multivariate geostatistical analysis with application to a
TRANSCRIPT
Edith Cowan University Edith Cowan University
Research Online Research Online
Theses : Honours Theses
2003
Multivariate geostatistical analysis with application to a Western Multivariate geostatistical analysis with application to a Western
Australian nickel laterite deposit Australian nickel laterite deposit
Ellen Bandarian Edith Cowan University
Follow this and additional works at: https://ro.ecu.edu.au/theses_hons
Part of the Geology Commons
Recommended Citation Recommended Citation Bandarian, E. (2003). Multivariate geostatistical analysis with application to a Western Australian nickel laterite deposit. https://ro.ecu.edu.au/theses_hons/338
This Thesis is posted at Research Online. https://ro.ecu.edu.au/theses_hons/338
Edith Cowan University
Copyright Warning
You may print or download ONE copy of this document for the purpose
of your own research or study.
The University does not authorize you to copy, communicate or
otherwise make available electronically to any other person any
copyright material contained on this site.
You are reminded of the following:
Copyright owners are entitled to take legal action against persons who infringe their copyright.
A reproduction of material that is protected by copyright may be a
copyright infringement. Where the reproduction of such material is
done without attribution of authorship, with false attribution of
authorship or the authorship is treated in a derogatory manner,
this may be a breach of the author’s moral rights contained in Part
IX of the Copyright Act 1968 (Cth).
Courts have the power to impose a wide range of civil and criminal
sanctions for infringement of copyright, infringement of moral
rights and other offences under the Copyright Act 1968 (Cth).
Higher penalties may apply, and higher damages may be awarded,
for offences and infringements involving the conversion of material
into digital or electronic form.
MULTIVARIATE GEOSTATISTICALANALYSIS WITH
APPLICATION TO A WESTERN AUSTRALIAN NICKEL
LATERITE DEPOSIT
BY
ELLEN BANDARIAN
A Thesis Submitted to the
Faculty of Computing, Health and Science
Edith Cowan University
Perth, Western Australia
In Partial Fulfilment of the Requirements for the Degree of
Bachelor of Science Honours (Mathematics)
June 2003
Supervisors: Associate Professor Lyn Bloom
Dr Ute Mueller
ABSTRACT
Although it can be time consuming and computationaJiy more expensive to
work with multivariate data, it is often desirable to exploit the relationships among
and between the variables sampled across a study region. In most cases the
availability of this secondary information can enhance the estimation of the primary
vmiable(s). The aim of this research is to des<::ribe and then demonstrate the use of
multivariate statistical methods in gcostatistical analysis. The methods will be
illustrated by application to a multivariate data set from an actual mineralisation. The
data suite is known as MM22D and comes from the Murrin Murrin nickel mine near
Laverton in Western Australia. Two four variable subsets of the MM22D data suite
are used in the application. The first is MM22DHC4 and consists of nickel, cobalt,
iron and zinc, chosen for their high correlations with each other. The second is
MM22DTOP4 and consists of nickel, cobalt, magnesium and iron, chosen for their
economic importance to the mining company.
This thesis presents the theory and the process of the modelling and
estimation of multivariate data. We demonstrate the use of principal component
analysis in a geostatistical environment, in particular for the detection of intrinsic
correlation. We illustrate the modelling and estimation of an intrinsically correlated
data set (MM22DHC4) and estimate the variables in this data set using ordinary
cokriging, principal component kriging and ordinary kriging. In addition we illustrate
the derivation of a general linear model of coregionalisation and estimate the
variables of this data set (MM22DTOP4) using ordinary cokriging and ordinary
kriging. The grade control data from the MM22D data suite, which were considered
reality, were used as a comparison and assessment of the accuracy of all of the
estimates. As the data used in this study were isotopic it was anticipated that there
would be little difference in the estimates obtained which was i"'deed the case.
2
DECLARATION
I certify that this thesis docs not, to the best of my knowledge and belief:
(i) incorporate without acknowledgment any material previously submitted for a
degree or diploma in any institution of higher education;
(ii) contain any material previously published or written by another person except
where due reference is made in the text; or
(iii) contain any defamatory material.
3
ACKNOWLEDGEMENTS
I wish to thank the Faculty of Computing, Health and Science, in particular
Professor Patrick Garnett for the honour of receiving the Dean's Scholarship in 1997.
I would like to express my appreciation lo Anaconda Operations Pty Ltd for making
the Murrin Murrin data available to me. I would also like to extend my thanks to The
Statistical Society of Australia Inc for the privilege of being awarded the Honours
Scholarship 2003. My thanks go also to my supervisors Associate Professor Lyn
Bloom and Doctor Ute Mueller whose patience, motivation and support made this
project possible.
To my Mum and Maryann I convey my sincere appreciation ior all of the
extra babysitting they have done. Thanks also to Dad and Odette for tbir past and
ongoing encouragement. Finally and most importantly my heartfelt thanks goes to
my husband Nann and my two beautiful daughters Katherine and Emily. Their
support, encouragement and enthusiasm for my studies has been and continues to be
unwavering.
4
TABLE OF CONTENTS
Abstract 2
Declaration 3
Acknowledgements 4
1 Introduction 7
1.1 Background and Significance 7
1.2 Objectives 9
1.3 Thesis Outline 10
1.4 Software 10
1.5 Notation 11
1.6 Acronyms and Abbreviations 13
2 Theoretical Framework 15
2.1 Principal Component Analysis 15
:2.2 Multivariate Random Function Model 18
2.3 ,Modelling the Coregionalisation 21
2.4 Kriging 27
2.4.1 Simple and Ordinary Kriging 28
2.4.2 Ordinary Cokriging 30
2.4.3 Principal Component Kriging 32
3 Data Analysis 34
3.1 MM22D Data Suite 34
3.2 Exploratory Data Analysis 36
3.3 Spatial Data Analysis 43
3.4 Principal Component Analysis 48
3.4.1 Principal Component Analysis ofMM22DEXP 50
5
3.4.2 Principal Component Analysis ofMM22DHC4
3.4.3 Principal Component Analysis of Mlvl22DTOP4
4 Ordinary Cok:riging
4.1 Variography and Cokriging ofMM22DHC4
55
58
62
62
4.1.1 Variography and the Intrinsic Coregionalisation Model 62
4.1.2 Cross Validation for Cokriging of M1vl22DHC4 67
4.1.3 Cokriging Estimates of M:M22DHC4 70
4.2 Variography and Cokriging of MM22DTOP4 74
4.2.1 Variography and the Linear Model of Coregionalisation 75
4.2.2 Cross Validation for Cokriging of MM22DTOP4 79
4.2.3 Coktiging Estimates of MM22DTOP4 81
5 Principal Component K.tiging 84
5.1 Variography of the Principal Components 84
5.2 Cross Validation for Ordinary Kriging of the Principal Components 89
5.3 Principal component Kriging Estimates 92
6 Ordinary Kriging 97
6.1 Variography of Nickel, Cobalt, Magnesium, Iron and Zinc 97
6.2 Cross Validation for Ordinary Kriging 104
6.3 Ordinary Kriging Estimates of Nickel, Cobalt, Magnesium and Iron 107
7 Comparison of Estimates 109
8 Discussion and Conclusion 126
References 131
Appendices 133
Appendix A Principal Component Analysis of Transformed Data 134
Appendix B Cross Validation Results 139
Appendix C ISATIS Parameter Files 152
6
1 INTRODUCTION
1.1 Background and Significance
Geostatistics enables us to analyse spatially dependent data, that is, data
where location as well as value is important. Such data arise naturally in the earth,
mining, petroleum and environmental sciences. Geostatistics can be described as a
set 0f statistical tools that allow us to describe, interpret and model the spatial
continuity that is a fundamental feature of many natural phenomena. Furthennore
once we have a mathematical representation of the spatial continuity of the data we
can use this model to estimate values at unsampled locations across the study region.
Geostatistical methods generally provide better estimates than traditional methods
and most importantly provide us with a measure of the accuracy of the estimates.
This rapidly advancing area of applied mathematics has its origins in the
mining industry in South Africa in the 1950's. Mining engineer D. G. Krige and
statistician H. S. Sichel were among the first to implement new statistical methods
that did not rely on traditional methods based on normally distributed data. In the
1960's G. Matheron further developed and formalised these innovative ideas and
introduced the concept of a regionalised variable which he defined as a spatially
distributed phenomenon that exhibits a particular spatial structure consisting of both
random and structured aspects (Matheron, 1970, p. 5).
It is often the case that data collection in the earth sciences consists of
observations of many variables, some more densely sampled than others. While it is
more cumbersome and computationally more expensive to work with multivariate
data the availability of auxiliary information can enhance the interpretation and
estimation of a primary variable. Multivariate geostatistics takes into account the
relationships between and among the variables as well as incorporating their spatial
distribution across a region. The relationships between these variables can be
identified and summarised using methods of multivariate statistical analysis.
7
One classical, and probably the most commonly used, multivariate statistical
technique is principal component analysis which dates back to the early 1900's, in
particular to the work of Pearson (1901) and Hotelling (1933). Principal component
analysis involves the construction of linear combinations of correlated variables into
(preferably) fewer :.mcorrelated factors that account for the majority of the variation
in the original data (Afifi & Clarke, 1996, pp. 330-331). More specifically, principal
component analysis is concerned with explaining the variance-covariance structure
of a set of variables by means of a few linear combinations of these variables
(Johnson & Wichern, 2002, p. 426).
From a geostatistical perspective principal component analysis is a
particularly valuable tool as it can be used for reducing the number of variables in a
data set and for the detection of intrinsic correlation. One of the benefits of reducing
the number of variables in a multivariate data set is the practical advantage of
modelling fewer semivariograms. Furthermore, if the orthogonaiity of the principal
components extends to any separation vector h we may proceed with an intrinsic
coregionalisation model. Under this assumption we may perform cl~ssical kriging
individually on the principal components. This technique is known as principal
component kriging and is computationally less expensive than other methods such as
cokliging (Goovaerts, 1997, pp. 233-234).
When the data are not intrinsically correlated we must consider not only the
spatial variability of the individual variables but we must also take into account the
joint variability 0f each pair of variables. The linear model of coregionalisation is a
mathematical model that characterises the spatial variation of a multivariate system
at different spatial scales. The requirement of a linear model of coregionalisation is
that all direct and cross semivariograms or covariances are jointly modelled and
share a common set of basic structures. This then calls for the inference and
modelling of N/.Nv+l)/2 direct and cross semivariograms. While this can be a
tedious procedure the problem lies in whether the model fits adequately in the
mathematical sense (Goulard & Voltz, 1992, p. 269), more precisely the
coregionalisation matrices need to exhibit positive semi-definiteness in order for the
model to be permissible.
8
1.2 Objectives
The aim of this study was to describe and demonstrate the use of multivariate
statistical techniques in geostatistical analysis. The theory of principal component
analysis in a traditional multivariate statistical environment and the use of this
technique as it applies to geostatistics are presented. We examine also the theory of
the general form of the linear model of coregionalisation as well a particular case
known as the intrinsic coregionalisation model. In addition we discuss the theory of
the estimation techniques to be used, that is, ordinary kriging, ordinary cokliging and
principal component kriging.
For the application of the theory we had four main objectives. Firstly we
aimed to exhibit the use of principal component analysis in a geostatistical
environment, namely to determine whether the data were intrinsically correlated.
Secondly we wished to show examples of both the intrinsic coregionalisation model
and the linear model of coregionalisation. Thirdly we aimed to demonstrate the
estimation techniques of principal component kriging and ordinary co kriging. Finally
we wished to include a comparison of these multivariate techniques to the univariate
estimation method of ordinary kriging as well as with 'reality'.
To demonstrate these objectives we analysed a mineralisation (MM22D),
which comprises three dimensional grade and thickness measurements on eight
variables: nickel, cobalt, magnesium, iron, aluminium, chromium, zinc and
manganese. These data came from the Murrin Murrin nickel mine near Laverton in
Western Australia This data suite consists of two data sets: a grade control data set
:Mlvl22DGC and an exploration data set MM22DEXP. The set MM22DGC
comprises grade accumulations for each of the eight variables and will be considered
reality for assessment and comparative purposes. The set MM22DEXP is a subset of
MM22DGC and will be used to perform the analysis. This data set is isotopic and
comprises accumulations for each of the eight variables jointly sampled at 125
locations.
As the data were measured on different scales and have vastly different
ranges, means and variances we used the standardised values for the majority of the
analyses. Initially we performed a principal component analysis on the MM22DEXP
data set and determined that the data were not intrinsically correlated We then
9
investigated two four variable subsets of MM22DEXP, denoted as rvr:M:22DHC4 and
:M:M:22DTOP4. The fanner subset was intrinsically correlated; hence we used an
intrinsic coregionalisation model and prefonned the estimation using ordinary
cokriging. In addition we individually modelled the principal components from this
subset and perfonned the estimation using principal component kriging. For the latter
subset the variables were not intrinsically correlated hence we used a linear model of
coregionalisation and perfonned the estimation using ordinary cokriging. The
estimates were then compared to the grade control data, lvfl\.122DGC, and the
estimates obtained from ordinary kriging.
1.3 Thesis outline
Chapter two of this thesis discusses the theoretical framework relevant to this
study. This includes principal component analysis from both classical and
geostatistical perspectives, the multivariate random function model and the various
methods of kriging, namely ordinary kriging, ordinary cokriging and principal
component kriging. Chapter three presents a detailed exploratory data analysis of the
MM22D data suite. In chapter four we demonstrate the application of ordinary
cokriging using both the MM22DHC4 and MM22DTOP4 data sets. Chapter five
presents the application of principal component kriging using the principal
components extracted in the eigenanalysis of lv!M22DHC4. Chapter six presents
ordinary kriging of rhe nickel, cobalt, magnesium, iron and zinc variables. Chapter
seven presents a comparison of the estimation techniques used and chapter eight
presents a discussion and conclusions of the research.
1.4 Software
There are various software packages available to assist in geostatistical
analysis. My study has used primarily the packages listed below. It is appropriate at
this point to mention one package in particular, ISATIS, which is recognised as an
industry stanrlard. While ISATIS a relatively new package in the workplace, we are
in the fortunate position of having it available at this university.
10
I have taken this opportunity to familiarise myself with the capabilities of
ISATIS, to the point where I would now be considered proficient in implementing
this package in the workplace. ISATIS offers the geostatistician analytical features
and capabilities not previously available. Some of these features include
simultaneous modelling of direct and cross scmivariograms, sequential cokriging of
numerous variables, alternative measures of spatial variability and alternative kriging
methods.
3PLOT (Kanevski et al, 1998): post plots of data and estimates
ISATIS (Bleincs eta!, 2000): estimation
MATLAB 5.3 (1999): matrix manipulation and calculation
MICROSOFr EXCEL (2002): graphical representation of data;
spreadsheet calculations
MINITAB 12.1 (1998): summary statistics, principal component analysis,
residual analysis
V ARlO WIN 2.2 (Pannatier, 1996): semivariogram inference and modelling
1.5 Notation
The notation used throughout this study is a combination of that used by
Goovaerts ( 1997), Wackernagel (l998b) and Johnson and Wichern (2002).
V: forall
A: study region
a: range parameter
M: coefficient of the basic covariance model c,(h) or semivariogram
model g1(u) in the linear model of regionalisation of the random
variable Z(u)
b~: coefficient of the basic covariance model c1(h) or semivariogram
model g1(u) in the linear model of coregionalisation of the random
variables Zi(u) and 2j(u)
B,: corcgionalisation matrix including the coefficients b~ of the basic
II
covariance model t'!{h) or semivariogram model Rl(h) in the
' corresponding linear model of coregionalisation C(h) = L,S1c1(h) 1~o
'· or f(h) = :Ln,g,(h) c-o
C(O): covariance value at separation distance I hI= 0
C(h): stationary covariance of the random function Zfor lag vector h
C(h): covariance function matrix of size Nv x N\'
C;J(h); stationary cross covariance between the two random functions 2j
and 2j for a Jag vector h
Cov {.): covariance
q(h): /th basic covariance model in the linear model of (co)rcgionalisation
E{·): cxpcctcdvaluc
E: is an clement of
f(h): scmivariogram function matrix of size N,. x Nl'
g1(h): /th basic scmivariogram model m the linear model of
(co )regionalisation
Y(h): stationary scmivariogram of the random function Z for lag vector
h
f;;(h): stationary cross scmivariogram between the two random functions
~and -0" for a Jag vector h
h: separation vector
Aa(u): kriging weight associuted to z--datum at location Ua for estimation
of the attribute z at location u
Aulu): cokriging weight associated to z--datum at location Uu for estimation
of the attribute z at location u
m: stationary mean of the random function Z(u)
m: vector of stationary means
m(u): expected value of the random variable Z(u)
N,.: number of variables Z;
n: number of data values available over the study region A
12
n(u): number of data values z(ua) used for estimation of the attribute z at
location u
,1,(u): number of data values z1(uu) used for estimation of the attribute z at
location u
Q: matrix of eigenvectors extracted in the principal component analysis
rv: linear correlation coefficient between variables Z1 and Zj
u: coordinate •:ector
Uu: datum location
Var{·}: vanancc
f.1 (u): A1h region ali sed factor corresponding to the (/+ 1) basic covariance
Z(u):
Z:
Z:
z:
z(u):
z(ua):
Z;(Ua):
' model in the linear model of coregionalisation C11 (h) = ~)~c1 (h)
generic continuous random variable at location u
univariate random variable valued function
multivariate random variable valued function
continuous variable (attribute)
t111e valw! at unsampled location u
z-da!um value allocation Ua
z1-datum value at location Ua
1•0
1.6 Acronyms and Abbreviations
The following is a list of acronyms and abbreviations used in the Figures and
Tables throughout this thesis.
AL:
CO:
CR:
FE:
ICM:
LMC:
MAE:
aluminium
cobalt
chromium
iron
intrinsic coregionalisation model
linear model of corcgionalisation
mean absolute error
13
MG: magnesium
MN: manganese
MSE: mean square error
NI: nickel
OCK: ordinary co kriging
OK: ordinary kriging
PC: principal component
PCK; principal component kriging
PCK2: principal component kriging (two retained principal components)
ZN: zinc
14
2 THEORETICAL FRAMEWORK
) .1 this chapter we discuss the theory of principal component analysis from
both classical and geostatistical perspectives. We then introduce the multivariate
random function mcdel and the linear mode! of corcgionalisation. Finally we discuss
the vnrious kriging algorithms employed in this study: ordinary kriging, ordinary
cokriging and principal component kriging, The development of the theory and the
notation used follows that of Goovaerts (1997), Lay (1997) and Wackemagel
(1998a).
2.1 Principal Component Analysis
Almost any kind of data collection in the earth sciences involves
simultaneous measurements on many variables. Suppose that we are dealing with Nv
variables, that is Nv dimensional data, measured at n locations u in the region A
While multivariate gcostatistics gives us the tools to incorporate this additional
infonnation, in practice one seldom considers coregionalisations of N., greater than
three. The reasons for this include notational and computational complexity and
difficulties of statistical inference and modelling of the cross covariance or cross
semivariograms (Joumel & Huijbregts, 1978, p. 173).11 is often desirable to exploit
the interrelationships among and between the N., variables by representing them
through a few linear combinations of these variables.
One multivariate technique used to achieve this is known as principal
component analysis. This is one of the most commonly used methods of multivariate
analysis. It is simple to implement and interpretation of the results is often
straightforward. In its simplest form a principal component analysis consists of
defining a linear transformation that maps a set of N., correlated vmiables into Nv
uncorrelated principal components (Wackemagel, 1998a, p. 127). From a purely
mathematical perspective Johnson and Wichern (2002, p. 426) describe a principal
15
component analysis as a means of explaining the variance-covariance structure of a
set of variables by means of a few linear combinations of these variables.
Each principal component is a linear combination of the original variables
and the amount of information conveyed by each principal component is measured
by its variance (Afifi & Clarke, 1996, p. 330). The principal components are
arranged in order of decreasing variance, thus the first principal component is the
most infom1ativc and the least informative is the last principal component. One can
then choose to retain only the first few principal components that account for the
majority of the variability of the original data, making the subsequent analysis
simpler. A principal component analysis can also be used to test for normality of the
data; if a selected principal component is not normal then neither are the original
data. Other classical uses of principal component analysis are to identify outliers and
reveal relationships among and br:cween variables that may not have previously been
identified. One of the most important ge.ostatistical applications of principal
component analysis is to detect intrinsic correlation. This is done by examining the
cross corrclograms or cross scmivariograms of the first few principal components
and will be discussed fu1ther in section 2.3, Modelling the Coregionalisation.
The basic features of a principal component analysis consist of the extraction
of the eigenvalues and eigenvectors of a square, symmetric covariance or correlation
matrix. Principal component analysis does not require multivariate normality though
the interpretation and application are enhanced when this condition is met.
Algebraically the principal components arc specific linear combinations of the
random variables Zi(u) for i = 1, ... , N,,. Geometrically the linear combinations
represent the choice of a new coordinate system by rotating the original system with
Z,{u) as the coordinate axes (Johnson & Wichem, 2002, pp. 426~427).
If the variables have widely differing ranges, if they are measured on
differing scales or if the units of measurement arc not consistent, it is advisable to
standardise the variables. A principal component an1lysis performed on the
covariance matrix can be severely affected by large or inconsistent variances. Hence
by standardising the original data (or equivalently using the correlation matrix as
opposed to the covariance matrix) we ensure that the assigning of the weights in a
principal component analysis is not innuenced by variables with large variances. It is
16
important to note that all interpretations of a principal component analysis based on
the correlation matrix must be in tenns of the standardised variables.
Let R be the correlation matrix of the Nv random variables Z;(u) fori= 1, ... ,
Nv. The spectral decomposition of the Nv x Nv syrr'TI.etric matrix R is given by
(!)
where Jl,,~, ... ,I\.N are the eigenvalues of R and epe2 , ... ,eN are the associated ' '
nonnalised eigenvectors with eJe; = 1, i = 1, ... , N,. and eie1 = 0 when i :f. j. In matrix
notation the spectral decomposition of R then is
R=QAQ' (2)
where Q is the orthogonal matrix whose columns are the corresponding unit
eigenvectors epe2, ... ,eN, (note QQT = I) and A is the diagonal matrix [ltd of
eigenvalues of R fur k = 1, ... , Nv such that?,;::~~ ... ;::AN,. This orthogonal change
of variables does not change the total variance of the data. Furthermore the
eigenvalues determine~.! in the spectral decomposition are the variances of the
principal components.
The first principal component then is the eigenvector corresponding to the
largest eigenvalue of R, the second principal component is the eigenvector
corresponding to the second largest eigenvalue of R and so on. The kth principal
co.nponent is given by a linear combination of the set of Nv original variables:
The set of all Nv principal components are linear combinations of the set of Nv
original variables:
N
Y,(u)= !,q"Z,(u) (3)
for k = 1, ... , Nv. In matrix notation, Y(u)=QTZ(u) and has variance matrix
QTRQ =A. The total variance of Yk(u), k = 1, ... , Nv, is equal to the total variance of
Zi(u), i = 1, ... , Nv, and is given by tr(R)=tr(A)=/\.1 + k + ... + 1\.N =Nv. Hence the
variance of Yk(u) is Ak and the fraction of the total variance that is explained by Yk(u)
is measured by 1\.k/ tr(R).
17
J
The set of Nv principal component scores is computed at each datum location
Ua in A for a= 1, ... , n as linear combinations of the standardised sample values
Zi(Uu) at that location multiplied by the loading of variable Z,{u) or the kth principal
component Yk(u)
(4)
• fork= 1, ... , Nv with m; and 0'; being the mean and standard deviation respectively of
the Zi(u) data and qk; obtained from the matrix of eigenvectors Q in equation (2).
2.2 Multivariate Random Function Model
One of the fundamental aims of gcostatistics is to characterise the behaviour
of the population of a sampled attribute over a study region A In order to do this we
are required to model the statistical characteristics of the population using only the
available sample data. In essence geostatistics uses a probabilistic approach to model
the uncertainty about how the attribute behaves between the sample locations.
Consider a set of n sample data values denoted by z(ua), where o:=l, ... , n.
These observations may be considered as a subset of a larger, possibly infinite,
collection of observations. The value z(ua) can be thought of as one possible
realisation of a random variable Z(ua). Similarly the value z(u) at an unsampled
location can be thought of as a particular realisation of the random variable Z(u) for
each location u in the region A. The characterisation of a random variable Z(u) is
detennined completely by the cumulative distribution function, that is
F(u; z)=Prob{Z(u):5z)
for all z.
(5)
The set of random variables Z(u), denoted by Z, for all locations u in the
region A., { Z(u), u E A}, is called a random function and is characterised by the
set of all its N-variate cumulative distribution functions
(6)
18
•
for any locations uk, where k=:l, ... , N. In general it is impossible to infer the entire
spatial law of a random function so it is necessary to obtain an approximate solution
to most of the problems encountered. In the most commonly used geostatistical
procedures we assume that the random function is stationary, that is, the
characteristics of the random function remain invariant under translation. A random
function is said to be strictly stationary if for any set of N points u1, .•. ,UN and for
any vector h, the two vectors of random variables [Z(u1), ... , Z(uN)l and [Z(u 1 +h),
... , Z(uN +h)] have the same multivariate cumulative distribution function
F(u1, ... , UN; Zll'"' ZN) =F(U1 + h, ... , UN + h; Z1 , ... , ZN)
for all locations u,, ... , UN and any vector h.
(7)
As it is not possible to verify strict stationarity from experimental data we
usually require only second-order stationarity where the first two moments (mean
and covariance) are constant (Armstrong, 1998, p. 18). Hence we require, for all
locations u, that the mean exists and: is constant
E{Z(u)}=m
and the two-point covariance exists and depends only on the separation vector h
C(h)= E{Z(u)·Z(u +h)}-E{Z(u)} · E{Z(u +h)}
(8)
(9)
In many cases this assumption is not appropriate so we assume intrinsic stationarity
where the increments Z(u +h)- Z(u) are assumed to be second-order stationary.
E{Z(u+h)-Z(u)} =0 (10)
Var{Z(u +h)- Z(u)} = E{ [Z(u +h)-Z(u)]'} =2y(h) (11)
The function y(h) is called the semivatiogram and is the basic tool for the
interpretation of the spatial variability of the attribute being investigated.
The probabilistic approach to a coregionalisation (a regionalised phenomenon
that can be represented by several intercorrelated variables) is similar to that of the
requirements of a single variable whose concepts can be easily broadened to
incorporate a multivariate random function. The vector of unsampled values of Nv
variables [z1(u), ... , zN. (u)] can be considered as a particular realisation of the vector
of Nv random variables [Z1(u), ... , ZN, (u)] for all locations u over a region A.
Similarly this vector of random variables may be thought of as one particular
realisation of a multivariate random variable valued function:
19
([Z1(u), ... , Zv,Cn)]; UE A) denotedbythevector Z.
In order to define a cross covariance function that depends only on the
separation vector h the direct (auto) and cross Qoint) covariance functions Cu(h) of a
set of Nv continuous random functions are defined in the framework of joint second-
order stationarity. That is, for all locations u over the study region A the mean of
each variable Zi(u) exists and is constant and the covariance of a variable pair Zi(u)
and ZJ(u)~:-:::sts and is translation invariant
E[Z,(u)] = m,
C,1 (h)= E[(Z, (u) -m,) · (Z1 (u +h) -111, )]
(12)
(13)
for all i, j;:::; 1, ... , Nv. The mean-value vector of a random function Z is defined as:
m(u)=E[Z(u)] (14)
Hence the cross covariance functions Cu(h) may be written as the covariance matrix
C(h) = E{ [Z(u)- m] · [Z(u +h)- m]T), that is:
[
C,.(h)
C(h)= :
Cv;{h)
(15)
As with the univariate case it may be that in practice the covariance function
between any two locations u and u + h does not exist in which case it is necessary to
weaken the joint second-order stationarity hypothesis to the joint intrinsic hypothesis.
The assumption here is that there is only weak stationarity of the first two moments
(mean and variance) of the difference of a pair of values located at u and u + h, that
is
E[ z, (u+ h)-z, (h )}o (16)
Cav[ z, (u +h)- z, (h ),Z1 (u +h)- Z1 (h)]=2r, (h) (17)
for all i, j = 1, ... , Nv.
Hence the cross semivariogram functions y;j(h) may be written as the
semivariogram matrix r(h) = V2E{ [Z(u)- Z(u ·:·h)] · [Z(u)- Z(u + h)]T), that is:
20
(18)
It is important to note that while the cross variogram is symmetric in (h, -h),
this is not necessarily so for the cross covariance function, that is Cij(h) t- Cii(-h) and
Cij(h) :f; Cji(h). Goovaerts (1997, p. 73) states however that in practice this assumption
in ignored as the usual tools for description of the spatial variability are symmetric
and the verification of the presence of a lag effect is generally not possible as a result
of insufficient data.
2.3 Modelling the Coregionalisation
The regionalised variable possesses a local, random, erratic aspect which
accounts for local irregularities as well as a general structured aspect which reflects
large scale tendencies (Armstrong, 1998, p. 15). Both of these aspects need to be
taken into account in the process of developing a representation of the spatial
variability of the regionalised variable. Our goal when modelling the semivariogram
or covariance is to obtain a suitable interpretation of the spatial structure that
characterises the association and causal relationships and main features of the
corcgionalisation. Our need for a model of the coregionalisation arises from the fact
that it is likely that for estimation purposes we will require a semivariogram or
covariance value for some distance and/or direction for which we do not have 11. value
(lsaaks & Srivastava, 1989, p. 371). Inference of the scmivariogram or covariance
model provides a set of functions that allov.r us to compute semivarior,ram or
covariance values for any possible separation vector h.
Only certain functions may be used to model the cross covariance ;md cross
semivariogram. Let Z/u) fori::: 1, ... , Nv be a set of intercorrelated random functions,
Ua for a= 1, ... , n be a set of 11 data locations andY be a finite linear combination of
the random variables Zi(ua), where Ua. is a sample location in the study region A and
i = 1, ... , Nv. The variance of Y must be non-negative and can be written as the linear
combination of cross covariance values
21
(19)
where Cij(h) denotes the cross covariance at a lag distance h. The variance in terms
of the matrix C(h) is
" " Var[Yl=L~>:C(u,-u1 )Jc1 <eO (20) a~I P=I
where A.., =[A.aP···•A.aN, ]T and A! denotes the transpose of the vector Au. Thus the
matrix of covariances C(h) must be positive semi-definite to ensure that the
variances of Yare non-negative.
Using the relation C(h) = C(O)-f(h), the variance in terms of the matrix r(h)
is:
n n n n
Var[Y}=C(O) :L<!:LJc,- LLl.!r(h)Jc1 <eO (21) a"'l fl=l a=l p .. i
When a semivariogram is unbounded and has no covariance counterpart the variance
of Y is defined on a condition that the vectors of weights Aa sum to the null vector.
" " " Var[Y}=- LLJc!r(h)l.1 <eO with LA• =0 (22)
a=! P"I e>.=l
Thus the matrix of semivariograms f(h) must be conditionally negative semi-definite
to ensure that the variances of Yare non-negative.
Recognising whether a function is positive definite or conditionally negative
definite is not easy, nor is it simple to te'lt for positive definiteness. Hence it is
common practice to model a coregionalisation by using only a few basic structures
that are known to be admissible. The following list is not exhaustive but includes the
most commonly used admissible models in their standarised fonn.
' Nugget effect model
{0 if h=O
g(h) = 1 otherwise
• Spherical model with range a
if h;:;a g(h) = Sph( ~) = J 11 s ~ - o.s( H l otherwise
(23)
(24)
22
• Exponential model with practical range a
(-3h) g(h) =I - exp -;;- (25)
• Gaussian model with practical range a
( 3h') g(h) =I - exp T (26)
• Power model
g(h) = hw with 0 < w < 2 (27)
These models are considered to be the 'basic models' and are expressed in
their isotropic form, that is they are independent of direction (h =I h I). These basic
models can be modified to incorporate anisotropy if required.
The nugget effect model is a bounded transition model and is characterised
by its discontinuous behaviour at the origin. The spherical and exponential models
are also bounded transition models whose behaviour at small separation distances
near the origin is linear. The spherical model reaches its sill at the distance a, also
known as the actual range. The exponential model reaches its sill asymptotically with
a practical range a, the distance at which the semi-variogram value is 95% of the sill.
The Gaussian model is a bounded transition model with quadratic behaviour at the
origin and also reaches the sill asymptotically with practical ra:1ge a, the distance at
which the semi-variogram value is 95% of the sill (Isaaks & Srivastava, 1989, pp.
373-375). Finally the power model is an unbounded model and has no covariance
counterpart. The behaviour of the power model at the origin is dependent on the
value of the parameter ru, that is, linear when w = 1 and parabolic as w approaches
two.
The linear model of coregionalisation is a mathematical model that
characterises the spatial variation of a multivariate system at different spatial scales.
