course, be the covariance
f ( U )= A U e - 2 ~ d 2 / a 2 ,
deduced from the transforrned experimental data, u(x,)= cPsL [zo(x,)],
VUE [--a,
+a], (VII. 18) Vx, E I, Although not rigorously, the inverse transformed simulation
where A* = 16 la^ J n z,(x)= cpG[u,(x)] will then, in practice, have the required covariance
C(r>.
In practice, this function is limited to the interval [-2a, +2a] and, on rthis Thus, the simulation z,(x) will reproduce not only the first two moments
interval, the random measure T(dr) is approximated by (8R + 1) discrete - expectation and covariance or variogram - but also the univariate dis-
random variables ?;,, corresponding to a separation interval b = 4a/8R = tribution or, more precisely, the experimental histogram of the data zo(x,).
a/2R, with R 3 20. If, in addition, the conditioning of the simulation to the experimental
data zo(x,) is desired, it is advisable first to condition the simulation y ( x )
Vll.A.4. Gaussian transform to the transformed data u(x,)= [zo(xp)] through relation (VII,3), i.e.,
data u(x,) and from the simulated values u,(x,) located at the same points two states, one highly mineralized and the other waste, cf. Fig. VII.4. Each
x,. These krigings are carried out using the covariance Cu(r) of the trans- of these two states considered separately may have a Gaussian spatial law,
formed RF U(x). but when considered together they definitely do not. A straightforward
By way of example, consider an experimental histogram of real grades simulaticln of such a process, using the turning bands method, would result
(zo(x,), VxmE I), to which a lognormal distribution can be fitted. The in a profile (the bold line on Fig. VI1.4), which, over large zones, may have
transformed variable which follows a normal distribution is: U(x) = log the expectation and covariance of the combined states of the real
Zo(x), and, thus, Zo(x)=exp [U(x)]. After determining the covariance phenomenon, but nevertheless would be a poor representation of this
Cu(r), which characterizes the spatial regionalization of the transformed reality.
grades {u(x,)=log zo(x,), Vx, E I), the turning bands method is used to In such a real case, the two states could be differentiated by structural
provide a simulation u,(x) which has a Gaussian histogram and the analysis, and they would then be simulated independently; the transition
imposed covariance C,(r). This simulation is then conditioned to the from one state to another being simulated separately. Alternatively, one
transformed data u(x,) by formula (VII.3). Upon taking the inverse trans- could consider a regularized phenomenon, taking averages within volumes
form z,,(x)=exp [u&)], the required conditional simulation is obtained. suficienrly large to include a great number of different states; this
In practice, however, the Gaussian tranform function Zo= pa(U) is not regularization would then provide moving average profiles with a more
always so simple or so obvious. The most general methods are a graphical continuous character open to simulation by the turning bands method.
correspondence between the cumulative histogram of the data zo(xa) and
the standard normal distribution of U(x), cf, section VI.B.3, Fig. VI.13, or
the approximation of function cp;i by an Hermite polynomial expansion, cf.
section VI.B.3, formula (VI.22) and following.
A fundamental remark
The above procedure reproduces the required univariate distribution of
the RF &(x), but this does not mean that the 2-, 3-, . . . , n-variate dis- Real o u t l ~ n e
tributions are also reproduced, i.e., the spatial law of the RF Zo(x) is not FIG. VII.4. The turning bands method is not suited for the straightforward simula-
reproduced in its entirety. As far as geostatistical applications are con- tion of a two-state process.
cerned, this is not very important since:
(i) most geostatistical techniques (kriging, estimation of recoverable Vll.A.5. Simulation of blocks
reserves and study of their characteristic fluctuations) require only Up to this point, it has been supposed that the support of the simulated
the univariate distribution of Zo(x) and its covariance or variogram, values i.,(x) or z,,(x) is the same as that of the available data. However. it is
function; often required to simulate the grades of blocks Z,(x), e.g., mining unit
(ii) the conditioning of the simulation z,,(x) to the experimental data grades, while the available information Zo(x) is defined on the support of a
zO(xa)induces a certain robustness with respect to the features of the core sample (this core sample being approximated to a quasi-point
reality zo(x), which are not specifically known or imposed on the support).
simulated model. This robustness increases with the amount of
conditioning data.
1. Averaging simulated points
However, it should be stressed that the turning bands method produces
simulations with a Gaussian spatial law and, as such, these simulations One method for simulating block values is to simulate point grades z,,(x) on
cannot correctly represent mineralized phenomena having spatial laws a very dense grid and then average all simulated point grades that fall
which do not have the very continuous character of the Gaussian spatial within the required block dimension. This will result in an approximate
law. Consider, for example, a non-stationary phenomenon consisting of conditional simulation z,,,(x) of the block grades. As the point simulation
512 MINING GEOSTATISTICS VII SIMULATION OF DEPOSITS 5 13
z,,(x) is conditional to the data zo(x,), the block simulation wdl alslo be The disadvantage of the method is that a conditional simulation of N
conditional to the same data. blocks requires a non-conditional simulation of n N points and, more
The block is approximated by n interior points (x,, i = 1 to n), and the significantly, the conditioning process requires kriging these n N points. If
true grade of the block there are N = 100 000 blocks and n = 10 points per block, then problems
could arise with storage and calculation time. Cuts in computer time can be
obtained, first by kriging only the N blocks, and second by simulating and
is approximated by the arithmetic mean of the n interior point grades kriging directly these N blocks, as shown in the following section.
Vll.A.6. Simulation of coregionalization {Uk(x),k = 1 to L) is simulated, and then the required coregionalization is
In mining applications, there are often several spatially intercorrelated obtained by applying the Gaussian transform:
variables of interest, e.g., overburden and mineralized thickness, and the z k ( ~ ) = c P k [ U k ( ~ ) ] , v k = 1 to L.
corresponding grades of saleable metals and of impurities.
The spatial coregionalization of the L stationary RF's {Zk(x), k = 1 to L} Remark 2 The conditioning of the experimental data simulation is
is characterized by the covariance matrix [Ckk(h)], cf. section II.A.4, with carried out, variable by variable, by direct kriging and not by cokriging.
c k k 8 ( h=) E{Zk(x)Zk.(x+ A ) } - rnkmk,, and mk = E{Zk(x)}.
which would be tedious and too costly:
z k . ; c ( x )= ~ E K ( x ) + [ z k , s ( ~ ) - z * , s K ( x ) ] ,
This covariance matrix must be positive definite and the lincar model
proposed for it in sectlon III.B.3, formula (III.25), is wherc zEK(.j)and Z ; . , ~ ( X ) are kriged from the real data z ~ ( x , ) and the
simulated data za,,(x,), respectively.
In this model, all cross- and direct covariances Ckks(h) are expressed as
linear combinations of H basic direct covariances K,(h). (It will be recalled V1I.B. THE PRACTICE OF SIMULATIONS
here that, in the above notation, i in the superscript position represents an
index and not a power.) VII.B.l. Case studies
Consider now n stationary RF's {Y,(x), i = 1 to n} with direct covari-
ances Ki(h), respectively, and all independent of each other. The L sta- 1. Conditional simulation of the Prony deposit
tionary RF's {Zk(x), k = 1 to L), defined by the L linear combinations The nickeliferous laterite deposit of Prony (New Caledonia) is a surface-
weathered subhor~zontal formation of basic rocks (peridotites). The
deposit was sampled by several successive campaigns of vertical drilling on
a regular horizontal grid; each core sample was analysed in I-m lengths for
have a covariance matrix [Ckk'(h)]given by nickel grade and the grade of impurity MgO. Global and local resources
and recoverable reserves were evaluated by geostatistical techniques after
each sampling campaign was completed.
The next step was to tackle some of the mining problems, such as
Thus, the simulation of the coregionalization {Zk(x),k = 1 to L ) with the deciding on the type of mining and the corresponding equipment, the need
(VII.20) covariance matrix can be reduced to the simulation of indepen- for blending, and the degree of flexibility of the mill. For this, it was
dent realizations of tr RF's Y,(x) with covariances K,(h). These tz realiza- necessary to predict the fluctuations at various scales of such parameters as
tions are then linearly combined by means of formula (VII.21) ~ 1 1 tthe h overburden, mineralized thickness and the recovered Ni and MgO grades.
following parameters: Thus, it was decided to simulate the deposit and test the various methods
considered for mining by applying them to the simulated deposit.
bLkt= akiak.8, Vi = 1 to n, and Vk, k t = 1 to L. (~11.22) The simulation was limited to the first production zone, and consists of
If the number of parameters ski, ak,i is insufficient to reproduce each of 17 600 simulated vertical drills on a square 10-m grid. The simulation was
the coefficients bik, of the linear model (VII.20), it is advisable to use the based on a previous geostatistical structural analysis of 176 real drill-holes
more complete linear combinations of formula (111.24). on a square 100-m grid within the first production zone, cf. Fig. V.24, and
was conditioned to these real data.
Remark 1 T o reproduce the univariate distribution functions of each of A conditional simulation involves three steps: structural analysis, which
the component RF's Zk(x), the covariance matrix of the Gaussian RF's provides a model of the deposit; conditional simulation itself; and, finally,
Uk(x)= (p&l [Zk( x ) ] should be considered. First, the coregionalization various verifications. Simulated deposits can be used for various purposes,
518 MINING GEOSTATISTICS VII SIMULATION OF DEPOSITS 519
depending on each particular mining project, and some of these typical drill-hole on each mining unit. Thus, the cut-off grade t, must be applied to
uses are presented in section VII.B.3. the pre:viously defined drift curve estimating the vertical profile of mean Ni
grades of the mining unit, cf. Fig. VII.5. For each value tc, a certain number
Driftand cut-offgrades The vertical profiles of the Ni grades measured of "service variables" are defined. These variables, regionalized in the
along the drill cores all have a similar shape, cf. Fig. 11.1. There is a two-dimensional horizontal space, are: overburden thickness Po, mineral-
progressive enrichment in nickel from the upper surface, followed by a ized thickness P, and the corresponding accumulations AN, and AMgo.The
rapid impoverishment, and, finally, a further enrichment at the bedrock simulation of the deposit thus consists of generating a realization (core-
contact. This is translated into geostatistical terms as a vertical drift of the gionalization) of the corresponding service variables on a 10 m x 10 m grid
Ni grades. The drift curve (or mean tendency) can be estimated by uni- for each value of the cut-off grade r,.
versal kriging from the vertical profile of the grades of the core sample
lengths, and it can be shown that this drift curve is an estimator of the Horizontal structural analysis At the time of this study, the first pro-
vertical profile of the mean grades of small panels centred on the drill duction zone was sampled on a systematic 100 m x 100 m grid, and by two
holes, cf. Ch. Wuijbregts and A. Journel (1972) and Figs.II.1 and VII.5. cross-shaped dr~llingsets spaced at 20 rn to detect small structures. This
information was used to characterize the horizontal regionalizations of the
previously defined service variables. In the actual study, four cut-off grades
were considered; six service variables for each cut-off. But only the results
obta~nedfor the mineralized thickness PL,the overburden Po, and the Ni
accumulation A N ,for the two cut-offgrades r, = 0 (no vertical selection) and
r, = 1.096 will be given here.
For proprietary reasons, all values given here have been multiplied by a
constant factor, which does not affect the relative curves (variograms
divided by the squares of the means), nor the relative dispersions.
Table VII.1 gives the means m and the relative dispersion variances
~ ' l r nof' the various service variables within the zone of interest. These
experimental values were deduced from the 176 drills covering the first
production zone on a 100 m x 100 m grid. Table VII.2 gives the correlation
coefficients between these service variables for each cut-off grade.
Figure VII.6(a), (b), (c) gives the experimental relative variograrns of the
variables Po for fc = 1.0 and PL and AN, for fc = 0. Figure VII.7(a), (b)
shows the scatter diagram for Po and PL and the corresponding cross-
variogram for t, = 1.0.
TABLE VII. 1. Experimenrul means and
dispersion variances
m - 23.7 26.9
The entire mineralized thickness intersected by each drill cannot be 0
economically mined. Only a certain mineable thickness, defined by a
D2/m2 - 0.08 0.17
cut-off grade t, in nickel, is retained within each mining unit of 10 m x
10 m. It is supposed that future selection will be made on the basis of
information equivalent to a 10 m x 10 m grid of samples, i.e., one central
MINING GEOSTATISTICS VII SIMULATION O F DEPOSITS 521
TABLE VII.2. Experinzen~ai cor- For I, = 0 there is no vertical selection and thus no overburden; the
relanon coeficrenrs thickncss PL and accumulation AN; for tC = 0 thus characterize the in situ
resources before any selection is made.
1, PPO-PL PPL-AN,
0 - 0.87
1.0 -0.45 0.98
FIG. VII.7. Correlation P,,- P, for t,= 1.0. (a) Scatter diagram; (b) relative cross-
variogram.
simulated values r,, which d o not satisfy these inequalities are replaced by parameters can then be determined by obtaining a good approximation
the corresponding kriged value z&, cf. the conditioning relation (VII.3). between ypo and y$o. . *-..
