Research article Geochemistry: Exploration, Environment, Analysis

Published online July 12, 2019 https://doi.org/10.1144/geochem2019-031 | Vol. 20 | 2020 | pp. 217–232

State-of-the-art analysis of geochemical data for mineral

exploration
E. C. Grunsky1,2* & P. de Caritat3,4
1
Department of Earth & Environmental Sciences, University of Waterloo, Waterloo, Ontario N2L 3G1, Canada
2
China University of Geosciences, Beijing, China
3
Geoscience Australia, GPO Box 378, Canberra ACT 2601, Australia
4
Research School of Earth Sciences, The Australian National University, Canberra ACT 2601, Australia
ECG, 0000-0003-4521-163X; PdC, 0000-0002-4185-9124
* Correspondence: [email protected]

Abstract: Multi-element geochemical surveys of rocks, soils, stream/lake/floodplain sediments and regolith are typically
carried out at continental, regional and local scales. The chemistry of these materials is defined by their primary mineral
assemblages and their subsequent modification by comminution and weathering. Modern geochemical datasets represent a
multi-dimensional geochemical space that can be studied using multivariate statistical methods from which patterns reflecting
geochemical/geological processes are described ( process discovery). These patterns form the basis from which probabilistic
predictive maps are created ( process validation). Processing geochemical survey data requires a systematic approach to
effectively interpret the multi-dimensional data in a meaningful way. Problems that are typically associated with geochemical
data include closure, missing values, censoring, merging, levelling different datasets and adequate spatial sample design.
Recent developments in advanced multivariate analytics, geospatial analysis and mapping provide an effective framework to
analyse and interpret geochemical datasets. Geochemical and geological processes can often be recognized through the use of
data discovery procedures such as the application of principal component analysis. Classification and predictive procedures can
be used to confirm lithological variability, alteration and mineralization. Geochemical survey data of lake/till sediments from
Canada and of floodplain sediments from Australia show that predictive maps of bedrock and regolith processes can be
generated. Upscaling a multivariate statistics-based prospectivity analysis for arc-related Cu–Au mineralization from a regional
survey in the southern Thomson Orogen in Australia to the continental scale, reveals a number of regions with a similar (or
stronger) multivariate response and hence potentially similar (or higher) mineral potential throughout Australia.
Keywords: geochemistry; analytical methods; compositional data; multivariate analytics; process discovery; process validation;
predictive mapping; machine learning; geospatial coherence; Melville Peninsula; Nunavut; Thomson Region; New South Wales
Thematic collection: This article is part of the Exploration 17 collection available at: https://www.lyellcollection.org/cc/
exploration-17
Received 29 April 2019; revised 6 June 2019; accepted 11 June 2019

What do geochemical data represent? the understanding of human health issues from natural contamination
of the source rock or the effects of the urban environment through
Geochemical datasets can be defined as geochemical data derived anthropogenic activities that result in local pollution.
from a range of media (e.g. soil, till, regolith, lake sediments, stream
sediments, bedrock) collected at a spatial scale consistent with the
geological/geochemical processes being investigated. Continental,
Survey area and density
regional and local scale surveys reveal increasingly detailed Sample density is a critical aspect of geochemical survey design and
processes ranging from the tectonic assemblage of continents to subsequent interpretation. Sample density, generally described in
hydrothermal veining, for instance. terms of the average area, or volume (in 3D) that each sample site
The intent of geochemical surveys is to provide a spatial represents, will have an influence on the detection and discovery of
geochemical description of the general geology or dominant geochemical/geological processes that have acted at a specific spatial
geochemical processes as manifested in the medium being scale. Local scale or high-density surveys have sample site densities in
sampled. For example, the geochemistry of glacial till or regolith the range of more than 100 sites per km2 to one site per km2. Regional
over an area may reflect the underlying geology, or it may reflect the scale geochemical surveys can vary from one site per km2 to one site
source material that has been transported. per 500 km2, and continental scale surveys can vary from one site per
500 km2 to one site per 5000 km2 (Geological Survey of Northern
Ireland 2007; Reimann et al. 2009, 2010, 2014; Caritat & Cooper
The value of geochemical datasets
2011; Smith et al. 2011). It is evident that high-density surveys are able
Geochemical surveys generally contribute to the economy, environ- to detect local scale processes, which can be associated with mineral
ment and society through supporting fact-based decisions in the deposits. As the density of a survey decreases, the likelihood of
following applications: mineral exploration, regional geological randomly sampling a site that is associated with alteration or
mapping, agriculture and forestry, environmental baseline monitor- mineralization decreases. Conversely, studying increasing larger
ing, environmental remediation, geohealth and general land use areas enables detection of large-scale geological processes such as
stewardship. For example, geochemical surveys can have an impact on continental accretion, collision, and major fault and shear zones.