The requirement of a linear model of coregionalisation is that all direct and cross
semivariograms or covariances are jointly modelled and share a common set of basic
structures. The linear model of coregionalisation consists of :1 set of intercorrelated
random functions 2i so that their corresponding semivariogram matrix r or
covariance matrix C is by construction admissible. This means that each random
23
function 2i is a linear combination of (L+l) spatially uncorrelated components Y},
for l = 0, ... , L. Each of these componerts acts at a particular characteristic spatial
scale and has a covariance function c, ami· zero mean (Joumel & Huijbregts, 1978, p.
172).
(28)
for all i = 1, ... , Nv with
• E{Z,{u)} =m1 (a)
• E[ Y/(u))=O forallk=l, ... ,n1 andi=O, ... ,L (b)
. {c(h) ifk = k'andl Cov { Y/ (u), Yj. (u +h)) = ' .
0 othenvrse
= I' (c) •
If follows then that the cross covariance function associated with the spatial
components Z;(u) and ZJ(u + h) can be written as the linear combination of the cross
covadance between any two random variables Yl1 (u) and yl;· (u +h).
(29)
As the random variables Y/ (u) in equation (28c) are uncorrelated (orthogonal)
except when I = r and k = k' simultaneously, equation (29) reduces to a linear
combination of (L+l) basic covariance models c1(h). Hence we define the linear
model of coregionalisation as the set of Nv x Nv direct and cross covariance models
c,,(h)
' CIJ(h) = Lb~c1 (h) (30)
1~0
for all i, j = 1, ... , Nv, where the sill of the basic covariance model c1(h) is mattix
valued and defined by
(31)
for all i, j = 1, ... , Nv and l = 0, ... , L. Similarly we can define the lineoar model of
coregionalisation in terms of the set of Nv x N.~ direct and cross semivariogram
models rv(h) such that:
24
' Y,(h)=~>:g,(h) (32)
'"' The coregionalisation matrix of size Nv x. Nv , Bt = [ b~ J, is by construction a
positive semi~definite matrix and is the variance-covariance matrix which describes
the multivariate correlation at each of the characteristic scales I for l = 0, ... , L
(Wackemagcl, 1998b, p. 27). Thus for a linear model of coregionalisation to be
admissible it must satisfy two conditions:
i) the functions c1(h) (gt(h)) are admissible covanance (semivariogram)
models and
ii) the (L + 1) corcgionalisation matrices 8 1 are positive semi-definite.
This second condition is readily checked by confirming that the eigenvalues of each
of the (L+ 1) coregionalisation matrices 8 1 is real and non-negative.
The multivariate nested covariance function model Cv(h) with positive semi
definite coregionalisation matrices B1 expressed in matrix notation is ,_
C(h) = LB,c,(h) (33)
'"' with B1 = A1Af and A1 = [a~ J. Correspondingly, the multivariate nested model
associated with a linear model of coregionalisation of intrinsically stationary random
functions expressed in matrix notation is
' r(h) = LB,g,(h) (34) ,., where Bt arc positive semi-definite matrices and g1(h) are the semivariogram models.
If the multivariate correlation structure of a set of Nv variables is independent
of the spatial correlation the multivariate correlation is said to b~ intrinsic. According
to Chiles & Dclfiner (1999, p. 337) Matheron introduced the intrinsic
coregionalisation model in order to validate the use of the correlation coefficient
from a gcostatistical perspective. The problem arises with the fact that variance (or
covariance) of spatially correlated data within a finite domain A, depends on A
(Chiles & Dclfiner, 1999, p. 337). The intrinsic coregionalisation model may only be
implemented when the correlation riJ between the random functions 2/ and ~does
not depend on spatial scale, that is:
25
(35)
The intrinsic corcgionalisntion model is a special case of the linear model of
corcgionalisation where the coefficients (sills) of any basic structure c1(h) or g1(h)
that constitute the model are proportional to each other. That is, b,~ =rp,1
·b1 for
i, j = I. ... .md I= 0, ... , L. The intrinsic coregionalisation model is the simplest
multivariate model used in gcostatistics as all direct and cross covariance and
semivariogram models arc prof}Qr.f.t'nal to the basic standardised covariance,
£ L
C(h) = Lii c1 (h), or scmivariogram, y(h) = Lb1 g1 {h), functions
'
c,1 (h)=~,1 C(h)
Y,,(h)=~,y(h)
(36)
(37)
for all i, j = 1, ... , Nv and L.b' =I. In matrix notation the intrinsic coregionalisation /=0
model is written
l'(h)=<l>y(h)
C(h)=<I>C(h)
(38)
(39)
where the matrix of coefficients 4> = [tpuJ is equal to the variance-covariance matrix
under the assumption of second-order stationarity (Wackernagcl, l998b, p. lO).
While the intrinsic coregionalisation model is more restrictive than the linear
model of coregionalisation it is of particular benefit as this model reduces the
inference of Nv (N1• + 1)/2 covariance or semivariogram functions to the inference of
only one covariance or semivariogram model and N,. (N,. + 1)/2 coefficients. One
way of detecting whether variables are intrinsically correlated is by examining the
codispersion between the variables. This can be done by checking graphically
whether the codispersion coefficients
Jy,(h)y g(h)
are constant and equai to the .sample correlation coefficient. If they arc the
correlation of each pair of variables does not depend on spatial scale. This can
26
however become time consuming when the number of variables is large. An
alternative method is to perform a principal component analysis on the data and
compute the cross corre!ograms of the first few principal components which account
for the majority of variability in the original data. If the cross correlograms of the
principal components are zero for any separation vector h this implies that the
orthogonality of the principal components is not dependent on spatial scale. In other
words for the data to be intrinsically correlated we require the cross correlograms of
the principal components to be zero for any separation vector h.
2.4 Kriging
Once we have a mathematical representation of the spatial continuity of the
attributes of interest in the fonn of our random function model we are able to proceed
with the estimation of those attributes at unsampled locations across the study region.
There arc many traditional point estimation techniques available, such as polygonal,
Dclaunay triangulation, inverse distance squared and moving average methods. The
overriding problem with these techniques is that the 'best' one is dependent on one's
choice of criteria as to what is 'best'.
In the 1950's South African mining engineer Danie Krige developed a
technique of interpolation in an attempt to more accurately predict ore reserves.
Based on Kiige's work, Georges Matheron developed the Theory of Regionalized
Variables in the early 1960's in which he combined Krigc's pioneering work into a
single framework which was coined "krigeage" in recognition of Krige's
contlibution to the field (Chiles & Delfiner, 1999, p. 150). Fonnally, kriging now
refers to a family of least-squares linear regression algorithms that share the
objective of minimi~:ing the estimation (eiTOr) valiance subject to the constraint of
unbiasedness of the estimator (Deutsch & Joumel, 1998, p. 14). Over the past several
decade~ kriging has become a fundamental tool in the field of geostatistics as it has
established itself as a superior method of estimation.
27
2.4.1 Simple and Ordinary Kriging
Let Z be a second-order stationary random function with mean m. The
estimator Z*(u) of Z is given by a linear combination of random variables Z(ucr:) with
weights lta(u) r.hosen such that the estimator is unbiased tl.nd the estimation variance
is minimised. The basic generalised spatial least-squares regression estimator Z*(u)
is defined as
n(u)
Z*(u)-m(u) = L,!."(u)[Z(u.)-m(u.)] (40) a~l
where the quantities m(u) and m(ua) are the expected values of Z(u) and Z(ua)
respectively. By making the assumption that the both the sample value z(ucr:) and
unknown value z(u) are realisations of the random variables Z(ua) and Z(u)
respectively we are able to define an estimation error random variable Z*(u)- Z(u).
In order for the estimator to be unbiased we require that the expected value of the
estimation error be zero:
E{Z* (u)-Z(u)}=O (41)
The estimation error variance is then given by
"; (u )= V6r{Z* (u)- Z (u)} (42)
and is minimised under the constraint of unbiasedness of the estimator.
If we assume m(u) to be known and constant across the study region A the
linear estimator Z*(u) is known as the simple kriging (SK) estimator z;K(u) where
"'"' [ "'"' l z;K(u):::: ~1\..~"(u)Z(&a) + 1- ~1\.:K(U) m (43)
and the .AsK(u) are determined such that the error variance is minimised. The simple
kriging estimator automatically exhibits unbiasedncss as the m~an error is equal to
zero. The simple kriging system then becomes, in terms of the covariance function, a
system of normal equations
n(u)
L,!.:' (u)C(u, -u,l = C(u. -u) /l~l
for all a= 1, ... , n(u). The simple kriging minimum error variance is given by
(44)
28
n(u)
"L (u) = Var{z;, (u)-Z(u)}=C(O)- LJ;' (u)C(u" -u) (45) a=l
In most cases however it is not acce~table to assume that the mean is known
and globally constant. An alternative approach is ordinary kriging which assumes
that the mean is unknown but locally constant. Ordinary kriging then limits
stationarity of the mean to the local neighbourhood of the location u being estimated.
The linear estimator (40) is then modified to incorporate the constant local mean
m(u).
(46)
In order to filter the mean m(u) from equation (40) we must impose the constraint
that the sum of the weights be equal to one. This yields the ordinary kriging
estimator z;x (u) which again must be solved to obtain the optimal kriging weights
such that the estimation error is minimised under the unbiasedncss constraint.
n(u} n(u}
z~, (u) = "'}c OK (u)Z(u") with >A OK (u) = I L..J " • ....~ "
(47) a=! a •1
The ordinary kriging system involves n(u) weights ).~K (u) and the Lagrange
multiplier tloK(n), which accounts fur the unbiasedness constraint. Expressed in
tenns of the covariance function this system is
n(u}
L ).~K (u)C(ua -Up)+ Pox (u) = C(ua -·u) 11=1 n(u}
L?c:' (u) =I P=l
(48)
for all a= I, ... , n(u). Unlike the simple kriging system the ordinary la.iging system
may also be expressed in tenns of the semivariogram as
n(u)
L"~' (u )Y(u" -u, )- ~0, =y(u" -u) /1=1
I:,l~K (u) =I
(49)
/1=1
for all a= 1, ... , 11(0). The ordinary kriging minimum error variance is given by
29
a;"' (u)= = Var{z;K (u)-Z(u)}= C(O)- ~,!~K (u)C(u, -u)-PoK (u) (50) a"' I
2.4.2 Ordinary Cokriging
The linear estimator (40) is readily extended to the multivariate case where
we have available Nv c'Jntinuous random variables Zi(U). We consider Zi(u) as a
realisation of the random variable Z;(u), i = 1, ... , Nv, with Zi(ua) being the set of n
sample data located at ua, a= 1, ... , 11. Let Z1(u) be the primary random variable of
interest with E{ZJ(U)} = lllJ(U), E{Z1(u,,)}= mJua) and Aa, and Aa, being the
weights assigned to z1 (ua) and Zi(u,) respectively. 1n order to estimate a primary ' '
variable with Nv-1 auxiliary variables the linear estimator (40) is extended to
incorporate the additional information.
Analogous to the kriging paradigm the cokriging algorithm generally only
retains the data closest to the location u. Again we wish to determine the weights A" '
and A", sur:h that the estimation variance
a; (u )= Var{ z; (u)- z, (u )} (52)
is minimised under the unbiasedness constraint that the expected error is zero.
E{z; (u)-z, (u)}=O (53)
The three most commonly used types of cokriging are
i) Simple cokriging where the mean mJ(u) is known and constant throughout
the study region A
ii) Ordinary cokriging where the mean mJ(u) is unknown but constant
throughout the study region A and
iii) Cokriging with a trend where the mean mi(u) is unknown and varies as a
function of the spatial coordinates u.
Pertinent to this study is ordinary cokriging which limits stationarity of the
mean to the local neighbourhood of the location u being estimated, In order to filter
the means m 1(u) and m1(u), fmm equation (51) we must impose the constraints that
30
the sum of the weights of the primary variable be equal to one and the sum of the
weights of the secondary variable(s) be equal to zero. This yields the ordinary
cokriging estimator
(54)
which again must be solved for the kriging weights such that the estimation error is
minimised subject to the unbiasedness constraints fori= 2, ... , Nv.
~) l, .l~" (u )=1
The ordinary cokriging system can be expressed in tenns of the direct and cross
covariances as
(55)
for a 1 = 1, ... , n1(u) and i = 1, ... , Nv. As for the ordinary kriging case the ordinary
cokriging system can be expressed in terms of the direct and cross semivariograms as
(56)
for Clj = 1, ... , n1(u) and i = l, ... , Nv. The cokriging vmiance is given by
"~oc• (u) = Var{zg[, (u)- z, (u)}
= C, (0)-pf" (u)-ff ~" (u )C,, (uo, -u) (57)
i:l "'"']
As the cokriging systems (55) and (56) can become unstable if the variances of the
primary and secondary variables differ by several orders of magnitude it is advisable
to use standardised variables if this is likely to be a concern.
In general the benefit of incorporating secondary infonnation is fully
exploited when the primary variablc(s) of interest is undersampled. In the isotopic or
31
equally sampled case the estimates obtained from ordinary cokriging are likely to be
similar to those obtained by ordinary kriging. However one advantage of cokriging a
set of equally sampled variables is that we preserve the coherence of the estimators.
That is, the cokriging of a sum of variables is equal to the sum of the cokrigings of
each of the variables. Another advantage of cokriging in both the isotopic and
heterotopic cases is that the estimation variance is less than or equal to that of the
kriging estimator. In the particular case of the intrinsic coregionalisation model the
ordinary cokriging estimates will be equivalent to those obtained by ordinary kriging.
2.4.3 Principal Component Kriging
As we have discussed previously, a principal component analysis transforms
a set of correlated variables into a set of components that are uncorrelated at I hI= 0.
The advantage of principal component kriging is that it reduces the estimation
problem of cokriging N1• variables to the kriging of Nv principal compr,;nents. The
overriding assumption of principal component kriging is that ~he principal
components are mutually orthogonal for any separation vector h
Cov{Y, (u),Y,. (u +h)}= Cw (h)= 0 (58)
for all k 'f:. e. If the data aw intrinsically correlated condit.ion (48) is then satisfied.
This means that there is no benefit in incorporating seco.1dary infonnation as the
kriging and cokriging estimates will be identical. Hence the principal components
can be kriged independently. One of the drawbacks however of principal component
kriging is that the data must be isotopic, that is only those data that are jointly
measured can be considered.
Having determined the N,. principal components we calculate the principal
component scores Yk(Ua) as per equation (4) and model theN,. semivariograms yu(h)
from these scores. We then estimate the principal components separately at each
unsamp!ed location u in A As the mean of each principal component is zero, the
ordinary kriging estimator of the kth principal component at location u is of the fonn
Y~~· (u) = ~lc~,' (u )Y, (u.) (59) a=!
32
where the kriging weights are obtained from the ordinary kriging system as displayed
in expression (48). The estimate of z1(u) is then recalculated as a linear combination
of the principal component estimates at each location plus the mean m1 of each
attribute.
K
z~b~(u) = Lak;Yh~·(u)cr; +m; (60) k-1
The coefficients ak1 are obtained from the matrix A= [aH] = Q-1 = QT where Q is the
orthogonal matrix of eigenvectors calculated in the principal component analysis.
Should sufficient variability of the original data be explained by P principal
components (where Pis ideally significantly less than Nv), it is possible to retain only
these and yet simultaneously estimate all Nv original variables without the loss of too
much information. In this case the estimates are obtained by modifying equation (60)
as follows:
' ""' "' ·''''() ZPcKP = ~ap1YoK u a;+ m (61) p=l
33
3 DATA ANALYSIS
In this chapter we introduce the IVIM22D data suite used in this study. We
then proceed with an exploratory data analysis of the grade control and exploration
data sets contained in the data suite. We next discuss the principal component
analysis of the :rvt:M22DEXP data set and assess whether the data are intrinsically
con·elated. In addition we consider the principal component analysis and subsequent
assessment of intrinsic correlation of two four variable subsets of MM22DEXP
called MM22DHC4 and MM22DTOP4. The first subset consists of variables that are
highly cmrelated with nickel and cobalt. The second subset consists of variables that
are considered to be economically the most important to the mining company.
3.1 MM22D Data Suite
The data to be us~d in this study come from Anaconda's Murrin Murrin
nickel mine near the town of Laverton in Western Australia. The data have been
collected from an area within this mine known as MM2. Murphy, Bloom and
Mueller (2002) explain thai in this region "the laterite deposits are of the dry-Climate
type and occur as laterally extensive, undulating blankets of mineralisation with
strong vertical anisotropy and near normal nickel distrib•Jtions." The data suite
MM22D comprises three dimensional grade and thickness measurements on eight
variables: nickel, cobalt, magnesium, iron, aluminium, chromium, zinc and
manganese. For the purposes of this study the data have been transfonned to
accumulations (average grade multiplied by total thickness) hence are treated as two
dimensional.
The :MM22D data suite consists of two data sets: the grade control data set
:rvt:M22DGC and the exploration data set r-.11vl22DEXP. The data 11I\1122DGC are
grade control accumulations for each of the eight variables jointly sampled at 1718
locations over an irregularly shaped study region with grid spacings of 12.5m by
12.5m as displayed in Figure 3.1. This data set is considered reality and will be used
to assess the accuracy of the results obtained from the estimation methods.
Mlvi22DEXP is a subset of :tv1M22DGC and comprises accumulations for each of the
34
eight variables jointly sampled at 125 locations across the study region as displayed
in Figure 3.2. Ninety-seven of these data are located at grid spacings of 50m by 50m.
In addition there is a local cluster of 24 samples at 12.5m spacings around the
location at easting 1098.03, northing 298.53. There are also four additional samples
along the line of drill holes extending southwards from easting 1302.22, northing
357.59 at 25m spacings. These local clusters are sampled for the purposes of
identification of the short range characteristics of the variables (Murphy et al., 2002).
550 500 450
400 0.0 350 • . e l ~ 300 • <:> z 250
200 i 150 100 • • • 50
550 650 750 850 950 1050 1150 1250 1350 1450
Easting
Figure 3.1: MM22DGC grade control sample locations
550 500 • 450 • • 400 • • • • • •
0.0 350 • • • • • • • • • • • • . e ·mH· •
~ 300 • • • • • • • • • • • • • • 0 • z 250 • • • • • • • • • • • • • # • • • •
200 • • • • • • • • • • • • • • • • • • 150 • • • • • • • • • • • • • • • • •
100 • • • • • • 50
550 650 750 850 950 1050 1150 1250 1350 1450
Easting
Figure 3.2: MM22DEXP exploration sample locations
.·
35
3.2 Exploratory Data Analysis
Two types of measurements have been used to assess the level of the
attributes. Nickel, cobalt, magnesium, iron and aluminium are measured in parts per
million (ppm-metres) and chromium, zinc and manganese are measured in percent
metres (%). The only erroneous data identified in the exploration data set are the
samples located at casting 900, northing 398 (denoted by a red marker in Figures 3.1
and 3.2). As the readings for each variable are incomplete, this record has been
removed from the both i.he grade control and exploration data sets. Tables 3.1, 3.2
and 3.3 display the descriptive statistics for the grade control, exploration and
declustered exploration data (using cell declustering) where SD and CV are the
standard deviation and coefficient of variation respectively.
Overall the exploration data have reproduced the grade control sununary
statistics well. In most cases summary statistics of the declustered exploration data
are very similar to the clustered values with manganese being the only mineral to
show "~'Y difference. As the differences are minimal the exploration data rather than
the declustered data will be utilised in this study.
Table 3.1: Summary statistics for MM22DGC grade control data
Ni% Co% Mg% Fe% AI% Crppm Znppm Mnppm
n 1717 1717 1717 1717 1717 1717 1717 1717
Mean 13.05 0.82 62.41 313.83 46.33 120053 2312.90 38874
Median 12.98 0.78 57.39 312.19 44.49 118127 2314.20 35411
Min 0.24 0.00 0.20 9.60 1.40 345 86.0 0
Ma. 42.64 4.90 378.93 813.79 163.90 372240 6831.99 236803
SD 6.64 0.56 52.73 144.53 25.91 67042 1054.3 34943
Skewness 0.20 1.26 1.51 0.08 1.18 0.33 -0.03 1.57
cv 0.51 0.68 0.85 0.46 0.56 0.56 0.46 0.90
36
Table 3.2: Summary statistics for Mlvl22DEXP exploration data
Ni% Co% Mg% Fe% AI% Crppm Znppm Mnppm
n 124 124 124 124 124 124 124 124
Mean 14.07 0.85 78.46 318.6 47.99 82471 2130.10 46354
Median 13.37 0.68 58.69 314.9 44.20 81357 1989.0 36279
Min 0.56 0.0?. 8.07 11.0 1.40 2119 88.0 1149
M" 42.64 2.97 378.93 778.99 146.40 217900 4918.10 236803
so 6.72 0.55 67.51 144.2 27.13 48238 944.3 43314
Skewness 0.95 1.24 2.06 0.22 LOS 0.42 0.33 1.69
cv 0.48 0.65 0.86 0.45 0.57 0.58 0.44 0.93
Table 3.3: Summary statistics for declustered MM"22DEXP exploration data
Ni% Co% Mg% Fe% AI% Crppm Znppm Mn ppm
n 96 96 96 96 96 96 96 96
Mean 15.28 0.91 87.00 354.0 49.82 91059 2330.8 50985
Median 15.09 0.82 68.65 331.3 45.05 91800 2182.2 44391
Min 3.83 0.14 8.07 98.6 8.41 2119 326.0 2220
M" 42.64 2.44 378.93 778.99 146.40 217900 4918.10 200600
so 6A7 0.50 72.57 128.4 28.06 46718 881.10 39234
Skewness 1.05 0.94 1.85 0.33 Ll4 0.41 0.43 l.l7
cv 0.42 0.55 0.83 0.36 0.56 0.51 0.38 0.77
37
Graphical summaries of the grade control data are in displayed in Figure 3.3.
Cobalt, aluminium, magnesium and manganese are all strongly positively skewed
(skewness coefficient greater than 1) for the grade control data.
~ ~ ~ ~ . ~ ~ ~ ~ I I I I I I I I I
I .. I
I .. I
:J. I
11 I
f
.lo .\. .lo .lo I I I I
161Ao Confidtnc! l!l"eNIII I or Mu
t::J::J .1. .1
I
MtloCDnh~lnl.,.,lll l or t.\1
c::::r:::::::l I
" I ., ,\ I
" I .. I
q I • I I I I I I I
,1g, .,....,;....__,,.........,..-...., I
Va~able: Nl
AndfUCII'M>Mng NomuVy Tnt
o\o5Quwed 1.812 P-Y•.... oooo
"'"" 9Dw Vtnance ....... " Kunoli.s
" ......... llto...M .. -'"'""'" .. .........
130410
""' •lo4.0917 0.111650!1 -!i.4E-01
1117
Q_2<fllJ . ..... 12.184CJ ,.., 42.1400
IS'KCON~IRII'I'f'llifO'"IN
12,7U7 1:i.3LU~
I$"Confidt"CtiiUN'tllo.rSign'll
1 4153 UTOO
191. Co,.llhncellt(~tllar Median
178720 13.4084
Variable: MG
Ancleucw~ . .o•r~~ NollN!Iry ff:!; t
A-Sqlllmf 9.404 P.V•..._ OOCQ .... -y ...... Sl~n ~otis • Mftmu~t
htOuartle ·lniDI••U•
"'•"""'
.,..,.. s1n• neo:o 1.51115 2.n18!J
1711
0.200 24.511 <48.616
""' 378930
iS'MoConlldluw::elnltrvtl for ~
59113 6481'6
IS" Collllclenu lruNII 101 S6grm
i l DU 5455-4
IS'K. COitidfl'll:t lntM'tl far.,ediH
4159 51 111
Varieblo:Al
AnO....OMnoHonnallv Tnt
A-Squated n1w P-v-.. a_ooo .... 413316 90w ,., .. VeiWK:e IIT U M 6k.W.ll 1. 18215 K.lftotis 2Ci31132
" 11 17
""""'m t .oiOO ISt OU:IItil: 29.5!M ..... • 1.810 ll'doartill 58.oi07 •-m 18U6C
IS'KCIIMide..:elrt.""'ll fDJ~
45 105 4l..5S.
tM~ C_.IOtfllt"e ltUf\oll l015$rN
.25..DIII 26.104
16K Cca'identfllrHNalfar"'edAn
40315 oiJ.2!M
vaneble: ZN
~ISort-0.-rilng NOifl'llllily Test
A-SQuattd 3.571 P·VIU. 0000
2l1U1 1"'28 111150& -U"'12 -ISE-Gt
UIT
... _ 11100
lsco.rta. ISOO WHW! 2313.00 lniOI.Jirlile 3tJ5.9.00 lotJII(rNPft tlttl1 81
16'K Confidence lnuwlll fot,.ll 77&401 23113.81
15~ Co,.IIHnc• lf•ttval tor stwna 10'10111 1000.18
IK~ CorldffiCelrteNIII f ar M~i.ln
'2'33000 1411100
I I I I 2§S llO j.OS dO I I I I
ts• Olnl•cl!ftet! lrtrv,." -" -' ·.,· .,.........,,
.lo ... .:.
as~ OlnfldtiK~ l rtsvll ror Mu
IS'!. CDnlldr:~l rtsvllfot llfu
t• I I
I"IDO l'rJCOO I I
I zmr ::a: t: sr»:: ts•COfttdtn:•lileNIIIfarWe.t•n
""" """ ci::::J I
es• Conlld~t~a tntttv• for Mect~t~
I II
Vanoble: co
MdwtDnOifingNom.Uy Tat
2Li70 .... IAUII 0 81 M53 ao.v 0558130 Vlri~e 0312171 s~u 1..2!MII l(uiD•s 3 .151.42
" 1117
"''*""" D.DaJOO I•C.,nle Q4l000 ..... O.ll:IID lMOYrie 1.10ta0
"'"""" 4.1mDD
tstlo CCWid!tl(:~rtrt8'VJI rw.., 071101 o..aao
UK C(Jihl'fttelt!UtVlllfOfS~""
054)115 Q_51tl)1
.,.CDnfiden::•lrterval. lor Medaf'l
01Ctl00 0 .78100
Va~ab le: FE
Md• mn.O.,i'lg Nomllty Tau
A-Sq~r.t Ul-i fWIII~ OJIIID
Nt•n SDiv v•--=• . ........ K1110th: • ..... ~ t•Oull1lle Med•n 31dQurie Wtwnun
313100 14U3l 1(81111
I.IIE42 -2..1E.OI
"" uoo 110152 lt8201 oi01211 81.3182
15,. Confld!rcel rtS'v .. for Nu
311111811 3:1)871
l mo\ Cftldenc;:elntervllllot!IQIN
191i11U 141$37
Z tlo Cenfrdtrn lrU:rvl!l forNMen
3 101114 318113
Variable: CR
....... tofttJ•U)g Na~.-lty Tat
~·- 5047 P.Vtl~. 0001
Mt~n nms:a !lOov """ Vwt.-.ce 4 49E:t0 1 S~u 03322-45 K111!1)1is .JOE41 N 1111
Minm1n1 ... I• Curie 11~10 Mltdr.n 117480 JldQ,urde llla:Jiilll .... JIW!IIIn 372'NO
liS'!. Oll'lfid!-rc:el rt rnl fot"' u••o 1237211
!5'A CII'Cid!:ntc lntuvd. tarSotm~
1473 IIIDIJ
15'K Confide.rc:t Irs eN• far WtODn
117120 121111
Venable: MN
AI\Oerson-Oalfrl; Notm11iry Ttsl
A-Sqund- 4U71 P.VaiUk OOCQ
Jlil87oi4 ,...,. I_,. 1-"311 3 47111
1111 - . lit O.... 11131 Mfdilft 308m )'dQJidik 54150 Umnun 23810:1
I$'K Coridtftt'e lnt.rt¥111 fat lll
nno 40528
IM4 CarM'iiMIKt lrternl far Sf;tm1
33812 38152
95'KCorj'ldeJ!ttlrteiVtlforMed.ln
29018 32800
Figure 3.3: Graphical summary of.MM22DGC grade control data
38
The skewness of the variables cobalt, aluminium, magnesium and manganese
1s also apparent in the exploration data as displayed in Figure 3.4. The exploration
data for iron and chromium are normally distributed with nickel near normal.
Variable: Nl Variable: CO
Ar!dtt~Of'ONIII'!Qt~TCSI AnteiS~N!rriUtyTI'• ...., ... ,,., ......... .... P.VIIIut o.o:n P.Va\le" '""' """ 140731
'···~ "''"' - 6_7191 """" ,....,., v ..... 4514B7 V..arct 01&QI&I
""""" OIH~~1 ........... I ""'
I - ,., .. I
..... .. 1ll""' .. , N I >I ,. N I >I
......, """' ......... 01121101
"'"""" u:~&s ,.~t OMIOJ ...... ,,... .... ~ ...... ........ '""" """"" 1111!0 M"Orttdl:f'Q...__b'WU ·~~
.,.., tw. ~Hen.lbMu """""' ....
115"0:1rf161nc•.lrt.,..,fo'.., ""'Cll'tim211*Mt b t.lu
I I ~ I ,.,.. 15.2tl15
I I I I IL7til1 OOWI
" " " ""Cbt1bnn h!MI ... Sqna .. " .. .. ""~tarwlbSi~ I I
517-'1 7.1781 I I
0.48552 '"""' M'KCotfidtrnoelfteN!IIflJ' Mt~bn UK ~~ trmw1 b MBhn 15~ CmlllefH W:fflll tar Mllbn 120l!D .. .,. ll!itl.~lrlfi\MbM.a~ 0617711 oamn
Variable: M8 Variable: FE
Andefson.ONI1119tbmJ!ityTut A~t«mdr-tT•• ........ ,,. ·- 021 P-V1h• 0000 p.v ... . .. ·- ,.,.... - 318111 - 8751J9 - ... 111 v ...... 455785 v-• ,., ... ......... 10511ol ......... 02tst8 - ,.,., ....... -llE-01
" "' N ... I I "'-•VI1 '"'" ......... OOCIUl - l«c:Md" '"" htDJril -znast ..... .. ... ..... 314ts9
:tlfo..o.tolf 11193 ~Q.Irie GI<B 15"-Qrllidtra herd b t.lu """"" 37UXI ~'K Oridlm:• lrl5'41 b !Au """""" 718892
Z%eo..tldttl<:elltfl'\&lfflhlu 9S"~Woldm:eldH<':IIbrl.tU
I ol I I ~. I ""' 90.458 ,/o I I .k .!o ,J., ,Jo
292 .9Sll "'"" .. .. " " "'' ~e\"'trMibr SIQI'D "" "' 95~0Yildtooe~~bSIIJnl
'""' 71141 I I I 1211191 IM153
{16-~ltttNlllktMfd-. ~~I'W..MIItlfiJotlbn liS" ~dett•"'mb Medii .. ..... 69374 15-.Conldn:.eW.II"ddfcrMniW'I
"'"' ,.,,.
Variable:~ Vaiable:CR
~fii'OIIi ... tmnlktTrsa 1-~Hom:Uylal - '"' - '"" P.V1fut 0000 P.v-..e oro
·- """' .... 82-4712 - 27.1252 - ..,.,, V"""'• 135n.:~ V•~ 2:nE•CIJ ·- I .... -- ....... """"" """'' I I I ' I """"" -UE~I N "' .... """' '"""' I""' """""
II "' l.lnrru n 1.<00 I
·~ 2119 ltlelMiM '"" 1-.Qlrie 45310 ..... 44.1(18 ""'"' ""' .. """ .. ..,.. JodO..r-1• . ~..,
IS'KCUltdencttuMibrt.IU ·- ...... II&Mo~Her.ettr Mu Mftmm "'"" li6~ Ccwtdtrnbt~kt t.lv 115'KQrir:lftnmfdbUu . .1, I I .,, .. ., ...
..loa ,.!., I ..lao """ ..... " .. " -~~b59n;o .... IWCMI~hrMib Sliflll
I I :.a us ..... I .... ""' e•~n.-truea;,., I!KClr*Sttlte~DMd!M ~0711d!tlceWt.wbt.ta:bn 31111 ..... ··~W.erwiiD't.IEd;n ..... """
Vllliable:ZN Variable: fVN
~Otiii'IQ~tyltu ~ruNo!m11t)'lm ......... !1.«169 """'""' , .,,
P-W~ ,,.. P-V~Ik ""' .... UD.13 ·- ....,., ..,., ... ,. ..,., qll43 ...... 891719 ...... l-......... 0:01859 -~ 183912 ·-- -1.3E.OZ I I I I I I
....... 1""" II .,. - """' t2mllO lmorD 2lOUlO :utiXI) N ,,. """""'
.,., """""' I Mil
Ill~ 1533_75 ··~ ""' ....... ...... ...... .,. ··-- ,.,..., .. """" 1110)7
~0711dlnc:e..,....A:Irt.lu ........ -491810 ""Qarfldsu h~ b Mu ·-- 2JOm
Z\lo~lf'UNIID'Uu ts%0wh:lerclr~bM\I
I I . I I I 11t171 ,., ..