Such an approximation will slightly modify the structural variability of the The details of these correspondance calculations yp,+ yr, are given in
simulation z,,, but is quite acceptable provided that the number of altered Addendum A. The structural model used in the simulation of the Gaussian
simulated values is small. variable U is a spherical model with range a , =260 m, a sill C, =0.215
and a nugget effect Co= 0.07. The theoretical model ypo deduced from this
Correlation PL-ANI. Except for their nugget constants, the direct and model is a good approximation of the experimental semi-variogram y'Yp,,cf.
cross-variograms on Fig. VII.6(b and c) can be deduced one from the other
Fig. VII.6(a).
by a multiplicative factor. This suggests a linear coregionalization model
with a single basic model, cf. formula (III.25), i.e.,
R e s ~ ~ loft s !he simulation The simulations corresponding to each value t,
of the cut-off grade were carried out independently. However, for a given
value of t,, the various service variables were simulated on the basis of their
coregionalization model.
The details of the calculation of the various parameters of the non-
conditional simulation are given in Addendum B. The kriging of the
with relative nugget constants Co = 0.04, a, = 1.5, a,, = I, and sill values
17 600 groups of service variables, required for the conditioning process,
pa52.67, Pap= 1.34. The values of these parameters ensure the positivity was described in section V.B.4, Fig. V.24. The remainder of this section
of the coregionalization model (VII.24), since, from formula (III.26),
consists of the verification of the final conditional simulations.
Tables VII.3 and VII.4 give the means, relative dispersion variances and
correlation coefficients of the 17 600 groups of conditionally simulated
The basic model y(r) consists of two nested spherical models: y(r)=
valuies. These results can be compared with those of the corresponding
GYI ( r ) f G Y Z ( ~ ) Tables VII.1 and VII.2 for the 176 real conditioning values.
with
Figures VII.8 and VII.9 give the histograms and isotropic semi-vario-
a l = 80 m, , - 0.015,
and sills ( C -
grams of the conditionally simulated values, along with - for comparison -
= 800 m, C2 = 0.036. those of the conditioning data values.
As the experimental histogram of the variable PL for tc = 0, cf. Fig. TABLEVII.3. Simulated means and disper-
VII.9, appears symmetric and Gaussian, PL can be simulated directly sion variances
without any prior Gaussian transformation. The same holds true for AN,;
because of the very high correlation between AN,and P, (pPL-AN,= 0.87)) fc Po PL
if the PL-histogram is satisfied, so will the AN,histogram. - 23.8 26.2
m
DZ/mZ 0 - 0.08 0.175
Regionalization of Po The experimental histogram c3f Po, given on Fig.
VII.8, is clearly asymmetrical. The simulation of Po will, therefore, require m 11.7 11.2 16.7
first a simulation of U, the Gaussian transform of Po; the required simula- DZ/m2 0.30 0.46 0.51
tion will then be given by Po = q,,,(U). The Gaussian transform function
cp,, can be fitted graphically to the experimental histogram on Fig. VJI.8. It TABLE VII.4. Simulated Cor-
can also be assumed that POfollows a lognormal distribution, in which case relation coeficients
q p o is given directly by formula (VI.8). The model y, for the Gaussian
variable U has been deduced directly from the structural model yp, of the
initial variable Po. Because U has to be simulated, a simple standard model
(spherical with a nugget effect)is adopted for y,, from which the model yp, is
deduced and compared to the experimental semi-variogram Y*p,: the
MINING GEOSTATISTICS VII SIMULATION OF DEPOSITS
Analysis of the results There is an excellent agreement between the There is a satisfactory agreement between the variograms of the simula-
global statistical characteristics, as can be seen from Tables VII.1 and tions and the theoretical models fitted to the experimental variograms, cf.
VII.3. The slight differences between the simulated and experimental Figs VII.8(b) and VII.g(b).
values is largely accounted for by the fluctuation variances of these
experimental characteristics on the zone considered.
There is a n equally good agreement between the strong correlations, cf.
Tables VII.2 and VII.4. O n the other hand, the weak correlation (pp,,- ,J =
-0.45), which was not taken into account (Po and P, were simulated
independently) is poorly reproduced.
Metres
Deiermination of the pararneiers For t, = 0, the two variables to be simulated are independent RF's Y, and Y2 on each of the 15 lines D and then taking the sum
PL and A,,, which are coregionalized according to the linear model (VII.24). Y = Y,+Y2:
From section lII.B.3, formula (llI.23), the linear model can be generaled by
simulating two independent realizations z and z' of the same RF Z with zero Y , = k,R, = k , ( ~+fi)
, and Y2 = kZR2= k,(~,*f,).
expectation and a semi-variogram y(r). The required realizations p, and a,, are The simulated range (2R + 1)u must approximate the real range and, thus, R is
then obtained by taking linear combinations of these two realizations z and z ' , plus an Integer equal to a/2u. For the first structure yl wlth a range a , = 80 m, R, =
a purely random variable that accounts for the nugget effect: 80/20 = 4 and the simulated range is, thus, a , = (8 + 1)lO = 90 m, which is a close
enough approxlmatlon of tnc real SO-m range. Similarly, for the second structure y,
with a range a, = SO0 m, It2= 40 and the s~mulatedrange is a ; = 810 m. Note that
the two ranges a , and a, can be reproduced exactly by altering the spacing of the
simulated points along the lrnes D. In practice, the approximations (a,, a ; ) and
The identification of the expectation E{PJ = m, and the relative semi-variogram
ypJd;= Co+y(r) leads to the following conditions: ... .,
(a,, a;) are largely justified, as can be seen by comparing the variogram on Fig.
7
VI1.II.
A:+A: = 1, and we take, for example, A , = 1, A, = 0, As the variables T, and T, are regularized by an odd weighting function f(x)=x
on the interval [-R. + R ) , the expectations of the RF's R l and 0, are zero:
E{E,} = 1 and Var {E,)= Co = 0.04.
The identification of the other terms of the coregionalization model results in the
conditions which entails
E{E,}= 1 and Var (8.1 = a,Co = 0.06,
A: +A:=p. = 2.67, The parameters k , and k, can be determined from the variances Var {Y,}and
A I A 3 + A 2 A 4= Pap = 1.34, Var {Y,]. For a range a = 2R, the regularized RF can be expressed according to
(VII.lO) as
and we take, for example, A, = 1.34, A, = 0,935.
Simulation of the purely random variables E For each of the 40 401 points of the
simulation zone, the subroutine RANDU provides two independent realizations to
and I, of the RV T evenly distributed on the interval (0, 11. The two purely random and, thus,
variables defining the two nugget effects are, thus, given by
v a r {R, = E{R~}= j
+R +R
- I? x d x l R y E{T(r+x)T(r+y)) dy.
Now,
E{T(r+x)T(t+ y)}= m $ + o $ . S,
where
The expectations of these two nugget effects are satisfied: E{eP}= E{E.} = 1 and, by
identifying the variances, the parameters k, and a can be determined: 1 if x = y,
m,=E{T}=O.5, Var{T}=1/12 and 6,(y)=(
0 if not.
Var { E , ) = k i Var {T,) = k:/12 = C,, i.e., k, = 0.69;
Thus, the variance becomes
Var { E ~=}k; Var {To}= Co;
Var (6,) = Var { ~ , } + aVar
~ {E,}= C o + a 2 C o =a,Co=0.06, i.e., a = 0.707.
In addition, the nugget constant aOpCOof the relative cross-variogram
y,+Jm,. m, can be verified:
where a = 2R. But the simulated range is a ' = 2R + 1, and, thus,
The three-dimensional simulation of the R F Z is then obtained by summing the This simulated data base represents a zone D consisting of 500 blocks of
15 independent realizations of Y on the 15 generatrng lines, yielding
dimensions 11 m x 11 m x 10 m, each of which conslsts of 11 X 11 x 20 =
( 2R + I)' 2420 ele~mentnryunits of dimensions 1 m x 1 m x 0 - 5 m, cf. Fig. VII.12. A
Var { Z )= 15k Z-----
144 ' vertical profile of 20 elementary data (grades, for instance) may represent a
This last relation can be used to determine k. For t, = 0 and for the structure y, profile of data from 24 pits analysed in 0.5-msamples. The data base thus
with a sill C, = 0.015, consists of 1 210 000 elementary values, from which one can extract the
vertical profile of grades of any of the 5 5 0 x 1 1 0 = 6 0 500 pits, or the
C,= 15ki (2R1 'I3 and k , = 1.405 x lo-'.
+
vertical profile of the true grades of the 20 samples of 0.5 m thickness of
144
any of the 500 component blocks.
For the second structure y, with a sill C2= 0.036,
D R 2 + 1 ) 3 and
c2=15k: k, = 0.806 x LO-).
144
Results of the non-conditional simulations As a verification, the semi-variograms
of the non-conditionally simulated variables were calculated in the two main
directions of the simulated zone. Figure VII.ll(a) and (b) give the simulated
semi-variograms and the imposed theoretical models for P, and A,, for r, = 0. For a
step size u = lorn, the first points of these simulated variograms are calculated
from 40 200 pairs of simulated values.
Zone 0
of a vertical profile of grades, cf. Figs 11.1 and VII. 13. The second term - CK and CB)of the two RF's K(u, v) and B(u, v) were chosen so that the
R(u, v, w) - represents the residual fluctuations around this drift. random drift D(u, u, w) accounted for a fixed part of the total variability of
, We impose the condition E{Z} = E{D}, which entails that E{R} =: 0. The the grade Z = D + R.
realizations d(u, v, w) and r ( y u, w) of the two RF's, drift and residual, are The positive constants Al, A2, A, fix the form of the cubic f ( w ) =
simulated independently. In this particular study, it was assumed that the h3w3+ h2w2+A1W, i.e., the general form of the drift.
elementary grades followed a lognormal distribution, and so a Gaussian
transform was used to ensure that the final realization z = d + r produced a Sir?udation of the residuals The three-dimensional regionalization of the
lognormal histogram. R F residual R(u, u, w ) is anisotropic with a more rapid variability in the
vertical direction. The model chosen to represent this R F is the nested sum
Simulation of the drift The vertical drift along each profile of grades
represents a progressive enrichment followed by an impoverishment, and,
finally, a further enrichment toward the bottom of the deposit, as shown on Qo is an isotropic three-dimensional nugget effect characterizing
Fig. VII.13. The following RF was chosen to characterize this drift: measurement errors and structures with ranges less than the smallest
D ( u , U, w) = K(u, v)[h3 w ~ ~ ( " +
*A" ) ~ w ~ +AI ~ . ~') 1,)
~ ( wB(u* ('VII.3 1) dimension considered, which in this case is 0.5 m.
Q, is a stationary RF characterizing an isotropic structure with a small
where K(u, v) and B(u, u) are two stationary and independent RF's range, in this case a spherical model with a range a l = 3 m.
defined in the horizontal plane (u, v), and A,, A,, A 3 are three positive Q2(u, v) is a stationary KF characterizing a horizontal structure with a
constants. larger range (a, = 30 m) corresponding, for example, to lenses in a sedi-
The cubic function f(w) = A3 w3 + w2+ A , w is, thus, modulated by the mentary deposit. The model can be refined by imposing a geometric
two RF's K(u, u) and B(u, v). The multiplicative factor K(u, v) defines the anisotropy on Q2, cf, section III.B.4, Figs 111.10 and III.ll(b), to represent
maximum value f(w,) of the function f(w), cf. Fig. VII.13(a) and Cb). The a pseudo-elliptical form of these lenses.
power B(u, v) defines the vertical coordinate wo of this maximum, cf. Fig. The three RF's Qo,Q,, Q2 are independent, and their expectations are
VII. 13(b), (c) and (d). taken as zero, to ensure that the expectation of the grade is equal to that of
The horizontal variabilities of the two independent RF's K(u, v) and the RF drift:
B(u, u) are characterized by isotropic spherical semi-variograms without
nugget effect, with identical ranges equal to a = 100 m and sills CK and Cg.
Thus, two sampling pits less than 100 m apart on a horizontal plane will Thus, on any given vertical grade profile, the residuals are correlated
have correlated drift curves d(u, v, w) and dl(u, v, w), while pits further only if they are less than 3 m apart. On any given horizontal plane, the
than 100 m apart will be independent. The expectations and variances (sills residuals are correlated only if they are less than 30 m apart. The horizon-
tal correlation of the grades Z = D + R for distances from 30 m to 100 m is
due only to the corre tation of the drifts.
in place of the unknown block grades profile (polygonal influence estima- method. Kriging is then used to condition the simulation to the real data
tion), then the resulting estimates of recoverable reserves will be biased, cf. values, and the Gaussian transform Z = cpZ(U) of this conditional simula-
the case study of section VI.A.4. Thus, a cut-o~ffgrade zo = 0.03% applied tion restores the required histogram. Finally, the simulation is verified by
to the central pit profile of the second block on Fig. VII.14 would result in means of a number of statistics, such as calculations of the direct and
a recovered thickness of 2 m with a mean recovered grade of 0.036%, cross-semi-variograms, histograms, scattergrams, etc.
whereas no block slice has a mean greater than 0.03%. The aim of kriging
is to provide an estimated profile as close as possible to the profile of the
Non-conditional Gaussian simulation
mean panel grades.