© 2019 Commonwealth of Australia (Geoscience Australia). This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0
License (http://creativecommons.org/licenses/by/4.0/). Published by The Geological Society of London for GSL and AAG. Publishing disclaimer:
www.geolsoc.org.uk/pub_ethics

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 218 E. C. Grunsky & P. de Caritat

Geochemical data glass bead or disk that can then either be analysed directly or taken into
solution with an acid, e.g. HNO3.
The following is a brief summary of the primary considerations that The resulting acid solution is then presented to an analytical
must be taken into account when obtaining, compiling and instrument after dilution as appropriate. Common technologies include
synthesizing geochemical data prior to statistical treatment and inductively coupled plasma optical emission spectrometry (ICP-OES)
interpretation. This is not intended to provide all of the details that and inductively coupled plasma mass spectrometry (ICP-MS). In these
are necessary when obtaining geochemical data. techniques, the acid digest is first aspirated into a chamber and
typically admixed with argon gas before being converted at high
Choosing the sample material temperature to a plasma. Subsequently an optical emission spectrum is
produced where each element has a unique emission spectrum (ICP-
The choice of sample media is a critical part of the strategy of any OES). Alternatively, a mass spectrometer can be used to separate the
geochemical survey. Sampling bedrock may reveal ‘in situ’ geochem- elements or molecules based on their unique mass signature (ICP-MS).
ical processes pertaining to the underlying geology. Sampling regolith Older but still current methods of instrumentation include atomic
that has been derived by weathering of in situ bedrock may present a absorption spectroscopy (AAS). Fire assay is the preferred method of
geochemical signature that reflects both the protolith and its weath- preparation in ore-grade materials for the determination of Au, Pt and
ering. Sampling transported material, such as glacial till, lake Pd. Methods such as X-ray fluorescence (XRF) and instrumental
sediments, stream sediments, overbank sediments, colluvial or alluvial neutron activation analysis (INAA) have the ability to analyse a sample
material may reflect varying amounts of transport and mixing of without a wet digestion thus delivering a true total analysis. The former
several processes, which may be desirable. It is important to recognize is routinely used for the determination of major mineral forming
the nature of the sample media and abilities and limitations of what can oxides (e.g. Al2O3, SiO2, etc.).
be interpreted from the derived geochemical data.

Quality assurance/quality control (QA/QC)

Choosing the appropriate size fraction
A critical component of geochemical analysis is the monitoring of all
In sample media that are comprised of a mix of mineral/organic matter, procedures that result in the analytical values. Quality control
the size fraction of the mineral grains or particles analysed can be measures include the use of blanks, duplicates and standards to ensure
important in distinguishing between geological/geochemical pro- that the results produced are fit for purpose. Two ‘quality’ parameters
cesses. In most sample media, a distinction between a coarse-grained are commonly determined: precision, which measures the repeat-
(typically >63 µm and <2 mm) and fine-grained (<63 µm) fraction is ability of measurements and accuracy, which quantifies how close the
commonly used. Coarse-grained material can be considered, in many obtained results are to the ‘real’ values. Blanks allow the detection of
cases, to represent locally derived particles, or minerals that have not contamination introduced at any stage of the process, from sampling
undergone weathering, comminution or chemical dissolution. The to analysis. Duplicates can be separated into two types on the basis of
geochemical signature from fine-grained mineral matter may represent their purpose: field duplicates (collected within a given distance from
minerals that have undergone weathering, comminution and chemical the original sample), which are used to quantify the total (sampling,
dissolution/precipitation. The fine-grained size fraction is generally preparation and analytical) precision; and laboratory duplicates (split
considered to reflect a greater range of geochemical processes, in the laboratory under controlled conditions), which test only the
although this is dependent on the source material and the nature of the analytical precision. Finally, standards also come under two guises:
subsequent processes that occurred. Another consideration to be aware firstly, internal project standards (IPSs), which can track drift in the
of relating to particle sizing is that certain sample analysis methods preparation and analysis steps within and between batches and
(e.g. fusion to prepare glass discs for XRF or LA-ICP-MS analysis) secondly, certified reference materials (CRMs), which compare
may require the sample to be ground or milled to a given specification results to certified analytical results and are used to establish accuracy.
(e.g. >X% of mass passing through a 60 µm sieve). If this is the case, There is a wide range of CRMs (rocks, soils, sediments, water and
the impact of breaking up mineral aggregates and/or litho-fragments at vegetal matters) available; those chosen should be similar to the
the sample preparation stage, as a fit-for-purpose strategy, must be material that is being analysed.
borne in mind when interpreting the results.