I I , ,J., ..loa '''" ..... .... .... , .. , ?1'0 ""' "" Ht\~n..,.. lbrS9'f'IJ , ... .... IS%CcnDEricelrl ti'\DIIDI'9!JTB I I I I . I
llall 107808 I ,..., -15Mo~lrtl!Nifftt lotedian 9.5'~ Cort..ftrlta lr'ttlrui ID!"Mfit;n ti~ Ccrictfonce lrltfMi t~~ t.4edl.n lUI In 22~78
g~~lrUMb'Medi21 '27S'16 .. ~.
Figure 3.4: Graphical summary of exploration data
39
From the summary statistics displayed in Tables 3.1 and 3.2 it is evident that
the range, mean and variance of the auributes vary greatly. In order to make the
attributes comparable the data need to be standardised by subtracting the sample
mean and dividing by the sample standard deviation. Tab!.'! 3.4 displays the
standardised summary statistics for the exploration data (with the erroneous data
removed) and Figure 3.5 shows the graphical summary.
Table 3.4: Summary statistics for standardised MM22DEXP exploration data
Ni% Co% Mg% Fe% AI% CrPPM ZnPPM MnPPM
N 124 124 124 124 124 124 124 124
Mean 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Median -0.10 -0.30 -0.29 -0.03 -0.14 -0.02 -0.15 -0.23
Min -2.01 -1.51 -1.04 -2.13 -1.72 -1.67 -2.16 -1.04
Max 4.25 3.89 4.45 3.19 3.63 2.81 2.95 4.40
SD 1.00 1.00 1.00 1.00 1.00 1.00 1.00 l.OO
Skewness 0.95 1:24 2.06 0.22 1.08 0.42 0.33 1.69
In many geostatistical and statistical procedures, we make the assumption of
second-order stationarity and normality. Although a principal component analysis
does not require normality of the data, the results are enhanced if this condition is
met. In many cases it is possible to transform the data in order to remove any trend,
stabilise the variance and achieve normality. However, one of the problems with
transforming data is that the estimation errors can become exaggerated when the
estimates arc back transfonncd. Hence, it is preferable to avoid transformation if
possible. Figure 3.6 displays the normal probability plots for each of the eight
variables. The plots for cobalt, magnesium, aluminium and manganese show strong
deviation from normality in the upper and lower tails. Figures 3.7 and 3.8 show the
lognormal and Weibull plots for these variables. While neither transformation
appears to have adequately removed the asymmetry of the raw data the lognormal
plots seem to have Jess deviation than the Wei bull plots. The decision of whether to
40
transform the data was further examined when we performed the principal
component analysis on the exploration data.
..:. I
I
I
~' I
-
I .0)3
..:.
I .. I
I I .02 00
I I
I D. I I
I
"
.~, I
I
"
.~, I
stendanised N ckel
A~Notmllcylut
~ OltJ P.v•w DOll - ...... - 1..., V•l3f"Ct- ltlOlO) 5~n O.MSI$1
"""'"' '''"" " 114
u ....... -201H4 1 ....... .-... u .. _
~~-........ ...... ........ 4251§1
amO:wlklerumrvllltrMu ~ITI7fl 0 17778
95'KQridcn:.e ......... ,..
o•13 tt.cn 8S1t~t ........ li:lfldtChll
.n:wn.1
stordortlsed Mlgno:Oun
Andc:n-cnOIII~fQn1kylta ~ ,.,. P-VM QDa)
..... ~"""' """' '"""' ·~ '"""' ......... 2~\W -- CI!I&M7
" "' MJnm.m -10i1Sl hiO..ll\l~ -OE&22 ·- .oam "'~ 021237 ......... ·-15%~h,..bMii
-0117'111 017Tll
IS'KQirl'ldence~tnbSI"""'
OEM13 1 1-4l77
t9K CcnlcPnoe lti{Nt lar ... ..._
.. _ '"' r.v•w .....
u ... ..... """ """" .... ~ '"""' 5ki'W'I!IU I Q&WI - ·· I!CXKl6
" 17<
MftnJm ·• r!J<IJ Itt CU.U. .A) 1on8 ~ -GilJel ~o...rtft 0 .. 5171 M~t<mJm :182812
iS" Ccr*krce ~ tw Mli ..om111 omn
M~IMIMI!w!iqnt
DEll 1 tq72
ts"~~lbtt.I«Urr ftM\)7
stlrdcrdsed Zinc
~~tam:nyTU
A·Sqvftd 04111 P.Vatue 024-;f ..... ....... ""'"' '"""' - I""" ......... 0.3Jit61 ... _
. , :E.-Qll
" "' M.rrm.m ~1 11257 l ltQnls!Je -tllliM ·- .. _ .. .,..,. Om>!
"""""" 201131
~Corf.dMcehH411bNv
.017776 01771!1
15~~1!11nt!r'1111fllr!lo'.t~ OIMIJ IIA2n
~ Ccrl\anc.-Jrtf!"'411tlt ... t6MI
-OVMD ~~
I .,
I
~' I
85%Co'IWetalttMIIiftrNu
E ~ ..II
I I
" " I I I
.~,
- --~
I
~· I
I
" I
stendcrdsed Cobalt
Ald!!SM~ Notnuity Tnt A.~ 2~11 P.v~ o.mo .... ~ ..... '"''' IJIOODIJ
""'""' lllllllll ......... 124262 -·· 171055 .. "' ....... -15\301 l$tcu.l .. ..07oQIO ·- ~-· ,..,..., ,,.,., """""' J895S7
95% Oridlrn lrtt!v.JI' b !.11.1
..o..tme otma IM5-~elrUMIIkrS!grN
0811113 ll4272
IS'K~t_,..b'Uftbrl
..0.40ll)4 ..0.0 15:1$"
stendlrdsed lrm
~~Tesa
A-Sq.J~ 0.38i p.v~..e 0 380 .. ....
'""""' '"""' G.lt!Dll .1fif41
110
Ri,.~r-.....~MII
..omra om7t 85% OJHidt!'Ce~ftr59ra
DBEilll 114212
ts'K~ ..... b~
stendaridsed OYOITiun
~~Test ........ 0470 P.V:i1L 01<1 . ... ~ .... so.. "'""" V~..-.c:e l(](l]OO] ,..,..,.,. 0 4 10047
"""'"' -19E..OI
" 12<
......... . . ...,. luQooM ~110311 ·- ~0Zl09
"'""" .. 0.81123
"'"""" 280753
95% Co(.Wm:e lrUMI b Mu
..o.m7& otm&
M~lllfHI:ftrs;g.n.
oa11n3 uczn ~~a.u..-Jbld.t<Url
A-S~d 442!1 P-VIIIIA!I """ y~ ODCXIOO
"""' I.IXDDD ....... """" ......... UHII
~· l.'fml N '" """"m -1.04368 1-R-OJM .. <l16104 ...... ~2mll
"""""" ...... """""" "''"' Z'tlo~tWtM'btUu
.otme oams A5!0 Griderc.•hlaMifaSIQII'D
DBIIU3 114212
~"carr-.. taM tlr t.l.e:brl
-.Q_UIJII -OG.m$
Figure 3.5: Graphical summary for standardised J.\.1M22DEXP exploration data
.·
41
Nonnal Probabi~ty Plot for N1 Normal Probability Plot for CO
..... .._ ..._r .. m .. .. - .. .· . - O.l.Ull -· ._ .. , .. _ .. ~ ... ;: .. . """"
,...,.,. .. " I
00 .. eo ..
E 70 E "' " 10 " .. e 50 e 50
" ..
" ..
D. .. D. " 20 20
10 " 5 5
:
· 10 20 " ... Data Data
Normal Probability Plot for MG Nonnal ProbabiiHy Plot for FE
.. ..... _ t.t. &lim'N .. ...... , ....
" ._ 3tUtJ _,
"-"'1 _, 1 .....
" .. .. 10 .. .. E 70 E 70 .. .. " .. e 50 e 50
" .. 0) ..
D. 30 D. 30 20 20
" " s ..
·10 0 200 '" 400 100
Data Data
Nonnal Probability Plot lor AL Nonnal Probability Plot lor CA
t.t.r. ....... ... .. _
.. .... .,_ _, 12471.2 - l7.015f -· ...... .. t 5 .. .. .. ..
E 70 E 70 .. .. " .. e 50 11 50
" ..
" ..
D. 30 D. 30 20 20
" " s .. .-!
... 100 I SO ...... Data Data
Normal Probability Plot l or ZN Normal Probability Plot lor MN
.. M.(SIRio!IM K.&tt.iM .. .. · Mt!M: 2:1)1),13 . . ·' .• .....,, ....... .. ..,, NOM~
_, f31 3t,3
•• ., .. .. .. : E 70 .. E 70
" ..
" eo e 50 e 50 .. •• "
.. D. .. D. 30
20 20
" " • 5
.-
·1000 2000 3000 ... 0 5000 · 100000 100000
Data Data
Figure 3.6: Normal probability plots of raw exploration data
42
•• 95 90
- 80 c: 70 Q) 60 e so ., •o a. ;~
10
'
co • MG
Data
# . ..
10 100 1000 \0000 100000 1000000
Data
Figure 3.7: Lognormal probability plots for cobalt, magnesium, aluminium and
0.0 1 0. 10 I 00
Data
manganese
10 00 100 00 1000 00
99
Bo
u 30
- 20 c: Q) 10 e ~
10 100 1000 10000 100000 1000000
Data
Figure 3.8: Weibull probability plots for cobalt, magnesium, aluminium and
manganese
3.3 Spatial Data Analysis
. ""'
The post plots for each of the eight variables for both grade control and
exploration data are displayed in Figures 3.9 and 3.10 respectively. The post plot for
the grade control nickel reveals a large concentration of high values across the lower
eastern half of the study region. These high values drop to moderately high across the
lower western half of the study region. Another smaller concentration of medium to
high nickel values is located at the north-western comer. There is a long band of low
nickel values along the northern perimeter of the study region and another along the
43
eastern perimeter. The exploration nickel data reflect the spatial distribution of the
grade control nickel very well with only the low values on the eastern perimeter
under represented.
The cobalt grade control values seem to be predominately high in small
pockets across the southern half of the study region, along the western perimeter and
in the north-western comer. The entire south-western comer appears to have medium
to high levels of cobalt. There are low values along the northern perimeter of the
region as well as in the south-eastern corner. The exploration cobalt values also
reflect the spatial distribution of the grade contro1 Jata very well, although the low
values in the south-eastern comer and the high values in the north-western comer arc
not well represented.
The post plot of the grade control magnesium shows small pockets of high
values in the north-western corner, central eastern region, eastern perimeter, south
eastern and the south-western comers. The concentratians of low values are located
predominately across the centre of the study area with pockets along the western and
northern perimeters. The spatial distribution of the exploration data is also very
representative of the grade control data with the only exception being an under
representation of low values along the south south-eastern and south south-western
perimeters of the study region.
The post plot of the grade control iron indicates that there is a large area of
high values across the central southern region extending across toward the south
western corner and up along the western perimeter. There is also a small pocket of
high values in the north-western corner. The low values extend in a long band along
the northern perimeter of the region with a large area of low values in the south
eastern corner. Overall the spatial distribution of the grade control iron is very well
represented by the exploration data, the only exception being an under representation
of low values in the south-eastern comer.
The post plot of the grade control aluminium shows two large areas of high
concentrations in the south-eastern and south-western regions of the study area.
There is also a pocket of high values in the north-western corner. The low values are
predominately located across the western side of the study region; in addition there
are small pockets of low values in the south-eastern corner. The spatial distribution
of the exploration data is also very representative of the grade control data with the
44
only exception being an under representation of high in the north-western corner of
the study region.
The post plot of the grade control chromium shows clear cut areas of high
and low concentrations with the high values along the south-eastern perimeter
extending across the central southern region into the lower western side. The low
values extend across the entire northem boundary extending down along the east
south-eastem perimeter. The spatial distribution of the exploration data is in this case
not very representative of the grade control data. The large regions of high
concentrations appear to be greatly under represented across the entire study region.
The post plot of the grade control zinc shows hi~h concentrations across the
entire southem half of the study region and in the north-western comer. The low
values are located along !he northem and south-eastern perimeters. The spatial
distribution of the exploration data is very representative of the grade control data
with all high and low values well represented across the entire study region.
The post plot of the grade control manganese shows a large concentration of
high values occupying the entire south-western region, extending east across the
lower southern half up into the south-eastern corner. There is also a pocket of high
values in the north-western comer of the study region. Three regions of low values
are evident, on the central north-western side, across the centre of the study area and
across the central south-eastern region. Again, the spatial distribution of the
exploration data is very representative of the grade control data with all high and low
values well represented across the entire study region.
45
MMZ20GC ~ N'cktl
.
~
:::.:~ . ... ==· =!:·· :::·· '· :• . ._;:: ... ·-
r '~ "" "~ ,_ .. ~ uml:t
"""
I. :: --"' --·-
1~-·,,.._ ··-------~·-
MM220GC ~ Cobalt
~
MM220GC ~ion
MM220GC ~ Manginttt
·""• ..
r=:.•
Figure 3.9: Post plots ofl\1M22DGC data
.... .. r t.UQ ·-·-'"' ·~ ·-·-... ....
1
.. ttl ..
··~ =-~~· _.,. -·Q .. .....
1
.. _ .... ·--·"u'~
UI:IOonCO ~.~ .. , .. " , ...... . 1* }_ •.. ~
1
.. ---.. ........ ...., .. ...... ----"""" --
""'
46
MM22DE>O' MM220EX!' ~t-icktl
r !:i Cobilt
r ·~ • .. .. '"' "" • • ·- • ,..,
..~ , _ •m • • • • .. .. • • • • • "" '"" "'' • • • • • • • • IUIO • • • • • • • • .... • • ~- • • '"' ~ I § • g • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • ,., • • • • • • • • • • • • • •
'" . • • • • a • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • •
:i n n • ., .. ~ - " .. " " 590 10l5 1490 ... 1035 1460
MIIZ2cexP MMZZOEJO' a Magnesium
r !:i~ ••
r "~ ...... 3- ....
• • ~-»= ,.,. • • • • ~~
1'\ • .zl • • • • • .. ... -·~ • • • • • • • • • ...
" • • • • • . .... It ··- • • • • • ~~
~ • • ~ . • • • • • • • • • • • • • • ~ • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • ] . • • • • • • • • • • • • • • • • • • • • • • • • .. • • • • • • • • • • •
g !" • .. • • g - - • ~ . ... 1036 ....so. ..., 1015 -IIM220EXI' IIM220EXI'
oAlunnwn ~Chromium ~
r r • "~ • -.. ~- --• • n.,. • • ---~ ..... ... 41M.
I• tii"·OS • • • • • '1101) • • • • UO'I'ul't
~·~ lj4hoCI$
• • • • • • • • • <>~- • • • • • • • • • Ull .. •<'f • • "., • , ...... ., .... • ~ • n • • . • • • • • • • • • • • • • • • • • • • . • • • • • • • • • • • • • • • • • • • • • • • • • • • .. • • • . • • • • I) • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • ~ • • • ~ h • ~ .. ~ .. " • ..
590 1035 , .. 60 "" 10:!5 14!!0
MM22DeXP MM220El<l'
~Zine ~ Manganm
r r • --.... .... • • • '"".m
~~-........ ......
• • • • _ ..
I • • • • ---- ·~-gfa -- • • . I • • -~-• • .. • • • • • • • -~~- • • • • --• § • • • • • • • • • • • • • • • • • • • • • • • • • • • • • I I • • • • • • • • • • • • • • • • • • • • • • • • • • • •
I ~ • • • • • • • • • • I • • • • • • • • • • • • • • • • • • • • • • • • • • ft • • • • • • • • • • • • g " " g .. ~ II ,. ~
••• 1035 1480 '"' 10lS 1460
Figure 3.10: Post plots of:MM22DEXP data
47
3.4 Principal Component Analysis
In a principal component analysis one assumes that the relationships between
pairs of variables is linear, hence it is desirable that the scatterplots between variables
exhibit linearity and the correlation coefficients are significantly large. Figure 3.11
displays the scatterplots between each of the variables. There appear to be moderate
to strong linear relationships between nickel-cobalt, nickel-iron, nickel-chromium,
nickel-zinc, nickel-manganese, cobalt-zinc, cobalt-manganese, iron-chromium, iron
zinc and zinc-chromium.
Table 3.5 displays the lower half of the correlation matrix and corresponding
p-values (in parenthesis) for the exploration data. The use of a two tailed hypothesis
test
Ho:p;:::Q versus H1:p:f.O
(where pis the correlation between a pair of variables) at a significance level of 0.05,
indicates that all pairs of variables, except for magnesium-chromium and
magnesium-zinc, are linearly related. However as a principal compoGcnt analysis
does not require all pairs of variables to have a strong linear correlation, rne.~ely that
there be some significant linear correlations, we conclude that the explora'.jon data
meet the assumption of linearity for a principal component analysis.
Table 3.5: Lower half of the correlation matrix for exploration data, with p-values in
parenthesis
Nl
co
NI CO
1 (0.00)
0.72 (0.00) 1 (0.00)
MG
MG 0.36 (0.00) 0.23 (0.01) 1 (0.00)
FE
FE 0.64 (O.IJO) 0.58 (0.00) 0.18 (0.04) (0.0C)
AL
AL 0.30 (0.00) 0.34 (0.00) 0.38 (0.00) 0.41 (0.00) 1 (0.00)
CR
CR 0.52 (0.00) 0.42 (0.00) 0.13 (0.16) 0.74 (0.00) 0.31 (0.00) I (0.00)
ZN
ZN 0.64 (0.00) 0.60 (0.00) 0.11 (0.21) 0.83 (0.00) 0.28 (0.00) 0.65 (0.00) I (0.00)
MN
MN 0.62 (0.00) 0.84 (0.00) 0.27 (0.00) 0.46 (0.00) 0.23 (0.01) 0.20 (0.03) 0.49 (0.00) (0.00)
48
~ • .
~ • • • •
32.12 ,.= ... . It:' . ... . liM:~·· ;ria; :·. Nl . ,: .. ~ ... · ,)Aii·. ..... • •• 11.08 • • • • •
~ • . • • • . . 2.23 - '" ,,
A: JiA .. ~~·· ... ,.: . . ()Q
;#;' c co ... ;.: ;.#.:Jw d - • •• • 0.76 • • ' • • "" • ..... . • • . . • • ..... 286.22 - • •• • • • • • . . .. .. { .. ;,: ... MG ) ... . :: .... • ,.6 • . ·-·~ . ·':-C/:l 100.79 --= J/lllrJ:" •
~ ., .. "'~ ... ·'·· 0 - • • -~ .. · ~ • -xn: • • •
" • • • • • • • .a 586.99 -~· ~~r -· :It' . ~·· 7~· 0 ;.• • • . . . FE .... ~, ... .. .: .. • :. .. .
~ 203.00 -"' • 0 • "" • • .. . • .. •• • • ~ ::r 110.15 -
J~'T.; .
tr~~·· • • • .; lit: .. . .I.' ..&!.. • • lr." ._ • • " . AL
~~·· ;9/tl!f:· •.. -:C' : . ~ , ... • •
37.65 - -t •• •• • • • \ • • . . .. • • • N 163955 - $:. $1·,· ,. ~ Jf.'f • . .~!!·~ • If" 0 • • . .
• .. CR ~ .. ~ 56064 - •• " .. II•
• • • N • • •
"tt ••
~ ~· . .. . . •
~: < 3710.58 - j#r ,. <I> •••• .. .
§. • • :1' •• •• ~·:" _.I ••: • ~:· ZN g. 1295.53 - • • - .. .. • " "' . . . . . • .
177889 - • • • • • • .. • • .·. . • • • • • • •
.-iii' #:·· •= jj;: .. 'iitff,.'l.. . : J.~ MN 60062 - ·:, • ..... . .... • ... . ., . • :11 • •
l l I I I I I I I I I I I I I I
(§' \'I- ()~I) <z-'1-'b \ <:P ~'!!'I-'If" 'J-'~- '1-<:Jb "'\'If" 9>9! '!>1~\\()~"' <f:l'i}< rf>"' .;'!> .;'0 <:P"'I- 'If"'!! \ \· '!>'~-' "' ,<SO \ 'fl"'' '!>1\ (). " ,11
""" \0
3.4.1 Principal Component Analysis of l\1M22DEXP
The results of the principal component analysis perfonned on the correlation
matrix of all eight variables are shown in Table 3.6. The principal components
extracted 53.7%, 14.8%, 13.1%, 7.6%, 4.1%, 3.5%, 1.9% and 1.4% of the total
variance.
Table 3.6: Eigenanalysis of the correlation matrix of :M:M22DEXP
Eigenvalue 4.29 1.18 1.05 0.61 0.33 0.28 0.15 0.11
Proportion 0.54 0.15 0.13 0.08 0.04 0.04 0.02 0.01
Cumulative 0.54 0.69 0.82 0.90 0.94 0.98 0.99 1.00
Variable PC! PC2 PC3 PC4 PC5 PC6 PC7 PCB
NI -0.41 0.08 0.13 -0.28 -0.40 -0.74 0.10 0.12
co -0.41 0.15 0.37 0.20 -0.26 0.23 -0.21 -0.69
MG -0.18 0.65 -0.37 -0.57 0.19 0.20 -0.06 -0.10
FE -0.42 -0.29 -0.17 -0.00 0.34 0.04 0.73 -0.25
AL -0.24 0.30 -0.57 0.70 -0.06 -0.15 -0.07 0.11
CR -0.34 -0.42 -0.33 -0.20 -0.51 0.46 -0.13 0.26
ZN -0.40 -0.32 0.01 -0.06 0.58 -0.17 -0.60 0.07
MN -0.35 0.31 0.51 0.16 0.13 0.31 0.17 0.60
As the variance of each standardised variable is equal to one, we seek
components whose corresponding eigenvalues are greater than one. This ensures that
the explanatory value of the component exceeds that of any single variable. The first
three components have eigenvalues greater than one and account for 81.5% of the
variability of the original data. However, consideration of the scree plot in Figure
3.12 reveals that the eigenvalues appear to level off after the fifth component. As the
associated eigenvalue for the fifth component is only 0.33 it would be of no benefit
to retain this component. If data reduction were the objective here we would be
required to retain the first four principal components of this data set. The decision of
50
how many components to retain in order to account for "enough" of the variability of
the original data is a subjective and strongly debated area (Afifi & Clarke, 1996). In
general however one would aim to retain as many components as required to account
for at least 85% of the variability. However some authors consider that as little as
80% (Johnson & Wichern, 2002) is adequate, whereas others debate that at least 90%
of the variability should be accounted for (Dunteman, 1984). In this case the first
four componevts would then account for 89% of the variability of the original data
which would certainly be "enough" for even the most stringent authors.
'
l
Component Number
Figure 3.12: Scree plot of eigenvalues for the principal component analysis of
MM22DEXP
The first principal component has the greatest weighling on nickel, cobalt,
iron and zinc with manganese and chromium following. The lowest weights are
associated with aluminium and magnesium. While all the weights are negative there
is little to be interprei:ed from this component. The second principal component is
dominated by magnesium and has positive weights for this variable as well as
manganese, aluminium, cobalt and nickel. The weight of nickel for this component is
only 0.08, so this variable barely contributes to this component. The negative
weights, in order of magnitude, are associated with chromium, zinc, and iron. The
51
third principal component is composed of positively weighted manganese, cobalt,
nickel and zinc. The contribution of zinc to this principal component is negligible
with a weight of only 0.01. The negative weights are associated with aluminium,
magnesium, chromium and iron. The fourth principal component is dominated by
aluminium, whose weight is positive as are the weights for cobalt and manganese.
The negative contributors are magnesium, nickel, chromium, zinc and iron, with the
latter two being almost zero. There does not seem to br any clear interpretation for
any of the principal components from this data set.
Figure 3.13 displays the scatter plots of the scores of the first four principal
components. The plots between PC1-PC2, PC2-PC3 and PC3-PC4 reveal potential
outliers. The corresponding sample data was checked and it was ascertained that
these data were genuine, hence they could not be treated as erroneous and were left
in the data set.
There does appear to be some slight curvature in some of the scatter plots, in
particular those between PCSI, PCS2 and PCS3, suggesting that the strongly skewed
variables cobalt, magner.ium, aluminium and manganese, may benefit from a
transformation. These variables were transformed logarithmically and the principal
component analysis repeated. The results were compared to those obtained in the
principal component analysis of the raw data and are displayed in Appendix A. The
eigenvalues in each are case very close in value with the cumulative propmtion of
variability almost identical by the fourth component. The eigenvalues of the first and
second principal components are somewhat higher for the transformed data while the
third eigenvalue is somewhat lower. There have been numerous small changes in
some of the elements of the coefficient vectors, though they are all small.
52
2.24
-2.64
1.89
-1.72
1.34
-1.51
1.70
-0.67
j -
-
-
-
-
PCS1
• A. . .. •
• .. ~ ~ ~·fit" ,., .I ••• • • • •• .. ····~~ . . . .. •• •
• ~:·. • • :to"
PCS2
. ., -If. ... • .., ..
• • •
• ~ •• • • -,- -,-
.. ~~~ ~-"' .. • ·-t·• • • ... ,,. ., . • • .. ... -.... ··~··· •• • • ,. , • • •
• • ... PCS3 : , . . - • • • •
·~ • • • • • •••
PCS4
-,-
Figure 3.13: Scatter plots of the principal component scores of MM22DEXP
Furthermore, while the scatter plots of the transformed principal component
scores (also displayed in Appendix A) do appear to be less correlated and circular in
shape, the improvement is not substantial. Given the results of the two principal
component analyses are very similar, the curvature in some of the plots in Figure
3.13 is not sufficiently marked and the dangers of exaggeration of estimation errors
we feel that the data do not warrant transformation.
The principal component scores were calculated usmg expression ( 4) in
Chapter 2. The omnidirectional cross correlograms of the scores were then computed
at 32 lags with a lag spacing of 12.5 metres. The value of the cross correlogram at
the first lag is significantly different from zero for PC1-PC2, PC1-PC4 and PC2-
PC4. In fact there is clear structure in the lower lags for PC1-PC2 and PC1-PC4.
There is also strong deviation from zero at the ninth and twenty-sixth lags of PC 1-
PC3. While the cross correlograms for PC2-PC3 and PC3-PC4 show no clear
structure there are significant non-zero values at lags two and four. Figure 3.14
displays the most problematic of the cross correlograms. Clearly the original data are
not intrinsically correlated.
53
p(O)-p (lhl)
0.2
0.1
0
-0.1
-0.2
-0.3
-0.4
0 50 100 150 200 h
250 300 350 400
Figure 3.14: Cross correlogram for the first two principal components of
MM22DEXP
In order to demonstrate the multivariate geostatistical techniques of cokriging
and principal component kriging it was decided that we would investigate the
possibility of creating two smaller subsets of MM22DEXP. Several subsets were
investigated according to the following criteria. As the data come from a working
nickel-cobalt mine these two variables were to be included in both subsets. We also
required that one of the subsets consisted of only variables that are highly correlated
with nickel and cobalt. We have chosen a highly correlated data set for two reasons.
Firstly, as previously stated, a principal component analysis assumes that the
relationship between each pair of variables is linear; hence high correlation
coefficients between variables are desirable. Secondly, the higher the correlation
coefficient between the primary and secondary variables, the greater the contribution
of the secondary information to the estimation of the primary variable. More
precisely the higher the correlation coefficient, the larger will be the kriging weights
of the secondary variables. The highly correlated data set is MM22DHC4 and
consists of nickel, cobalt, iron and zinc. The choice of the second subset was based
on the fact that the data are from an actual working nickel-cobalt mine; hence we
considered it appropriate to include the variables considered to be economically most
important to the mining company. This subset is referred to as MM22DTOP4 and
consists of nickel, cobalt, magnesium and iron.
54
3.4.2 Principal Component Analysis of MM22DHC4
The results of the principal component analysis performed on the correlation
matrix of nickel, cobalt, iron and zinc are shown in Table 3.7. The principal
components extracted 75%, 14%, 7% and 4% of the total variance. Only the first
component has eigenvalue greater than one and accounts for 75% of the variability of
the original data. The first two components account for an impressive 89% of the
variability of the original data. The scree plot displayed in Figure 3.15 shows that the
principal components level off at the third and fourth components. If data reduction
were the objective here we would be required to retain the first two principal
components of this data set. These would then account for 89% of the variability of
the original data.
Table 3.7: Eigenanalysis of the correlation matrix of nickel, cobalt, iron and zinc
from MM22DHC4
Eigenvalue 3.00 0.55 0.27 0.17
Proportion 0.75 0.14 O.D7 0.04
Cumulative 0.75 0.89 0.96 1.000
Variable PC! PC2 PC3 PC4
NI -0.50 -0.42 0.76 0.04
co -0.48 -0.60 -0.64 -0.06
FE -0.51 0.50 -0.02 -0.69
ZN -0.51 0.47 -0.12 -0.71
55
2 3 4
Component Number
Figure 3.15: Scree plot of eigenvalues for the principal component analysis of
MM22HC4
The first principal component has almost equal weighting on nickel, cobalt,
iron and zinc. As the weights are all the same sign and of almost equal magnitude it
is possible to consider this component to be an overall grade component. The second
principal component has its largest negative weight associated with cobalt followed
by nickel. Zinc and iron are both positive and are fair! y equal in magnitude. There is
a clear grouping of nickel-cobalt and iron-zinc. While the actual sign of the
coefficients is not relevant, the fact that the groupings display opposite signs is. This
suggests that this component is a contrast of nickel-cobalt and iron-zinc. The third
principal component is dominated by nickel which in this case is positive. Cobalt,
iron and zinc have negative weights however iron barely contributes to this
component with a weight of only -0.02 and the weighting on zinc is also small.
Hence, as the nickel and cobalt weights are of opposite sign this component appears
to be a nickel-cobalt contrast component. The fourth principal component is
dominated by negative weights on zinc and iron with cobalt and nickel barely
contributing. As the weights are the same sign and of similar magnitude this
component appears to be a zinc-iron component. Further interpretation of these
principal components requires knowledge of their geological significance. This is
56
beyond the scope of this research and would require the consultation of a geologist
with the relevant expertise.
The scatter plots of the principal components are displayed in Figure 3.16. As
outlined earlier it has been decided that the data do not warrant transformation. There
is some curvature in the plot of PC1-PC2 however the remainder are sufficiently
uncorrelated. There do appear to be some outlying values, particularly in the plots
including PC4. The corresponding sample data was checked and it was ascertained
that these data were genuine; hence they could not be treated as erroneous and were
left in the data set.
Figure 3.17 shows the cross correlograms of the first three principal
components from MM22DHC4. Despite an initial high value at the first lag of the
cross correlogram of the first and second principal components the remainder of the
lags of this and the other graphs appear to be sufficiently close to zero. In addition
there is little to no structure in the first plot and clearly no structure in the subsequent
plots. The cross correlograms then give an indication that it would be possible to use
an intrinsic coregionalisation model for the variables in the MM22DHC4 data set.
1.86
-2.25
1.73
-0.80
1.71
-0.32
0.77
-1.06
i PCS1
• • -
:..~il-- •• • • - •• ... - . ..
• •••• •
- .'J ... •
'
• • !of • ·' • +. • • • •
PCS2
•
.~(: . • • •
•• . ,.. .... • • •
••• • -· ... J •• \. I • • • •
~ • • • ·~
• • • • • • • • •
PCS3 ·-· • • . . .:. • • .. ., • PCS4
• •
'
Figure 3.16: Scatter plots of the principal components from MM22DHC4
57
p(O}-p(lhD
-0.2
-0.3
p(O)-p (lfll) 0.4
0.3
0.2
0.1
-0.1
-0.2
-0.3
·0.4L...~-~~-~~-~~~~ -0.4 L....~-~~-~~-~~~--'-.1
0 • 100 1m - - - - - 0 • 100 1m - - - - -I~
PC1-PC2
p(O)-p ClhD 0.4
-0.1
-0.2
-0.3
lhl
PC1-PC3
-0.4'--~-~~-~~-~-~~
0 • 100 1m - - - - -~I
PC2-PC3
Figure 3.17: Cross correlograms of the first three principal components from
MM22DHC4
3.4.3 Principal Component Analysis of MM22DTOP4
The results of the principal component analysis performed on the correlation
matrix of nickel, cobalt, iron and magnesium are shown in Table 3.8. The principal
components extracted 61%, 22%, 11 %and 6% of the total variance. Only the first
component has eigenvalue greater than one and accounts for 61% of the variability of
the original data. The first two components account for 83% of the variability of the
original data. The scree plot displayed in Figure 3.18 shows that the principal
components do not level off at all. If data reduction were the objective here we
would be required to retain the first three principal components of this data set. These
would then account for 94% of the variability of the original data, however one
would question the utility of reducing a four variable data set to a three variable set.