Only the non-conditional simulation by the turning bands method will be
?resented here, since the other steps in a conditional simulation all use
programs that were given in preceding sections, such as GAMA (section
III.C.4) for structural analysis and KRI3D (section V.B.3) for kriging. The
problem of non-conditional simulation can always be reduced to one of
simulating a number (NS) of independent realizations of Gaussian RF's
with given covariances. This latter simulation can be broken down into the
following three steps.
(i) The Gaussian variables U,(x) that are to be non-conditionally
simulated are linear combinations of NS Gaussian components
Vk(x)which are independent of each other.
Program SIMUL If the bias due to limiting the regularization interval to 4 a (instead of
infinity since the true range is infinite) is disregarded, then the covariance
Program SIMUL below handles NS independent Gaussian components of Yiis
defined in spaces of the same dimension (one, two or three). Each time a
space of different dimension is required, the program should be rerun.
However, the program can be easily modified to accept components lJk(x)
that are defined in spaces of different dimension.
Each of the NS components Vk(x) has an elementary covariance model
that is either spherical or exponential. The nugget effectis represented! by a
zero range covariance, and is treated separately.
whcre A = 1-c-~"'. e. second term, which represents the additive bias,
is negligible and each value yi can be corrected by the multiplicative factor
Correction of the discrete approximation CK to obtain the imposed variance c(''(o)= 1. Thus,
The first step is to generate the one-dimensional simulations along the 15
lines of the regular icosahedron. The one-dimensional elementary covari-
ances on these lines are obtained by regularization of a stationary random
measure, cf. formula (VII.9). This random measure is obtained in a discrete
form T,, by random drawings from a uniform distribution on thc interval
[-0.5, +0.5], which has a zero expectation and a variance of one twelfth. Operations of progratn SIMUL
This discrete approximation of the random measure on a grid b causes a The three-dimensional network of points to be simulated (the same for all
slight bias in the one-dimensional covariance - cf. formula (VII.12) - which the realizations) is defined by its number of rows, columns and levels, and
is corrected as follows. by the grid dimensions (parallelepipedic in rectangular coordinates).
First, for a spherical covariance with range a, the random variable T, is The grid UN of the discrete approximation on Iines (i.e., the thickness of
regularized on the interval [-a/2, +a121 by the algorithm (VII.ll): the turning bands) is taken as the smallest dimension of the three-dimen-
sional simulation grid. The number NGMAX of simulated points per line is
such that the length of the simulated part of the line is greater than the
main diagonal of the parallelepipedic network to be simulated.
with b = a/2R. The RV Y, has a covariance given by A spherical covariance is indicated by a range entered with a positive
sign; an exponential covariance is indicated by a parameter a entered with
a negative sign (the practical range is, thus, 3a). The values DX on the lines
and a variance given by are obtained by regularization of a random measure. There are KD values
of this random measure for each range. The regularization is carried out
over KD = 2*NR+ 1 values for the spherical model and KD = 8*NR+ 1
Each of the values y, can then be corrected by the multiplicative factor CK values for the exponential model, with N R = AA/(2*UN), and NR>20,
to obtain the imposed variance c(')(o)= 1, and, thus, m2 = 1/Ct'(0). AA being the range of the spherical model or the distance parameter a of
Second, for an exponential covariance with parameter A = l / a and, thus, the exponential model.
a practical range of a ' = 3 4 the discrete random variable T, is regularized Each point of the parallelepipedic network to be simulated is defined by
on the interval [O,4a] by the algorithm (VII. 11): its three integer coordinates (row, column and level) with respect to the
origin located at the centre of the network. These three coordinates cor-
respond to the trihedron formed by the first three of the 15 lines. The four
other trihedrons, corresponding to the 12 remaining lines, are deduced
with b = a/2R. from the first by the rotation matrix SX.
538 MINING CEOSTATISTICS VII SIMULATION OF DEPOSITS 539
The projections XI onto each of the 15 lines of each point of the C CLiDCIDN S I Z E OF S I P U L L T E D G P I D A L O h E L I N E S *
parallelepipedic network are determined. The three-dimens~onalsimula- c C C L U H h S ANC L E V E L S
tion is then obtained by summing the 15 one-dimensional simulated values C NkAS I N I T I P L I S A T I O N PARAHETEF FOR RAhCCp
C N U C B E R G E N E R A T I O N - T b K E 6 N Y NUMEER
DX at these 15 locations XI on the 15 lines. C CIV=16**-8 I S A G I V E h CONSTAH1
The values simulated on the 15 lines are stored in the matrix DX C I U T U N I T NUMBER OF O U T P U T F I L E
(NGMAX, 15, NS). C THE S I F U L A T I C H S ARE STCPEC I N SbCCESSICN:
The three-dimensional simulated values are stored in a direct access file C ( I (YSl IiJvL)tI=1iNLI )tJ=ltNCC.)sL=l th1V)t
C KtlthS
in the following manner.
C UStNS),CS(NS*NS) M E A N S A h D C C Y A R I A N C E S CF
r" SIPULATED LPLUES
(i) Each of the NS independent simulations correspond to N P =
C G ( N S * 2 8 1 SEMI-VARIOGRAW A L C h E COLUMNS
NLI*NCO*NIV values in the file. The simulations are stored in
sequence starting at address 0, so that simulation no. 1 corresponds
L
C - *COMMONS
to the first NP values, simulation no. 2 to the ncvt NP values, etc.
(ii) Each of the simulations can be read as a matrix YS(i, j, k ) with three
C
C ....~......~........~*..~....-~*.....-..--
IOUT L I N E P R I N T E R U N I T NbMBER'
SUBROLTINE S1lUL(hS,AthLIthCCihIVtCLiCC,Ch [:
I~NHAS~IUT~YICX,US~CS,G) CCMHCh [ C U T
c
C
C
.......*.................I.....*-*
ThPEE-OIMEhSIONAL S I P l L P T I O h
............... ECUIVdLENCE(S(l11,1)tSl(l ) l s ( S X ( l C ) t ~ ( l c l i l ~ ~
D A T A S l / C . 5 ~ - C . B C S 0 1 7 ~ 0 . ~ O E . O ~ ?t C . @ C S c 1 ? *
10~309017~-0~5rC~309017tC i Cm
. B5O S O l 7 /
C SUBRCUTINE O F NS I N D E P E N O E h T G d L S S I L N L A F I A T E S L
C I h O R P A L O i l ) W l T k E X P O N E N T I A L GR S P H E P I C P L C S U B R C U T I N E G E N E R A T E S I T S Chh P A h C C P h l F B E P S
C S C k E n f S ( I S O T R O P I C I N THREE C I h E k S I C h S ) . . R E S U L T S 1:
G ARE STOREC ON P C I R E C T - A C C E S S F I L E . . CATA C I V / Z 3 9 1 C O O O O / , I A R / - 1 1 3 7 4 6 7 %
C
C
c
Y
...
OPARACETEFS
k O R K I N G VEETCR:
1.
C
lI@R/8SC000001/
COMPUTE P O T A T I O N P A T R I X SZ O F TbE 15 B A N D S
C 1 - C O h T A I N S T h E RPNDCM V b L l E S T C E E t:
C C I L L T E D C h KC TERFS.I!KE A DIPEhSIOh cc 1 u + i , a
C GREATER T C A N 8 * N R + l 1 SXlKl=C.
C hR=RANGE/SUBDIVISIOh C h BAhDS SX(l)=l.
C S U B C I V I S I O N I S E O L A L 10 H I N - CF CL,CC,Oh S X ( 5 1.1.
C 2 - C C h T A I N S TI-E V A L U E S O F TPE k S S I l l L A T I C h S S X ( 9 1-1.
C F C R A C C L U M N OF N L I T E R H S : Y ( h L I * h S ) CO 10 Kz2.4
C hS NUMBER OF I N C E P E N D E N T S I F L L A T I C N S KO=K-I
C CX(hGMAX*lS*NS I M O R K I N G VECTCR C C h T A I N l h G T H E DO 11 I s 1 1 3
C V A L U E S O N Tt-E T U R h I h E E A h D S : DO 11 J 1 1 1 3
C hGMAX IS E Q U A L TO T t E NUMBER CF P C I N T S S 4 I v JlK)=O.
C SIMULATE0 ON A TURNING BAND=CIbEChAL 0 0 12 f l ~ l r 3
C CF TOTAL P A R A L L E L E P I F E E h L I l h C C , h I Y 1 2 S(I. J , K ) = s ( I,J,KJ+S( I.I~,KC)*S(I~~J~~)
C C O U h T E R I N NUMBER O F S U R O I V I S l O h S +6 11 C O N T I h U E
C A(NS) RANGES: >O. S P H E R I C P L SCHECE 1 0 CONTIhUE
C SO- hUGGET EFFECT c --
C to. E X P O N E N T I A L SCHECE C; I N I T X b L I Z E ANC COMPUTE S I v l L A T l O N P A P A P E T E P S
C hLltNC0,hIV NUMBER O F L I h E Z t C G L . U H N S , L E b E L S O F c AND suao IVISION OF B A N D S
C SIWULATED GRID c
530 MINING GEOSTATISTICS VII SIMULATION O F DEPOSITS
1H = I P R * N h A S + I B R 3 2 FIK)=(l.-XTI*EXP(-XT)
LS=O C1=1.-EXF(-2.*EPS)
CO 2 0 0 I S = l , N S CK=SORT( l Z . * C l * C l / ( C l - E P ~ * E X P ~ - 2 ~ * E P , S ) 1 )
L S ( I S)=O. IND=1
CO 2 0 0 J S = l , N S c
LS=LS*l 3 3 DO 3 4 I D = l t l S
2 0 0 CS(LS)=O. LO 3 4 0 K x l t K C
C IH=IAP*IH+IBR
UN=APlhl(DLtDC,DN) 340 Y ( K ) = I H * o I V
YRITE(IOUTv2000) UNINLIvCL~NCOiCC~hIb~DN C
Ct=DL/UN C NEW P O I N T : O I L U T I G N O h KO R P N C O P V A L U E S
DC=DC/UN C
ON=Dti/UN 00 3 5 ~ ~ ~ ~ N G H A X
NGMAX=SQRT(NLl*NLI*OL*CLthCC*NCO*CC*CC+ IP-I P+l
1 N I V * N I V * D N * D N )+6 AD-0
W R I T E ( I O U T t 2 0 0 1 J k G M A X . kS KK=HOCl(J-1)*IPASlKD)
C
DO 3 5 0 K = l , K D
C S I H U L I T I C N GN 1 5 R E F E R E h C E B A N C S IF(KK.EO*KD)KKsO
C KK.J(K+l
1P=O
CO 3 I S = l t N S 3 5 0 IO=AD+F( K I * Y (KK)
IF(A(IS).NE.O.)GO TO 3 C C TRANSLATE RANCCH H E A f l R E
C hUGGET E F F E C T DO 3 5 1 # = l u I P A S
bA=O. 1FIKK.EO.KDIKKmO
IPAS=:! KK=KK +1
NR=l IH-IAR*lh+IBR
3 5 1 Y(KK)=IH*DIV
EPS-0-
IO=AC*CK
PAS=Uh/IPAS
US(ISl=US( IS)+AD
GO T C 3 0 2
OX(1 PI=AD
C
DO 3 5 2 JS=LIIS
30 bA=A@SLA(ISI)
IJS=IS+NS*(JS-11
FiR=AI/(Z*UN)
JP=IP+NGWAX*15*( J S - I S )
IPAS=l
3 5 2 CSIIJSI=CS(IJS)+AO*DX(JP)
IF(NR.GE.2O)GC T@ 3 0 1
NR=MAXOL 1 1 N R )
35 CONT XHUE
34 C O N T I N U E
CO 3 0 C 1 P A S = 2 y 2 0
IF('INC.EO.I~GO TO 36
IF(NP*IPbS.GE.2O)GO TO 3 C 1
Y R I T E ( I O U T , 2 0 0 2 ) I S v A A i P A S 1 EPS, I P A S 9CK
3 0 0 CONT I h U E .