58
Table 3.8: Eigenanalysis of the correlation matrix of nickel, cobalt, magnesium and
iron from tv1M22DTOP4
Eigenvalue 2.43 0.88 0.43 0.26
Proportion 0.61 0.22 0.11 0.06
Cumulative 0.61 0.83 0.94 1.00
Variable PC! PC2 PC3 PC4
NI -0.58 -0.03 -0.20 0.79
co -0.55 -0.20 -0.58 -0.56
MG -0.30 0.94 0.11 -0.16
FE -0.52 -0.29 0.78 -0.20
2.0
~ 1.5
l iii 1.0
0.5
2 3 4
Co1t1Jonent f\lurrber
Figure 3.18: Scree plot of eigenvalues for the principal component analysis of
MM22DTOP4
59
The first principal component has almost equal weighting on nickel, cobalt
and iron. While the weights are all the same sign magnesium is somewhat smaller in
magnitude than the others. Hence this component is not as easily interpreted as for
the previous case, although it could still be thought of as an overall grade component.
The second principal component is almost totally dominated by magnesium. Nickel
bare! y contributes to this component whereas cobalt and iron have similar weighting.
Again this component is not as easily interpreted, other than perhaps being a
magnesium component. The third principal component is dominated by iron. There is
a positive grouping of magnesium and iron, though the weights are by no means
close in value. Nickel and cobalt are both negative for this component, once again
the weighting is not equal. However as the weightings on nickel and magnesium are
quite small we could consider this a contrast component of cobalt and iron. The
fourth principal component is dominated by nickel with all the other weights being
negative. Magnesium and iron have fairly equal weighting but are relatively small in
magnitude. Hence this could be considered a nickel-cobalt contrast component. Once
again further interpretation of these principal components requires knowledge of
their geological significance and is beyond the scope of this research.
1.40 i -2.51 i PCS1 #· ...... ::~ ... ... ~s:
• • + •• I • • • • ..___.:._ __ ____, •• 1
• +\ • • • •
,---------,
2.79
-0.10 -
1.55-
-0.54
1.70-
-0.44-
• • •
:· ·;.,.,. • • 4 • :J.. PCS2 •• j v..~ .,· ••• • •
• • "··.-------, P==.~.~===l • • •
• .. ~fliL- k..~.: • • PCS3 • ~·.11 • ... ~~ ;P.! ~ + .• + ~~~:~·======~.---.---~L~·~·---~
.,....., ··.•:t .. ~.
• ' ' -·". • • •• .. , .
• '
PCS4
Figure 3.19: Scatter plots of the principal components from MM22DTOP4
60
The scatter plots of the principal components are displayed in Figure 3.19. As
outlined earlier it has been decided that the data do not warrant transformation. As
for the previous case there is some curvature in the plot of PC1-PC2 however the
remainder are sufficiently uncorrelated. There do appear to be some outlying values,
particularly in the plots including PC3 and PC4. The corresponding sample data was
checked and it was ascertained that these data were genuine; hence they could not be
treated as erroneous and were left in the data set.
Figure 3.20 shows the cross correlogram of the first two principal
components from MM22DTOP4. The first two lags have significantly non-zero
correlogram values. In addition there is definite structure at the lower lags which
indicates that the data are not intrinsically correlated. For this data set we will be
required to proceed with the linear model of coregionalisation and perform the
estimation using cokriging.
p(O)-p {lhl) 0.2
0.1
0
-0.1
-0.2
-0.3 L_ _ _j_ __ _j__ _ __l_ __ _j_ _ __l __ _l_ __ L___ _ ___j__~
0 50 100 150 200 h
250 300 350 400
Figure 3.20: Cross correlograms of the first two principal components from
MM22DTOP4
61
4 ORDINARY COKRIGING
In this chapter we demonstrate the estimation technique of ordinal"'; cokriging
using two data sets, Mlv122DHC4 and MM22DTOP4. To begin with v. e discuss the
variography of the MM22DHC4 and :tvnv122DTOP4 data sets. '.!.he former was
modelled using an intlinsic coregionalisation model; the latter was fitted with a linear
model of coregionalisation. All direct and cross semivariograms were calculated with
32 lags at a lag spacing of 12.5 metres and a lag tolerance of 6.25 metres. The
directional semivariograms were calculated with an angular tolerance of 45°. All
directional models were assessed for fit in the intermediate directions (with angular
tolerance set to 22.5°).
All models were cross validated using ordinary cokriging and performed
satisfactorily. The cokriging was perfonned using the software package ISATIS.
This package treats each variable nominated for estimation as the primary variable in
turn, using the other variables as secondary variables. 1717 estimates were obtainr:d
for each variable directly at the locations of the grade control data. The parameter
files for all kriging procedures are displayed in Appendix C. Comparisons of the
estimates obtained from the cokriging of each data set were made with the
MM22DGC grade control data.
4.1 V ARIOGRAPHY AND COKRIGING OF MM22DHC4
4.1.1 Variography and the Intrinsic Coregionaiisation Model
Recall that in the generic intrinsic correlation model the sills,b~, of any basic
structure g1(h) that constitute the model are proportional to each other. That is
b~ = fAj · b~ for all i,j and 1. As we are dealing with standardised data we require that
' Lb' = 1. Under the assumption of second-order stationarity the matrix of /•0
coefficients 4l = [!pij] is the positive semi-definite variance-covariance matrix C(O),
which in our case is the correlation matrix:
62
1 0.72 0.64 0.64
0.72 1 0.58 0.60 R= =<1>
0.64 0.58 1 0.83
0.64 0.60 0.83 1
The direct and cross semivariogram surfaces for nickel, cobalt, iron and zinc
are displayed in Figure 4.1. The semivariogram surface of nickel exhibits the
presence of two anisotropic structures, one short range and one long range. The
direction of maximum continuity of the short range structure is east-west (azimuth
90°). The direction of maximum continuity of the long range structure is
approximately azimuth 70°. Cobalt, iron and zinc also exhibit anisotropy with the
direction of maximum continuity being east-west. The east-west anisotropy is also
apparent in the cross semivariogram surfaces. Hence the semivariograms were
modelled with the direction of maximum continuity being east-west.
The semivariograms were fitted with an intrinsic correlation model consisting
of three basic structures, a nugget effect and two spherical structures. The first
spherical structure is omnidirectional with a range of 100 metres; the second is
modelled with east-west as the major direction with a range of 200 metres and an
anisotropy ratio of 0.5. This model can be expressed as:
1
0.72
0.64
0.72 0.64
1 0.58
0.58 1
0.64
0.60
0.83
0.64 0.60 0.83 1
where h'= = [1 0 ][ cos90' sin90'][h·] [ 0 1][1!·] 0 0.5 -sin90° cos90° hY -D.5 0 hY
Figures 4.2 and 4.3 display the direct and cross semivariograms fitted with this
model.
63
Nickel Cobalt y(h) y(h)
o= D "" .... ,.., r • ... ... . .. ._. .... ...
.... ·"'
0 .. .. Iron Zinc
y(h) Ylfll
0" 0" ~·
... .. .... ;r- •
'·' O. M .. ... o.o . ..
.... - 0 .. . .. ... ., .. ... . .. "' ,.,
~ .., .. ..
Nickel-Cobalt Nickel-Iron y(n) 'f(h)
u· 0" •n ... ... ...
,?- • r ... "' ... . ... " .
_..,
.. .. Nickel-Zinc Cobalt-Zinc
'I (n) 'I (h)
G" o· ... u '·'· .r • r .. uo
0 0 .,, . 0
..,. .... .... .. 0 - ... ... ... ... .. .. Iron-Zinc Cobalt-Iron
""' y(h) y(h)
0" u ... .... .... . ..
,?- • z-.. .., .,. .. .
Figure 4 .1: Direct and cross variogram surfaces for MM22DHC4
64
Nickel Cobalt YIP>D Direction 0 YIP'D Direction 0
2.1 2.4
1.8 2.1
1.5 1.8 .. 1.5 .. 1.2 .. --~ 1.2
0.9 .. ... . 0.9 ~ 0.8 . . . . . .. .. ..
0.8 0.3 0.3
50 100 150 200 250 300 "' 400 50 100 150 200 250 300 "' 400 1>1
~· YIP>Il Direction 90 YIP'D Direction 90
2.1 2.4
1.8 2.1 1.8
1.5 1.5 . .. •
. ... • -. . 50 100 150 200 250 300 350 400 50 100 150 200 250 300 "' 400
1>1 ~·
Zinc Iron YIP>Il Direction 0 y <P>D Direction 0
2.4 2.4
2.1 2.1
1.8 1.8
1.5 1.5
1.2 .. .. . • • . . . . . . . .. . . . . . -- -- . . ...
50 100 150 200 250 300 350 400 50 100 150 200 250 300 "' 400
~· ~· YQhD Direction 90 YIP'D Direction 90
2.4 2.4
2.1 2.1
1.8 1.8
1.5 . .. . ... .. . .. .. 50 100 150 200 250 300 "' 400 50 100 150 200 250 300 "' 400
1>1 ~·
Figure 4.2: Fitted direct semivariograms for MM22DHC4
65
YlltO
'' '0 ,., '' ' . .
. .
Nickel-Cobalt
. . Direction 90
. . . . 100 150 200 250 300 350 400
I~
'0
'' ''
Direction 0
'~~b~ 1 ••• 0·' c"'r"~·"·~----~·s .. ~c:-c g:.: ::: ••• : . .. . . " .
0 ' ~
'~
'·' ,., '' 0.0
0.0
0.,
0 50 100 150 200 250 000 ow ~
I~
Nickel-Zinc
. . ..
Direction 90
..... 0~-=--~~~~~=--=~~~~ o • a 'w ""' • ooo "" •
, .. '' ,.0
'' 0.0
'~ '0 ,., '·' ' O.o 0.0 OA 0.,
'~ LO , .. " ' o.• 0.0 OA 0., 0
.. "'
--w
...
I"
. .
Direction 0
. . . . . . 100 150 200 250 300 350 400
I~
Cobalt -Zinc Direction 0
. . . . .. ... '"" 'w ""' 'w ""' ow ""' I~
Direction 90
. · ... 100 150 200 250 300 350 400
"
'~ " LO
" '' 0.0
0
, .. '' '' '' '' 0.0 0.0
...
Nickel-Iron Direction 0
.. ...... ..
Direction 90
.. .. 0.,
0 ----------------~~~~-~--~-
Cobalt-Iron YlltO
'' '' '' 0.0 . . . . .
Direction 0
... . . .
"' '"" "" ""' 'w •oo •w ""'
Yl!h()
'' '' ,.,
I.
Direction 90
0.9 •••
O.o 1--__,~~--"'.-'--''-~~------0.0 ...
0 ,._•.!---------------- .. ---·--~--O.s o'--".--,"oo:--:-,w~-~~. co-c,",c---cooo~-c,.,~---'-'•
YOlO
" ,., ,., '' 0.0
0.0 ...
I"
Iron-Zinc
. ... .. Direction 0
. . .... .. . o_~ -·~--------- ---------------
-0.3 0'--:,:--:,:00;:--o,.:o-o-~-:;;--:,o.,;:---;=~-o,.;;;--.,~o· I~
YOlO
'' '' ''
Direction 90
1.2 •• • • •• • •• ~:: r--., 3'---'-__,.~'--c.-'-'"-c:-~---~ 0.3 • • • •
0 --------------------------
-O.s o'--",---,,"oo:--:-"~o-ooo=-c,ce.,:---cooo~-'c""~-.,c'c'o· I"
Figure 4.3: Fitted cross semivariograms for MM22DHC4
66
4.1.2 Cross Validation for Cokriging of Ml\U2DHC4
/One method of assessing the suit~bility of a model is to use cross validation.
This technique allows us to compare estimates calculated from our chosen model
using the specified kriging algorithm with the actual data at each sample location.
The cross validation procedure removes one sample data from the data set, estimates
it using the specified model, replaces it then repeats the procedure for all other
sample data in tum. An analysis of the residuals is then carried out to check for
nonna1ity of the errors, presence of Outliers and suitability of the neighbourhood
parameters.
The cross validation was carried out for all of the variables in this data set
using ordinary cokriging to perfonn the estimation. We present here the results for
nickel, those for cobalt, iron and zinc are presented in Appendix B. There was little
difference in the cross validation results of several various search neighbourhoods
investigated. Figure 4.5 shows the cross validation residual analysis for the search
parameters displayed in Table 4.1 and Figure 4.4. The histogram of residuals appears
roughly nonnally distributed with a few extreme high and low values apparent.
These extreme values are also apparent in both the plots of the true values versus the
fits and the standardised estimates. The location of these values is exhibited in the
base map. These values were veritied and as no justification could be made for their
removal, they were left in the data set. The plot of estimates versus the true values is
reasonably well clustered around the 4SO bisector, once again with some deviations.
Finally the mean error and standardised error variance displayed in Table 4.2 are
sufficiently close to the optimal values of zero and one respectively (Bleines et. al,
2001, p. 126). All other neighbourhood search parameters tested failed to further
improve any of these results. Overall the model appears to have reproduced the
sample (exploration) values successfully for all variables.
67
H (])
.j.) (]) ;.:
><
Error
St. Error
Table 4.1: Neighbourhood search parameters
Number of angular sectors 4
Minimum number of samples 4
Optimal number of samples per sector 4
Search radius (major axis of ellipse) (metres) 220
500.
400.
300.
200.
100.
Search radius (minor axis of ellipse) (metres)
Test Window X (Meter)
600.700.800.900l0001100120013001400.
X
X X
X X X X
X X X X X
600.700.800.900l0001100120013001400.
X (Meter)
120
500.
400.
300.
200.
100.
Figure 4.4: Search neighbourhood for MM22DHC4
Table 4.2: Mean and variance of cross validation residuals
Nickel Cobalt Iron
Mean Variance Mean Variance Mean Variance Mean
0.02 1.00 O.D2 1.24 0.01 0.66 0.02
0.01 2.39 0.02 2.57 0.01 1.55 0.02
>-<:
:;:: I]) rt I])
"
Zinc
Variance
0.64
1.50
68
500.
400.
300.
200.
100.
0.15
• • ·o u 0.10 " • il • u •
0. 05
0.00
Cross Validation
X (Meter)
600 700.800 900 ~0001100120013001400
+ + + I + +-t r++•+ ' -t l-t+ 11-t++l>l++lltlh't-i •+ + f t I + f I + + + -~.-+ + !' ~ • 0 t f I ++++++t++++ l-+ ++tll-tl-f-tl++!l +-t
·1-+ + + +t 600.700 800.900 ~0001100120013001400.
X (Meter)
{Z*-Z) IS*
(Z*-Z) /S*
4.
3.
2.
500. 1.
400. 0.
300. -1.
200.
-2. 100.
5. 0.15
" "' ow 0.10 ~ ~ . " 0.
' ' u • e-" . -•
0. 05
-5.
0. 00
Z* : STNI (Estimates)
-2. -1. 0. 1. 2. 3.
•
++ + •
•
-2. -1. 0. 1. 2. 3.
Z* : STNI (Estimates)
Z* : STNI (Estimates)
-1 0
• 1- I I i+ I +
... -i"i'A:i':lt>~~"'~ t ' I •1- ,:'.11-'! ~ '"f + + +'t'lt,_l.l"'N+ +
+f_ ''tt-t+ + +4" I ·t+- .:,·+ + . . . +
•
• • -1. 0. 1. 2.
Z* STNI (Estimates)
4.
4.
• •
Figure 4.5: Cross validation residual analysis for MM22DHC4 nickel
4.
3.
2.
1.
0.
-1.
-2.
5.
0 •
-5.
69
4.1.3 Cokriging Estimates of MM22DHC4
Ordinary cokriging of the nickel, cobalt, iron and zinc variables was
performed directly at the grade control coordinates using the search neighbourhood
parameters presented earlier (Table 4.1). As all vatiography, modelling and
estimation was performed on the standardised data it was necessary to transform the
estimates by multiplying eaGh value by the sample standard devir~tion and addir,g to
this the sample mean. Figure 4.6 shows post plots of the estimates for each of the ' ·variables along with the post plots of the grade control data. The smoothing nature of
the cokriging algorithm is clearly apparent here. The variability of the grade control
data has not been reproduced particularly well though o· 'era\1 the regions of high and
low values have been well represented. The lack of precise representation of the full
variability of the exhaustive data is however a drawback of any regression estimation
technique.
For all four variables the overall appearance of the post plots of the estimates
is consistent with the grade control post plots. In particular the estimates along the
north-western, western and southern perimeters arc representative of the grade
control values. The values along the eastern perimeter of t1,Ic study region have been
particularly poorly estimated. However most of these values lie outside the region
defined by the locations of the exploration data hence have been estimated using
values that are in some cases at least 75 metres away. Nickel seems to have been the
most affected by this as the last line of exploration drill hole!: consists of high values
only. Zinc is the least affected with pockets of high and low values still well
f4!presented in this region.
70
Nlcktl ~ Cradt Co~ol
... 1480
. . . . . . .. ~· ~ · .. . :~~:~~ j ::\1~1::
1
.. ··~ ·~ ·~ ··""
I" ~-·-·~-...... ~·,,._ ~·-
Cobalt Estimates
~riC=M~·M~~~~~---------------------------------,
• I'
1
.. '""' "~
tt.ato
-~·
1
.. 0.110 ·"" ·~ .... ·.... ·-
1~-'*·* ·-:·~,., .. ,. .... ··~""'
Figure 4.6: Cokriged estimates - Intrinsic Coregionalisation Model MM22DHC4
71
The residual post plots displayed in Figure 4.7 reflect the poor estimation
along the eastern perimeter with some of the highest residuals occurring here. In
general the area of greatest underestimation was the far north-western corner of the
study region. Iron appears to have been affected the most in this area with the
greatest underestimation occurring at the far north-western perimeter and the greatest
overestimation occurring just to the east of this region. In most cases the largest
errors were associated with the areas containing extreme high or low values. Figure
4.8 displays the normal probability plots and histograms of the residuals for nickel,
cobalt, iron and zinc. The histograms of residuals for nickel, iron and zinc are
reasonably symmetric and appear to be approximately normally distributed which is
supported by the linear shape of the normal score plots. The histogram of cobalt
residuals however is negatively skewed. Despite one clear outlying residual, the
normal plot shows an approximately· linear shape that is consistent with a normal
distribution.
Rtslduab ~Nt:kti· ICM
r _,,.. .,., ~ ...
~!..o
·= 11«0
~=
1
--~- ' ~-.. ·~"" ~.... --
Residual a
~ Cobalt ·~CM
~
Residuals
~ Zinc · ICM
Figure 4.7: Residual post plots for MM22DHC4
..
r 4 -_, .. 4 ... _,,.. . .., .. ... "'"
14~--UlOOOO ·1~.w\l
•U).. , _ 09= ·=--
72
250 20
200
"' '" 10 u !50 0: " " .,
" 100 ... 0 .,.. " " .t 00:
-10 50
0 -20 0
Residual Nickel Normal Score
2 350 0
300 I
"' 250
'" 0 u 0: 200 " -I "
., " !50 ... .,..
" -2 " 100 00: .t 50 -3
0 -4 • -4
Residual Cobalt Normal Score
250 500 • 200
"' u !50 '" 0:
" " " ., 0 .,.. 100 ~
"' " .t 50 00:
0 -500 0
5 0 Residual Iron N annal Score
300 3000 0
"' 200 2000
u 1000 0: '" " " 0 "
., .,.. 100 ~ -1000 " " .t 00: -2000
0 -3000 -4000
00
3000 Residual Zinc Normal Score
Figure 4.8: Residual histograms and normal probability plots for MM22DHC4
73
4.2 VARIOGRAPHY AND COKRIGING OF MM22DTOP4
4.2.1 Variography and the Linear Model of Coregionalisation
As the variables in the MM22DTOP4 data set arc not intrinsically correlated
it is necessary for us to jointly model the direct and cross semivariograms and build a
linear model of corcgionalisation. Goovaerts (1997, p. 114) identifies some practical
guidelines for selecting the basic structures in a linear model of coregionalisation.
These include that every basic structure incorporated in the cross semi varia gram
model 11_-fh) must be present in both direct semivariogram models }1;(h) and Jt(h).
Furthermore if a basic structure is not present on a direct semivariogram model it
cannot be incorporated on any cross semivariogram model involving this variable.
However it is not necessary for a cross semivariogr:im model Yu(h) to include every
structure that appears in both direct semivariogram models J1;(h) and i1J(h).
The procedure used for constructing our linear model of coregionalisation
was to model the direct semivariograms for each of the five variables then use these
basic structures to obtain the best possible fit for the cross semivariograms. In
addition to the direct and cross variograrn surfaces for nickel, cobalt and iron
displayed previously in Figure 4.1 those for magnesium are displayed in Figure 4.9.
The direct surface for magnesium is isotropic as is the surface for iron-magnesium.
i'-li£:\,·1·rr,agnesium and cobalt·magnesiurn are anisotropic with the direction of
maxhP·Fo continuity being east-west.
74
Magnesium y(h)
0: ... ... ... 0
....
Cobalt-Magnesium
Nickel-Magnesium
r •
Iron-Magnesium ~------ -----
., (h)
0·~ ., . .. . ... . .. ~·
y (h) y(h) ,,.
u on .. .r • :: r
~ .. . .. -tlD
0 .. ..... ... . .. . ... ~·
.. - ... ... ...
Figure 4.9: Magnesium direct and cross semivariogram surfaces
Having obtained a suitable fit on all the direct and cross semivariograms it
was necessary to check the coregionalisation matrices for positive semi-definiteness.
The criterion used to check for this condition is that a matrix whose eigenvalues are
greater than or equal to zero is a positive semi-definite matrix (Datta, 1995, p. 22).
Furthermore a symmetric diagonally dominant matrix with positive diagonal entries
is positive definite. This second characteristic allowed us to modify the
coregionalisation matrices in tum to ensure the positive semi-definiteness condition
was met. This meant a combination of increasing the values on the main diagonals
and decreasing the values on the off diagonals until a positive semi-definite matrix
was achieved. Having deduced one matrix, the others were modified such that the
sum of the coefficients for each basic model equalled the total sill for that
semivariogram. The process of creating a diagonally dominant matrix was then
repeated for the next coregionalisation matrix and so on until all three matrices were
positive semi-definite. While it is recognised here that these matrices may not be
optimally positive semi-definite (as for the minimised weighted sum of squares
algorithm outlined in Goovearts (1997)) with only four variables it was possible to
deduce the positive semi-definite matrices manually in order to provide an adequate
fit.
75
The final model selected consists of three basic structures, a nugget effect and
two spherical structures. The eigenvalues corresponding to each coregionalisation
matri:.> ar~ displayed in Table 4.3 confinning that the model is permissible.
Table 4.3: Eigenvalues corresponding to each of the coregionalisation matrices
Nue.et Svhericall Svherical2
0.02 0.09 O.oi
0.05 0.13 0.13
0.07 0.30 0.70
0.18 0.76 1.60
The first spherical structure is isotropic and has a range of 75 metres. The second
spherical structure is modelled with the direction of maximum continuity in the east-
west direction and has a range of 200 metres and an anisotropy ratio of 0.7. This
model can be expressed as:
0.1 0.05 0.03 0
0.05 0.09 0.05 0 g, (b)
0.03 0.05 0.08 0.02
0 0 0.02 0.05
0.27 0.17 0.03 0.19
0.17 0.51 0 0.17 Sph(ihi) 0.03 0 0.1 0.04 75
0.19 0.17 0.04 0.4
0.63 0.5 0.3 0.45
0.5 0.42 0.18 0.4 Sph( ~) 0.3 0.!8 0.08 0.12 200
0.45 0.4 0.12 0.55
where h'= = [I 0 ][ cos90° sin90°]["'] [ 0 ']["'] 0 0.7 -sin90° cos90° hy --Q.7 0 hy
76
Figures 4.10 and 4.11 display the direct and cross semivariograms fitted with this
(permissible) linear model of coregionalisation.
Nickel Cobalt y(lhD Direction 0 Ylihll Direction 0
2.1 2.4
1.6 2.1
1.5 1.6
•• 1.5 .. 1.2 .. 1.2 --~ .... .. . ... .. .. .. 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400
lhl lhl
yllhll Direction 90 Yllhll Direction 90
2.1 2.4
1.6 2.1
1.5 1.6 1.0 • 1.2 . .. 1.2 ~ .... . -..
--.~ ------'--~
50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 lhl I~
Magnesium Iron YI~D Direction o Ylihll Direction 0
2.4 1.6 2.1 1.4 .. 1.6 1.2 .. • • ~ 1.5 .. •• ~ 1.2 .. . . . . . . . . . •
50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400
~I ~I
Yllhll Direction 90 Y(lhD Direction 90 2.4
1.6 2.1 1.4 1.6 1.2 1.5
1 1.2
. • . . ..
0.6 .. .. . • 0.6 .. . • •• . 0.4
, 0.2 ...
0 0 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400
lhl lhl
Figure 4.10: Fitted direct semivariograms for MM22DTOP4
77
Yll~
'' '·' '' '·' '
'' '' '' ''
Yilt(! , ,, '' '' '' '' ''
"
Nickel-Cobalt Direction 90
.. . . .. ~
'"' '"' roo "" .,
'"' ~0
I~
Direction 0
.. .. ..
\00 150 200 250 300 350 400 I~
Nickel-Iron Direction 0
. .. ..... ~. . . . ..
0 _.__!------ - - - - - - - -------- -- -
~','---:.,c--,".,c---:,".,:---:,c,:--,c,=---:.,cCc-c,".,:-~~:· I~
vu .. , '' '' '' ..
Direction 90
.. . . . 0 ----------------~-~--~----
~.0 L_~OLO -~WLO-~,,~-c,::--c,.,::-~.,::-~,~OO::--.,.,c4-l~
Cobalt-Iron ,a,. Direction 0
'' " .., '' . ... ''
. '·' . . . . .. .
~------------------------~.,
0 "' '"' '"" = "" .,
"" ""' I~
ra,. Direction 90
'' '' '' '·' .. . '·' '·' ...
__. __ '!.._-
~' 0 "' '"' '"" "'' '"' = '"" "" I~
Y(ll(l
0.81 0.72
'·" 0.64
'·"
y(ll(l
'·' 0]
'·' '·' '' '·' 0.2 -
. .
Nickel-Magnesium
. .. . . Direction 0
.. · .
----------------------~
50 100 150 200 250 300 350 400
I"
Direction 90
. . . . .. . . . . . -~------------------------
100 150 200 250 300 350 400
I~
Cobalt-Magnesium Direction 0
.· . . . .
0 ----·--.--------------------
~.' "--~,L,--,",c,-c,~.,:---:,.,c0:--,c00c---:000cCc-c,".,:-~.,.,~ I~
'·' '·' '·' ~'
Direction 90
Iron-Magnesium Direction 0
.. - _._ ~--- !_ -------------------
~ 'L_---cc----c:c----::c---::c---::c---::c---::c----:~ 50 100 150 200 250 300 350 400
I~
Yllhf) Direction 90
''
~.,
~· L_-=--=--;O;--:-;O;---;::;:---;:'cc->;0;:-~~ 50 100 150 200 250 300 350 400
"
Figure 4.11: Fitted cross semivariograms for MM22DTOP4
78
4.2.2 Cross Validation for Cokriging of MM22DTOP4
The cross validation was carried out for all of the variables in this data set
using ordinary cokriging for the estimation. We present here the results for nickel,
those for cobalt, magnesium and iron are presented in Appendix B. Several
neighbourhood parameters were investigated with little difference in results. Figure
4.13 shows the cross validation residual analysis for the search parameters displayed
in Table 4.4 and Figure 4.12. Note that the minimum number of points in this case
was set to three. While this may appear to be too small, however, there was a group
of nine locations at the extreme northem perimeter of the study region that could not
be estimated using a minimum of four points. An increase in the size of the search
neighbourhood was considered. However as the affected points were outside the
region of sample data, hence extrapolated, it was deemed more appropriate to reduce
the number of points to three. In this way the extreme locations could be estimated
without affecting the satisfactory perfonnance of the search parameters for the
remainder of the region.
The cross validation residuals appear to be roughly nonnally distributed and
the plot of estimates versus the true values is well clustered around the 45° bisector,
with only two obvious outliers. The mean errors displayed in Table 4.5 are
sufficiently close to zero and the standardised error variance is sufficiently close to
the ideal value of one respectively. Overall the model appears to have reproduced the
sample values successfully for all variables.
Table 4.4: Neighbourhood search parameters
Number of angular sectors 4
Minimum number of samples 3
Ootimal number of samples per sector 4
Search radius (maior axis of elliose) (metres) 220
Search radius (minor axis of elliose) (metres) 100
79
H
"' .w
"' ~ ><
500.
400.
300.
" 200.
100.
0.15
" l " 0.10 ' • g • " •
0. 05
0. 00
Test Window X (Meter)
600.700.800.900~0001100120013001400.
500. X 500. X
400. X X X X 400. XX 9 X
xlmx 300. X X X X X 300.
200. XX XXX X 200. X
xxxx * X X 100. X X X X 100.
600.700.800.900~0001100120013001400.
X
Figure 4.12: Search neighbourhood for MM22DTOP4
Cross Validation
X (Meter)
600 000 800 900 ~0001100120013001400.
+ • + I + + +
+++++•+ • ++f+ +·+++++·!·1·++,+++ :++ ++++++++++- ++-f •• ++++++++++++++ ++ + + ++++++ +++··!··++""'+ +
+ + + + t 600.700.800 900 l000110012 0013 001400.
X (Meter)
(Z*-Z) /S*
{Z*-Z)/S*
' ' e
' ' ' 0 500.
t H
400. "~ 300.
i .. ~-' 200. ~
100.
0.15
~ 0.10 ~
0.05
! e·
' •
,. STNI (Estimates)
-2. -1. 0. 1. 2. 3.
4.
3.
+'+ I· 2.
1. +
0.
-1. • -2.
-2. -1. 0. 1. 2. 3.
,. ' 'TN' {Estimates)
Z* STNI (Estimates)
-1. 0 1
5.
0 •
-5. • •
1. 0. 1.
Z* ; STNI (Estimates)
><
:;:: <D rt <D ~
4.
4.
• •
2.
Figure 4.13: Cross validation residual analysis for MM22DTOP4
4.
3.
2.
1.
0.
-1.
-2.
5.
-5 .
80
Table 4.5: Mean and variance of cross validation residuals
Nickel Cobalt Ma esium Iron
Mean Variance Mean Variance Mean Variance Mean Variancl~
Error 0.02 0.87 0.02 1.15 -0.02 0.76 0.01 0.64
St. Error 0.02 1.85 0.02 1.96 -0.02 1.48 O.DI 1.45
4.2.3 Co kriging Estimates of MM22DTOP4
The ordinary cokriging of the nickel, cobalt, magnesium and iron variables
was perfonned using the linear model of coregionalisation discussed earlier. The
estimates were calculated at each of the 1717 grade control sample locations using
the search neighbourhood.. parameters displayed in Table 4.4. The parameter files for
this cokriging procedure are located in Appendix C. As previously, in section 4.1.3,
it was necessary to transform the estimates into their unstandardised form by
multiplying each value by the sample standard deviation and adding to this the
sample mean. Figure 4.14 shows post plots of the estimates for each of the variables
along with the post plots of the grade control data. Once again the smoothing nature
of the cokriging algorithm is clearly apparent. The estimates of n:...:k.d, cob<llt and
iron appear to be similar to those obtained from the cokriging using ~he intrinsic
coregionalisation model. As the discussion of these variables is analogous to that in
section 4.2.3 we omit it here.
The overall appearance of the post plots of the magnesium estimates is
consistent with the grade contwl values. The estimates along the north~western,
western and eastern perimeters are representative of the grade control values. The
values along the southern perimeter of the study region have been poorly estimated.
While the areas of high magnesium concentrations have been well represented the
concentrations of lower values have not. This is a result of an under representation of
low values in the southern most line of exploration drill holes.
81
Nicktl
~GradeCo~ol
Magnesium o Grade Control
Iron a Grade Control (;j • • •
I" '"" "~ ,.., ..... ·""" ···,.~,0 ·-
1
... ··~
"" ·-
Magnesium Estimab:s:
~ LMC • MM22DTOP4
Iron Estimates
~ LMC • MM22DT0~4
....
I., .. ,_ ... ... ··~HOI
1400
Figure 4.14: Cokriged estimates - Linear model of Coregionalisation MM22DTOP4
The residual post plots for nickel, cobalt and iron are displayed in Figure 4.15
and are similar to those obtained in the previous section. The residual plot of
magnesium shows a small pocket of overestimated values at the eastern edge of the
study region and a small pocket of underestimated values at the south-eastern
perimeter. Otherwise the residuals are of fairly small magnitude across the entire
study region. The histograms and normal score plots of residuals for nickel, iron and
zinc are similar to those presented in the previous section. Those for magnesium are
displayed in Figure 4.16. The histogram of residuals is reasonably symmetric and
82
looks approximately normally distributed. The normal scores plot however deviates
from a straight line along the middle section.
Res idual' 1jl Jfckti•LMC
Residuals
~ Magnesium : ~~C
300
!>.. (.) 200 Q <1.1 ;::l 0" 100 11)
~
0
Ruiduab
~rc'=b~~·L=M~c --~--------------------.
1 ... . ..., , .. ..... ·•= .... ... ·~ ··'"" ·•• ••il
1
=-·l~b)(l
... ~ .. •w..,
-~--'"""
:< .: ...
Figure 4.15: Residual post plots for :MJ\122DTOP4
300
200 (0
100 ;::l "t3 .... "' 0 11)
0::: -100
-200 •
Residual Magnesium Normal Score
1
_ ... -"· •x:e.oct ... 00 • . .... '""" s.:,)~ -==
•
Figure 4. 16: Residual histogram and normal score plot for magnesium
..
83
5 PRINCIPAL COMPONENT KRIGING
In this chapter we demonstrate the estimation technique of principal
component kriging using thr. principal components extracted in the eigenanalysis of
the MM22DHC4 data set. As the orthogonality of the principal components extends
to any separation vector h we were able to individually model them. All of the
semivariograms were calculated with 32 lags at a lag spacing of 12.5 metres and a
lag tolerance of 6.25 metres. The directional semivariograms were calculated with an
angular tolerance of 45°. All directional models were assessed for fit in the
intenncdiate directions (with lag tolerance set to 22.5°).
Once the spatial variability of the principal component scores had been
modelled the estimates at unsamplcd locations were calculated using ordinary
kriging. All models were cross validated using ordinary kriging and perfonned
satisfactorily. Once again the software pack<lge used to perform the kriging was
ISATIS. 1717 estimates were obtained for each principal component directly at the
locations of the grade control data. The parameter files for all kriging procedures are
displayed in Appendix C. Having obtained the estimates of the principal component
scores it was necessary to recalculate the nickel, cobalt, iron and zinc estimates using
the coefficients obtai!Jed in the eigenanalysis of the MM22DHC4 data set. These
estimates were then compared to the MM22DGC grade control data.
5.1 Variography of th.l! Principal Components from MM22DHC4
For the first principal component (PCl) the semivariogram surface exhibits
anisotropy with the direction of maximum continuity being east-west. This is further
supported by the standardised semivariogram surface also displayed in Figure 5.1.
The final model chosen consists of three structures, a nugget effect and two spherical
structures. The first spherical structure is isotropic with a range of 90 metres. The
second spherical structure is anisotropic and is modelled with cast-west as the major
direction with a range of 200 metres and with an anisotropy ratio of 0.45. This model
can be expressed as:
84
y(h) = 0.1g 0 (h)+ 0.4 7 Sph (M) + 2.4 7 Sph (~) 90 200
Figure 5.2 shows the directional experimental semivariograms fitted with this linear
model of regionalisation.
Y(hl Ys (hl
0:, u ... .l" • "'
~· ... IJ ... .., .. -
..... .. Figure 5.1: Semivariogram and standardised semivariogram surfaces for PC 1
~ (lhl) Direction 0 8.1 7.2 • 6.3 5 .4 4.5 •
• • • • ••• • 3.6 • • • _...__.
'''~·· . .
1.8 • • • •
0 .9 •• 0
0 50 100 150 200 250 300 350 400 I hi
y (lhl) Direction 90 8.1 • 7.2 6.3 5.4 • • • 4.5 • • • • 3.6 • • •
''V - • • 1.8 ... •
• • 0.9 • 0
0 50 100 150 200 250 300 350 400 I hi
Figure 5.2: Fitted directional semivariograms for PC1
85
Figure 5.3 displays the semtvanogram surface for the second principal
component (PC2). The surface does not exhibit any anisotropy which is further
supported by the standardised semivariogram surface also displayed in Figure 5.3.
The final model chosen consists of two structures, a nugget effect and an isotropic
spherical structure with a range of85 metres. This model can be expressed as:
r(b)= 0.08g0 (h)+047Sphn;J
Figure 5.4 shows the omnidirectional experimental semivariogram fitted with this
linear model of regionalisation.
..
y(h)
0::: . .. .r- 0 .. , ·~ •..
- - - - 0 - - - - -.. Figure 5.3: Semivariogram surface for PC2
~ (lhl) Omnidirectional 0.72 •
• • 0.64 - • • • • • • • 0.56 • ... • - • t---- ... • • •• • • 0.48 • • • • •• •
0.4 • 0.32 •
0.24
0.16
0.08
0 . 0 50 100 150 200 250 300 350 400
I hi
Figure 5.4: Fitted omnidirectional semivariogram for PC2
Ys(h)
Figure 5. 5 displays the semivariogram surface for the third principal
component (PC3). This component exhibits anisotropy with the direction of
maximum continuity at azimuth angle 65°.
86
y(h)
DOG
00 .... .... 0 .. ...
Figure 5.5: Semivariogram surface for PC3
Figure 5.6 displays the directional semivariograms fitted with the final model. It
consists of three structures, a nugget effect and 2 spherical structures. The first
spherical structure is omnidirectional with a range of 50 metres; the second is
modelled in the direction of maximum continuity (azimuth 65°) with a range of 140
metres and an anisotropy ratio of0.7. This model can be expressed as:
r(h) = o. o 1g0 (h)+ o. o5Sph(MJ + o.215Sph(~) 50 140
where h' = [1 0 ][ cos6SO sin65°][hx] = [ 0.42 0.91][hx] 0 0.7 -sin65° cos65° hY -0.63 0.30 hY
lv <lhD Direction 25 0.4
0.35 • •• • • • 0.3 ••• • 0.25 v···.· • • • • 0.2 •
• • 0.15 • •• 0.1
0.05 • 0 0 50 100 150 200 250 300 350 400
lhl
lv (lhD Direction 115 • 0.4
0.35 • • • 0.3 . . . .
0.25 v· • .. • • • 0.2 • . • ••• •
0.15 • . •
0.1 •• • 0.05
0 0 50 100 150 200 250 300 350 400
lhl
Figur~ 5.6: Fitted directional semivariograms for PC3
87
The semivariogram surface of the fourth principal component (PC4)
displayed in Figure 5. 7 exhibits both a short range isotropy and a long range
anisotropy. From the semivariogram surface the direction of maximum continuity
appears to be at azimuth angle 65°, whereas the standardised semivariogram surface
suggests north-south (azimuth 0°). Investigation of several directional
semivariograms, such as those displayed in Figure 5.8, revealed that the anisotropy in
various directions was not sufficiently marked, hence it was possible to treat the data
as being isotropic.
y (h)
0:: ·-·~ ~ D
O.iU ...
Js (h)
Figure 5.7: Semivariogram and standardised semivariogram surfaces for PC4
r <lhll r(lhll
0.8 Azimuth 90° 0.8 Azimuth 45° 0.6 0.6
0.4
0.2
40 80 120 160 200 240 280 320 360 40 80 120 160 200 240 280 320 360 lhl lhl
r<lhll r(lhll
0.8 Azimuth 0° 0.8 Azimuth 315° 0.6 0.6
0.4
0.2
o~~--~-L--L--L~L__L __ ~_.
0 ~ 80 w - 200 ~ ~ ~ 360 40 80 120 160 200 240 280 320 360 !hi lhl
Figure 5.8: Directional semivariograms for PC4
The omnidirectional model chosen is displayed in Figure 5.9 and consists of three
structures, a nugget e:tfect and two isotropic spherical structures. The ftrst spherical
88
structure has a range of 100 metres and the second has a range of 250 metres. This
model can be expressed as:
y(h) =0.018g0 (h)+ 0.065Sph(11) + 0.092Sph(_M_) 100 250
~ (lhl) Omnidirectional 0.32 • 0.28 • 0.24 • • • • • 0.2 • • '------- • 0.16 • -. • • • •• • • 0.12 • • • • • 0.08- • • • 0.04 g
0 I I I I
0 50 100 150 200 250 300 350 400 I hi
Figure 5.9: Fitted omnidirectional semivariogram for PC4
5.2 Cross validation of the Principal Components
Cross validation was carried out individually using ordinary kriging for each
of the principal components. Various search neighbourhoods were investigated with
little difference in the results. The final search neighbourhoods used are displayed in
Table 5.1. We present here the results for the first principal component, the
remaining cross validation results for PC2, PC3 and PC4 are displayed in Appendix
B. Figure 5.10 displays the search neighbourhood for the first principal component
and Figure 5.11 displays the residual analysis of the cross validation. The residual
histogram shows that the errors are roughly normal! y distributed and the estimates
versus true value plot is reasonably close to the 45° line with the exception of a few
outliers. In all cases the model provided successful estimates. Table 5.2 shows the
mean and variance of the errors and the standardised errors. The mean errors are
sufficiently close to zero and the standardised error variances are sufficiently close to
one.
89
Table 5.1: Neighbourhood search parameters
PC1 PC2 PC3 PC4
Number of angular sectors 4 4 4 4
Minimum number of samples 3 3 3 3
Optimal number of samples per sector 4 4 4 4
Search radius (major axis of ellipse) (metres) 220 120 190 250
Search radius (minor axis of ellipse) (metres) llO 120 100 250
Rotation (azimuth) 90° NA 65° NA
Test Window X (Meter)
600
500. 500.
400. 400. e< H (j)
+' 300. (j)
~ X X 300.
::;:
" rt
" >< 200. 200. >i
100. X 100.
013001400.
X (Meter)
Figure5.10: Search neighbourhood for PC1
90
500.
" 400.
• " • "
300.
" 200.
100.
0 .1'5
• • ·K 0.10 u
' • il • u • 0. 05
0. 00
Error
Cross Validation
X {Meter)
600.700. 800.900 .1000110012 0013001400. .,+ + +
+ +++-1· +++++++ -·1- .++++ +++++++++-r - ++··--· + ++++++++++' +++···· ++++-t+++++++-t-+1+-t ++-t+++++++-t++-tt++
H + + j 600.700 800. 900 .10001100120013 001400.
X (Meter)
{Z*-Z)/S*
(Z*-Z) /S*
500.
400.
300 .
200.
100.
. 0.15
0.10
0. 05
0.00
.. 3.
.. "
2. rl
1. • , •
0. " ~ rl -1. u
" " ~~
-2.
-3. " • "--4.
-5.
5.
" K ' • • ~~ • " 0 •
' ' 0 • K· N • -•
-5.
Z* : pel (Estimates)
-5. -4. -3. -2. -1. 0. 1. 2. 3. 4.
•
•
+ +
4 .
3 .
2.
1.
0.
-1.
-2.
-3.
-4.
-5'" _:..._-f, _-_--,3'-. ~-21-.-_+1.--,iO-. ...,, ... _ -2+.-i-3 _ ___,, ....... _ -5.
Z* : pel (Estimates)
,. pel (Estimates)
3. 2. 1. 0. 1. 2. 3.
• 5.
• • + + +++ + + ,. -1~ .
-Jr\-t't+-..it-t ++** -~! * ++-f,yt-+~~-~+4-+ -tlr~·
0. + + it+. -i.l· + -H· + + + -tt-''' ~~+ +
-5.
3. 2. 1. 0. 1. 2. 3.
,. pel (Estimates)
Figure 5.11: Cross validation residual analysis for PCl
Table 5.2: Mean and variance of cross validation residuals
PC1 PC2 PC3 PC4
Mean Variance Mean Variance Mean Variance Mean Variance
-0.04 2.27 0.00 0.70 0.01 0.34 0.00 0.15
St. Error -0.02 2.20 0.01 1.74 0.02 2.35 0.01 1.85
91
5.3 Principal Component Kriging Estimates
The ordinary kriging of each of the principal components was performed at
the 1717 grade control data locations using the neighbourhood search parameters
displayed in Table 5.1. The parameter files used are displayed in Appendix C.
Having obtained the estimates of the principal component scores it was necessary to
calcnlate the individual variable estimates using equation (60) which we recall is
where Oi and m; are the sample standard deviation and mean respectively of the ith
variable. The coefficients aki are obtained from the transpose of the matrix of
r'igenvectors calculated in the principa: cornpn1ent analysis of !vflv:[22DHC4.
-0.50 -Q.48 -0.51 -0.51
A=(a.,]=Q' = -0.42 -0.60 0.50 0.47
0.76 -0.64 -0.02 -0.12
0.04 -0.06 -0.70 0.71
These calculations were performed using an EXCEL spread~heet.
Recall f1om section 3.4.2 we determined that if data reduction were the aim
of performing a principal component analysis on the W.122DHC4 data set that we
would retain only the first two principal componenls. These two components
accounted for 89% of the total variability of the original data. In addition to
calculating the estimates of nickel, cobalt, iron and zinc using all four principal
components we also estimated these variables by retaining only the first two
principal components. The estimates of the variables then were obtained by
discarding the estimates of the scores of both the third and fourth principal
components and the last two columns of the matrix of eigenvectors Q. The estimates
from oniy the retained {.>rincipal components are then calculated using equation (61)
which we recall is
92
where Ol and mi are the sample standard deviation and mean respectively of the ith
variable and P is the number of principal components retained. The coefficients apt
for the truncated estimates are:
A- a -[ ] [-0.50
p- pi - -0.42 -0.51 -o.51 -0.51] -0.60 0.50 0.47
Figures 5.12 shows thr post plots of the nickel, cobalt, iron and zinc grade
control data and principal component kriging estimates (all four principal
components). Overall the estimates have reproduced the grade control samples well,
once again though the smoothing nature of the kriging process is apparent. The same
areas of over and underestimation are apparent as for the previous kriging methods.
In fact the post plots are almost identical to those we have alrearly seen. Figure 5.13
shows the post plots of the grade control data and the estimates obtained using only
the first two principal components. Clearly the grade control data have been
reasonably well reproduced with very little difference to those estimates obtained
using all four principal components. The histograms and normal score plots for the
residuals from both types of principal component kriging are very similar to those
presented earlier, hence are omitted here. Figures 5.14 and 5.15 display the residual
post plots for both type~: of principal component kriging. Both sets of plots are once
again very similar to those of the other kriging methods.
93
Nick~
g GrtdtCo~ol
Cob all
hn ~ Grade Conbot
·: . . .. :: 'i\ ..
·r!• ...
1
.. .,~ ,,., 11;1» ..... ,._ 1 ....
1~-···~ ·-
Zlnc;l!slinalts ~ ~rlncipal Compontnt ~glng
Figure 5.12: Principal component kriging estimates
1
.. ·.... ·..... ~2 .•
:::: ... " ·-
1
.. .... ... ····'"" ··-
,,._
1
.. ..... ~-.. , . ...... -·· :::
1~-·-.... _ ··----''""" ·-~ ...
94
Cobalt
~ Grade C~n~ol
Iron ~ Grade Control
Zinc:
~ Gradt Control
.,;, : ~ : :·· .... . · ·!:-
:::· :::i::~~ni~::· ... ··:::·::::::::·::.::.
::. · .• ;; : .. : · ·~: !; ::::.: ........ .. g • : :=~== :: ;;~:; :·:;Eii1::: i;n;::::i~i~hi ~ ~ :: : · · ::~?;,; · ;=::
• .::"!!! . ' ; .... : .:· . : :;::::; ~ · ~: • .=-: .,; :.r :~ ~· .:; -~: ~ .. =·:::::: .. • ::.·.: :.f2 : E:·;~ ;~~~~~=~f:: ~ : · :~~;: -!:~::~:~~h ::
.. .. ,. .. ::; .. : .·=·.:· :.·i • • ":'f '.: ·· :: : .... :; .. . · :.:~· ::·:. . . . . ·:.::.~=~ :: :.: .. :~ • : !• :·· • • • . .•. ···. •· • ··· · .: ::··
1
.. ''"" ·-,,~ (100'
ZI Z!.tm
_, ~·~w-
,...ckel Estimates
~PCK2
Cobalt Estimates
~1'012
Iron Estimates
~ 1'012
I" .... "~ ·,.., ! UIO ~~
1480
,_, U;lo<Q:>Q. ,.,.,
1
.. -~0-.~ ... :l!IUOOO
Figure 5. 13: Principal component kriging estimates using two retained principal
components
.·
95
Rtsidu~ls
~ ~ckoi-PCK
Reskfuals
1---~-··~.. '""' M..... --
Residuals
~ Cobillt -PCK
Residuals
~Zinc - PCK
Figure 5.14: Residual post plots for principal component kriging
r ·U-·~ ~-.. ·-,.., W#.'O
1
---., ... --·1(01100 .. '""" .... .. .., ... .,
~sidua!s
~ Cobilt·PCKZ
g
Residuals
~Zinc·~ . •
1
.,.,.., <~•:10 0100
'""" .. "~-"'"' --
r ,.,,
··--H W
-·-.... ·-,, ·-
lmM_.-.,,.,fi<ICI ,,.,_ .. '""' .....
Figure 5.15: Residual post plots for principal component kriging using two retained
principal components
.·
96
6 Ordinary Kriging
In this chapter we demomtrate the estimation technique of ordinary kriging
using the variables nickel. cobalt, magnesium, iron and zinc from the MM22DEXP
data set. The data were left in their raw (unstandardised) fonn for this section. In
addition to the grade control data for comparison the estimates obtained from
ordinary kriging were used to assess whether the multivariate techniques enhan-::ed
the estimation, the discussion of which we defer until the next chapter. For the
variography and modelling the same parameters were used as for the previous
methods, that is, all semivariograms were calculated with 32 lags at a Jag spacing of
12.5 metres and a lag tolerance of 6.25 metre.!>. The directional semivariograms were
calculated with an m1gular tolerance of 45.
The cross validation using ordinary kriging for each of the variables showed
that the models all performed satisfactorily. As previously IS A TIS was used for the
estimation. Again 1717 estimates were obtained for each variable directly at the
locations of the grade control data. The ordinary kriging parameter files are also
displayed in Appendix C. The accuracy of the ordinary kriging estimates was
assessed via comparison with the MM22DGC grade control dato..
6.1 Variography of Nickel, Cobalt, Magnesium, Iron end Zinc
The semivariogram surface for nickel is displayed in Figure 6.1. As expected
it displays the same anisotropies identified earlier in Chapter 4, that is the presence of
two structures, one short range and one long range. The direction of maximum
continuity of the short range structure is east~wcst (azimuth 90°), which is confirmed
by the standardised scmivariogram surface, also displayed in Figure 6.1. The
direction of maximum continuity of the long range structure is approximately
azimuth 70°. Detailed examination of the directional experimental scmivariograms,
including those shown if Figure 6.2, revealed that the anisotropy is mo.- l pronounced
in the east~west (major) and north~south (minor) directions.
97
y(h) Ys(hl
0~ 0" ' "
,:- . ... .. ·-.,
Figure 6.1: Semivariogram and standardised semivariogram surfaces for nickel
Y (IIIII y(lhl)
Azimuth 90° 210 Azimuth 45° 180 180 150 150 120 120 90 90
60 60
30
50 100 150 200 250 300 350 400 60 100 150 200 250 300 350 400 lhl lhl
y(lhl) y(lhl)
210 Azimuth 0° 210 Azimuth 315° 180 180
150 150
120 120
90 90
60 60
30 -~ 30
0 50 100 150 200 250 300 350 400 0 50 100 150 200 2~0 300 350 400
I hi lhl
Figure 6.2: Directional semivariograms for nickel
The model selected consists of a nugget effect and two spherical structures.
The first spherical structure is isotropic with a range of 90 metres. The second
spherical structure is modelled with east-west as the major direction with a range of
160 metres and an anisotropy ratio of0.56. This model can be expressed as:
r(h) = 5 g0 (h)+ 12.15Sph(MJ + 28.3 5Sph (~) 90 160
Figure 6.3 shows the d~rectional semivariograms fitted with this model.
98
<I•D Direction 0 100 - • 80
• • 60 ••
?'~· • ••
40 • • • •• • • • • 20
• 0
0 50 100 150 200 250 300 350 400 lhl
<I•D Direction 90 100 • 80 • • • 60 • • • • • • v.· 40 ""T •
• • • • • • 20 •• 0
0 50 100 150 2<;;;,
250 300 350 400
Figure 6.3: Fitted nickel directional semivariograms
The semivariogram and standardised semivariogram surfaces for cobalt are
displayed in Figure 6.4. The direction of maximum continuity of the anisotropy
clearly looks to be in the east-west direction. The model for cobalt consists of a
nugget effect and two spherical structures. The first spherical structure is isotropic
with a range of 100 metres. The second spherical stmcture is modelled with east
west as the major direction with a range of 200 metres and an anisotropy ratio of 0.5.
This model can be expressed as:
y(h)=0.012g0 (h)+0.16Sph(M_)+0.13Sph(~) 100 200
[1 0 ][ cos90° sin90°][hx] [ 0 l][hx] where h'= = 0 0.5 -sin90° cos90° hY -0.5 0 hY
Figure 6.5 shows the directional semivariograms fitted with this model.
99
Ys(h)
0: .. >M
'M ,_
' ·'
Figure 6.4: Semivariogram and standardised semivariogram surfaces for cobalt
IY (lhl) Direction 0 0.72 0 .64 • 0.56 0.48 •
0.4 • • • • • '"7 • . • 0.24 • •• •• • • • 0.16 • • •
0.08 •• 0
0 50 100 150 200 250 300 350 400 I hi
y (lhl) Direction 90 0.72 • • 0 .64 0.56 0 .48 • • •
0.4 • • • • • • •
:::/ • • • . . •
0.16 • • • .. 0.08 •
0 0 50 100 150 200 250 300 350 400
I hi
Figure 6.5: Fitted cobalt directional semivariograms
The semivariogram surface for magnesium, displayed in Figure 6.6, exhibits
a short range isotropy and a long range anisotropy. The direction of maximum
continuity is azimuth 330°. This anisotropy is more pronounced in the madogram
surface, also displayed in Figure 6.6 .
..
100
- t i!ltl ..., 0 ttl ltlO ~ :na -.. Figure 6.6: Semivariogram and madogram surfaces for magnesium
M(ll)
0:. ~·
"" ... ,
The model for magnesium consists of three structures, a nugget effect and two
spherical structures. The first spherical structure is isotropic with a range of 140
metres. The second spherical structure is modelled in the direction of maximum
continuity azimuth 330° with a range of 300 metres and an anisotropy ratio of 0.5.
This model can be expressed as:
r(h) = 13 5 g0 (h)+ 23oosph(!LJ + 2162Sph(~) 140 300
where h' = [ 1 0 ][ cos330° sin330°][hx] = [ 0.87 -0.5][hx] 0 0.5 -sin330° cos330° hY 0.25 0.43 hY
Figure 6. 7 shows the directional semivariograms fitted with this model.
101
y (lhl) Direction 120 8100 7200 • 6300 • 5400 •• • •
:::~ ·· •
• 2700 • •• • • • • • • • • • • 1800 • • 900 • •
0 0 50 100 150 200 250 300 350 400
I hi
(I hi) Direction 30 8100 • • • • 7200 6300 5400 • • • :::::v • • •
• • . • • •
• • 2700 • 1800 900 • • •
0 0 50 100 150 200 250 300 350 400
JilL
Figure 6.7: Fitted semivariograms for magnesium
The semivariogram surface for iron is displayed in Figure 6.8. The data are
clearly anisotropic with the direction of maximum continuity being east-west.
""' , .. 120 .. "" .. .. 0
.... ... .... . , .. • :zoo - - - ~ ~ 0 ~ ~ ~ ~ -..
y (h)
Figure 6.8: Semivariogram surface for iron
The model for iron consists of three structures, a nugget effect and two spherical
structures. The first spherical structure is isotropic with a range of 75 metres. The
second spherical structure is modelled with east-west as the direction with a range of
185 metres and an anisotropy ratio of0.54. This model can be expressed as:
102
y(h) = 1 030g0 (h)+ 31 OOSph(MJ + 16790Sph (~) 75 185
where h'= · = · [1 0 ][ cos90° sin90°][h~J [ 0 1][hx] 0 0.54 -sin90° cos90° hY -0.54 0 hY
Figure 6.9 displays the directional semivariograms fitted with this model.
r !lhll Direction 0
~5000 ~0000 • • ~5000 poooo • • • •• • • ~5000 •• • • • • • P,ooo~.· • • 15000 •• 10000 • • •
5000 •• 0 .
0 50 100 150 200 250 300 350 400 I hi
1 (lhll Direction 90
~5000 • ~0000 • psooo ~0000 • ~5000 •• • • • •• • poooo ~ --•• • .. .
• • • '"00 / • • 10000 • 5000
0 0 50 100 150 200 250 300 350 400
I hi
Figure 6.9: Fitted semivariograms for iron
The semtvarwgram and standardised sennvanogram surfaces for zinc are
displayed in Figure 6.10. The anisotropy is clear here with the direction of maximum
continuity being east-west.
Ys(h)
0 .. ~ 0.1
Figure 6.10: Semivariogram and standardised semivariogram surfaces for zinc
103
The final model selected consists of a nugget effect, a spherical isotropic short range
structure of 85 metres and an anisotropic long range structure of 200 metres
modelled with east-west as the major direction and an anisotropy ratio of 0.4. This
model can be expressed as:
y(h) = 17700g0 (h)+ 274239Sph(M)+ 601500Sph(~) 85 200
[1 0 ][ 90' h h ' cos were = 0 0.4 - sin90'
Figure 6.11 shows the directional semivariograms fitted with this model.
y(lhD 100000
1800000
1500000 •
Direction 0
1200000 • • • • • • •
I-- -~-=10~·~·--~·~·~·~·--·~·~ 900000 - • -
600000~· ••• • • 300000 • • OL_ __ L_ __ l_ __ ~.--~--~--_L---L---~~
0 50 100 150 200 250 300 350 400
Y(lhD 100000
1800000
1500000
1200000
300000
•
•• • • •
lhl
• • • • • •
• • • •
Direction 90
• • •• 900000/ 600000
• 0~--L_ __ l_ __ ~--~--~--_L __ _L __ _L., 0 50 100 150 300 350 400
Figure 6.11: Fitted directional semivariograms for zinc
6.2 Cross Validation of Ordinary Kriging
As for the previous data sets cross validation was carried out for all the
variables for various search neighbourhoods with little difference in the results. The
final search neighbourhoods used in the cross validation and subsequent kriging are
displayed in Table 6.1 with those for nickel displayed in Figure 6.12. We present
104
here the cross validation results for nickel, those for cobalt, magnesium, iron and
zinc are displayed in Appendix B. Overall, in all cases, the models have estimated
the sample data values well. In particular for nickel, as displayed in Figure 6.13, the
residual histogram appears to be roughly normally distributed with the exception of a
few values at each tail and the estimates versus the true values plot is reasonably well
clustered around the 45° bisector. The standard error of the mean is close to zero and
although the variance of the standardised errors is relatively high, as displayed in
Table 6.2, it is not high enough to be of concern.
Table 6.1: Neighbourhood search parameters
Nickel Cobalt Magnesium
Number of angular sectors 4 4
Minimum number of samples 3 3
Optimal number of samples per sector 4 4
Search radius (major axis) (metres) 180 220
Search radius (minor axis) (metres) 110 120
Rotation (azimuth) goo goo
Test Window
500.
400. H ID
'"' ID e 300.
>< 200.
100.
X (Meter)
013001400.
X
X X X
X X X X
XX XXXXXXXXXXX XX X
X (Meter)
4
3
4
300
160
330°
500.
400.
300.
200.
100.
Figure 6.12: Neighbourhood search parameters for nickel
Iron Zinc
4 4
3 3
4 4
210 220
110 110
goo goo
"' :;: ro " ro "
105
500.
;; 400.
~ • 3 300.
" 200.
100.
0.15
~ 0.10
l ~
0. 05
Error
St. E
Cross
X (Meter)
600.700.800.900 ~000110012001300.1400.
' 500.
+ + + + + + 400.
.... +,++ • ++r ++1·+++++++,+++ .. •+ 300. +-f-·1 ++++-1--t-+··+-t--fl •• +++-~+-+++++++++""'++ 200. ++++++++++++++--++
++ + + +t 100.
600.700 .800. 900 10001100120013001400.
X (Meter)
(Z*-Z) /S*
0.15
0.10
0.05
(Z*-Z)/S*
Validation
-2.
4.
• '. ' H : 2. • ' ~ 1.
"~ 0.
Iii .. ~" 0
-1.
!!. -2.
_,.
_,
5.
0.
-5.
,. ' STNI (Estimates)
-1. 0. 1. 2. '.
+
•
-1. 0. 1. '. '. ,. ' STNI (Estimates)
Z* : STNT (Estimates)
'
0. 1.
Z* : STNI (Estimates)
Figure 6.13: Cross validation residual analysis for nickel
4.
4.
3 • " ffi
2. " " 8 1. 2 0
0. ' • ' -1. ~
-2.
4.
2.
5.
0.
-5.
'.
Table 6.2: Mean and variance of cross validation residuals
Nickel Cobalt Magnesium Iron Zinc
Mean Variance Mean Variance Mean Variance Mean Variance Mean Variance
0.02 0.92 o.oz 1.24 -0.02 0.77 0.01 0.69 0.02 0.65
0.02 1.85 0.02 2.54 -0.02 1.85 0.01 1.96 0.02 1.72
106
6.3 Ordinary Kriging Estimates of Nickel, Cobalt, Magnesium, Iron and Zinc
Ordinary kriging of the nickel, cobalt, magnesium, iron and zinc variables
was performed directly at the 1717 grade control coordinates using the search
neighbourhood parameters presented earlier (Table 6.1 ). Figure 6.15 shows post plots
of the estimates for each of the variables along with the post plots of the grade
control data and the residuals. As previously the smoothing nature of the kriging
algorithm is clearly apparent here. Overall the estimates have reproduced the grade
control values well . The same areas of over and underestimation are apparent as for
the previous kriging methods. In fact the post plots are once again almost identical to
those we have already seen. Figure 6.14 displays the residual post plots of each of the
variables. Once again we omit the residual histograms and normal score plots as they
are almost identical to those presented earlier.
R.,ldu~s
r ~ Cobillt·OK .. _ ... ·-.... .. "" ·-..... .... ""' ~
·- ... Rukfulls
~Iron ·OK_
r ·11~.000
-tOO .tOO --.. "'"" ·~-1110000 ...... .., ... ...... ....... ....... -
1480 .,
..... ......... .... ... . .:·: . ... . -:!:.:: .
. ;;· .. : ;.:~~z:::::u·=::~-: ··: ;;:-:;·:;~=~~t · ::= ~ : i:.
I~;. · <~ :~~~~:::::::: •:!~ ::;:::::~:= ; ;~=~~! : : ..... ~E=-~ :~: .. :: :~:: ... ~~ :; ~: : : ~~ ~ = · ·::::·: .. "i!::f;~z;! ·i=!:· o.::=.:::=::i-:i
£ .. ... '!':·· .•••••.•• ·:·· ······ •• • _ •.•.••. • - •• . ., 10)5
1···-----.. ..... ----
Figur.e 6.14: Residual post plots for ordinary kriging
r :·-. .. ·-. .. . .... _, ... ·-.. ~= ·-·-
. ...
r ...... --...... ·•to-.. ··-----: '::
107
Cobalt ~Grade Control
Iron
~ Grade Control
Zinc o Grade Control
,.
1
.. a.•• ··~ ·~ ··'"'
1480
1
.. ,..,.,a : t21DCZI ----=·-:111!1.«')
Magnesium Estimates
g Ordinuy Kriging
Zinc EsUmztes ~ Ordinary Kriging
Figure 6.15: Ordinary kriging estimates
1
.. ~·= '""'"' t ru~m .... -~ llllilDXl
1490
108
7 Comparison of Estimates
The post plots shown earlier in Figures 4.6, 4.14, 5.12, 5.13 and 6.15 show
that, as expected, the ordinary cokriging and ordinary kriging techniques used have
resulted in very similar estimates. The principal component kriging estimates are also
very similar to the other estimates. In addition the principal component estimates
using only the retained principal components arc very similar to all other estimates.
This indicates that very little infonnation has bee;n lost by retaining only two
principal components. Figures 7.1 to 7.5 and Tables 7.2 to 7.6 display the descriptive
statistics for each of the variables. Overall all of the kriging techniques used have
reproduced the grade control and exploration descriptive statistics reasonably well. In
all cases the mean and median of the estimates has over estimated that of the
exploration and grade control data. The only exception is zinc where the mean and
median of the estimates arc close in value to those of the grade control data. The
values obtained for the mean and median for each variable arc very close in value for
each kriging method.
The standard deviation of the estimates in all cases is considerably lower than
that of the grade control and exploration data, yet another indication of the
smoothing of the variability due to the kriging process. This is further supported by
the coefficient of variation which in all cases is lower for the estimates than that of
the grade control and exploration values. The standard deviation and coefficient of
variation obtained for each very is very similar for each kriging technique used. The
skewness of the histograms of the grade control data for each variable has been well
reproduced for all variables except cobalt. While the grade control cobalt is strongly
positively skewed the estimates are not, in fact the skewness of the estimates is
around one third of that of the grade control data.