60 T O 3
3 0 1 PAS=Uh/IPAS
36 W R I T E L I O U T ~ Z 0 0 . 3 )I S , A A , P A S s E P S t I P b S , C K
EPS=PbS/bA
3 CONT I l k U E
NR=NR*IPbS URITE(IOUT,2004)( IS* I S ~ l r h 5 )
I F ( A ( IS).LT.O.iGO TO 3 1
00 3 7 I S - 1 , N S
C S P H E R I C A L SCt-EPE
IIS=IS+NS*(IS-1)
3 0 2 CK=SCRT(36./NF/(NR+l)/ ( 2 * h F t1) )
C S I I I S ) - ( C S t 11s)-US( I S ) * L ' 5 ~ I S 1 / ~ 1 5 * F c G R P X ) ) /
KO=Z*hR+l
l(15*NGMAX)
DO 3 0 3 K = l v K C
U S ( I S )=US( I S ) / ( l S * N G W A X 1
3 0 3 FlK)=K-NR-1
00 3 7 0 J S - 1 . I S
IND=O
IF(JSeEO-ISIGO TO 3 7 0
GO T C 3 3
IJS- IS+NS*Y J S - 1 )
C E X P G N E N T I A L SCHEME
JJSIJS+NS*IJS-11
31 KD=8*hR+l CS(IJS1-lCS(IJS)/(15*NGHbXl-US(IS)*USf JS))/
DO 3 2 K = l , K D
lSQRT(CS1 IIS)*CS( J J S I )
XT=EPS*(K-1)
370 C O N 1 I N U E
542 MINING GE0STATISTIC:S
VII SIMULATION OF DEPOSITS 543
WRIT EtlQUT*2005J I S , U S ( IS89 ( C S f I J S I~ I2=LK+NGPAX*(KO-1 l+NGHAX*l5*( IS - 1 )
lfJS=I~,IIS,NS
Il=I+hLI*(IS-1)
37 CONTINUE C . J
*
L
43 Y(Il~=YfIll+DX(I2J
41 CONTIhUE
C THREE-GIMENSIGNAL S I C U L A T ICN ON N I b LEVELS DO 4 4 I S = l * N S
-3
CO 5 I S = l , N S
I I S x IS+NS*( IS-1)
C S ( 1 I S ) = ( C S ( 11s)-US( I S ) * L S ( I S l / h P l / N P
US(IS)=US( IS)/NP
DO 5 1 J S = l , I S
1FLIS.EO.JS)GC TO 5 1
JJS=JS+NS*( JS-1)
IJS=IS+NS*(JS-1)
CS(I JSI=(CS( IJS)/NP-US(I$I*US( JS) ) /
lSQRT(CS( lIS)*CS( JJS)
5 1 CONTIhUE
W R I T E ( I O U T ~ 2 0 0 5 1I S , U S ( I S ) , ( C S ( I J S ) r
llJS~15111SrNS)
5 C O N 1 IhUE
IWP=(KS-l1/1e+l
ISM=FLOAT(NS) /FLOAT( I M P )+C -9999 Vll.B.3. Typical u s e s of simulated deposits
DO 5 2 I M = l , I H P
ISl=l+ISP*(IM-11 A simulated deposit represents a known numerical model on a very dense
I S ~ ~ M I N O ( N SI SVF * I M )
MRITE (IOUT92OOB) (15, I S x I S l , I S 2 1
grid. As the simulation can only reproduce known, modelled structures,
DO 5 3 K z l v K V the simulation grid is limited to the dimensions of the smallest modelled
IK1= IK-1 )*NS+ISl structure. Various methods of sampling, selection, mining, haulage, blend-
IK2= IK-l)*NS+IS2 ing, etc., can be applied to this numerical model, to test their efficiency
0 0 54 I K x I K l , I K 2
NC=N I V * N C O * ( N L I - K
before applying them to the real deposit. The great advantage of such tests
5 4 G( IKl=O.S*GI I K I / N C is that they do not require the large investment or the delays involved in
O=K*UK*DL setting up a pilot operation; they can be carried'out for the cost of
IF(K-GT.20)D=UN*OL*~lO*K-180) comruter time, and delays are limited to the response time of the cornput-
WRITE(IOUT~2007)K~D~hC~~G~IK)~IK~IKl~IK2) ing equipment used.
5 3 CONTINUE
52 CONTINUE Some exan~plesof typical uses of simulated deposits are as follows.
C Comparative studies of various estimation methods and various
2000 fORMATtlHlrr CARACTEPISTICS OF TFE ' 1
1 ' S I M U L A T I O H '/1H r 5 X q l 4 t ' * ' ) , l X , ' * * *** ', approaches to mine planning problems, including geostatistics, cf. P.
210('*')//1H 1' S U B D I V I S I C h O h BAhDS='rF7.3, Dowd and M. David (1975), A.Marechal (1975d).
3 1 0 x 9 ' S I Z E OF G R I D : ' 9 1 4 9 ' L I N E S ,
D L I' Studies of the sampling level necessary for any given objective, cf.
4F7.3/1H '
9 5 6 x 1 14. C O L U M N S DC x ' r F 7 . 3 / 1 H 9
R.L. Sandefur and D. C. Grant (1976). In this particular work, the
556x1 1 4 7 ' LEVELS ON 'rF7.3)
2 0 0 1 F O R M A T ( 1 h O q ' KUMBER O F P a I h T S S I M L L A l E C / ' numerical model consisted of the true data from a very dense
l * ' T U R N I N G BAN0 ',IS,' h L Y B E R OF I k C E P E h D E N T ' sampling grid over a given zone.
2.' S I C U L A T I O N S a '*I2//) A study of the influence of the amount of information available at
2 0 0 2 FORNAT(1H t1*SIHULATIOh*',13r' S P H E R I C A L ', the time of selection on the recovery of the in situ resources, cf. 0.
1' S C H E M E R A N G E ~ ' t F 8 - 4 , ' L A G = l , F E . ? ,
2' P R E C I S I O N = ' r F 5 - 3 r ' LAGS/SUBO. 14. Bernuy and A. Journel (1977) and section (VI.A.4).
3' FACTOR CK r ' r F 8 . 4 ) A study of the size of the minimum selection unit on the recovery of
2 0 0 3 FORMATllI' r'*SIMULATION*'r13r' EXFCNEhTIAL ' 9 the in situ resources. A study of the influence of spatial concen-
llSCHEWE RANGE = ' t F 8 . 4 r ' LAG ='*Fe.?, trations of rich and poor panels on this same recovery and on the
a1 P R E C I S I O N a ' ~ F 5 - 3 , ' L A G S / S U B D . = ' , 1 4 , dilution of the recovered ore. All these studies are involved in
3 ' FACTOR CK t ' r F 8 - 4 )
2004 F O R M A T ( l k O 1 ' S T A T I S T I C S CF R E S L L T S Ch BANDS' problems of choice of mining method, equipment size, type of
l ' ' ( C O R R E L A T I 0 N H A T R I X ) ' / I H * 2 X 9 1 0 ( '-'Ir
2' -- ------ - ----- 'Ilk t'SLHULAlICh'/lH 9
haulage, etc., cf. J. L. Rebollo (1977),I. Clark and B. White (1976),
P. Switzer, and H. Parker (1975).
546 MINING GEOSTATIST1C:S VII SIMULATION OF DEPOSITS 547
(v) A study of blending from various stopes to stabilize the tonnage and (iv) an isotropic regionalization of the grades characterized by a spheri- -
quality of production (grades, hardness, mineralogy). A study of the cal model with nugget effect, and a range of approximately 70,m.
requirement for blending stockpiles, cf. J. Deraisme (1977), whose
work is partially presented in the following case study. Figure VII. 15 also shows the histogram of the block values kriged from
(vi) A study of the determination of production rater ~ n cut-off
d grades, the central drill core intersections on a horizontal 18 m x 18 m grid. This
which vary in time, and their impact on recovered benefits, cf. P. histogram of kriged values has a mean m*= 1% Cu and a variance
Dowd (1976). D;* ((v G) = 0.185O/0~;the kriging variance is cri = 0.032%~.
It must be understood that results obtained from simulated deposits will
apply to reality only to the extent to which1 the simulated deposit
reproduces the essential characteristics of the real deposit. Thus, the more
the real deposit is known, the better its model will be, and the closer the
co#itional simulation will be to reality. As the quality of the conditional
simulation improves, not only will the reproduced structures of the vari-
ability become closer to those of reality, but so also will the qualitative
characteristics (geology, tectonics, morphology, etc.) that can be deter-
ministically introduced into the numerical model. Once again, it must be
stressed that simulation cannot replace a good sampling campaign of the
real deposit.
not significantly different from his prediction. The short-term (e.g., hourly) 18 m x 18 m x 5 m blocks per day. The resulting daily production of 35 000
fluctuations of the mill feed grade are the responsibility of the metallurgist tons is sent directly to the mill.
and the process controller. The mobility of the shovels means that the eight blocks to be extracted
, It is assumed here that grade is the primary variable affecting mill can be chosen so as best to stabilize the daily mean grades over the
operation. Any other regionalized parameter (impurity grades, crushing fortnight. All the ore of any given mining front (450m long) must be
indexes, etc.) can also be simulated and used in an analogous manner. extracted before beginning on the next, and the total moving distance of
each shovel is to be kept to a minimum.
Projected nilning methods Three difierent mining methods were consid-
ered to meet the required degree of honiogenizat~onof daily mean grades
over a fortnight's production. These methods are all rtorz-selcctrve: every
block is eventualljf sent to the mill.
Method I . Zone G is mined by two 10-m-high benches, cf. Fig.
VII.l6(a). From each bench. a big shovel of limited manoeuvrabll~tycan
extract two contiguous 18 m x 18 m x 10 m blocks per day. The two shovels
advance parallel to the mining front and cannot reverse: they must extract
all the ore along an) given front (18 x25 = 450 m long) before beginning
on the next front. All extracted ore is sent directly to the mill. Slnce the
motion of the sho\els is predetermined in t h ~ smethod, there is no choice
for the mining engineer, who must rehnquish his part In the homogenlza-
tion process.
Method 1'. Mining is identical to the previous method, but some of the
mined ore can be sent to a stockpile where homogenization can take place,
cf. Fig. VII. 16(b).
The stockpile consists of two subpiles. one rich, the othcr poor. The Mill
maximum capacity of the total stockpile is two day's product~on('70 000 (0) t
tons), which is the equivalent of eight large blocks. A block will be sent to M~ll
the stockpile when there is a risk that its mean grade afTects the day's mean (b)
feed grade. When a block is sent to the rich subpile, the equivalent of a
block is taken away from the poor subpile to meet the daily relquired FIG. VII.16. View of the mining fronts (blocks l o r n high). (a) Method 1;
tonnage (35 000 tons). (b) method 1'.
In addition, the existence of the stockpile providcs some degree of
flexibility to the mining operation, since on certain days, three b1ocl.s may The homogenization obtained by blending the production coming from
be extracted from one bench and one on the other, provided that equilib- the four working levels will result in higher mining costs.
rium is restored within eight days; alternatively. five blocks may be
extracted within any given day, provided that this over-production is Only method 1, for which the path of
Simulation of the mining processes
balanced by an equivalent under-production within the next eight days.
the big shovels is entirely determined, is completely simulated by
But mining is expected to adhere to the norm as often as possible, i.e., daily
computer. For methods 1' and 2, the daily decisions such as which block to
production of two blocks per bench.
mine, and whether to send it to the mill or to the stockpile, are made by an
operating mining engineer;just as they would be in a real situation. The
Method 2. The same zone G is mined by four 5-m-high benches, cf. Fig. decisions were made through visualization of the state of the fronts and the
VII.17. From each of these benches, a mobile shovel can extract two small stockpiles, taking the past and immediate future mill feed grades into
VII SIMULATION OF DEPOSITS 55 1
account. The decision of the mining engineer is certainly not always opti-
mal, but the object of this study is precisely to reproduce reality (including,
in particular, human errors) and not an absolute optimum which'is often
inaccessible in practice. .%.j
The engineer makes all his decisions on the basis of kriged estimates of
the mining blocks, the blocks being kriged from the central drill core
intersections. However, it is the true block grades that are delivered to the
mill, cf. section VI.A.3.
As for the type of blending performed within the subpiles considered in
method l', it is assumed that the true mean grade of the unit taken from a
subpile is a normally distributed random variable, with the foilowing
characteristics.
(i>l An expectation m, equal to the true mean grade of the subpile at the
time considered. This instant mean grade varies around 1.654 Cu for
the rich subpile, and 0.6% Cu for the poor subpile.
(ii) A fixed relative standard deviation crs/ms equal to 0 . 2 5 % ~for the
rich subpile and 0.3%' for the poor. These values were chosen so
that the confidence interval [ms*2vs] for the rich subpile, for
example, includes all individual true block grades sent to this partic-
ular subpile during the course of one year.
This model of the blending effect of the stockpile is very simplistic and
pessimistic, but is enough to give an order of magnitude of the influence of
the stockpile on the homogenization of the mill feed grades. In a real case,
it would be a simple matter to replace this model either by a deterministic
system or a more precise probabilistic model.
Results For the three simulated methods of mining zone G, Fig. VII.18
shows the variation in daily mean grades during the first two months (60
production days). J. Deraisme's study was carried out for a year's pro-
ductjon (280 days) but there is not enough space here to show all the
results.
The full-line curves show the variations of the true daily grades (i.e., the
homogenization actually achieved), while the broken-line curves represent
the variation of the estimated daily grades (i.e., the homogenization pre-
dicted on the basis of the kriged block grades).
The mill operator will obviously prefer methods 1' and 2 to method 1,
which the mining engineer may prefer because of its lower mining cost.
Thus, there is a real problem involved, which can only be solved by
assigning costs to the mill recovery losses caused by method and to
balance these costs with the ones caused by stockpiling (method 1') or by
the more flexible mining operation of method 2.
..
552 MINING GEOSTATISTICS VII SIMULATION O F DEPOSITS 553
Table VII.5 gives the different dispersion variances observed at different estimation procedures (including kriging) are irrelevant for predicting
scales, and, thus, gives an overall picture of the effect of each mining these variabilities, for they take no account of the destructuring effect of
method on the fluctuations of the mill feed grades. mining, trucking, stockpiling, ctc.