Table 7.1 displays the mean square errors and mean absolute errors for each
variable and each kriging method used. The lowest mean square CITOr for each
variable is coloured red and the lowest mean absolute error is coloured blue. Overall
for the kriging methods used for each variable the values fo!" these measures are quite
similar. For nickel the lowest mean square error and mean absolute error were
109
associated with the principal component kriging estimates. The lowest mean square
error for cobalt is associated with the principal component kriging estimates using
only two principal components. Correct to two decimal places cobalt had identical
values for the mean absolute error for all methods. However correct to four decimal
places the principal component kriging estimates had the lowest value for this
measure. Both the mean square error and mean absolute error for magnesium were
lowest for the estimates obtained from cokriging. The iron mean square error and
mean absolute error were the lowest for the principal component kriging estimates
obtained using only two principal components. For zinc the lowest values for both
error measures were associated with principal component kriging.
Table 7.1: Mean square error and mean absolute error for each method
Nickel Cobalt Magnesium Iron Zinc
MSE 41.71 0.34 NA 20489.95 847999.31 OCK-ICM
MAE 4.88 0.43 NA 107.91 705 .32
MSE 41 .46 0.33 3269.59 20003 .96 NA OCK-LMC
MAE 4.86 0.43 39.51 106.59 NA
MSE 40.50 0.33 NA 20497.71 838231 .21 PCK
MAE 4.81 0.43 NA 107.92 699.99
MSE 42.51 0.32 NA 18568.06 884728.72 PCK2
MAE 5.05 0.43 NA 105.29 733 .77
MSE 41.56 0.34 3612.69 20365.44 843825.91 OK
MAE 4.87 0.43 41.09 107.32 701.78
,·
110
' ' " "' ., ' ' '
96tlConfldencelll'tiMIIforWu
I I I I I I I I I •z• 12.1 111 t:lJ t:s.o 11.1 t l..2 u' ''"
I I I I I I I I I
9511. Conftdtnn lnteMIIor MJ
' ' .,
I I I I I I I I I 141.5 t.c.tS 14M 1$0$ 11515 t5.7S 1$35 IH5 tU6
I I I I I I I I
r:~;ag~====~·~:::::::::J 95"JioCCWl~-MIIforM!dan
' ' I ' ' " ,.
" .. " ' ' ' -~".lio CcnlldtncelnltMII f« M#
' ' ' ' ' ,J, ' ' "' ... ... ... I" '" ... ' ' ' I '
»Sft ConftdencelnteMI for MK!t.n
I'ICKELGC
Nlden:o...Oatllng Notmaltt-(Tut
A.$quaf1!d I 122 P-Vatue. 0.000
-htClulttillt l.lrii;~n 3nf0t.i.mile Ma.limum
13.D470 11407
44 0117 0111!05 .f4E--O'l
1711
02400 ...... 12.1140 17AS70 41&400
SS%ConftdencetlteMI IorNu
12.1321 13.3113
95"- Cont.dli'~~ee .,.,_, h»f $itm• 1415l lllOO
t5'4Coftld•nu~IDfMidllft
n~t770 n<~AIW
I'ICKEL OCK ICM EST
llfld'llrsM-DarllngNonnallt{T-st
ASqu-.d 1713 FWIIIue. 0000
...... ISOSSS
""'"' ..... VarialK.e ,. ... , ,._, .. , 038:Htl
""'""" 10786t N t7H
Mnunum 05070 tstOw~ 119002 Maclan 151481 JrdOu•"- 179030 M!IJiimum 426N4
95~ Conlldon<e lnttMII« ldi
14.8403 152707
iS'Io CoMdence iniMWI for S9n1 4.3;88 4 7033
05~ Con~cSM~c•M*WitorMidla
14.8121 154841
NICKEL PCK EST
Jnderson-DorlingNorrnlflt,oTH t
ft.Scf.Jnd 1788 P-Va111.: 0000
...... 150821 - ·~ ""''""' 206153 -... 0414107 K"""h 100005 N 1717
Mnlmum 0 567:! 1<>10uertilt 11.&462 f>1adi~n 15007G lrdOUartile 1711038 M;t)l!mum 42.0332
85-J. Conlidlncelnr.tv.al lot,.,
t• 8A72 IS 2770
~ ConicNftc.ln ..... lor 59n• 4383S • 6876
95~ Confr.denw lnllfWf for Mtlhn
147270 153383
I I I I I 141 15.0 IS I 151 15.3
' 959f.Conldance tt•MIIQfMf(IJen
I I I I 12 II • t4 15
'c==='=r::~:':::::::~' 85'4 CoMff:na "*N.IIII!Dt t.t~~:ddn
' " " ' ' - ' ' ,. .. ' '
' ' . .,
I I I I I I 1 I I !Hill t<~~7t t•to 1.,1 1508 HUI 1$.2CI \5. ISC
I ~I=~·;;=.=' =;~' :;;~'::::;' =:'J I tiiJo CUII4tfKe ....... loiMdM
I I I I I I I IC85 1U5 15015 1515 15.25 1!35 1545
I I I I I I I
' 15_4
NICKEL PCK2 EST ,tnden;on-O.ilffin:o Normality Tall
ASQu..-d. 1 053 P..v.tue· 0009
15JI911 .. 1715
17,4581 0 111155 0 182415
1717
Mnlm~.m 1.9002 ISI OUI!Irtllt 12.0981 Ml dlao 15 0059 :SRIOUMit 17.9-4S1 Mmtnum 34.2800
as. Contdtnot rnteMtiiOf t.t~
14.8a43 ts.28tt
OS'4Conld..wt nlen'M IDr Sigml
4 0433 4 .32)2
OS!Mo Contderu ltUMI f« Mliditn
148386 15.3818
NICKEL EXP
AndtBon-OIIfttiQ NotmtlitfTtsl
,A,.Squtttd. Dill P.YI IUt 0021
We•n -.,.._ 14.0131 17111
4!.1411 $kt.Mes.t Kur!osis
o ... ,.,, 200151
" ,,. .......... 1$10ulfllt WediJn lrdCk.IJ1111t MI""Um
. .... ••oes
133110 IUOSS 41.1400
115 '4C:Otlftcfrnceln"rwiiDrWw
128711 15.2815 95'1iConlidtncetnll,.,..fofSiQtna
~..1741 1a11l
t 5'4 Cont4tnq '-'* w Mtdilln ll.Olll 110021
NICKEL OCK LMC EST
l<ndl>rson..o.ttngNcwmlljWTest
A-Squwtc~ 2360 P.Valua: 0.000
15.01M3 -43112
181723 .,_. I.OS941
1711
05676 11.0001 150810 17.8358 .,.,..
95% Conllt»ncalnltMllfot MJ
1-48807 1$.2181
95"'Confidtneal~ lorS9fta .tlt813 •U107
9~Con6dinct .,.,.lot ... cl ..
, .. _, .. so ts.ct60
NCKELOKEST
kldlnon-Daltlno Norm~~ Test
ASqu•ed I 744 PNIIut. 0000
Mnim1.1m 1S10ultfilt Mtdan 3tdQJ«t~la
""""""m
1500311 4 41117
19..5338 03151126
105288 171 7
0.5676 119902 15.1481 f1 8M4 426194
9St.4 Conac~tnc. NMl ror !vlJ
1• 1746 151i30
es~ C'oflldenw ll:l..,...lclt Sl9ft• •.2 181 4 5711
95"JG.ConlldtftwlnW't'IIIIOf~.a
1<4 871M 15 4169
Figure 7.1 and Table 7.2: Nickel descriptive statistics
GC EXP OCKICMEST OCKLMCEST PCKEST PCK2EST OK EST
n 1717 124 1717 1717 1717 1717 1717
Mean 13.05 14.07 15.06 15.09 15.06 15.09 15.08
Median 12.98 13.37 15.15 15.08 15.01 15.10 15.15
SD 6.64 6.72 4.55 4.32 4.54 4.18 4.42
Skewness 0.20 0.95 0.38 0.37 0.41 0.20 0.36
cv 0.51 0.48 0.30 0.29 0.30 0.28 0.29
Ill
'
' ...
.... '
' ... '
' lOS
' ' 3.30 ' .... ' ....
' -
' ... '
' 075 ' '"' '
2.2
DS'I.ConfMitncelnlltMlffofMJ
J ' ' ' '
'
'
j ' ... '
' "
' ·- .... 0815 ·= .... . .... ' ' ' ' ' 1::
IS'-Coni~ lni!MWI fOf Mlclan
' ' 22 ·a
U'll Confidence lnteMIIIorMJ
I I I I 080 011 Oil Oi'l
I I I I
j 05" Conlidtttc.lnleMII lor Mtdian
'
' .... '
'
o.lo
COBALT <X:
A""HMrson-a..t~ntNonnthtyTttl
~ 21970 P-Vtlut 0000
..... OltD-453 - 0551730 .......... 0312178 ........ 1.2SM9 ......... 315142 N 1117
M ....... 000200 11tou.tllt 042000 ....... 0 73200 3ntW .... 110400 .......... .......
85~ConfcMnct ~tiM' M.l
0 78301 014500 1
115~ Confldonc. t'lteMIII b' Sign• D I
0..54065 0 57807 I
85'M.Contellr'!Ctii'UMI torMtdien
0.70000 0.78800
COBALT OCK ICM EST
MellniOCI·O•hngNormllrtvTtll
ASQutttd 3 050 P-VtiUt 0 000
...... S10ov V•l.nce Sktwnen KurtOsis N
0123812 0.37110&8 0139920 0.42415$5 0,389970
1711
Mnlmum 0 02168 1s10Urie 06'2111 Mild... 0.80$30 3td ~-- 1 18418 Mwmum 2 t7f»'2
IIS .. ConfidWittln~IOtMJ
0..$10&11 0~15l
15'1. CoAMtnc:U'IIitMI f« Sto"na ' •• '
1 uo
'
' 07
' ' 0.8
'
' ,. '
' ,.
' OJ
95't. Conftdence lnteMIII f« 1\\llfan
' .... ' •• 1 ' . ., ' '
' .., '
' "
' ... ' c - ;::c 0 31tte 0.38702
IS. ConW.nc.tnse,_ fofh'ediM
COBALT PCK EST Jrndtr~on-OartingNomuMyTut
A-5Qvarecl 3 742 p.y.....,. 0000
...... o•134oc SIOov 0.3~007 -" 0 tl42U s.:.wneu 0444113 KurtoSlt 048'2135 N 1717
Mnlrnum 002171 tstOuettllt 0 &3850 ....... 090322 31d0Uiftllt 118366 M:IJimum 2 06845
95'4 Con~dencelntervel ror Ml
0.90513 0 ~088
1)5'1f. Conl!dancaln._MI lOt Sl~t
03$715 031187
ass Con~n<alm.Mtl ror Mtdr.n 0 88318 0 02224
ol1 a.b '
' 0 175
.:.. '
' ....
a5tloConldance hll:lmlllorMtlfan
' 22
95~ Conldance W.el\/81 b rM;
M"'Contdtnoe ~ t«Mld•n
COBALT PCK2 EST
Andenoon-051ing Normafit'/Tnl
ASIJiariiCl 1.366 P-Yafut_ 01102
..... 092113<4 SOO.v 0_3465'J8 \lllrianca 0120129 -... .. 0.320382
""""" 0~16S3 N 1111
Minimum ....... lstOUelfte 067556 ,.den 01U9S5 3u•OU•"IIe 1.15030 Mbllmum 2.&4319
95'!1o Coni"Kiance lntt!MIIIor t.t.1
0.90473 0 .9375-4
' "
D5'Jio Confidence Interval lor stgme
033538 O.!St59
QS'IIo Conrid$11(tlnterval fof t.;•den 089912 0 .938f8
' ....
CCBALTEXP
MdoBon·o.MgNonnelityTesf
.... Sqvttlld 2 948 P.V.tu. 0000
... .. -V«llnCI Sbwntu Kurtosis N
Mrumvm tstCkuwilt
""""' 3nt()J ...
""""'""'
0..02:)4 0548063 0.288184
12426? 178355
124
002000 04t100 ....... 111!iJ50 2te800
05'Ao~~in!wt111forMI
0 74817 0»4330
05'4 Confldtn<tln~MII tot Sgmt
0~552 062400
85t. Conlldlnc.tn•,.,., lot Mtdi-
COBALT OCK LMC EST
Mdtrson-Dtlf11ng Normel?lyTtst
A&lutnul 5.188 P·VtiUI 0.000
...... , SIOov v .. .... Slot .... . t<orsoslt N
0822813 0.381703 0 131553 0448700 038S73CI
1711
Mnlmum 0021S8 1st Ouertite O-'l327 Mtcl., 0 80030 lrdOu.... 117t3S Muiml.fl'l 297042
SIStloConld ..... ~tH'MJ
0805$-t 013998
a5'1Ct Conldtnte rnteMII b' S9'n a 035008 0375'21
IS..Cont~ Htnel .,..Midlln
0.18445 092814
COOALT OK EST
A'tdelSon~Normlfityl'est
~d am p.vatut 0000
.... .. 0123024 - OS74481 y..,.,ct 01 .. 023e SMawnen 0421$~ Kurto111 0373$57 N t717
Mnlmum 0.02158 l stO...-Itlla 061781 Mtd .. 0 00030 3ASOI.IIttllt 11~79 Mt»Mum 287().42
85 .. ContdtnOI ~ttMII br ~
Olil0590 0;.1135
OStlo Confl"ntt \'!!eMf tr Sfgme
038238 0387-'5
SIStloContdtnctlrCtP'o'ft t)rMtdtan
0 87900 0 .1211•
Figure7.2 and Table 7.3: Cobalt descriptive statistics
GC EXP OCKICMEST OCKLMCEST PCKEST PCK2EST OK EST
n 1717 124 1717 1717 1717 1717 1717
Mean 0.82 0.85 0.92 0.92 0.92 0.92 0.92
Median 0.78 0.68 0.91 0.91 0.90 0.92 0.91
SD 0.56 0.55 0.37 0.36 0.37 0.35 0.37
Skewness 1.26 1.24 0.42 0.45 0.44 0.32 0.42
cv 0.68 0.65 0.40 0.49 0.40 0.38 0.46
112
' "
' 10
'
' ..
I I I I no 210 J2o :no
' ' '
' .. OS" Contdtnct ht.rval b' Modian
I I I I 731 18D DJ 310
' ' '
95~ Conft.c»nu tnleNIII for"'
' "' I
.:, '
MAGES1UMGC
Pndtrson-O.tngNormaM)'THI
.A.Squ.-ed S\)..404 P..V&Iut 0.000
Mttn 62.4090 sc... 52.7288 Valiance 2760.33 Skewness 1.511715 Kurtosis 2 .7278g N 1111
MnimLm 0100 1S10tutl111t 24.518 MJdi"lllt 48616 3tdou.r... 83~9 Ma:lomum 378.930
ts«.Cofltct.nc..-arvaJbrJ,t,, 5991:! G4.905
15,. ContMnce blel\'liJ tlrSigme
51022 54.5~
9S,.Conldtnce nt.MlforMt<Jian
•ruse su1e
MAGESIUM OCK lMC EST Mderscn.o..ung Nonnality Test
A-Squertct 78.145 P-V•tur 0000 ..... 85.6754 ..,.., 50902< -" 2SOI.tl$
""""'"' 170070 ........ 3.30107 N 1717
M n!Mum 1243 1A0u ..... 53.8\.4 Mod'" 71704 3fdOu~· 100734
""""'""' 378883
05'11. Conlldencelntarval for" t..tJ
83.288 88085
115% Confidtnc.lntff'llelfor Sigma
40255 5:ues
OS,._ ConlldtnttlnltMII tor M!tcfian
159.004 73.730
' .. ' .. '
' .. 1~ ~ ' '
9SW. Conldtnc. hteM! lcf Medum
~ ' • ' ' .I "' ' - ' ' 230 2111
I I
OS'M!Contden~ lnteMI brMJ
' ...
' 15
'
' ' ,,. ... '
CCl I
10 I
o::J ' "' '
os• Conldtnc• heM! tlt Mtdl"an
' .. I
MAGNESIUM EXP
Jtndefton-OarlingNonnabtyTest
..-..squ~~n~d. 1.334 P.V•kJe 0000
~an 7U556 StOliY (17,!5118 V.arianc:e 4557 as SkeWnt$$ 2.0811<4 KurtoSIS 4 .G52.57 N 12•
Mninnm 8.070 ts.t Quaflllt 33.545 Mtdian 58 888 lrd 0'*"- 92 713 Ms:JOmum 378..930
OS,. ContdtrM w.rvetlbf ~
66455 1NJ458
os• Contdet'ICt Htrvat * stom• 60027 17.147
111514 Conld•nce ht•rval for Mfdltn 47404 69.374
Magnesh.m ~ Estimates .And•JSon-Oartlng NonnalltlfT•st
ASquartd sa 863 P-Vatue. 0000 ..... 17'5752 ..,, ...... , ......... 3138.02 Sbwnns 170588 KudOS!$ 2.07187 N 1111
Mnimum 8070 bt0Jartile 52075 Mtdtan 1 \ 000 3rd0Ualtil• 102525 Mmmum 378 030
OS~ Conldenea htorval l:lr ~ 84 g, 90.226
OS" COt'lld¬ hi& Milot Sigma
~ 188 '57.93t
95" Conldoncal'lleMI br~_, Ul-40 72.128
Figure 7.3: Magnesium descriptive statistics
Table 7.4 : Magnesium descriptive statistics
GC EXP OCKLMCEST OK EST
n 1717 124 1717 1717
Mean 62.41 78.46 85.68 87.58
Median 48.62 58.69 71.70 71.00
SD 52.73 67.51 50.90 56.00
Skewness 1.51 2.06 1.70 1.71
cv 0.84 0.86 0.59 0.64
113
' "'
I
"'
I
"' I
,J, I
' I ,.. ... I I
I ... I
' ... '
I
"" '
95"4 Car\~tnce lntlrvallor MJ
I ... :J. I
IS~ConAOtnu: ln_....ku'Midt.en
' ... I
' '"
I
I ... I
,.,
I
"'
I 40
I
I
IRONGC kldftt:on.()ert~ngNonnllfity'Tes1
'"~·- 731-4 P-Y.We 0000 ..... ..,.... YIIIMCe Sl..•wn•ss Kurtosfl N
313.830 1-4-4.513 20889.9
8.19E-02 -28E-01
1717
Mr~lmum 9600 1st OUirile 210 852 Mechl'l 318.208 3rd Ou•ntte 41107.298 h'e»mum 813.792
85'11!>Cofti!dtnc.ln~larM.I
306.889 310671
ts'I'COflftdentelnlllmllforSigm•
131 156 140.SS7
9S'4Confldtnc.lnt«'4*-'r~"' ~,n.....,. ,.~,:\
IRONOCK ICMEST Mdtrlon.O.tllng No~malilyTo~t
A.$o.llfed. 1.067 P-YIIIue. 0008 ..... SlOe• V.tlenn Sbwntst ........ N
t.~lmum
I"OU-* -·· 3tdOuerwe ...........
348.488 97.072
9423.01 -1.8&02
0..2G7661 1717
11.521 284015 350336 413774 ns.540
05'1. Confidence Inti,.... I« MJ 343193 353033
li:SIJ. Confldenceltdorval rorSigma
8U31 100.433
$5'4 toniiO.nC'O intervllfort.ioeOIM
S48.011 357.545
IRONPCK EST Mct.rton-OerHng NonnefilyTest
~~!- ·~~ 348.680 17618 ......
-IUE-02 0.1508&11
1117
MnlmuM 104SI9 111 Ou~tt~l• 281.875 M$d'tn 352.486 Jrd Outllllt <1 18281 M;)idmum 178 647
95" Confidec~ttlnleMII for M..1
344 056 351.303
9S"' Confldenc.lnttMII fof ~a t.4SI7 1010W
IS .. Conadenc.tlnbiMII fDf Mediaft
348 .tSt 358399
' ,., I
I I 040 740 I I
85~ Conl dtnct lnlttvel br Medlen
0 I ... 1111 I
IRON PCK2 EST
.tnttmM-O•IIngNonnaldyTest
A-Squamlf· 1.332 P-Vtlue. 0.000
...... 347.079 SIOev 93139 Verience 8674.95 S~IUSS - 1.3f-01 Kurtosfs 0.352736 N 1117
.,.n~mum 6632 1st au.rtil• 290032 Median 348111 ~Ouar\Je -408.133 MlliiMUm 747.648
8S,Contdence IJUMI'I b"MJ
3<42170 3514t7
85"' Conldenu hltMII tor Slgm•
ao.12s a&,3S4
95' ConM•nc• lh~MIII br Median
3<43 658 353.536
IRONEXP Mderson.O,..,.gNOfl'l'lacittTetl
A.Squ•ed 0..381 P.VIlte· 0.!80
...... SOw
""'"''" S<twnu'l Kultosis N
318.617 144.176 20nea
0.215086 -U£-01 ...
Mnlmum 11.000 lstOUarote 221.851 ~i:an 3 14899 ltd ou.rru ~8.501 MD~moo. nu92
95'4Conld«tcelntMvattotMJ
202..18i 344246
1511. ConfldMce III'IRtNII tor Signa
128 191 1&4.763
15'4Conldttltto Nth!~~ lOt MlldiM
288271 "33481&
IRON OCK LMC
Hldtl'tOn.01111ing Norma 4ft Tut
A.Squate<J· 1.228 P-VI!ue: 0003 ..... SOw -" Sbwness Kilnosis N
311a.nc i<I 4 7S
mssa -7.4£02
0295871 1111
11..522 286.19!il 3503JG 413.774 na.S40
85!. Contdenctln!VfVall« MJ
344 302 353.245
t5"'Conll6ence lnteMli iOJSiO'fll
91<418 97.7<48
OS~ Confidence Interval lor Mtdl.n 344 5e9 3!5&.9&2
IRON OK EST
lnct.Ron--Oeding NonnUt(TMt
AScp•W 2.407 P.Vtklt 0.000
347659 978:11 .... ..
-1.1E-01 01D5378
1711
Mnlmum 11 522 Ill 0ul!lfDie 281.132 tw\)dan 354.6&1 3rd0uartilt 415216 Ml:ofmum ns 540
OS' Ccnlidlnce lntetv.!l lor Ml
30029 352-280 t511o ConldtnctlnleMif lOt SIOMI
suss 101208
8S'Io Con6dtnc-.lnteMIIIof ~lltl 3o48t84 360429
Figure 7.4 and Table 7.5: Iron descriptive statistics
GC EXP OCKICMEST OCKLMCEST PCKEST PCK2EST OK EST
n 17 17 124 1717 1717 1717 1717 1717
Mean 313.83 318.62 348.49 348.77 348.68 347.08 347.66
Median 312.19 315.00 350.34 350.34 352.49 348.11 354.66
SD 144.53 144.18 97.07 94.48 97.68 93.14 97.82
Skewness 0.08 0.22 -0.03 -0.07 -0.09 -0.01 -0.01
cv 0.46 0.45 0.28 0.27 0.28 0.27 0.28
114
I I """ .. ,. I I
,-. 24~ I I
1- I
I I .,.. .... I I
riD .Jm ~ Jo 2~0 23110 ~ ~ ,k
I I I I I I I J
I I • 20o 4150 I I
ZINCOC
~·l•d 3.571 P~Yifue. 0.000
'"'"" '2313.91 - 1054.28 V.l'ian(e 1111509 SktwnHS -3.1£..02 KLirtosls -1~-01 N 1717
Mnlmum 8600 1st I:Aialtll• 1589.00 MedlM 238300 !kdau.... 305901 ~um 1831.99
85'4 Conld<Mc. lrfth31 roc MJ
2204.01 2363.81
85ft ConW.nu lntflmlllot sqna 1020 18 1090.78
951. COnedenctfn~ foth4t<lfln
2J)() 00 2442-01
ZJNCOCKICM
N!Otft Oft.O.IIU'IO Norme!Jiyltsl
A.$q,l•ed. 0363 P-YIIu.- 0.441
...... ..,, -· .......... -~ H
Mnlmum htOU ... ....... 3«10Ualllla Ms.limum
2300SS ... ,. 444227
88<E-02 -3..5E-02
1717
90.-42 1856.28 228122 2762..82 4915.84
$5'1. ConlldoncalnteMIIIor Ml
2.27500 2338 10
05'11. Coftlldtnte rnlfii'Val tor Sigma
844 »3 68958
85'1 ConlldencetntaMiforMN&n
126133 2331.88
ZINC OK EST
lndefs~o.lng Nonntittlest
ASquand 0305 P."AAtH 0567 - 1:30774 - 66607 ......... -443&116 Sktwn.st 5S3E-02 tQ.Inosls -lf£-02 N 1711
Mnlmum 9042 l iii t OUel1ilt 1851.56
""'"" 2309.55 3f'd0Uif01t :U8170
""""''m 4 915.84
05'1tConMenc.lttteM!lfforMJ
227$.22 233V2l
IS 'I. Coftlidtnct IM.....W b Sqna
144.51 8&913
f5"' COMdtn<'t kl.,. fofMfdan D81.3J 234732
IJ.o -d.. I I
~0 22.~0 2~ mo ~0 73110 TJ~ mo ~ Jso
I ... I
I
""' I
I I
I
'""'
06 .. COnfidence SnlliiMI IOt MJ
I I I I I I I -""' 2315 ""' ""' 1345 2355 I I I I I I I =-- -~~-
I
""' I
ZINCEXP
/ftdtfSOftoo.lng Notmet.ty Test
AScaua""ct OMQ P.Yalua· 0.:2<4-4
..... so.. V.1Janc. SktwnHS t<tutosls N
Mnlmum tst OJaltlle -3rdOJri• -.um
2130.13 04431
891719 0.331 859 -1.3E.02
ll< .... 153375 1089.00 2859.50 .. 918. 10
OSII.CoMdMce~tort.\.1
1ge227 12i1.00
15,., Conldl-.lnlert'ill b-Si1Jn• 8l0.S1 107008
9s-4Confid..,('fln!01n'tJIOt._VIil
1869 &7 2287 78
ZINC PCK EST
Andlf&On-04111nQ NotmelltfT"I
A.$(JJ•ed. 02o48 P.V.rue 0.151 - 2308.20 ..,., &6581 -. ... .. 33011 Skewness 2.18E·D2 ........ -8.3E-02 N 1111
Mn.mllfl'l .... 1.510Jde 1862.17 -M 23091& ltd0UtJ'I1• 111.4111 Malmum .41912.69
95" Confidence Interval fOf MJ
2274.74 2337 71
os-.. Conftdonce lnuuval lor Sigma
&11427 688.86
95'4 Confidence kllent'lllor Mldi10
1261.87 :2347.16
ZINC PCK2 EST Mdtnoo.o.ting Nonne~a!r Test
ASQu.,.d' 1.122 p.v.tue 0000
..... -Vanence .......... Kurtosis N
hfnimum l t lOUaniil• ........ 3rd0u.-11le Mamum
2310.29 600.30 370137
·U!;.-01 028S42o4
1711
82.69 1~5.00 232611 17:2031 ~1.88
DS-1. ConMencelnteJWtlor Ml
2217 50 n45.09
iS'~. CGnldlnceMtw'tlltor SH}m•
51170 61845
ts,..Centctenulntttwt torMtOten ?111 :)11 ?:\57AA
Figure 7.5 : Zinc descriptive statistics
Table 7.6: Zinc descriptive statistics
GC EXP OCKLMCEST PCKEST PCK2EST OK EST
n 1717 124 1717 1717 1717 1717
Mean 2312.91 2130.13 2306.55 2306.26 2316.29 2307.74
Median 2383.00 1989.00 2281.22 2309.16 2326.11 2309.55
so 1054.28 944.31 666.50 665.81 608.39 666.07
Skewness -0.03 0.33 0.09 0.02 -0.02 0.06
cv 0.46 0.44 0.29 0.29 0.26 0.29
115
•
Next we consider the question of conditional bias by assessing the Q-Q plots
of the grade control data and estimates of each variable for each of the kriging
techniques used. These plots are displayed in Figures 7.6 to 7.10. It is clear from
these that the overe~timation of the mean for the nickel, cobalt, and iron variables has
resulted from significant overestimation of the low values and underestimation of the
high values. Overall nickel and iron appear to be the least affected by conditional
bias. As all of the kriging techniques produced similar estimates for each variable, as
was expected, there is little differellce in the Q-Q plots .
The underestimation of high values is not too marked for nickel for all
techniques except the principal component kriging where only two principal
components were retained. This technique has grossly underestimated the
particularly high values. The full principal component kriging has the best
performance with regard to underestimation of the high values. There has been
significant overestimation of the low values, particularly mound the 5111 to 201h
percentiles. There is little difference between the techniques in the overestimation of
low values although after the 251h percentile principal component kriging performs
marginally better than the others.
For cobalt the underestimation of the high values is significant regardless of
the kriging technique. In this case the principal component kriging (two principal
components) was the worst offender. There has been some overestimation of low
values for this variable, again particularly around the 5th to 20th percentiles. Overall
principal component kriging is marginally better than the other techniques here,
although all perform similarly.
Magnesium has suffered overestimation across all values although the
cokriging estimates perfonn somewhat better than the ordinary kriging estimates
from the 801h percentile. Magnesium was particularly difficult to model with very
erratic semivariograms. Little benefit was gained from examining the other measures
such as the madogram and pairwise relative semivariograms, though these measures
did assist in better detining the ranges of the model. The poor appearance of the Q·Q
plot for this variable indicates that the model has not ade.J.uately captured the erratic
behaviour of this vmiable.
The underestimation of high iron values is not as marked as for the other
variables discussed so far. In this case the principal component kriging method
116
obtained using ~mly two principal components is the worst offender, though certainly
not as significant as seen previously. Ordinary kriging has the best perfonnance from
the 95th percentile. The overestimation of lower values for iron follows a similar
pattem to nickel and cobalt with little difference in techniques.
For zinc the underestimation of the high values is significant regardless of the
kriging technique. Again the principal component kriging (two principal
components) was the worst offender, though only marginally. The best perfonner for
this variable was the ordinary cokriging using the intrinsic correlation model. There
has been some overestimation of iow values for this variable, particularly around the
5th to 15th percentiles. The only technique that differs from the others at the lower end
of the plot is principal componem kriging (two principal components) with
mru.ginally higher overestimation from the lOth to 35tl' percentiles.
117
...... ....... 00
5 10 15 20 25 30
Nickel Grade Control Data
-+- OCKICM
- OCKLMC
- PCK
- OK
--PCK2EST
35 40 45 50
......
...... \0
"l'1 ~-'"I (1)
-....)
-....)
10 I
10 "t:) -0 ..... Cl)
~ '"I
8 CJ E.. .....
~ .. ~
.§ .. rl.)
~ .. -~ ,.Q
= u
6
5
4 -+- OCKICM
--. QCKLMC
......_ PCK 3 --11 - OK
--+- PCK2 EST
2
1
0 ~------~------~----~------~-------,------,
0 1 2 3 4 5 6
Cobalt Grade Control Data
u ~ ~
~ 0
0
T
f
0 0
0 0
0 0
0 0
0 0
~
0 ~
0 ~
0 ~
0 ~
~ ~ ~ ~
N
N
""""" """""
s~ltnnps1[ mn!s~u~uw
Figure 7.8: Q
-Q plots for m
agnesium
0 0 ~
8 ~ 0 0 ~
0 0 N
0 0 """""
0
:3 =
Q -~ =
Q
u ~
"C
f: ~
~ .... riJ
~ ~ =
~ 120
d u ~
-~
f2 ~
~ ~
~ 8 0
0 ~
~
t ' f T
f
0 0
0 0
0 0
0 0
0 0
0 0
0 0
0 0
0\
00
t"'--
\0
l/') ~
M
N
S~JUUI!JS:i[ U
O.IJ
Figure 7.9: Q
-Q plots for iron 0
0 0 ~
0 0 0 ~
0 0 00
0 0 \0
0 0 N
0
121
8000
7000
6000 f'-2 ~
"Tj ~ 5000
~- a ·-.., -Cll ~ 4000 :--l - (J
0 = /:) N 3000
I
---+- OCK ICM
---....-- PCK
- OK
- PCK2EST
/:)
'§- I - 2000 [/}
0> .., I 1000 ~-g
0
0 1000 2000 3000 4000 5000 6000 7000 8000
Zinc Grade Control Data
-N N
Finally we examine the estimation variance for each of the kriging methods
as displayed in Figures 7.11 and 7.12. For comparative purposes we display the
estimation variance for ordinary kriging using the standardised exploration data. The
estimation variance is clearly highest in the far north~westen comer, along the north
western boundary, in the far south-eastern comer and along the southern boundary of
the study region. These areas all fall outside the region bounded by the exploration
data, 1· _. we are extmpolating rather than interpolating in these areas. The
estinuion variance is lowest around the locations of the exploration data, and is, as
expected, zero at these locations {kriging is an exact interpolator). The estimation
variance was identical for all of the variables estimated using the intrinsic
coregionalisation model. The estimation variance of the principal components gets
successively smaller with each principal component.
While the estimation variance is expected to be lower for the cok~ging
estimates this was true in our case for nickel and iron only for the intrinsic
coregionalisa!ion model and nickel only for the linear model of coregionalisation.
Overall the estimation variance for nickel was lower around the locations of the
exploration data, however as we move further away from these points the estimation
variance becomes the same as the ordinary kriging estimation variance. The
estimation variance of the intrinsic coregionalisation model is lower still than the
other methods. For cobalt the ordinary kriging estimation variance was the same as
that for the intrinsic coregionalisation model, both of which were belter than those of
the linear model of coregionalisation. For magnesium and iron the ordinary kriging
estimation variance was lower overall than the other methods. For zinc the intrinsic
coregionalisation model yielded the lowest estimation variance.
123
Estimation Variance
~ t-ac:kri·LMC
" . ·
!l '' " ....
E!tilmtionVJrfance ~ Cobalt·LMC
I ,.
EstlllllGon'hllanet o MagllfSIUm • LMC ll . .
Estimation Variance ~lrOn·LMC
:..
- ~ 1o
:;: .. . .. . ,.,
Estimation Variance E; Zinc-OK
•· ....
....
.. ··= · ..
l" w ···... .... ,,.,., .... ·...
. ...
,. --~ --.,
1480
Estimation Variance ~ lron·OK
...
Estimation Variance
!!.
.~ -
5: ...
·.:. : ::. ~ :::: . ~. . .: ..... . ~~ ~ : .. ..... .
1
.. '·'"" ... ··-···.... ~ . .. ...
~.
~ -----·~ ..
.. ..,
1
.. .. ~ ··... ·... .... ·-
1
.. · ·~ ... ... . .. . ... ·.. ~ . ..
....
1
.. .. .. ··... ···-·'""
Figure 7.11 : Post plots of estimation variance for ordinary co kriging and ordinary
kriging
124
&Jtimlllion v.rtonn ~ PC1
eslimotion ""''"" ~ PCJ
r -·-·-"" ·~ ·-"~ ·-·-
....
r .. ~ ·-... ·-·-·-'"" ... "~ ·-
1400
Eltimalion VarilflCI
~ PC2
I . g.
~ 500
EsUmltion variance
~ PC4
~
'""
1
.. .. ~ ··,.,. ... ·... ·... "~
1490
Figure 7.12: Post plots of estimation variance for principal component kriging
.·
125
8 DISCUSSION AND CONCLUSIONS
D~ta collection in the earth sciences is rarely limited to just one variable at
each location. Such is the case in mining, where it is common that each drill core
taken will be sampled for not just the primary variable of interest but also for other
attributes. These secondary variables are often spatidly cross correlated with the
primary variable as well as each other. Hence each of the variables sampled may
contain useful information about the others. Multivariate geostatistics allows us to
exploit the relationships among and between the variables with the aim of improving
their estimates at unsampled locations. This is particularly the case when the primary
variable of interest is undersampled in relation to the secondary variables.
This thesis presented the theory and the process of the modelling and
estimation of spatially dependent multivariate data. These methods were illustrated
by application to a multivariate data set from the Munin Murrin nickel mine near
Laverton in Western Australia. We have demonstrated the use of the classical
multivariate statistical technique principal component analysis in a geostatistical
environment. We have exhibited the development and application of three models of
spatial contin11ity, namely the intrinsic coregionalisation model, the linear model of
coregionalisation and the linear model of regionalisation. We have used each of these
models to estimate values of the attributes at unsampled locations using ordinary
cokriging, principal component kriging and ordinary kriging.
The data suite MM22D used in this study came from an actual mineralisation
and consisted of two data sets, a grade control (exhaustive) data set and an
exploration (sample) data set. As the data ai"e isotopic we did not have the case where
the primary variable is undersampled in relation to the secondary variables. This
would be unrealistic in a mining situation when all data are coming from the same
drill core sDmples. The data suite comprises three dimensional grade and thickness
measurements on eight variables: nickel, cobalt, magnesium, iron, aluminium,
chromium, zinc and manganese. For the purposes of this study the data were
transfonncd to accumulations, hence were considered two dimensional. In order to
demonstrate the different types of models of spatial variability yet remain within the
126
scope of an honours thesis we decided to limit the analysis to two four variable
subsets created from the exploration data. One subset consisted of highly correlated
variables (nickel, cobalt, iron and zinc), the other subset was chosen for the
economic importance of the variables (nickel, cobalt, magnesium and iron).
A detailed exploratory data analysis was carried out on each of the variables
in the MM22DGC and M:M:22DEXP data sets. Overall the exploration data
reproduced the grade control summary statistics well. Cobalt, aluminium,
magnesium and manganese were all strongly positively skewed for both the grade
control and exploration data. It was however decided that a transformation of these
variables was not wmTanted, A principal component analysis was performed on the
MM22DEXP data set and it wus determined (by examining the cross correlograms of
the principal components) that the data were not intrinsically correlated. We then
considered the principal component analysis of the two four variable subsets
M1v122DHC4 and MM22DTOP4. It was determined that the former subset was
intrinsically correlated. In both cases the principal components extracted could be
reasonably well interpreted.
The first estimation technique we investigated was ordinary cokriging. In
order to implement the cokriging algorithm we required a permissible model of the
spatial variability of the variables being estimated. The first model that we
demonstrated was the intrinsic coregionalisation model using the MM22DHC4 data
set. The intrinsic coregionalisation model is a particular example of the linear model
of coregionalisation where all of the sills of the basic scmivariogram models are
proportional to each other. Under the assumption of secondMorder stationarity the
matrix of coefficients was for our data set the correlation matrix. As we were dealing
with standardised data we required that the sills of the basic models summed to one.
The actual modelling process then for this data set was relatively simple in that it
consisted of identifying the basic structures that captured the spatial variability of the
variables and determining their corresponding sills. The conditions for this type of
model though are restrictive and in practice the intrinsic model rarely fits
experimental corcgionalisations (Gooevaerts, 1997, p. 117).
The second model we investigated was the more general linear model of
coregionalisation which was demonstrated using the MM22DTOP4 data set. The
difficulty with this type of model lies not only in the fact that each of the
127
Nv(Nv + 1)/2 (10 in our case) direct and cross semivariogram models must be jointly
inferred, but that each of the coregionalisation matrices must be positive semi
definite. In order to ensure that the linear model of coregionalisation we were
building was indeed permissible it was necessary to first detennine jointly the basic
set of stn:ctures that best captured the characteristics of the direct semivariograms
(nickel, cobalt, magnesium and iron). It was then necessary to restrict our choice of
structures for the cross semivariograms to those used for the direct ones. Having
obtained a model with a suitable fit on all direct and dross semivariograms it was
necessary to make the coregionalisation matrices positive semi-definite. As we were
dealing with fairly small coregionalisation matrices we did this manually using the
property of diagonal dominance. This process can indeed be tedious, time consuming
and frustrating, and in addition may not be optimal, so for larger coregionalisation
matrices an iterative procedure, such as that outlined in Goovaerts (1997) is
recommended.
The estimates obtained from cokriging using both models were compared to
the grade control data, which we considered to be reality for this study. We
compared the postplots of the estimates with those of the grade control data. In both
cases the cokriging estimates captured the overall spatial distribution of the grade
control data. In addition we compared the descriptive statistics of the estimates
obtained for each variable with those of the grade control data. In all cases the mean
and median of the estimates has over estimated that of the exploration and grade
control data. The only exception is zinc where the mean and median of the estimates
were close in value to those of the grade control data. The standard deviation of the
estimates from both cokriging cases was considerably lower than that of the grade
control and exploration data, which is an indication of the smoothing of the
variability due to the cokriging process. The residual analysis for both cokriging
methods showed the residuals to be approximately nonnally distributed with the only
exception being cobalt. In both cases the histograms of the cobalt residuals were
negatively skewed.
The second estimation technique we investigated was principal component
kriging. The overriding assumption of this technique is that the data are intrinsically
correlated. Hence we used the principal components extracted from the MM22DHC4
data set for this method. In order to perfonn the ordinary kriging of the principal
128
component scores we were required individually model the spatial variability of the
principal component scores using a linear model of regionalisation. We then used
ordinary kriging to estimate the principal component scores at unsampled locations
and recalculated the estimates of the original variables using the corresponding
coefficients from the matrix of eigenvectors. In addition to estimating the variables
using all four principal components we krigccl only the first two (which accounted
for 89% of the total variability of the original data) and calculated our estimates
using only these.
The estimates obtained from both forms of principal component kriging were
compared with the grade control data. Overall, for all variables, the post plots of the
estimates in both cases have reproduced those of the grade control values reasonably
well. In addition the descriptive statistics of the estimates were compared to those of
the grade control data. Once again the mean and median of the nickel, cobalt and iron
estimates have overestimated those of the grade control data while those for zinc arc
close in value. The standard deviation is considerably lower for all four variables
estimated than for the grade control data. The residual analysis confirmed that the
residuals were approximately nonnally distributed with cobalt once again being the
only variable to not conform this. What is most significant here is that the estimates
obtained by using only the first two principal components are remarkably close in
value to those obtained using all four.
The ordinary kriging estimates were also obtained. for the nickel, cobalt,
magnesium, iron and zinc variables. As for the individual principal components this
involved obtaining a linear model of regionaHsation for each of the variables
individually. As the data we used were isotopic it was anticipated that the estimates
obtained in the ordinary kriging method would be similar to those obtained from the
other methods. This was indeed the case with all estimates for each variable being
almost identical. In particular in the case where the data are intrinsically correlated
the ordinary cokriging and ordinary kriging estimates are expected to be equivalent.
This was especially reflected in the Q·Q plots where the dark blue ordinary cokriging
(intrinsic coregionalisation model) line is completely covered by the orange ordinary
kriging line.
The only significant difference between methods was for magnesium where
the ordinary cokriging estimates slightly improved on the overestimation experienced
129
across all values for this variable. The only other major difference between methods
was that estimates obtained from the principal component kriging using only the first
two principal components consistently underestimated the higher values to a greater
extent than the other methods.
In conclusion then the estimates obtained from the various kriging methods
have confirmed that in the case of isotopic data, in particular when the data are
intrinsically correlated, there is little practical benefit to be obtained from cokriging.
The true benefits of cokriging are only fully appreciated when the primary variable is
undersamplcd in relation to the secondary variables and those secondary variables
are well correlated with the primary variable. A combination of technological
advances has lessened the practical and computational disadvantages of multivariate
geostatistical methods. There have been marked improvements in the development of
geostatistical software that allows interactive modelling simultaneously of both direct
and cross semivariograms. In addition there are now available automated iterative
procedures, such as those offered in the software package AGROMET, for
determining the positive semi-definite coregionalisation matrices. With increased
processing power the computational expense solving of large kriging matrices is
becoming less of an issue.
This study also raised the question as to the benefits of using principal
component kriging. As this technique relies on the data being intrinsically correl~ted
we are able to individually model and krige the principal components. However if
the data arc intrinsically correlated then there is no benefit in using cokriging as these
estimates are equivalent to the kriging estimates. Either way we are required to
individually model and krige either Nv variables or Nv principal components. From
the estimates obtained in this study using principal component kriging with both four
and two principal components there was little difference in the results. This would
suggest then that the renl benefit in using principal component kriging is when we
are able to reduce the number of principal components retained, effectively reducing
the number of variables to be modelled and estimated. To be of real benefit however
one would want the number of principal components to be retained to be
significantly less than the number of original variables, yet still account for the
majority of the variability of the original data.
!30
REFERENCES
Afifi, A. A., & Clarke, V. (1996). Computer-Aided Multivariate Analysis. Boca
Raton: Chapman & Hall.
Annstrong, M. (1998). Basic Linear Geostatistics. Berlin: Springer-Verlag.
Bleines, C., Deraisme, J., Geffroy, F., Jeannee, N., Perseval, S., Rambert, F., Renard,
D., & Touffait, Y. (2001). !SA TIS: Software Manual (3rd ed). Paris:
Geovariances.
Chiles, J.-P., & Delfiner, P. (1999). Geostatistics: Modelling Spatial Uncertainty.
New York: Wiley-Interscience.
Datta, B. N. (1995). Numerical Linear Algebra and Applications. Pacific Grove:
Brooks/Cole Publishing Company.
Deutsch, C. V.,& Journel, A. G. (1998). GSLIB: Geostatistical Software Library and
User's Guide (Seconded.). New York: Oxford University Press.
Dun ternan, G. H. (1984). Imroduction to Multivariate Anolysis. Beverley Hills: Sage
Publications, Inc.
Goovaerts, P. (1997). Geostatistics for Natural Resources Evaluation. New York:
Oxford University Press.
Goulard, M., & Voltz, M. (1992). LinearCoregionalization Model: Tools for
Estimation and Choice of CrossVariogram Matrix. Mathematical Geology, 24
(3), 269-286.
Hotel ling, H. (1933). Analysis of a complex of statistical variables into principal
component;;, Joumal of Educarional Psychology (24), 417-441.
Isaaks, R. M. & Srivastava, E. H. (1989). An Introduction to Applied Geostatistics.
New York: Oxford University Press.
131
Johnson, R. A. & Wichern, D. W. (2002). Applied Multivariate Statistical Analysis
(5th ed.). New Jersey: Prentice Hall.
Journel, A. G. & Huijbregts, C. J. (1978). Mining Geostatistics. London: Academic
Press Inc.
Kanevski, M., Chernov, S. & Demyanov, V. (1998). 3PLOT [Computer Software].
Moscow: ffiRAE.
Lay, D. C. (1997). Linear Algebra and its Applications. Reading: Addison-Wesley.
Matheron, G. (1970). The Theory of Regionalised Variables arzd its Applications.
Fasiscule 5, Les Cahiers du Centre Morphologic Mathematique, Ecole des
Mines de Paris, Fontainebleau.
Murphy, M., Bloom, L. M., & Mueller, U. A. (2002). Geostatistical Optimisation of
Mineral Resource Sampling Costs for a Western Australian Nickel Deposit.
IAMG 2002: Proceedings of the 8 11 Annual Conference of the International
Association for Mathematical Geology. (Volume I, pp. 105-110).
Pannatier, Y. (1996). VARIOWIN: Software for Spatial Data Analysis in 2D
[Computer software]. New York: Springer-Verlag.
Pearson, K. (1901). On lines and planes of closedst fit to systems of points in space.
Philosophical Maagazine, 6(2), 559-572.
Wackernagel, H. (1998a). Multivariate Geostatistics. Berlin: Springer-Verlag.
Wackemagel, H. ( 1998b ). Principal Component Analysis for autocorrelated data: a
geostatistical perspective. (N-22/98/G). Fontainebleau: Centre de
Geostatistique. Retdeved 5 November, 2002 from
http://citeseer.nj.nec.com/cache/papers/cs/5260/ftp:zSzzSzcg.ensmp.frzSzpub
zSzhanszSzpcaeof.pdf/wackcmagel98princip&l.pdf
132
APPENDIX A: PRINCIPAL COMPONENT ANALYSIS OF
TRANSFORMED DATA
Here we consider the appropriateness of the transfonnation of the strongly
skewed variables cobalt, magnesium, aluminium and manganese. We firstly review
the eigenanalysis of the original data as covered in section 3.4.1
Table Al: Eigenanalysis of the correlation matrix of MM22DEXP
Eigenvalue 4.293 1.!84 1.047 0.607 0.326 0.279 0.154 0.109
Proportion 0.537 0.!48 0.131 0.076 0.041 0.035 0.019 0.014
Cumulative 0.537 0.685 0.815 0.892 0.932 0.967 0.986 1.000
Variable PC! PC2 PC3 PC4 PC5 PC6 PC7 PCB
Nl -0.409 0.082 0.133 -0.277 -0.401 -0.739 0.096 0.121
co -0.407 0.153 0.368 0.201 -0.258 0.233 -0.213 -0.685
MG -0.180 0.650 -0.365 -0.571 0.186 0.195 -0.063 -0.095
FE -0.416 -0.288 -0.168 -0.004 0,343 0.036 0.730 -0.253
AL -0.243 0.301 -0.565 0.698 -0.062 -0.154 -0.073 0.106
CR -0.340 -0.422 -0.331 -0.201 -0.510 0.459 -0.130 0.261
ZN -0.404 -0.320 0.007 -0.055 0.582 -0.172 -0.599 0.065
MN -0.351 0.305 0.507 0.162 0.130 0.308 0.169 0.600
134
4-
3-<J) ::l
Cij > c 2-<J) O'l [j]
1-
0 I I
~ ~ ~ ~ ~ ~ 1 2
Component Number
Figure AI: Scree plot of eigenvalues for the principal component analysis of
MM22DEXP
We then performed a principal component analysis on the correlation matrix
of the raw nickel, iron, chromium and zinc data together with the log transformed
cobalt, magnesium, aluminium and manganese data. The results of this principal
component analysis are very similar to those obtained using the raw data and are
displayed in Table A2. The principal components extracted 58.2%, 15.7%, 8.7%,
6.6%, 4.0%, 3.8%, 1.9% and 1.0% of the total variance. Only the first two
components have eigenvalues greater than one and together account for only 74% of
the variability of the original data. By incorporating the third principal component
this value is increased to 82.6% and by the fourth principal component we have
accounted for 89.3% of the variability of the original data. As with the untransformed
case the scree plot in Figure A2 reveals that the eigenvalues appear to level off after
the fifth component. Again the eigenvalue for this component is small, only 0.322,
hence it would be of no benefit to retain it. As for the untransformed case if data
reduction was the objective only the first four principal components (which account
for 89.3% of the variability of tbe original data) would be retained.
135
Table A2: Eigenanalysis of the correlation matrix of MM22DEXP with log
transformed cobalt, magnesium, aluminium and manganese
Eigenvalue 4.659 1.259 0.694 0.530 0.322 0.306 0.150 0.080
Proportion 0.582 0.157 0.087 0.066 0.040 0.038 0.019 0.010
Cumulative 0.582 0.740 0.826 0.893 0.933 0.971 0.990 1.000
Variable PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8
NI -0.388 0.071 0.230 -0.203 -0.645 0.564 0.031 0.123
TRCO -0.412 0.198 0.141 0.377 0.212 0.178 -0.108 -0.735
TRMG -0.234 0.609 -0.131 -0.667 0.094 -0.293 0.053 -0.124
FE -0.386 -0.366 -0.022 -0.121 -0.119 -0.359 -0.746 0.041
TRAL -0.295 0.086 -0.869 0.307 -0.175 0.019 0.111 0.117
CR -0.327 -0.450 -0.137 -0.364 0.587 0.397 0.173 0.089
ZN -0.382 -0.325 0.215 0.061 -0.191 -0.523 0.619 -0.074
TRMN -0.368 0.369 0.300 0.356 0.325 -0.058 -0.046 0.633
5
4-
(]) 3-::J (ij > c (]) Ol w 2-
1-
0-I I I l ~ 1 2 3 5 7 8
Component Number
Figure A2: Scree plot of eigenvalues for the principal component analysis of
MM22DEXP with log transformed cobalt, magnesium, aluminium and manganese
136
------------------------- -
We next consider the viability of the transfonnation of the highly skewed
variables cobalt, magnesium, aluminium and manganese by assessing the effect on
the principal component analyses. The eigenvalues in each are case vcv close in
value with the cumulative proportion of variability almost identical by the fourth
component. The eigenvalues of the first and second principal components are
somewhat higher for the transfonned data while the third eigenvalue for this data set
is somewhat lower. There have been numerous small changes in some of the
elements of the coefficient vectors, namely in the first analysis nickel, cobalt, iron
and zinc dominate the first principal component, all with fairly equal weighting. In
the second analysis, the transfonned cobalt receives the greatest weight while nickel,
iron, zinc and transformed manganese have lower, fairly even values. Similar
changes have occurred in the subsequent principal components, though they all
appear to be small.
Figures A3 am! A4 display the scatter plots of the scores of the first four
principal components for each analysis. There does appear to be some slight
curvature in some of the plots in Figure A3, in particular those between PCS I, PCS2
and PCS3, suggesting that the data may benefit from a transfonnation. By
comparisor., the scatter plots in Figure A4 do appear to be less correlated and circular
in shape, suggesting that the transfonnation has somewhat improved the analysis.
However given that the results of the two principal component analyses are very
similar, the curvature in the plots in Figure A3 is not sufficiently marked and the
dangers of exaggeration of estimation errors we feel that the data do not warrant the
transfonnation.
137
2.24
-2.64
1.89
-1.72
1.34
-1.51
1.70
-0.67
j
-
PCS1
0 ... . .. 0
0 .. "' "' ..... ~ ,., .I ...
• 0
• ... .... , . ... .. •• •
• •~·. .,.. PCS2
. ., -Aif.,
• .., .. • •
•
~-• •• •
4 ~0 "" .0 . . ;_., •• •<t• 0 "' 0 0 • oO ,. 0" •
• 0
-~ -~0 .... . •. •• • • ~ ,. • 0 0
• • •• PCS3 : , • • • • • •
0 . .... • • • • • •••
PCS4
'
Figure A3: Scatter plots of the principal component scores of MM22DEXP
~ • • •
4.66 ·J· ••• .: •.. PCS1 -1.16 . .... . .. .t• • • • • •• # ••• . :• . • •
·~ ·~· 0.50 -·· . •: . ~ • • • • • • PCS2 • •• -2.54 # • ~· • • •
., . . • •• • - . • • • • • • • • • •
1.28 --· ·~·· • ... • *
• • ·* l! .. PCS3 .i •. · -1.01 - ~ ...... • • •• +# ~ • • •• • • • • • •
• • • • • • .. -- ... • 0.81 • • •• ol . .,. . PCS4 #. •• -·! ,.. ~ . • . ....... ~. -1.29 .... ••• • • •
' ' ' ~.~ro 1J..roro ~().
,?,· \)~\) ~5J~ ~ './.-'0 ~'./.-~ '0~ \).
Figure A4: Scatter plots of the principal component scores of MM22DEXP with log
transformed cobalt, magnesium, aluminium and manganese
138
APPENDIX B: CROSS VALIDATION
Cokriging Cross Validation of Cobalt, Iron and Zinc from MM22DHC4
500.
400.
300.
200.
100.
0.15
0.10
0.05
0. 00
Isatis data/stexp
Variable
Cross Validation
X (Meter)
600 700. 800 900 ~0001100120013001400
+ ' .
+++++•~ + ++i·+-+ + r + r + + + + +il+ ... , r • + ++++++r+e ++~-t:t.•• +++++++++••+,+tl + +++++++·+++1 ~+ .... ,++
+ + • + t 600.700.800 900~0001100120013001400.
X (Meter)
{Z*-Z) /S*
(Z*-Z) /S*
" ""
500.
400.
3oo.
200.
100.
0.15
0.10
0. 05
0. 00
• " ' ' w ~::: ' " ' ' 0 ' e-" '~ '
4.
3.
2.
1.
0.
-1.
Z* : STCO (Estimates)
·'r·--~-~1-----"ot·----1~----"'r·----3~-----•'r·,
•
•
• ••
+
4.
3.
2.
1.
0.
-1.
-2 --2,'".----,1".--~0~-----~1".---,2"-----~3-.--~.·-.J-2 . Z* : STCO (Estimates)
Z* : STCO (Estimates)
-1. 0. 1
• 5. 5 .
• • •
0. 0.
• • I •
-5. -5 .
• 1. 0. 1. 2. 3.
Z* ; STCO (Estimates)
Ellen M. Bandarian
Variable " "co May 13 2003 13:29:45 - Variable " STFE - Variable " STZN Standard Parameter File for Model: hc4 Standard Parameter File for Neighborhood: neighborhood Cross validation statistics based on 124 test data
Me= Variance Error 01820 1.24065 Std. Error 01552 2. 56868
Cross validation statistics based on 112 robust data Mean Variance
Error 0.12837 0.63143 Std. Error 15833 1.25607
A data is robust when its Standardh;ed Error lies between -2.500000 and 2 500000
Figure B 1: Cross validation of cobalt
ellen
139
500.
" 400.
' " ~ 300.
" 200.
100.
I.satis
Cross Validation
X (Meter)
600 700 800. 900 .10001100120013001400.
' + +· + + + ·+ +++++++ -1- ++1_+-+ + + + + + -1- + + +:utli+ ++ + + ++++++++++·~~W++-4c •e. + + + + +++4111++++ ++ ··:+ + ++++++++++·~+++··-~·+
t+ + ' ++ 600.700.800 900 U001100120013001400.
X {Meter)
(Z*-Z)/S*
(Z*-Z)/S*
o;
' " • ' ~
500. /! • • 400. '~
300. " ~-0
200. ~
100.
0.15
0.10
0. 05
0. 00
Z* STFE (Estimates)
-2. -L o. L 2.
3.
2.
1.
0.
-1.
-2.
Z* STFE (Estimates)
Z* STFE (Estimates)
-2. -1. 0. L
5.
0.
-~H :· + ~ -f .~-~~~4!~-~ ++ ++
-5.
• -2. -1. 0. 1.
,. STFE (Estimates)
3.
3.
" 2.
• 8
L " " 0.
~ 0
• " -1. ' ,._
-2.
3.
-5.
2.
data/stexp Variable " STNI
Ellen M. Bandarian
Variable " STCO - Variable #3 STFE - variable " STZN Standard Parameter File for Model: hc4 Standard Parameter File for Neighborhood: neighborhood Cross validation statistics based on 124 test data
Error Std. Error
Mean 0.01315 0.01033
Variance 0. 65802 1. 54736
Cross validation statistics based on 118 robust data
Error Std. Error
Me~
0. 00384 0. 01486
Variance 0. 45980 0. 87059
A data is robust when its Standardized Error lies between -2.500000 and 2.500000
Figure B2: Cross validation of iron
May 13 2003 13 28:53
ellen
140
Cross Validation
X (Meter)
600 700.800 900 ~0001100120013001400.
500. t 500.
+ ;; 400. + + + + + 400.
• +H·++++ ' ++~+ " 300. + + + + + + + + + +~~t ++-: + + 300. ! +++++++++·I· ++-P--• 0 +++++++•++++++-++ ,. 200. 200.
+++++-~+·f-t+++++ ++ 100. +• + + +r 100.
600.700.800 900 ~0001100120013001400.
X (Meter)
(Z*-Z) /S*
0.125 0•.125
0.100 0.100
• • ., 0
0. 075 0. 075 ' • • • " • 0. 050 0. 050
0.025 0.025
0. 000 0. 000
(Z*-Z)/S* Isatis data/ste:xp - Variable " STNl - Variable " STCO - Variable " STFE - Variable " STZN Standard Parameter File for Model: hc4 Standard Parameter File for Neighborhood: neighborhood Cross validation statistics based on 124 test data
Error Std. Error
Mean 0. 01876 0.01526
Variance 0. 64322 1.50370
Cross validation statistics based on 119 robust data
Error Std. Error
Mean 0. 03477 0.04257
Variance 0.50105 0.97769
A data is robust when its standardized Error lies between -2.500000 and 2. 500000
-2.
3.
• 2.
" rl • ' • 1.
' ~ • 0.
" "~
" -1. ~" ' ~
-2.
-2.
5.
" " ' ' -~::: 0 • ' " ' '
"'" 0 '
""" + '~ ' +
-5.
,. STZN (Estimates)
-1. o. 1. 2. 3.
3.
• +
* 2. + + ·!'- + + ++ • o~*-Wt.t 1. + -1-- ++ :j::
-t:t:'nr-P~" ++ J,t ·1: 'I+ -1+
·> -~-4~f-~ + 0.
+ +· - . 'f- :t •
t '/'+, -1. ++11 * "+ ++
++ • ,. + +
" .. -2.
-1. 0. 1. 2. 3.
,. STZN (Estimates)
Z* STZN (Estimates)
1. 0. 1. 2
5.
• •
+ +· + + ++ + + +
+· r++-:tbt+ ++ + + t- '~ ~~-ttf.i'f ·t:l- + + .,
++ t "'!: '•+~; +/p + .. + -tfl.-~+ .. 1 + +~. ~~ --»t-+t * + +
+· +
o.
• '
. -5 .
-1. 0. 1. 2.
STZN (Estimates)
Ellen M. Bandarian
"
-8
" z 8 8 ' ' ' ' ~
" ' ' "
May 13 2003 13;29:18
ellen
Figure B3: Cross validation of zinc
141
Cokriging Cross Validation of Cobalt, Magnesium and Iron MM22DTOP4
Cross Validation
.. ' " • ,
X (Meter) • 600.700.800 900 j_Q0011001200 13001400. ' ~
500. 500. 0 + u
;; 400. + ' ' 400. ... ~
'. + + I·+· t+ • + + + + • .. " 300. ' l- + + + ·1- + + + +tl+ I ,+ • + -f. 300. ~ ~-++ + f + + + +. + + + + * i +. 0
" 200. ~+++++-f-++0+-f-++ ++ 200. ~ t+t+++++++++++,++
100. t• • + +t 100.
600.700 800. 900 j_QQ01100120013 001400.
X (Meter)
(Z*-Z)/S*
0.15 0.15
" • • 0.10 0.10 " . ·o 0 w
" ~::: ' • 0 " & ' ' • " e·" " 0 ~ • "
0. 05 0. 05
0. 00
(Z*-Z) /S* Is a tis data/stexp - Variable " '™' - Variable " STCO - Variable " STMG - variable " STFE Standard Parameter File for Model: top4 Standard Parameter File for Neighborhood: neighbourhood top4 Cross validation statistics based on 124 test data
Error Std. Error
Mean 0. 02044 0. 02028
Variance 1.14637 1.95856
Cross validation statistics based on 115 robust data
Error Std. Error
Mean 0.14843 0.17417
Variance 0. 60231 0.97074
A data is robust when its Standardized Error lies between -2.500000 and 2.500000
. . 3.
2.
1.
0.
-1.
-2
5.
0 .
-5.
-2 .
-2.
,. ' STCO (Estimates)
-1. 0. 1. 2. 3.
•
• .-+
.I":~+ + +
+ij.. '1{. '' +-++· ift.fH+
.;t t+ * + + +"' .+· +• + ... ;t;+~1i"* -· . +/It- + +
~-·ij<l- 4+ -t~~rf- ++
-1. 0. 1. 2. 3. ,. ' STCO (Estimates)
Z* : STCO (Estimates)
1. 0. 1. 2.
1-+ + + + t-·lf ·,;~1f-t "t ·tlf~~:t·l~. -~ .. . -tt +
?t•tot *{ + Ji,; ++-t "*+
+ t- ~.tt +
••• • • •
• 1. 0. 1. 2.
Z* : STCO (Estimates)
4.
4.
3.
•
. . 3.
2.
1.
0.
-1.
_,.
5.
"
-" n 0
;:; 2 0
• ~
' ~
N
' ' 0. ~
-5 .
Ellen M. Bandarian
May 19 2003 10 54:53
ellen
Figure B4: Cross validation of cobalt
142
Cross Validation Z* STMG (Estimates)
;;
' " • , X (Meter)
' 600 700 BOO 900 ~000H001200l3001400. ' ~ + 500. 500.
+ 0
• . 400 . "~ H H + ·• + • •+•· i: .. +• + + + + + + 1·-mtl+ + + t +e 300. ~" ++ . + • + . + +++-t,.l·+ ' ++ + • + ' • + +++++. 200 . ~
;; 400.
' " 300. ~
" 200. + ++++ • + + + ' + • + ,. t •
t + +; 100. 100.
600.700.800 900 ~0001100120013001400.
X (Meter)
(Z*-Z)/S*
0.15 0.15
' • ' 0.10 0.10 " . " ' m g ~::: ' ' " i\ ' ' 0 •
' e-N
" '~ • ' 0. 05 0. 05
0.00 0.00
(Z*-Z)/S* Isatis data/stexp
Variable " STN' Variable " STCO
. Variable N3 STMG - Variable .. STFE Standard Parameter File for Model: top4 Standard Parameter File for Neighborhood: neighbourhood top4 Cross validation statistics based on 124 test data
Error Std. Error
··~ -0.02439 -0.02089
Variance 0. 76237 1.48314
Cross validation statistics based on 116 robust data
Error Std. Error
Mean 0.02676 0.04236
Variance 0. 32645 0.76904
A data is robust when its Standardized Error lies between -2.500000 and 2. 500000
5.
4.
• 3.
8. .++
++ +
1.
;
0.
-1.
8' ' STMG
,. ' """G
1. 0.
5.
4.
3.
8. t + +-* + +
1. ·l:FI·~·W++
0. ~;;~!+ * +f~"'l.ll't
-1. "'+ ++ -2. ++ + +
-3. •• -4.
-5.
1. 0.
,. ' STMG
Figure B5: Cross validation of magnesium
5. . . "
3. m
~ " 8. ;:; 8 ' 1.
' ' ~ ' • 0. ~
• -1.
(Estimates)
(Estimates)
1. 8.
5. .. • • 3.
2. +
+ 1. 'i 0. ' ~
+ + -1. ~
+ -2.
-3.
-·. -5.
1. 2.