The deviation between the dispersions resulting from the different min-
ing methods increases as the time interval decreases. Thus, whatever
method is used, the amount and location of the ore extracted during a
given fortnight will be approximately constant, and the dispersion vari-
ances of a fortnight's mean grade over one year will also be approximately
the same. On the other hand, methods 1' and 2 are particularly efficient in
stabilizing the daily grades within any week or fortnight.
The dispersions of estimated grades are always less than those of true
grades. This is a result of the classical smoothing property of the kriging
estimator, cf. relation (VI.1) or (VI.2): the dispersion of estimators is not a
good estimate of the real dispersion. For method 2, even though the
FIG. VII.18. Fluctuations of the daily mean grades at the mill-feed.-, True homogenization of the estimated grades is nearly perfect, there is still an
grades; - - - -. estimated grades. irreducible fluctuation of the true grades, due to the estimation variance of
the daily grades, cf. Fig. VII.18.
The following remarks can be made about these results. The experimental estimation variances of the daily grades for each
There is a considerable variation in the fluctuations of the daily grades method are:
sent to the mill, depending on the mining method used. The dispersion
variance of the daily grades in 15 days for method 1 is six times greater method 1, a$ =0.0089~/0*;
than for method 2. These orders of magnitude of the mill feed grades method l', & = 0.0131%2;
variabilities at various scales can be predicted only by simulation. Standard method 2, a$ = 0.0061 ?L2.
554 MINING GEOSTATISTICS
Note that for methods 1 and 2, the dispersion variance 0' of the true
grades can be deduced from that of the estimated grades D:* by the
approximate smoothing relation (VI.2):
D'= D$* -I-&.
This relation is not applicable for method 1' because daily production does
not come only from the mine but also from the stockpile.
For one year of simulated production, 17% of the production from
method 1' passed through the stockpile. This represents a considerable
proportion for a stockpile with a maximum capacity of two days' pro-
duction. In fact, to obtain the homogenization shown on Fig. VII.18,
Y ntroduction to Non-linear
almost 50% of production days involved departure from normal pro-
duction (two blocks per bench). One solution would be to increase the
Geostatistics
capacity of the stockpile by, for example, adding a third subpile of average
quality ore. In any case, it would appear that the capacity of the stockpile
should not be greater than three or four days' production. Method 1'
produces dispersions of the same order of magnitude as those of method 2
and since its mining cost is lower (10-m benches instead of 5 m) a feasibility SUMMARY
study would probably result in method 1' being chosen.
The estimators that have been considered until now are linear combinations of the
Globally, there is no selection involved in any of the methods considered available data, and their prerequisites are limited to only the moments of order 2 of
in this study, However, in his more complete study, J. Deraisme (1977) the IZF Z(x). If more is known about the R F Z(x),for example, if its I-, 2- or
studied various selection criteria for method 2 and compared the results of k-va.riate distributions can be inferred, then non-linear estimators can be built.
applying these criteria on the simulations to the results that could be These estimators are more accurate than the linear kriging estimators, and allow
predicted on the basis of grade histograms alone. The differences were the estimation of unknown functions which are not simply linear combinations of
unkriown values.
significant, and illustrate the well-known effects of concentration and dilu- The word "kriging" designates the procedures of selecting the estimator with a
tion of reserves: a rich block with a grade above the cut-off will not be miniirnum estimation variance within a given class of possible estimators (linear or
selected when it is located in an overall poor zone; conversely, a poor block non-linear). This estimation variance can be viewed as a squared distance between
with a grade below the cut-off will be selected when located in an overall the unknown value and its estimator; the process of minimization of this distance
rich zone. The prediction made on the basis of the grade histogram (cf. the can Ithen be seen as the projection of the unknown value onto the space within
which the search for an estimator is carried out. In terms of these projections,
various techniques outlined in section V1.A) are unable to take account of section VII1.A introduces the various non-linear kriging estimators such as the
this spatial localization of rich and poor grades; more generally, they conditional expectation or "disjunctive kriging".
cannot take the technological conditions of the mining operation into Disjunctive kriging is developed in section VII1.B. together with its application
account. Once again, a solution is afforded by the: simulation of the mining to the estimation of non-linear functions, such as the so-called "transfer function"
operation on a numerical model of the deposit: th~esimulations can be used defined as the proportion of unknown values (e.g., grades of support u ) greater than
a given cut-off within a given domain V.
to evaluate the effect of these technological conditions on the actual
recovery of the reserves.
On the other hand, the prerequisites were strengthened by requiring the VIII.2, with 8"c 8'c8,and the corresponding kriging estimators Z? and
knowledge, not only of the covariance or variogrann, but also of the 2- or Z*': as 8' includes 8", the projection Z*' is nearer to the unknown than
k-variate distributions of the RF Z(x). Non-linear estimators could then be Z*",i.e., in terms of estimation variance .*ir
FIG. VIII.l. The kriging estimator defined as the projection of the unknown Z(xo)
onto the subspace 8' c 'e.
The neutral element 0 = (A = 0). Z(x) of %' is the random variable almost
certainly null.
The vector space 8 is provided with a scalar producr equal to the
non-centred covariance (which is not necessarily stationary):
manifold C c The restrictions of to various linear manifolds C the L parameters al being unknown. The unbiasedness condition
guarantee the unbiasedness of the estimator Z * , as will be shown. (VIII.1)can then be written as
Conditions of unbiasedness
This relation is satisfied, whatever the unknown parameters a,, if
Consider the expectation of any element Z* = Ao+Ca AaZa E and only if
A. = 0 ,
which amounts to the restriction gn c %,,+,, and
The element Z*, considered as an estimator of Z ( x o ) is unbiased if and
only if
Ao+x Aarn(x,) = rn(xo). (WII.1) The last L conditions amount to a restriction of %, to the linear
a manifold CL, of dimension ( n - L).
Various cases are distinguished according to whether the expected Note the inclusions CLc CL-I.. . c C1c %,,c C&,,+]. The wider
values m(xo).m(x,) are known or not, and, in the second case, whether the set onto which the projection is made, the nearer the correspond-
m ( x )is stationary or not. ing projected value Z* will be to the unknown Z ( x o ) ,cf. Fig. VIII.2.
Each of these projection sets gives rise to a particular kriging process.
1. All the expectations are known (it does not matter then whether they
are stationary or not). The unbiasedness is then characterized by the
single condition (VIII.l),i.e., 1. Linear kriging with known expectations ("simple kriging")
The kriging estimator Z;, is of the form
ZgO= A O + C A~oaZa
a
2. The expectation of the RF Z ( x ) is stationary but unknown, i.e., and is the projection of the unknown Z ( x o )onto the vector space cf.
E { Z ( x ) }= m, Vx. The condition (VIII. 1) is then satisfied if and only if Fig. VIII.3. This projection is unique and is characterized by the orthog-
onality of the vector Z(x0)-Z;, to each of the (n + 1) vectors generating
Ao=O and C A , = ~ . %+I:
0
G(x0)- zi,, 1) = 0 ,
} (V111.2)
The first condition A. = O amounts to restricting the set of the p~ssible
( Z ( x o ) -Z;,, Z,) = 0, V a = 1 to n.
estimators to the vector subspace %,,c generated by the linear
combinations xaAaZa of the n data only. The second condition
1, A, = 1, amounts to restricting g,,to the linear manifold Cl defined
by the condition Ca A, = 1 on we~ghtsA,.
3. The expectation m ( x )is neither stationary nor known. We are then at
a loss to express the unbiasedness relation (VIII.l). It is necessary to
provide the form of the expectation m ( x ) , for example, m ( x ) is an
unknown linear combination of L known functions f(x), cf. formula
(V.9): /&A, /
FIG. VIII.3. Simple kriging defined as the projection of the unknown Z(xo)onto
%"+I.
562 MINING GEOSTATISTICS VIlII INTRODUCTION TO NON-LINEAR GEOSTATISTICS 563
Considering the centred covariance = wap- m(x,). m(xp),the pre- 2. Linear kriging with unknown stationary expectation ("ordinary kriging")
vious system (VIII.2) can then be expanded into
The unbiasedness condition (VIII.l) thus restricts the search for a liiieir
[ h= m(xo)-E A G p ( x s ) , estimator to the linear manifold Cl c gn defined by the two condi-
8 tions A. = 0 and x, A, = 1 , on the weights. The kriging estimator is th%n
defined as the projection of the unknown Z(xo)onto the linear manifold
C l ,cf. Fig. VIII.4.
Note that
m(x)= E { Z ( x ) )= ( Z ( x ) ,1).
The first equation is simply the unbiasedness condition (VIII.l),the n
last equations constitute a system of n linear equations with n unknowns,
called the "system of simple kriging":
FIG. VIII.4. Ordinary kriging defined as the projection of the unknown Z(xo)onto
The kriging estimator is, thus, written as the linear manifold Ci.
The centred estimator [ZR, - m(xo)]thus appears as a linear combina- (b) The vector Z ( x o ) - Z g , is orthogonal to each of the n vectors Z, - Y
tion of the n centred data [Z, - m(x,)]; more precisely, [ Z $ -m(xo)] is generating C1,with Za, Y EC1.As Y is an arbitrary element of Cl
the projection of the unknown [Z(xo)-m(xo)]onto the vector space 8; (for example, the mean value Y = ( l / n ) C , Z a ) , the previous
generated by the linear combinations of the n centred data. orlhogonality is characterized by the n relations
This system is simply the "ordinary" kriging system (V.l) most cornmonly 1 This linear kriging process in the presence of a drift is called "unbiased
used, at least in mining practice.
E{[Z(xo)- Z&12) = lIZ(xo)- Zg, 112, is found to be Kriging variance The c~~rresponding
minimum squared distance is writ-
1
i
ten, according to (V.13). as
2
1
u~~ = uxoxo -C* A ~ m c a r o + l~lf/(xo).
C
I
(VIII.8)
Note that the system (VIIIS), as well as the expression (VIII.G), can be * Once again, note that thc relations (VIII.7) and (VIII.8) can be written in
restated by replacing the non-centrcd covariance by the ccntrcd covariance
{ tcrms of the ccntrcd covariance a&or the semi-variogram ymp.
a:, = a,, - rn2. The unknown value nt2 is climinated from these equations,
thanks to the unbiasedness condition 1,A K l a = 1. Similarly, these equa-
Prerequisites to obtain Z:,. The form of the drift m(x), i.e., the L
tions can also be written by replacing the covariance function a,, by the
functions f/(x), is assumed to be known, and the statement of the kriging
semi-variogram function y,, defined by system (VIII.?) requires the knowledge of the covariance function (centred
o r not) or of the semi-variogram. In practice, since only one realization of
the non-stationary R F Z ( x ) is available, problems arise in the inference of
Prerequisites to obtain ZZ, The expected value needs no longer to be the nor]-stationary covariance E{Z(x)Z(y)}, cf. section V.A.2, G .
known, but is assumed to be stationary. The statement of the kriging Matheron (197 1, pp. 188), P. Delfiner, (1975).
system (VIII.5) requires the knowledge of the covariance function (centred
o r not), or of the semi-variogram.
Vlll.A.2. Conditional expectation
3. Linear krigirig in the presence of a drift ("uniuersal kriging") If {Z,, cu = 1 to n) are the n available data, the most general form for an
estimator Z * is a measurable function f ( Z l , Z,, . . . , Z,) of these data. The
The expectation of the RF Z ( x ) is neither stationary nor known, Dlut is of set H, of all these measurable functions of n data is a vector space which
known form: contains, in particular, the vector subspace 8" generated by all the linear
L
E{Z(x)} = rn(x)= C alfi(x). combinations 5,AaZ, of the n data. As H,, is the widest space where the
/=1 search for an estimator can be carried out, the projection of the unknown
The unbiasedness condition (VIII.1) thus restricts the search for a linear Z(xo) onto Hn is the best possible estimator of Z(xo) that can be deduced
estimator to the linear manifold C, c gndefined by the L following condi- from the n data Z,, cf. Fig, V1II.S. By definition, the best possible estima-
tions on the weights: tor is simply the conditional expectation of Z(xo), given the n data, cf. J.
Neveu (1964, pp. 1 16). i.e.,
c ~,fi(x,) = fi(xo),
a
vl= 1 to L.
The kriging estimator Zk, = C,AKLaZ,, is then defined as the projection of This projection is characterized by the orthogonality of Z(xo)-En& to
the unknown Z(xo) onto the linear manifold CL.This projection is charac- any vector Y = f ( Z 1 , . . . , Z,) belonging to H,,, i.e., by the relation
terized, cf. section V.A.2, by the system (V.12) of ( n + L ) linear equations
with (n + L) unknowns, i.e.,
Prerequisites for obtaining E,Zo In order to build the projection E,,Zo,
i.e., to find the n-variable function E,Zo=fK(Z1,. . . , Z,,)E H,, it is neces-
sary to be able to express scalar products of the type
( Z ( X O ) - ~ KJ ), = E { [ Z ( ~ O ) - ~ K ( .~.I,>Z .n ) l . f ( Z ~ ,.. . , Z,,)), Vf E H,.
566 MINING GEOSTATISTICS VIII INTRODUCTION TO NON-LINEAR GEOSTATISTICS 567
To do this, the distribution of the ( n + 1) variables {Z(xo), Z(xl), . . . , En&= U& and a variance given by the kriging variance cr& provided by
Z(xn)) must be known. In the general case, whlen the information is limited relat~on(VIII.4). This conditional distribution is, thus, completely deter-
to a single realization of the RF z ( x ) , such an inference is not possible and mined, and it is easy to obtain the required conditional expectation of the
the conditional expectation is inaccessible. initial variable Z(xo) = pZ[U(xo)], i.e,,
Note that, because the transform function p, is generally not linear, the
conditional expectation EnZois not linear with respect to the initial data
(ZI, , 2 2 9 . . > Zn).
for any /3 = 1 to n and any single-variable measurable function g. This Similarly, considering the projection of 26, onto do:
condition is satisfied if and only if Z(xo) and ZgK admit the same condi-
tional expectation upon each of the Zp, /3 = 1 to n, taken separately, i.e.,
The preceding condition (b) can, thus, be written as
(EpZo,g(Zp))=(EpZ;~,g(Zp)),
Vg(Zp)~dp.
'
This is satisfied if and only if
-
EOZo EoZcK, V p = 1 to ,I,
(The proof, although classical, is given later on.) which is the system (VIII.ll).
Y
MINING GEOSTATISTICS
C ( h ) =m2[eC'(h'- 1 ] + ~ 2 = m'!(e"'2-
1).
}
(VI11.13)
The n weights AKon arc determined from the simple lincar kriging system experimental estimation variance) than the classical linear kriging estima-
(VIII.3), and the variance E{[ lJ(xo)- Enyo]'} is simply the corresponding tors (simple or ordinary) that can be built directly from the initial data 2,.
kriging variance a& given by expression (VIII.4).
Consider now the conditional variable Z(xo)/Z1,. . . , Z, =
exp [ Y ( x , , ) ] / Z ,.,. . , Z,. This variable is lognormally distributed, and its
expectation EnZois deduced from the parameters of the distribution of
Y(xo)/ Y1, . . . , Y , through the formulae (VII1.13), i.e.,
I V1II.B. DISJUNCTIVE KRlGlNG AND ITS APPLICATIONS
Let us recall the results already obtained in section VIII.A.3. The dis-
junctive kriging estimator of an unknown Z(xo) from n available data
{Z,, a = 1 to n } is the sum of n measurable functions of each datum taken
The r~on-linear estimator Z: =exp [Y:, + a & / 2 ] thus appears as the separately, i.e.,
conditional expectation E,Zo which is the best possible estimator of the
unknown Z(xo) from the n available data Z,. The corresponding estima-
tion variance is expressed as
The solution functions f, are given by system (VIII.12):
E{[z(x,)- E,,Z~]~}
=t ~ e"'
' (1 - e-"zo). (VlII.15)
(The proof is immediate by expanding the square [z(xo)- E,,Z,]~.)
Note that disjunctive: kriging can be used to estimate any measurable
Prerequisires lo obtairl E,Zo The RF Z ( x ) must be stationary and rnulti- function p(Zo) of the unknown 20= Z(x0). This unknown value q(Zo) is
variate lognormal. and the expressions (VIII.14) and (VIII.15) require the projected onto the space D, providing the estimator &, =C, f,(Z,)
knowledge of the expectation and covariance (of either Z ( x ) or Y(x)= characterized by the system
log Z(x)).
In practice, it has been noticed that the estimator Z $ = exp [ YE, -+- cr&/2]
does not always fulfill the non-bias condition, i.e., the arithmetic mean of
the estimated values Z $ can difier noticeably from the expectatton tn
estimated from the available data. A solution then consists in considering
1 The statement of these disjunctive kriging systems requires calculation of
conditional expectations of the type E(f(Z.)/Zp), i.e., the knowledge of
the estimator
t the bivariate distributions of the RF Z ( x ) .
Ii VIII.B.1. Bivariate Gaussian distribution
g = KOexp [YE, + ff&/2],
Z (VIII.16)
the corrective factor KObeing determined by equating the arithmetic mean Let Y(x) be a stationary F G with univariate and bivariate Gaussian dis-
of the estimated values Z$ to the expectation m. tributions. This entails that:
This divergence of the estimator (VIII.14) with regard to its theoretical
non-bias is due to the lack of robustness of the exponential expression 1. all the random variables Y(x), Vx, have a Gaussian distribution
(VIII.14) with regard to the multivariate lognormal hypothesis. Even (assumed to be centred with unit variance), i.e.,
though the univariate distribution of Z ( x ) can be fitted to a lognlormal
distribution, it does not necessarily follow that its multivariate distribution G(du) = Yrob {u s Y(u)-s u +du} =- e-142/2du;
is also lognormal. And, in practice, it is not possible to test the multivariate a
lognormal character of Z ( x ) from a single realization z(x), particularly 2. all the bivariate clistributions of Yi = Y(x,) and Y, = Y(x,) are Gaus-
when one is limited to a few data points z,. sian, i.e.,
However, experience has shown that, if the univariate distribution
G ( d u , , d u , ) = f ' r o b { u , ~Y , S u , + d u , ; u,s Y , ~ u , + d u , }
of Z ( x ) is clearly lognormal, the preceding corrected estimator (VIII. 18)
Z $ = K Oexp [YE, +cr&/2] is usually better (in the sense of a lesser = @(U,,u,). G(du,). G(du,).
574 MINING GEOSTATISTICS VIII INTRODUCTION TO NON-LINEAR GEOSTATISTICS 575
It can be shown that the density 0(y ui) can be expanded in a series of Vlll.El.2. Equations of disjunctive kriging
Hermite polynomials qk(ui): - *I r.
Thus, the disjunctive kriging process very simply reduces to the solving The last n equations (for k = 1) are the simple linear kriging system of the
of K systems { k = 1 to K ) of simple linear kriging type. The kriging matrix unknown &, cf. (VIII.3).
of each system is deduced from the correlation matrix [ p a p ] by raising each Thus, the disjunctive kriging estimator is written as
element pap to the power k.
Kriging uariance Hence, under the hypothesis of univariate and bivariate Gaussian dis-
Taking account of the orthogonality of the two vectors Z o - Z ~ Kand tribution, the disjunctive kriging estimator is identical to the simple linear
Z;SK,i.e., E{[Zo-Z & ] Z g K } = 0 , the corresponding estimation variance kriging estimator. Kf the multivariate distribution of Z ( x ) is assumed to be
or kriging variance is written as Gaussian, the twc ~.:stimatorsZgK and Z$ are identical to the conditional
expectation E,,Zo,which is the best possible estimator.
Disjunctive kriging provides an estimator P& which only requires the Let g(u)= (1/J(277)) e-"'" be the Gaussian density and G ( y )the cor-
assumption that Y ( x ) is bivariate normally distributed. Consider the responding distribution function. Taking account of the definition ( ~ 1 . 2 l ) o f
function the Hermite polynomials H k ( x ) ,the parameters fk can be expressed as
where
FIG. VIII.7. Disjunctive kriging of the tramsfer function and the disjunctive kriging estimator is written as
P = E(I,,[Y(x)l/Y , , . . Y J . a .
Abbreviations
C G Centre de Gtostatistique, 3.5 rue St HonorC, 77305 Fontainebleau,
France
ENSMP Ecole Nationale Supkrieure des Mines de Paris, 60 Boulevard
Saint-Michel, 75006 Paris, France.
* The asterisk indicates the works that, in our opinion, are a must in a
somewhat complete geostatistical library.
"Geostat 75" NATO A.S.I. Congress "Geostat 75", Rome, Italy. The
proceedings are edited by M. Guarascio, M. David, and Ch. Huijbregts,
and published by Reidel Publishing Corporation, Dordrecht, Netherlands.
582 MINING GEOSTATISTICS BIBLIOGRAPHY 583
Report no. 2352. Marichal, A. (1976). Calcul numkrique des variograrnmes rigularisks, Internal
David, M. (1975). The practice of kriging, in "Geostat 7'5", pp. 31-48. report N-467, CG, Fontamebleau.
*David, M. (1977). Geostatistical Ore Reserve Estimation, Elsevier, Amsterdam. *Matern, B. (1960). Sparral Variation. Almaenna Foerlaget, Stockholm.
5 84 MINING GEOSTATISTICS t BIBLIOGRAPHY 585
I
t Agterberg, F. (1975). New problems at the interface between geostatistics and
*Matheron, G. (1962-1963). Traitk de Giostatistique Appliquie, Vols 1 and 2 ,
Technip, Paris. I geology, in "Geostat 75", pp. 403-421.
I
Matheron, G. (19636). Principles of geostatistics, Economic Geology, 58. 1246- Albuisson, M. and Monget, J. M. (1975). The structure function as a criterion for
1266. data quality assessment of oceanographic t~me-series,Mode Hof Line News, no.
: *Matheron, G. (1965). Les Variables Rigionalisies er leur Estimation, Masson, 79.
Paris. I Alfaro, M. (1 976). Aplicacion de la Teoria de las Funciones Aleatorias al recono-
Matheron, G. (1967). Kriging, or polynomial interpolation procedures?, C I M M cimiento forestal, in Trabajos de Esiadistica y de Inuestigacion operativa,
Transactions, 7 0 , 240-244. 6
Internal Report, ETS de Ingenieros de Minas, Madrid.
*Matheron, G. (1968). Osnovy Prikladnoi' Geostaiistiki, Mir, Moscow. Alfaro, M. (1977). Reconnaissancc htttrogkne et gCostatistique, in Proceedings of
*Matheron, G. (1969). Le krigeage universe/, Les Cahiers du Centre de Morpholo- the Pribarn Mining Congress, Prague, Ocr. 1977, pp. 123-134.
gie Mathtmatique, Fasc. 1 , CG, Fontainebleau. Alfaro, M.and Huijbregts, Ch. (1974). Simulation of a subhorizontal sedimentary
*Matheron, G. (1971). The Theory of Regionalized Variables and its Applicarions, deposit, in Proceedings of the 12th Inierr~ationalA P C O M Symposium, Golden,
Les Cahiers du Centre de Morphologie Mathtmatique, Fasc. 5, CG Fontaine- Colorado, April 1974, pp. F65-77.
bleau. Bernuy, 0. and Journel. A. (1977). Simulation d'une reconnaissance stquentielle,
*Matheron, G . (1973). The intrinsic random functions and their applications, Revue de l'lndusirie Minirale, Oct., 472-478.
Advances in Applied Probability, 5 , 439-468. Berry, P., Guarascio, id.and Sciotti, M. (1974). Analisi geostatistica del grado di
Matheron, G. (1974). Effet proportionel et lognormalitt ou le retour du :serpent fratturazione della rocca per le previsione del rendimiento in blocchi in cave di
de mer, Internal report N-374, CG, Fontainebleau. tufo litoi'de, in Proceedings of 1st Convegno Internationale sulla coltivazione di
*Matheron, G. (1975a). Le paramktrage technique des reserves, ~nternaireport pietre e mit~eralilitoi'cii. Torino, Otto. 74, pp. 1-35.
N-453, CG, Fontainebleau. Blais, R. and Carlier, A.. (1968). Applications of Geostatistics in ore valuation, ore
*Matheron, G. (1975b). Le Choix des modkles en gtostatistique, in "Geostat 7 5 " , reserve estimation and grade control, in Special Volume no. 9, CIMM,
pp. 11-27. Montreal, pp. 48-61.
*Matheron, G. ( 1 9 7 5 ~ ) .A simple substitute for conditional expectation: dis- Brooker, P. I. (1975a). Avoiding unnecessary drilling, Australasian IMM. Pro-
junctive kriging, in "Geostat 75", pp. 221-236. ceedings, no. 253.
*Matheron, G. (1975d). Forecasting block grade distributions: the transfer Brooker, P. I. (1975b). Optimal block estimation by kriging. Australasian IMM.
functions. in "Geostat 75". DD.237-251. Proceedings, no. 253.
*Matheron, G. (1977). Le choix'et I'usage des modkles topoprobablltstes, Internal Brooker, P. I. (1976). Block estimation at various stages of deposit development,
report, CG, Fontainebleau. in Proceedmgs of the 14th Inrernational A P C O M Symposium, Pennsylvania
R o ~ l e ,A. G. (1975) A Practical Introduction to Geostatrst~cs,Mining Sciences State Universiiy, pp. 995-1003.
Department, University of Leeds, Leeds. Bubenicek, L. and Haas, A. (1964). Method of calculation of the iron ore reserve
*Serra, J. ( 1 9 6 7 a ) Echantillonnage et estlmatlon locale des phtnomtnes cle tran- in the Lorraine deposit, A Decade of Digilal Computing, AIME special volume,
sition mmiers, Doctoral Thesis, IRSID and CG, Fontainebleau. AIME, New York. DD. 179-210.