(Estimates)
Ellen M. Bandarian
May 19 2003 10:55 27
ellen
143
Cross Validation ,. STFE {Estimates)
-2. -1. 0. 1. 2. 3 .
• 3. 3.
' 2. {+ + 2. , H
~ ¥+* ~
++·i t + ~ X (Meter)
' 1. + +* - 1. " 600 000 800 900 ~0001100120013001400 ' "\'+ ;:; t 1!!11 +
500. + ' 500. + + 2 + • 0. ++~.l" • 0. ' + • + +
" 400. + ++++ 400. "~ +-4r:lt ~
+f+++H· • ++rl- + -~+ + e
' .. ' " 300. ++++++++++,+++ ++ 300. -1. + + ~ -!--+ + -1. ,.. ' ~" !'; ++++++++++ + -1--t:< •• ' #* + • " 200. ·i··++++++ID++++++ -++ 200. !!.
+ + + +--+ + + + -I-++ + + +. ~ + + -2. l· +· -2. + 100. ·i· ·I· + • + -i: 100.
600. 700. 800. 900..1000110012 0013 001400. -2. -1. 0. 1. 2. 3.
X (Meter) ,. STFE (Estimates)
(Z*-Z)/S* STFE (Estimates)
-2 - 1 1
' 5. • 5.
0.15 •
+ -!+- + +1-• ' " ~ 0.10
+ :If.~;-+ -{+ -itJol-+tf-iij:+
-I-+ Jl·~+tt;tttP.t+ + .. " ' ' N
!; ~
0.05
0. 00
Isatis data/stexp
Variable #1 STNI Variable #2 STCO
- Variable JlJ STMG - Variable #4 STFE
(Z*-Z) /S*
standard Parameter File for Model: top4 Standard Parameter File for Neighborhood: neighbourhood top4 cross validation statistics based on 124 test data
Error Std. Error
Mean 0.01046 0.01025
Variance 0. 63882 1.45372
Cross validation statistics based on 119 robust data
Error std. Error
Mean 0 02007 0 03548
Variance 0. 46943 0. 82923
A data is robust when its standardized Error lies between -2.500000 and 2.500000
0. ;f ·f·
-5.
2.
0. ~· +'''+ + .;iilili..)\.> } -f+
+ -11- '-1-+ + + + ~t- ·I· . •
- 5. . 1. 0. 3. 2
STFE (Estimates)
Ellen M. Bandarian
May 19 2003 10:55 45
ellen
Figure B6: Cross validation of iron
144
Ordinary Kriging Cross Validation of Principal Components
Cross Validation ,. pc2 (Estimates)
-2. -L 0. L 2.
3.
• .. 2 .
' " • X (Meter) >
' L
600 700.800 900 ~0001100120013001400. ' 500. • 500. ~
0. + + M 0
-,; 400. + ++++ 400. "" ' +++++· + I I f f f w~ " 300. +++++ f ' ' 1-+lt •+ + t 300. -L g; +·+•+ ·I " ++ 1+ +t-t- ~1 ' >< 200. ++++++ 1·0+ ++ +++t++ 200. "- • +·1·--f-++++++++ + + + . ·+ + -2. • 100. <-+ + + +t 100.
600.700. BOO 900 ~0001100120013001400. -2. -L 0. L 2.
X (Meter-) ,. pc2 (Estimates)
(Z*-Z) IS* Z* : pc2 (Estimates)
-L o.
5.
• 0.15 0.15
" . " ' w ~ '
0.10 • ;; 0. , ' 0 . e-" '-'
' ' '0 0 c 0.10 ' il ' k •
+ + . + ++) + + ' +1- t'\1'1+ + 'Ul+ j;\;;Jt:;; .. . . Jt+ ~~+ 'if., -~+ +"it + ... "!= + .. + +--~'f~-t- + +
+ 0. 05 0. 05 • •
• -5.
• 0. 00
L 0.
{Z*-Z) IS* pc2 (Estimates) Isatis
3.
3.
2.
L
0.
-L
-2.
3.
L
5.
+
0.
-5 .
L
" ~ ;; 2 ' ' e c ~
~ ' "
w .
data/pcahc4 - Variable il : pc2 Ellen M. Bandarian
Standard Parameter File for Model: pc2 Standard Parameter File for Neighborhood: pc2 Cross validation statistics based on 123 test data
Error Std. Error
Mean 0. 00107
00337
Variance 0.69695 1. 73859
Cross validation statistics based on 116 robust data
Error Std. Error
Mean 0. 06321 0.10530
Variance 0.41468 0. 97307
A data is robust when its Standardized Error lies between -2.500000 and 2.500000
Figure B7: Cross validation of PC2
May 14 2003 14 15:39
ellen
145
Cross
X (Meter)
600 700.800 900 ~0001100120013001400.
500. T ., 500.
+ +
" 400. + +· +· ++ 400.
• + f· +·++++ + ++ •• " 300. ++ ++++++++.,!-++.++ 300. • 1l + + ++++++++ +++ +$-
" 200. I· -1- + + + +· + + + • + + +·+. + + 200. + + + + ., f++++++·++ ; + +
100. '
·f· f· + ,+ •· ·f 100.
600.700 800.900 ~0001100120013001400.
X (Meter)
(Z*-Z) /S*
0. 20 0.20
• 0.15 • ·o 0.15
" ' • • • 0.10 H 0.10 •
0.05 0. 05
0. 00 0. 00
(Z*-Z) /S* Isatis data/pcahc4 - Variable #1 : pc3 Standard Parameter File for Model: pc3 Standard Parameter File for Neighborhood: pc3 Cross validation statistics based on 124 test data
Error Std. Error
Mean 0 01006 0 01740
Variance 0. 34312 2. 35414
cross validation statistics based on 114 robust data
Error std. Error
Mean 0. 01059 0. 03801
Variance 0.14802 0.93526
A data is robust when its Standardized Error lies between -2.500000 and 2. 500000
Validation
-1. 3.
2. • ' rl
' ' • 1.
' ~ • ~
><A 0.
~N " ' ~ -1.
-1.
5.
• " ' ' m
~ ' ' " 0.
' ' 0 ' e·" : ~
,. ' pc3 (Estimates)
0. 1. 2.
•
+
•
.. • •
0. 1. 2.
,. ' po3 (Esthuates)
Z* : pc3 (Estimates)
0.
++ + +"!j;l},
. .. +
... +
+ t
1.
3. 3.
2. " " ~
1. 8 2 ' ' 0. ' ' ~
-1.
3.
5.
0.
•• -5. -5.
• 1. 0. 1.
Z* : pc3 (Estimates)
Ellen M. Bandarian
May 19 2003 12:ll:U
ellen
Figure B8: Cross validation of PC3
146
Cross
X (Meter)
600.700.800 900 j_QQ01100120013 001400.
500. ! 500.
" 400. + ·f·+ + + 400.
• +++++++ + ++I+ " 300. ++++ +++++·J,t++ ··+ + 300 . • 1l +++++++++·+- 1-+-t:·:+·+ " 200. ++++++++++++++ ++ 200.
++o+++++i-+++++ ++ 100. -t+ + + +-i- 100.
600.700 800.900 j_Q001100120013001400.
X (Meter)
(Z*-Z) /S*
0. 20
0.15
• • " g • 0.10 il • " •
0. OS
0.00 0.00
(Z*-Z}/S* Isatis data/pcahc4 - Variable Ill : pc4 standard Parameter File for Model: pc4 standard Parameter File for Neighborhood: pc4 Cross validation statistics based on 124 test data
Error Std. Error
Mean 0.00316 0 01022
Variance 0.15351 1. 85810
Cross validation statistics based on 117 robust data
Error std. Error
Mean 0. 01325 0.04842
Variance 0. 08061 0.99782
A data is robust when its Standardized Error lies between -2.500000 and 2. 500000
Validation ,. ' po4 (Estimates)
~2. ~1. 0. 1. 2. 2. 2.
•
.. 1. 1. "
' " ' " , 2 •
~ 0. 0.
~ ' • • u ' "" '
~"' ~1. ~1. ' + ".
rt • ~ ~2. ~2.
~2. ~1. o. 1. 2.
,. ' po4 (Estimates)
Z* : pc4 (Estimates)
~s. ~s.
• -0.5 0 0 0. 5
Z* , pc4 (Estimates)
Ellen M. Bandarian
May 19 2003 12:11:38
ellen
Figure B9: Cross validation of PC4
147
Ordinary Kriging Cross Validation of Cobalt, Magnesium, Iron and Zinc
Cross Validation z• ' co (Estimates)
0. 1. 2. 3.
3. • 3 .
• ., •• • " ~ 2. + 2 . • • n
X (Meter) > 0
• 'C+ ++ ;:; 600.700.800 900~0001100120013001400. , +•+ " .,. " 500. 500. "' + .[t:f:: ' + m + 0 + -it- -It + ' "
400. + . 400 . "0 1. ++_+t- + 1. ~ • + + + H>~ + ++t+ +t ++++·t •+ ' " 300. + +_ +_ I +_ ~ + + + +. + +_ .... 9 + 300. ~N +_· +t-6 -tJ • ~
#! + • " ·1+. oft-.tl· -l:l. ++++1++10•++++i•• m • " 200. +++++++++e++++ ++ 200. !!. +f•IJ!f ++
++++++++++++++,++ 0. +-~-~- +
0. 100. ++. ++-( 100.
600.700.800.900~0001100120013001400. 0. 1. 2. 3.
X (Meter) z• co (Estimates)
{Z*-Z)/S* Z* : CO (Estimates)
0. 2
0.15 0.15
• 5. 5 .
• • • • 0.10 0.10 " • " . ·< m ~ u ~::: " • m " o. g ,
' n + • ,. " " m~ • •
. rJ::~ -~1- +-r.:W·
+~4 1..:-!f'l ,#t,~c +' + J.:'l\ :t++ "" + +.;- '+. lfl-¥-_p- ++ +1r+:1f.++
'! 0 . N
;; •
0.05 0.05 • • . ~ • -5. . 5 .
• o.oo o.oo
0. 1. 2
(Z*-Z) /S* Z* : co (Estimates) I sa tis ok/exp Ellen M. Bandarian - variable nl : co Standard Parameter File for Model: co May 27 2003 13' 35' 15 Standard Parameter File for Neighborhood: stco Cross validation statistics based on 124 test data
Error Std. Error
0.01039 0.01632
Variance 0.37485 2. 63127
Cross validation statistics based on 112 robust data
Error Std. Error
Me= 0. 07014 0 15773
Variance 0.19056 1.27670
A data is robust when its standardized Error lies between -2.500000 and 2.500000
Figure BlO: Cross validation of cobalt
ellen
148
Cross Validation
X (Meter}
600 700. 800.900 ~0001100120013001400.
500. + soo. +
" 400. ' ' 400.
+ +·. + •+ • • • + + + " 300. +• + + + + +&+ + ' t •+ • ' 300. 1l t. + + + + + + + + + f + -!'" ,. 200. ., + + + + + + ' + + ++ t+e 200.
' ++++ + + + + + + + ·t t + 100. + • ,, 100.
600.700. 800.900 ~0001100120013 001400.
X {Meter)
(Z*-Z) IS*
0.125 0.125
0.100 0.100
• • ·rt
~ 0.075 0. 075
il • " • 0. 050 0.050
0.025 0. 025
o.ooo 0. 000
(Z*-Z) /S*
Isatis ok/exp - variable #1 ' MG Standard Parameter File for Model: mg Standard Parameter File for Neighborhood: stmg Cross validation statistics based on 124 test data
Error Std. Error
Mean -1.63389 -0.01960
Variance 3629.50520
1.94218
Cross validation statistics based on 113 robust data
Error Std. Error
Mean 3. 43792 0.11234
Variance 1353.50795
0.83542
A data is robust when its Standardized Error lies between -2.500000 and 2.500000
,. MG {Estimates)
400. 400. +
• 300. • 300. N
" • ~ • • ~ > • ~ + " 200. + 200. ;; " " t. • + ' •
~~ ' ~ ' ~~ 100. 100. "-• rt •
" • 0. o.
,. MG (Estimates)
Z* : MG (Estimates)
0. 100. 200. 300.
5. • 5 •
•
+., •. +
~ 0. N
• . + ?'~~-~- * " c"'t . , • 00 .:!·:+ -:lf-1- + ~ ~ +
• ;; o. . . '+ R ' 't~\ll)*t '+ , + P· N
·~ + ,_ • + t+
• .. • • -5. -5.
0' 100. 200. 300.
Z* : MG (Estimates)
Ellen M. Bandarian
May 27 2003 13:34:44
ellen
Figure Bll: Cross validation of magnesium
149
Cross Validation
800.
700.
• 600.
" ~ • 500. X {Meter) > • 600.700.800.900 .10001100120013001400. " 400. " 500. + soo. !0
+ + " 300. 400. + + + •• 400. ". ;;
+++++++ • ++t+ • *N 200. " • 300. ++++++++++,+++ ·+-+ 300. e + + + --1- + + + + + +' ·f· + .jlc ::·· • • " • 100.
" 200. ++ + + +++e+··i-++++ .... + + 200. !'. ++++++++++++++-·++ 0.
100. ++ + -1- +-t- 100.
600.700.800 900.10001100120013001400.
X (Meter)
(Z*-Z)/S*
0.15 0.15 5.
• • ' " . ·rl 0.10 u
' • il 0.10 • ~
~:: • N 0.
" ' 0 •
' e· N
" • -• • 0.05 0. 05
-5.
o.oo 0.00
(Z*-Z) IS* Isatis ok/exp - Variable #1 : FE
Z* : FE {Estimates)
0. 100. 200. 300. 400. 500. 600. 700. 800.
•
t +
• • +
+ + • +
• + +
0. 100.200.300.400.500.600.700.800.
Z'
100 200
#''+' + ++
' FE {Estimates)
FE (Estimates)
300 400. 500. 600.
•
• +
• •
• 100. 200. 300. 400. 500. 600.
z• FE (Estimates)
800 .
700.
600 .
soo.
400.
300.
200.
100 .
0.
5.
0 .
-5.
Ellen M. Bandarian
N
• " ;:; " ' • ' ' " ' ,._
N . ' N
standard Parameter File for Model: fe May 27 2003 13 35:57 Standard Parameter File for Neighborhood: stfe Cross validation statistics based on 124 test data
Error Std. Error
Mean 1. 72506 0.00942
Variance 14008.38280
1.65939
Cross validation statistics based on 117 robust data
Error Std. Error
Mean 2.82988 0.03734
Variance 9206.43755
0.89172
A data is robust when its Standardized Error lies between -2.500000 and 2.500000
Figure B 12: Cross validation of iron
ellen
150
Cross Validation ,. ' STZN (Estimates)
-2. -1. 0. 1. '· 3.
3. • 3.
2. + ++ 2. " • + + ~ • • + ++ + ++ • p " X (Meter) • 1. +· :\\ -l~kt-t;t 1. ~
600 800. 9 00 ~0001100120013 001400. ~ _+· -++
700 !¢ +~++ * 8
± H 500. 500. 0. 0. ' + " " •""Jl•+ + ro
" +, 400. + f· I·++ 400. '~ "'' ++ ' -;; • ' +++++++ • ++;+ + * "'++ " • 1i .. -1. -1. ' " 300 . +++++++ + ++,+++ _++ 300. +·~ +it~ 'lt- ++ !'. • ro"
"' " ++. + + • ++++++++++ ++·tl---·· ro
" 200. +++++++0++++-~+--++ 200. !!. ++ ++++++++ +++-+++---:--1---+ -2. -.~ -2.
100. ++ + + ++ 100.
600.700.800.900~0001100120013001400. -2. -1. o. 1. 2. 3.
X (Meter) ,. STZN {Estimates)
(Z*-Z)/S* ,. STZN (Estimates)
-2. o . 2.
0.15 . 0.15
• 5. 5.
•
* ;;; ++
o. ' "
• 0.10 0.10 " ' H • ·H ro • ;I u ~:: ' • ro" 0 • fr " ' 0 •
~ • H· N
" ro-• • 0.05 0. 05 •
-5. -5.
• 0. 00 o.oo
2. 0 2.
{Z*-Z) /S* ,. STZN {Estimates) I sa tis data/stexp - Variable #1 : STZN Ellen M. Bandarian
Standard Parameter File for Model: stzn May 19 2003 10:42:51 Standard Parameter File for Neighborhood: stzn Cross validation statistics based on 124 test data
Error Std. Error
Me~
0. 02430 0.02022
Variance 0. 65195 1. 72399
Cross validation statistics based on 119 robust data
Error Std. Error
Mean 0.03916 0.04470
Variance 0.50497 1. 00775
A data is robust when its standardized Error lies between -2.500000 and 2.500000
Figure Bl3: Cross validation of zinc
ellen
151
APPENDIX C: PARAMETER FILES
Ordinary Cokriging ofMM2DHC4
Kriging procedure
===---====================================================== Data File Information:
Directory = data
File =stexp
Variable(s) = STNI
Variable(s) = STCO
Variable(s) = STFE
Variable(s) = STA."'
Target File Information:
Directory =data
File = hc4cokrig
Variable(s) = stniest
Variable(s) = nisd
Variable(s) = stcoest
Variable(s) = cosd
Va.riable(s) = stfeest
Variable(s) = fesd _
Variable(s) = stznest
Variable(s) = znsd
Type =POINT (1717 points)
152
Model Name = hc4
Neighborhood Name= neighbourhood hc4- MOVING
Successfully processed= 1717
Writtentothedisk = 1717
Neighbourhood parameters
===================
Type= MOVING
X-Radius
Y-Radius
Z-Radius
=
=
220.00m
120.00m
228.12m
Rotation angle around Z = 0 degrees
Rotation angle around Y = 0 degrees
Rotation angle around X = 0 degrees
Minimum# of information = 4
Optimum# of information = 4
Number of angular sectors = 4
Automatic Sorting on
X-Mesh of sorting grid = 228.12m
Y -Mesh of sorting grid = 228.12m
Z-Mesh of sorting grid = 228.12m
No Heterotopic Search
No Minimum Distance used between Data Points
No Maximum Distance without Samples
153
No Maximum Number of Consecutive Empty Sectors
Model : Covariance part
======================= Number of variables = 4
-Variable I : STI'J!
~Variable 2. srco
~ Variable 3 : STIT
~Variable 4 : STZN
Number of basic structures= 3
S l : Nugget effect
Variance~Covariance matrix :
Variable l Variable 2 Variable 3 Variable 4
Variable l 0.0500 0.0360
Variable 2 0.0360 0.0500
Variable 3 0.0320 0.0290
Variable 4 0.0320 0.0300
S2: Spherical- Range= lOO.OOm
Variance-Covariance matrix:
0.0320 0.0320
0.0290 0.0300
0.0500 0.0415
0.0415 0.0500
Variable 1 Variable 2 Variable 3 Variable 4
Variable 1 0.5000 0.3600 0.3200 0.3200
Variable 2 0.3600 0.5000 0.2900 0.3000
Variable 3 0.3200 0.2900 0.5000 0.4150
154
Variable 4 0.3200 0.3000 0.4150 0.5000
S3 : Spherical - Range= IOO.OOm
Directional Scales = ( 200.00m, IOO.OOm, IOO.OOm)
Local Rot (mathematician)= ( 0.00, 0.00, 0.00)
Local Rot (geologist) = ( 90.00, 0.00, 0.00)
Variance-Covariance matrix:
Variable 1 Variable 2 Variable 3 Variable 4
Variable I 0.4500 0.3240 0.2880 0.2880
Variable 2 0.3240 0.4500 0.2610 0.2700
Variable 3 0.2880 0.2610 0.4500 0.3735
Variable 4 0.2880 0.2700 0.3735 0.4500
!55
Ordinary Cokriging ofM,~22DTOP4
Kriging procedure
============================================================ Data File Infonnation:
Directory = data
File = stex.p
Variable(s) = STNI
Variable(s) = STCO
Variable(s) = STMG
Variable(s) = STFE
Target File Infonnation:
Directory = data
File = top4cokrig
Varlable(s) = stniest
Variable(s) = nisd
Variable(s) = stcoest
Variable(s) = cosd
Variable(s) = stmgest
Vnriable(s) = mgsd
Variable(s) = stfeest
Variable(s) = fesd
Type =POINT(I717points)
Model Name = top4
Neighborhood Name= neighbourhood top4- MOVING
156
Successfully processed= 1717
Writtentothedisk = 1717
Neighbourhood parameters
=======================
Type= MOVING
X-Radius
Y-Radius
Z-Radius
= 220.00m
= IOO.OOm
= 228.12m
Rotation angle around Z = 0 degrees
Rotation angle around Y = 0 degrees
Rotation angle around X = 0 degrees
Minimum# of information = 3
Optimum # of information = 4
Number of angular sectors = 4
Automatic Sorting on
X-Mesh of sorting grid = 228.12m
Y -Mesh of sorting grid = 228.12m
Z-Mesh of sorting grid = 228.12m
No Heterotopic Search
No Minimum Distance used between Data Points
No Maximum Distance without Samples
No Maximum Number of Consecutive Empty Sectors
!57
Model : Covariance part
======================= Number of variables = 4
-Variable 1: STNI
-Variable 2: STCO
- Variable 3 : STMG
- Variable 4 : STFE
Number of basic structures = 3
Sl :Nugget effect
Variance-Covariance matrix :
Variable 1 Variable2 Variable3Variable4
Variable 1 0.1000 0.0500 0.0300 0.0000
Variable 2 0.0500 0.0900 0.0500 0.0000
Variable 3 0.0300 0.0500 0.0800 0.0200
Variable4 0.0000 0.3000 0.0200 0.0500
S2 : Spherical - Range= 75.00m
Variance-Covariance matrix:
Variable 1 Variable 2 Vmiable 3 Variable 4
Variable 1 0.2700 0.1700 0.0300 0.1900
Variable 2 0.1700 0.5100 0.0000 0.1700
Variable 3 0.0300 0.0000 0.1000 0.0400
Variable 4 0.1900 0.1700 0.0400 0.4000
S3: Spherical- Range= 140.00m
158
Directional Scales= ( 200.00m, 140.00m, 140.00m)
Local Rot (mathematician)= ( 0.00, 0.00, 0.00)
Local Rot (geologist) = ( 90.00, 0.00, 0.00)
Variance-Covariance matrix :
Variable 1 Variable 2 Variable 3 Variable 4
Variable 1 0.6300 0.5000 0.3000 0.4500
Variable 2 0.5000 0.4200 0.1800 0.4000
Variable 3 0.3000 0.1800 0.8200 0.1200
Variable 4 0.4500 OAOOO 0.1200 0.5500
159
Ordinary Kriging of the Principal Component Scores
Legend: pel (pc2) (pc3) (pc4)
Kriging procedure ·
Data File Information:
Directory = data (data) (data) (data)
File = pcahc4 (pcahc4) (pcahc4) (pcahc4)
Variable(s) =pel (pc2) (pc3) (pc4)
Target File Information:
Directory = pclest (pc2est) (pc3est) (pc4est)
File =pel est (pc2est) (pc3est) (pc4est)
Variable(s) =pel est (pc2cst) (pc3est) (pc4est)
Variable(s) = pclsd (pc2sd) (pc3sd) (pc4sd)
Type = POINT (1717 points) ( 1717 points) (1717 points) ( 1717 points)
Model Name = pel (pc2) (pc3) (pc4)
Neighborhood Name= pel (pc2) (pc3) (pc4) - MOVING
Successfully processed = 1717 (1717) ( 1717) ( 1717)
Written to the disk = 1717 (1717) (1717) ( 1717)
Neighbourhood parameters
Type= MOVING
160
X-Radius
Y-Radius
Z-Radius
220.00m (120.00m)(190.00m) (250.00m)
120.00m (120.00m) (1 OO.OOm) (250.00m)
228.12m ( 28. J 2m) (228.12m) (228.12m)
Rotation angle around Z = 0 degrees (?5 degrees) (0 degrees) (0 degrees)
Rotation angle around Y = 0 degrees (0 dcgrees)(O degrees) (0 degrees)
Rotation angle around X = 0 degrees (0 dcgrces)(O degrees) (0 degrees)
Minimum# of information = 3 (3) (3) (3)
Optimum# of information = 4 ( ) (4) (4)
Number of angular sectors= 4 (4) (4). (4)
Automatic Sorting on
X-Mesh of sorting grid
Y-Mesh of sorting grid
Z-Mesh of sorting grid
No Heterotopic Search
228.12m (?'2R 12m) (228 12m) ("'28 12m)
228.12m ( 28. 12m) (228.12m) (228 .12m)
228.12m ( 28 12m) (228 12m)(228. 12m)
No Minimum Distance used between Data Points
No Maximum Distance without Samples
No Maximum Number of Consecutive E mpty Sectors
Model : Covariance part
N umber of variables = 1
-Variable 1 : pel
Number of basic structures = 3
S 1 : Nugget effect
161
Sill = 0.1000
S2 : Spherical- Range= 90.00m
Sill = 0.4700
S3 : Spherical- Range= 90.00m
Sill = 2.4700
Directional Scales= ( 200.00m, 90.00m, 90.00m)
Local Rot (mathematician)= ( 0.00, 0.00, 0.00)
Local Rot (geologist) = ( 90.00, 0.00, 0.00)
Model : Drift part
Number of drift functions = 1
-Universality condition
Model : Covariance part
Number ofvariables = I
- Variable I : pc2
Number ofbasic structures= 2
S 1 Nugget effect
Sill = 0.0800
S2 : Spherical - Range= 85.00m
Sill = 0.4 700
Model : Drift patt
162
Number of drift fimctions = 1
- Universality condition
Model : Covariance part
Number of variables = 1
-Variable l : pc3
Number of basic structures = 3
S 1 : Nugget effect
Sill= 0.0100
S2 : Spherical- Range= 50 OOm
Sill= 0.0500
SJ : Spherical -Range= 98.00m
Sill 02150
Directional Scales= ( 140.00m, 98.00m, 98.00m)
Local Rot (mathematician)= ( 25.00, 0.00, 0.00)
Local Rot (geologist) = ( 65.00, 0.00, 0.00)
Model : Drift part
Number of drift functions = 1
- Universality condition
Model : Covariance pat1
Number of variables = I
163
- Variable I : pc4
Number of basic structures = 3
S I : Nugget effect
Sill = 0.0 180
S2 · Spherical - Range = I OO.OOm
Sill = 0.0650
S3 : Spherical - Range = 250.00m
Sill 0.0920
Model : Drift part
Numher of drift function'
- Universality condition
164
Ordinary Kriging of Nickel, Cobalt, Magnesium, Iron and Zinc
Legend: nickel (cobalt) (magnesium) (i11Jn) (zinc)
Kriging procedure
Data File Information:
Directory = ok (ok) (ok) (ok) (ok)
File = exp (exp) (exp) (exp) (exp)
Variable(s) = ni (co) (mg) (fc) (zn)
Target File Information:
Directory = ok (ok) (ok) (Clk) (ok)
File = niest (coest) (mgest) (lccst) (znest)
Variable(s) = niest (coest) (rngest) (fecst) (znest)
Variable(s) = nisd (co~d) (mgsd) (fesd} (znsd)
Type = POINT (1717 points) ( 1717 points) (1717 points) ( 1717 points} ( 1717
points)
Model Name = ni (co) (mg) (fc) (zn)
Neighborhood Name= ni (co) (mg) Ue) (zn) - MOVING
Successfully processed= 1717 (1717) (1717) ( 1717) (1717)
Written to the disk = 1717 (I 717) (1717) ( 1717) ( 1717)
Neighbourhood parameters
Type =MOVING
165
X-Radius
Y-Radius
Z-Radius
180.00m (220.00m)(300.00m) (210 00m)(220.00m)
llO.OOm (120.00m) (160.00m) (I 10.00m)(110.00m)
228.12m (228.12m) (228.12m) (228.12m)(228.12m)
Rotation angle around Z = 0 degrees (0 degrees) (120 degrees) (0 degrees) (0
degrees)
Rotation angle around Y = 0 degrees (0 degrees)(O degrees) (0 degrees) (0 degrees)
Rotation angle around X = 0 degrees (0 degrees)(O degrees) (0 degrees) (0 degrees)
Minimum# of information = 3 (3) (3) (1) (3)
Optimum# of information = 4 (4) (4) (4) (4)
Number of angular sectors = 4 (4) (4) (4) (4)
Automatic Sorting on
X-Mesh of sorting grid
Y -Mesh of sorting grid
Z-Mesh of sorting grid
No Heterotopic Search
228.12m (228 12m) (228 12m) (228. 12m)
228.12m (:28. l2m) (228.12m) (228. 12m)
228.12m (228. l2m) (228.12m) (228.12m)
No Minimum Distance used between Data Points
No Maximum Distance without Samples
No Maximum Number of Consecutive Empty Sectors
Model : Covariance part
Number ofvariables = 1
- Variable 1 : NI
Number ofbasic structures= 3
S 1 : Nugget effect
Sill = 5. 0000
S2 : Spherical - Range= 90.00m
166
Sill = 12.1500
S3 : Spherical -Range= 89.60m
Sill = 28.3500
Directional Scales= ( 160.00m, 89.60m, 89.60m)
Local Rot (mathematician)= ( 0.00, 0.00, 0.00)
Local Rot (geologist) = ( 90.00, 0.00, 0.00)
Model : Drift part
Number of drift functions = 1
- Universality condition
Model : Covariance part
Number ofvariables = 1
-Variable I : CO
Number of basic structures = 3
S I : Nugget effect
Sill= 0.0120
S2 : Spherical- Range= IOO.OOm
Sill- 0.1600
S3 : Spherical- Range= 1 OO.OOm
Sill = 0.1300
Directional Scales = ( 200.00m, 100.00m, I OO.OOm)
Local Rot (mathematician)= ( 0.00, 0.00, 0.00)
Local Rot (geologist) = ( 90.00, 0.00, 0.00)
Model : Drift part
Number of drift functions = I
-Universality condition
167
Model Covariance part
Number of variables = 1
- Variable 1 MG
Number of basic structures= 3
S 1 · Nugget effect
Sill = 135.0000
S2 : Spherical - Range = 140 OOm
Sill = 2300 0000
S3 : Spherical- Range= 150.00m
Sill = 2162.0000
Directional Scales= ( 300.00m, 150.00m, JSO.OOm)
Local Rot (mathematician)= ( 120.00, 0.00, 0.00)
Local Rot (geologist) = ( -30.00, 0.00, 0.00)
Model Drift part
Number of drift functions = 1
- Universality condition
Model : Covariance part
Number of variables = I
- Variable I : FE
Number ofbasic structures= 3
S I · Nugget efl'ect
Sill = 1030.0000
S2 : Spherical - Range = 75.00m
Sill - 3100.0000
S3 Spherical- Range= 99.90m
Si11 - 16790.0000
Diii.!Ciional Scales= ( 185.00m, 99 90m, 99 90m)
Local Rot (mathematician)- ( 0.00, 0.00, 0.00)
168
Local Rot (geologist) = ( 90.00, 0.00, 0.00)
Model : Drift part
Number of drift fi.mctions = I
- Universality condition
Model : Covariance part
Number of variables = 1
-Variable I : ZN
Number of basic structures = 3
S I · Nugget effect
Sill = I 7700.0000
S2 : Spherical - Range= 85.00m
Sill - 274239 0000
S3 : Spherical- Range = 80.00m
Sill= 60 I 500 0000
Directional Scales= ( 200 OOm, 80.00m, 80.00m)
Local Rot (mathematician)= ( 0.00, 0.00, 0.00)
Local Rot (geologist) = ( 90.00, 0.00, 0.00)
Model : Drift part
Number of drift functions = 1
-Universality condition
169
Date: l21h November 2002
David Selfe Anaconda Operations Locked Bag 4 Welshpool Delivery Centre Pilbara Street Welshpool, WA
Ute Mueller School of Engineering and Mathematics Edith Cowan University, WA
Dear Ute,
Re: Anaconda Multivariate Data Set
Anaconda
I refer to the e-mail sent by you on the jlh November 2002. Anaconda gives its permission for publication of research results on the data supplied to you by Mark Murphy on the following basis:
• Raw data is not published, • Original coordinate data will not be published and any published data will be made
anonymous by changing locations. • Anaconda will be informed and given a copy of any publications, • Anaconda will receive due acknowledgement ofit's support of the work, and it's
permission for the publication of the results, • This permission is only applicable to the studies of Ms Ellen Bandarian, any further
use of the data will require separate permission ~o be granted from Anaconda.
Yours Sincerely
David Selfe
Chief Production Geologist
An:u.::umJ~ Opcr~liuns Pry Llli ABNIACN 4J 076 717 505 L:vcl J 2, Qu~ys1lie Bldg 2 Mill Stn:el, Penh WA 6000
TeL 61892128500 Fa~: 61 89212 8501
M:Lil~ger or rhe Murrin Murrin Nickel CobJlt Project for
Anacondo Nickel l!d :Lilli Glencon: lnlcm,lional AG 1'0 Bo~ 7.'i 12. Perth W A 6850
l'crmi~soon lcllcr 112002.doc