Serra, J (1968). Les structures gigognes: morphologie mathkmat~queet Inter- Clark, I. and W h ~ t e ~. . ' ( 1 9 7 6 ) Geostatistical
. modelling of an orebody as an aid to
pritation mktalloghque, Mrneralrum Deposrta, 3, 135-154. mine planning, in Proceedings of the 14ih International A P C O M Symposium,
Sichel, H. S. (1951-1952). New methods In the statistical evaluat~onof mine Pennsylvania Stale Unruersiry, pp. 1004-1012.
sampling data. Transactrons of the Institutron for Minrng and Metallurgy, March. Chauvet, P. (1977). Example d'application de la gtostatistique non stationnaire:
le cokr~geage,in Proceedings of the Pribram Mining Congress, Prague, OC.
S~chel,H. S. (1966). The estimation of means and associate confidence lim~tsfor
small samples from lognormal populations, in Proceedrngs of the 1966, 1977. OD. 225-234.
. Chiles, J. P. (1975). Kriging, a method for cartography of the sea
c h a u v e t , ~ and
Symposium of the South African Institute of Mining and Metallurgy, Johannes-
floor. Internatronal Hydrographic R e v ~ e w52, , 25-41.
burg. Chiles, J P. (1975). How to adapt krigmg to non-classical problems: three case
*Switzer, P. and Parker, H . (1975). The problem of ore versus waste discrimination
stud~es,In "Geostat 7 5 " , pp. 69-89.
for individual blocks: the lognormal model, In "Geostat 7 5 " , pp. 203-2'18. Chiles, J. P. and Delfiner, P. (1975). Reconst~tut~on par krrgeage de la surface
topographique B partlr de divers schkmas d'tchantillonnage photogram-
mttrrque, Sociitt Francaise de Photogramme'trie. Bulletrn, 57, 42-50.
Chiles J. P. and Matheron, G. (1975). Interpolation optimale et cartographie,
C. PRACTICAL APPLICATIONS OF GEOSTATISTICS Annales des Mines, Nov.. 19-26.
Dagbert, M. and D a v ~ d ,M. (1977). Predicting vanishing tons before production
starts, or small blocks are no good for plann~ngin porphyry type deposits, in
Agterberg, F. (1970). Autocorrelation functions in geology, in Geostatistics, a
Proceedrngs o f the A ! M E Annual Meefrng, Atlanta.
Colloquium (D. Merriam, ed.), Plenum Press, New York, pp. 113-141.
586 MINING GEOSTATISTKCS BIBLIOGRAPHY 587
Darnay, J. (1974). L'ingBnieur des mines et la recherche operationnelle - Les Formery, Ph. (1963) ~ t u d geostatistique e du gisernent d'Ity Mont Flotuo, Inter-
applications du groupe SLN-Peiiarroya dans le domaine de la decision, in nal CG report.
Proceedings of the 8th World Mining Congress, Lima, Peru, Oct. 74, in Revue * F r a n ~ o ~ s - B o n g a r ~ oDn , and Marechal, A. (1976), A new method for open-pit
Indusrrie Minirale, 57 (3). 127-133. design and parametrizat~onof the final pit contour In Proceedrngs of he 14th
Damay, J. (1975). ~ p p l i c a t i b nde la GCostatistique au niveau d'un groupe minier, Intc~rnationalA P C O M Symposium, Pennsylvanra Stare Unrversiry, Oct. 1976,
in "Geostat 75", pp. 313-325. pp. 573-583.
Damay, J. and Durocher, M. (1970). Essai de traitement par mithodes giostatis- Guarascio, M. (1974). Valutazione dei Giacimenti Minerari: l'approcc~ogeo-
tiques du p r o b l h e de I'Cstirnation et de la repartition des teneurs dans les statist~co,L'lndustrra Minerarra, March, 1-20.
gisements en amas, Rapport interne Pennaroya, January 1970. Guarascio, M (1975). Improving the uranium deposit estimations (the "i I ovazza
David, M. ( 1 9 7 0 ~ Geostatistical
) -
ore estimation A step by step case study, in case), in "Geostat 75", pp. 35 1-367.
Decision-making in rhe Mineral Industry Special volume no. 12, CIMM Guarascio, M. and Raspa, G. (1974). Valuation and production optimization of a
Montreal, 185-191. metal mlne, In Proceedrngs of the 12th Inrernatronal A P C O M Sympos~um,
David, M. (19706). The geostatistical estimation of porphyry type deposits and Goklen, Colorado, April 1974, pp. F50-F64.
scale factor problems, in Proceedings of !he Prrbram Mining Congress, Prague, Guaraxio, M and T u r c h ~ A , . (1976) Exploration data management and evalua-
Oct. 1977, PD. 91-109. tlon techniques for uranium mining projects, In Proceedrngs of rhe Jlth Inter-
David, M. (19j'a). Grade tonnage curve: use and mtsuse in ore reserve estlma- natronal A P C O M Sympostrtm. Pennsylvania Stare University, Oct. 1976, pp.
tion, Transacrions o f the Insrirutron for Mining and Metallurgy, 81, 129-132. 45 1-461.
David, M. (1973). Tools for planning variances and condit~onals~mulations,in Guibal, D. (19730). Les fonct~onsauxiliaires B deux dimensions pour le schema
Proceedings of the 1 l t h A P C O M Symposium, Univer.rrry of Artzona, Tucson, sphirlque, Internal, Report N-347, CG, Fontainebleau.
April 1973, pp. D10-D23. Guibal, D. (19736). L'est~mat~on des okoumis du Gabon, Internal Report N-333,
David, M., Dowd, P. and Korobov, S. (1974). Forecasting departure from plan- CCI, Fontainebleau.
ning in open pit design and grade control, in Proceedings of 12th A P C O M Gu~bdl,D (1976) Elementos de geostadistlca apllcada, Apuntes dei Seminario
International Symposium, Golden, Colo., pp. F13 1-F1 S3. de Mmero-Peru, June 1976, Internal report, Mmero-Peru, Lima,
Delfiner, P. (19736) Analyse objective du geopotentiel et du vent gkostrophique Haas, A. and Joussel~n,C. (1975). Geostat~st~cs in the petroleum industry, in
par krigeage universel, Notes internes, Etabllssement d'Etudes et de Recher- "Geostat 75". pp. 333-347.
c h e ~Mitiorologiques, Paris, no. 32 1. Haas, A, and Mollier, M (1974). Un aspect du calcul d'erreur sur ies riserves en
Delhomme, J. P. (1976). Application de la theorie des variables rigionalisies dans place d'un glsernent: l'lnfluence du nombre et de la disposition spatiale des
les sciences de I'eau, Doctoral Thesis, Centre d'lnformatique Geologique, pu~ts,Revue de 1'1F.P., 29, 507-527.
ENSMP, Paris. Haas, A. and Vlalli~,J. R. (1976). Krlgeage applied to geophbsics, Geophysical
Delhomme; J. P. and.Delfiner, P. (1973). Application du ltrigeage B I'optimisation Prospecting, 24, 49-69.
d'une campagne pluviornitrique en zone aride, Proceedings of D E W A R P I D Hu~jbregts,Ch. (1971a). Courbes d'isovariance en cartographic automatique,
Symposium, Madrid, June 1973, pp. 191-2 10. Scrences de la Terre, 16, 291-301.
Deraisrne, J. (1977). Moddlisation de gisement et choix d'une mithode d'exploit- Hu~jbregts,Ch. (1971b) Reconstitution du variogramme ponctuel B partir d'un
ation, Revue de I'lqdustrie Minirale, Oct. 483-489. variogramme expkrlmental rigular~se,Internal Report N-24.1, CG, Fontaine-
Deraisme, J. (1976). Etude gCostat~stiquedu glsement d e Nickel de Mea, CMM bleau.
confidential report. Huijbregts, Ch. (1973). Regionalized variables and application to quantltatlve
Deraisme, J. and Guarascio, M. (1973). Estimation geostatistique d'un gisernent analys~sof spatial data, in Proceedtngs of N A T O A.S.I. "Dlsplay and Analysrs
de zinc, Revue d e I'lndustrie Minirale, Suppl. M ~ n eno. 5, 310-316. of Sparral Data", W ~ l e yand Sons, London and New York, pp. 38-53.
Dowd, P. (1971). Applications of Geostatist~cs,Internal Report, Zinc Corpora- Huijbregts, Ch, ( 1 9 7 5 ~ Est~mation ) of a mass proved by random diamond drill-
tion, N.B.H.C., Broken Hill, Australia. holes, in Proceedings of the 13th A P C O M Sytnpostron. Clausthal, Wesr
Dowd, P. (1975). Mine planning and ore reserve est~mationwith the aid of a Germany, Oct. 1975, pp. A.1, 1-17.
digigraphic console, C I M M Bullenn, 68, 39-43. Huljtregts, Ch, and Journel, A. (1971,) Estimation of lateritic t!pe ore bodies, in
Dowd, P. (1976). Application of dynamic and stochastic prograrwning to optimize Proceedings of the 10th International APCOMSymposium, Johannesburg, Aprii
cut-off grades and production rates, Transactions of the Inslrrution for Minmg 1972, pp. 207-212.
and Metallurgy, 85, A21-A3 1. Huijbregts, Ch. and Segovla, R. (1973). Geostatistics for the valuation of a copper
Dowd, P. (1977). Geostatistical applications in the Athabasca tar sands, in Pro- deposit, In Proceedings of the l l r h International APC0,Cf Sjmpusrurn, Uni-
ceedings of the 15th ~ n l e r n a r i o n a i ~ ~ ~ ~ ~ ~ ~ m Brisbane, pp.
p o s i u~rtst;alia,
rn, vers~tyof Arizona, Tucson, April 1973, pp. D24-D43,
235-242. Journel, A (1969). Etude sur l'estrmation d'une variable re'gronahsLie: apphca-
Dowd, P, and David, M. (1975). Planning from estimates: sensitivity of mine lion Li la cartographie sous-marine. Serv~ceHydrographique de la hlarine,
production schedules to estimation methods, in '"Geostat 75", pp. 163-183. Paris
588 MINING GEOSTATISTICS BIBLIOGRAPHY 589
Journel, A. (1973b). Geostatistics and sequential exploration. Mining Engrneer-
ing, Oct., 44-48. Matheron, G. (1955) Applicat~ondes Mkthodes statistiques i I'estimation des
Journel, A. ( 1 9 7 4 ~ ) Grade
. fluctuat~onsat various scales of a mlne ouiput, in grsenlcntr, Annnics dr M~nes,Dcc , 50-75.
Proceedings of the 12th International APCOM Symposium, Golden, Colorado, Mathcron, G (1959). Rcn~arqucssur la 101 de Lasky, Chronique des Mines
April 1974, pp. F78-F94. d'Ourre-Mer, Dec.
Matheron, G. (1970). Random funct~ ns, and thelr applications in geology,
I
Journel, A. (1975d). Ore grade distributions and conditional simulation~j- two
geostatistical approaches, in "Geostat 75", pp. 195-202. "Geostar~st~cs - A Colloquiu~~" K
(D, erriam, ed.), Plenum Press, New York,
*Journel, A. and Sans, H. (1974). Ore grade control in subhorizontal deposits, pp. 79-87
Transactions of the Institution for Mining and Metallurgy. Bulietin 812, A74- M~nguet, J M and Rebollo-Alcantara, J L. (1973). Aplicacion de 10s
A84. Ordenadores a la Mrnerra tie C~eloAbrerro, ETS de Ingenieros de Minas,
Journel, A. and Segovia, R. (1974). Improving block est~mationsat El T-niente Madrld.
copper mine (Chile), Internal Report N-371, CG, Fontainebleau. Monget, J M. (1971). A new stat~sticaltreatment of gravity data, Journal of the
Inrernatlonal Assocrarrotr of geodes)^, 102, 451-466.
Krige, D. G. (1966). Two-dimensional weighted moving average trend sur- Monget, J M (1972) The varlogram as an ald to study the structure of two-
faces for ore valuation, in Proceedings of the Symposium on Mathematics, dlmenslonal field data, in Proceedrngs of the Inrernat~onalCongress of Geology,
Statistics and Computer Applications in Ore Valuatron, Johannesburg, 1966. Monrreal., 1 9 7 2
South African Institute of Mining and Metallurgy, Johannesburg. Newton, M J (1973) The appllcat~onof geostatlstics to mine sampling patterns,
Krige, D. G. (1973). Computer applications in investment analysis, ore valuation in Proceedings of the I Irh International APCOM Symposium, University of
and planning for the Prieska copper mine. Proceedings o f the 11th International Arrzona, Tucson, Aprrl 1973, pp D44-D58.
APCOM Symposrum, University of Arizona, Tucson, April 1973, pp. G31- Olea, R (1972). Appl~cation of reg~onal~zedvanable theory to automatic
G47. contouring, Special report to the American Petroleum Institute, Research
Krige, D. G . (1975). A review of the development of geostatistics in South Africa, Project 131. .
in "Geostat 75", pp. 279-293. Olea, R. (1972). Application of regionalized variable theory to automatic
Krige, D . G. (1976). Some basic considerations in the application of geostatistics contouring, Special report to the American Petroleum Institute, Research
to gold ore valuation, Journal of the South African Instrtute of Minr,ng and Project 131.
Metallurgy, May. Orfeuil, J. P. (1975). GCostatis,tique et analyse des donntes, in "Geostat 75", pp.
Krige, D . G. and Rendu, J. M. (1975). The fitting of contour surfaces to banging 423-434.
and footwall data for an irregular orebody, in Proceedings of the 13th Inter- Parker, H. M. (1975). 'Ilie geostatistical evaluation of ore reserves, using condi-
national A P C O M Symposium, Clausthal, West Germany, Oct. 1975, pp. C.V. tional probability distribution. A case study for the Area 5 project, Warren,
1-12. Maine, - Ph.D. Thesis, Stanford University, California, USA.
Lallement, B. (1975). Geostatistical valuation of a gold deposit, in Proceedings of Parker, H. M. and Sandefur, R. L. (1976). A review of recent developments in
the 13th International APCOM Symposium, Clausthal, West Germany, Oct. geostatistics. in Proceedings of the A I M E Annual Meeting, Las Vegas, Nevada.
1975, pp. I . IV. 1-15. Parker, H. M . and Switzer, P. (1975). Use of conditional probability distributions
Le Du, R. (1976). Etude geologique et gtostatistique de la mintralisalion en in ore reserve estimation, a case study, in Proceedings of the 13th International
molybdbne du gisement de Mont Copper, MCmoire M.Sc.A., Ecole. Polg- APCOM Symposium, Clausrltal, West Germany, Oct. 1975, pp. M. 11. 1-16.
technique, Montreal. Rebollo-Alcantara, J. L. (197'7). Etude de I'exploitation du gisement de Fuentes
*Marbeau, J. P. (1976). Geostatlstique forestikre, Doctoral Thesis, CG, Fontaine- Rosas B I'aide de la simulation non-conditionelle de blocs, Revue de ['Industrie
bleau. MinPrale, Oct., 479-482.
Martchal, A. (19756). Forecasting a grade-tonnage distribution for various panel Regazzaci, J. D. and Royer, J. J. (1975). Analyse structurale et modble statistique
sizes, in Proceedings of the 13th International APCOM Symposium, Clo'usthal, des vents i 200 mb en rCgion tropicale. Note interne, Service MBt6orologique
West Germany, Oct. 1975, pp. E.I. 1-18. National, Paris, Sept.
Martchal, A. ( 1 9 7 5 ~ )GCostatistique
. et applications minibres, Annales des Mines, Rendu, J. M. (1970a). Some applications of geostatistics to decision-making in
Nov., 1-12. exploration, in Decision-making in the Mineral Industry, special. volume no.
Martchal, A. (1975d). Selecting mineable blocks: experimental results oblserved 12, CIMM, Montreal, pp. 175-184.
on a simulated orebody, in "Geostat 75", pp. 137-161. Rendu, J. M. (19706). Geostatistical approach to ore reserve calculati~n,
Martchal, A. and Roullier, J. P. (1970). Etude gtostatistique des g~sementsde Engineering and Mining Jo,urnal.
bauxite F r a n ~ a i sRevue
, de ['Industrie Mintrule, 52, 492-507. Rendu, J. M. (1975). Bayesian decision theory applied to mineral exploration and
MarCchal, A, and Shrivastava, P. (1977). Geostatistical study of a lower pro- mine valuation, in "Geostat 75", pp. 435-445.
terozoi'c iron orebody in the Pilbara region of Western Australia. Proceedings Rendu, J. M. (1976). An attelnpt to solve the problem of optimization of sample
of the 15th International APCOM Symposium, Brisbane, Australra, June 1977, spacing in the South Africa,n gold mines, Journal of (he South African Institure
pp. 221-230. of Mining and Metallurgy.
MINING GEOSTATISTICS
Royle, A. G.(1977). HOWto use geostatistics for Ore reserve classification, World
Mining, Feb. 52-56.
Royle, A. G., Newton, M.J. and Sarin, H. K. (1972). Geostatistlcal factors in
design of mine sampling programmes, Transactions ofthe Instrtution for Mining
and Metallurgy, 81, A81-A88.
Royle, A. G. and Hosgit, E. (1974). Local estimation of sand and gravel reserves
by geostatistical methods, Transactions of the Instirurron for Minrng and Metal-
lurgy. 83, A53-A62.
Rutledge, R. W. (1973). Geostatistics, an introduction, in Proceedings of the
Index A.
A N Z A A S Meeting, August 1973, Perth, Australia.
Rutledge, R. W . (1975). The potential of geostatlstics in the development of
Geostatistical Concepts
mining, in "Geostat 75", pp. 295-31 1.
Sabourin, R. (1975). Geostatistical evaluation of the sulphur contents in Lingan
coal mine, Cape Breton, in Proceedings of the 13th International APCOM
Symposiitm, Clausrhal, West Germany, Ocr. 1975, pp. I . 11. 1-16. Cokriging 324, 335, 340
Sandefur, R. L, and Grant, D. C. (1976). Preliminary evaluation of uranium Combining terms of estimation 311,
deposits, A geostatistical study of drilling density in Wyoming solution fronts, Absence of correlation, see also Pure 44 1
in Proceedings of the 1976 I A E A Meeting, Vienna, Austria, nugget effect 36, 152 Conditional non-bias 458, 462
Sichel, H. S. (1972). Statistical valuation of diamondiferous depos~ts,in Pro- Accumulation (regionalized vari- Conditional expectatlon 457, 565
ceedings of the 10th APCOM Symposrum, Johannesbrirg, 1972 pp. 17-25. able) 199, 281 Conditional posltive definite
Sinclair, A. J. and Deraisme, J. (1974). A geostatistical study of the Eagle copper Accuracy of a selection 458, 463 function 35, 161
vein, N. British Columbia, CIMM Bulletin, June, 1-12. Anisotropy 37, 175 Conditional simulation, see also Simu-
Sinclair, A. f . and Deraisme, J. (1975). A two-dimensional statistical study of a A prion variance 31, 37, 42, 78 lation of deposits 492, 494
skarn deposit, Yukon Territory, Canada, in "Geostat 75", pp. 369-379. Authorized linear combinat~on 36 Confidence interval 14, 50, 57, 60
Serra, J. (19676). Un crittre nouveau de d6couvertes de structures: le vario- Auxiliary funct~ons 108, 117 Convolution product 504
grarnme, Sciences de la Terre, 12, 275-299. Coregionalization 40, 47, 171, 253,
Serra, J. et Al. (1974). Laws of linear homogenization in ore stockyards. Pro- 236, 516
ceedings of the 11th International mineral Processing Congress, Caglrarr, Italy, Behaviour at infinity (of a Correlation coefficient 41, 256
April 1975; Report N-387, CG, Fontainebleau. variogram) 39 Correlograrn 33
Tulcanaza, E. (1972). Personal communication. Behaviour near the origin (of a vario- Covariance 31. 34. 36, 194
Ugarte, I. (1972). Ejemplos de modelos de estirnacion a corto y largo plazo, gram) 35,44, 232 Covariogram, see Geometric covario-
Boletin de Geoestadistica, 4 , 3-22. gram, or Transit~vecovariogrnm
Bias ( o f an estimation), see also Non-
Waiton, J. (1972). Resumen de un estudio de estirnacion de bloque grande de blas condition 49 Cross-covariance and variogram 40,
block-caving, Boletin de Geostadistica, 4. 43-54. Univers~dadde Chile. 171, 260
Blvarlate distribution of true and esti-
Zwicky, R. W. (1975). Prelimmary geostat~stical investigations of tar-bearing mated values 457, 465
sands, in Proceedings of the 13th APCOM Symposium, Nausrhal, \Vest Bivarlate Gaussian distribution 573
Germany, Oct. 1975, pp. I. 111. 1-15. Border effect (on the dispersion) 61
Border term (of the globai estimation Data file 356, 361
Deconvolution 90. 1-3 1, 270
variance) 437, 442 Discontinuity at the origin of a vario-
gram, sez also Nugget effect 39
Discrete approximation of a block 96,
Cauchy algorithm (for integral cal- 511
culation) 99, 109 Direct variogram, see Variogram
Cauchy-Gauss method (for calculation Disjunctive kriging 567, 573, 575
o f mean value 7) 98, 106
Charts (for auxiliary functions) 116,
123
49 1
Dispersion variance
-
Dispersion of grades 68, 70, 74, 297,
61, 67, 157, 246,
Class of angles and distances, see Grou- 297,450
ping into- Drift 40, 271. 313, 331, 500, 530
592 INDEX A. GEOSTATISTICAL CONCEPTS INDEX A. GEOSTIATISTICAL CONCEPTS 593
E Hole effect 168, 247, 403 Least square estimation methori 333, 0
Estimation error 46, 57, 60, 396, 398 Homogenization 297, 548 343 Ordinary kriging 304,563
Estimation variance 48, 52, 57, 398, Line term (of a global estimation vari- Optimum sampling grid 181,308
, 410 ance) 422,441
Estimation variance of a Linear estimator 51, 559
variogram 192 Icosahedron approximation (for !;imu- Linear equivalents (for a logarithmic
Expectation 31, 320, 500 lation) 503 model) 115 Permanence of distribution
Exponential model 116,163,240,507 Independence principle of elementary Linear kriging processes 559 hypotheses 468,481
Extension variance 54 errors 409, 412, 419, 441 Linear model of Positive definite-type function 35,
Indicator function (of a volume or sur- coregionalization 171, 257, 516, 161, 171,260
face) 98, 429 521 Prerequisites for the construction of a
F Inference of the moments of order 2, see Linear model of regionalization 113, kriging estimator 562, 571
Fitting a variogram model 230, 236 Statistical inference Product of two regionalized vari-
Fluctuations of the experimental Influence polygons (methods of estima- Linearity- of geostatistical
- ables 424, 443
moments 192 tion) 57, 330, 343, 448, 462 operators 150, 184 Proportional effect (direct or
Fluctuation variance 192 Information (available for selec- Local estimation (of a deposit) 20,304 inverse) 186, 247, 250,278
tion) 21, 449,484 Local variogram 192 Pure nugget effect 39, 152, 154, 157,
In siru resources 18, 303, 444 Logarithmic model of variogram 94, 229,312
G Intermeshed structures, see Nested 113, 166,244,292
structures Lognormal distribution 69, '74, 188, Q
Gaps (Kriging the -) 352 Internal independence hypothesis 437 468,480,525 Quality error 412, 435
Gaussian distribution 50, 466, 468, Intrins~c coregionalization 174, 326,
508,573 Lognormal kriging 570 Quasi-stationarity 33,43, 186, 345
Gaussian model (of variogram) 165, 427 Quotient of two regionalized vari-
Intrinsic hypothesis 33, 35, 43 ables 424
241,508 Invariance property of kriging 335
Gaussian transform 468, 472, 475,
Inverse (squared) distance (method of Massive deposits -391 R
508
Geometric anisotropy 177, 262 estimation) 57, 330, 333, 343 Mean value 7 or C 95, 123
Inverse lognormal distribution 70, 75, Measurement error 148, 197, 340 re model of variogram 165,244
Geometric covariogram 99, 430
Geometric error 412, 428, 435 188 Models of variograms 161, 230, 236 Random function 10, 29, 491
Irregular grid 416 Moment of order 1, see Expectation Random kriging 352,400
Global estimation (of a deposit) 18, Isomorphism of two random Random stratified grid 415
Moment of order 2, see A priori variance
410,439 functions 494 Random variable 10, 29,528
Global estimation variance 323, 410, Moving average 56, 505
Isotropy 37, 175, 228 Multivariate distribution 30, 509 Range (of a variogram) 36, 178
44 1 Realization (of a random variable or
Grade control (at production) 293, function) 17, 29,491
299 Recoverable reserves 20, 444, 462,
Grading over constant- thickness 85, 480,483,554
94,123,281 Neighbourhood of estimation 344,
Krige's relation of additivity 67, 302 399 Reduced difference (regionalized vari-
Graphical transform 476 Kriging process and estimator 2 1,304,
Grouping data into classes of Nested structures 148, 159, 166, 176, able) 427
343,557 183,291 Regionalization and Regionalized vari-
angles 211, 220, 272 Kriging matrices 307, 319, 357
Grouping data into classes of dis- Non-bias condition 56, 305, 317, 325, able 10, 27, 199,494
Kriging plan, 344, 385 559 Regular grid 209, 413
tances 210, 220 Kriging system, see matrices Regularization (process of) 77, 89,
Kriging variance 306, 398, 496 Non-linear geostatistics 556
Normal distribution, see Gaussian dis- 155,231
Kriging of the transformed vari- tribution Residual (fluctuation around a
able 566 Nugget constant 152, 159, 232 drift) 314, 530
Hermite polynomials and expan- Nugget effect 39, 151, 1 5 4 , 3 1 0 Resources, see I n situ resources
sion 472, 476, 481, 574 Numerical model of a deposit, see also Robustness with respect to the structural
Histogram (experimental) 69-77,204, Simulation of a deposit 24, 491, model 167,233,245
445 Lagrange procedure 306, 317 546 Robustness of a simulation 491,510
594 INDEX A. GEO'STATISTICAL CONCEPTS