The compositional nature of geochemical data

Choosing the appropriate analytical method – digestion
Geochemical data are, by definition, compositional in nature.
and instrumentation
Elements or oxides of elements are generally expressed as parts per
The choice of analytical method, which includes the method(s) of million ( ppm), parts per billion ( ppb), weight percent (wt%, or
sample digestion and subsequent instrumentation for determining simply %) or some other form of ‘proportion’. When data are
elemental abundances, is critical in the interpretation of the results. The expressed as proportions, there are two important limitations: first,
choice of sample digestion is generally the most important. Several the data are restricted to the positive number space and must sum to a
types of acid digestion, including four-acid (HF-HCl-HNO3-HClO4), constant (e.g. 1 000 000 ppm, 100%), and second, when one value
aqua regia (HCl-HNO3) and numerous weak/partial extractions, will ( proportion) changes, one or more of the others must change too to
(preferentially) dissolve specific mineral, organic and amorphous maintain the constant sum. This problem cannot be overcome by
phases, or target certain physical sites (e.g. adsorbed ions, exchange- selecting sub-compositions so that there is no constant sum. The
able cations). Four-acid digestion is a ‘near-complete’ digestion that ‘constant sum’ or ‘closure’ problem results in unreliable statistical
dissolves all but the most resistant minerals (e.g. monazite, zircon). The measures. The use of ratios between elements, oxides or molecular
use of aqua regia is useful for dissolving sulphide/oxide type minerals components that define a composition is essential when making
while leaving most silicate minerals unaffected. Weak acid leaches comparisons between elements in systems such as igneous
(extractions) tend to dissolve the coatings on mineral grains and/or fractionation (Pearce 1968). The use of logarithms of ratios, or
adsorbed species that are associated with alteration/mineralization simply logratios, is required when measuring moments such as
processes. Other methods of sample preparation include the use of a variance/covariance (Aitchison 1986; Egozcue et al. 2003;
total fusion (rather than digestion) whereby finely ground sample Buccianti et al. 2006; Pawlowsky-Glahn & Buccianti 2011;
material is melted with a flux (e.g. Li2B4O7) to form a homogeneous Buccianti & Grunsky 2014).

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 Advances in geochemical data analytics for mineral exploration 219

Fig. 1. Regional basement rock type map of Nunavut, Canada, showing the location of the Melville Peninsula study area.

The relationships between the elements of geochemical data are demonstrate geospatial coherence likely reflect an important
controlled by ‘natural laws’ (Aitchison 1999). In the case of geological/geochemical process.
inorganic geochemistry that law is stoichiometry, which governs Following process discovery, ‘process validation’ is the step in
how atoms are combined to form minerals, and thereby defines the which the patterns or associations are statistically tested to
structure within the data. Geochemical data are not the only type of determine if these features are valid or merely coincidental
data to exhibit structure. associations. Patterns and/or associations that reveal lithological
variability in surficial sediment, for instance, can be used to develop
training sets from which these lithologies can be predicted in areas
Methods where there is uncertainty in the geological mapping and/or paucity
Philosophy of outcrop. Patterns and associations that are associated with mineral
deposit alteration and mineralization may be predicted in the same
To effectively interpret geochemical data, a two-phased approach is way. In low-density geochemical surveys, where processes such as
suggested: initial process discovery, followed by process validation. those related to alteration and mineralization are generally under-
This strategy identifies geochemical/geological processes that exist sampled, it may be difficult to carry out the process validation phase
in the data but may not be obvious unless robust statistical methods relating to these processes.
are utilized. The process discovery phase is most effective when A multivariate approach is an effective way to start the process
carried out using a multivariate approach. Linear combinations of discovery phase. Linear combinations of elements that are
elements related by stoichiometry are generally expressed as strong controlled by stoichiometry may emerge as strong patterns, whilst
patterns, whilst random patterns and under-sampled processes show random patterns and/or under-sampled processes show weak or
weak or uninterpretable patterns. If the process discovery phase uninterpretable patterns. This approach was successfully used by
provides evidence that there is structure in the data, then models can Grunsky et al. (2012) using multi-element lake sediment geochem-
be built and tested using the process validation phase. If groups of ical data from the Melville Peninsula area, Nunavut, Canada, and by
observations are associated with specific processes (bedrock, Caritat & Grunsky (2013) using continental scale multi-element
alteration, mineralization, groundwater, weathering, gravitational catchment outlet sediment geochemical data from Australia.
sorting), then the observations can be assembled into training sets in Processes are recognized by a continuous range of variable
which the uniqueness of these groups can be tested. responses and an associated relative increase/decrease in element
concentrations. The presence of data that are reported at less than the
lower limit of detection (LLD), referred to as censored data, can affect
Process discovery and process validation
the derivation of associations in the process discovery stage. Using the
A two-step approach is recommended for evaluating multi-element detection limit, or some arbitrary replacement value (e.g. ½ LLD), as
geochemical data. In the first ‘process discovery’ step patterns, replacement values for censored data, although commonly per-
trends and associations between observations (sample sites) and formed, may bias any statistical (especially multivariate) calculation.
variables (elements) are teased out. Geospatial associations are also Treatment for censored data has been studied within the medical
a significant part of process discovery. Patterns and/or processes that epidemiology community for a long time and was recognized as a

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 220 E. C. Grunsky & P. de Caritat

Fig. 2. Geological map of the southern part of the Melville Peninsula, Nunavut, Canada, with lake sediment sample sites (shown as black dots) of Day
et al. (2009).

problem for geochemical data in the 1980s (Chung 1985; Campbell composition within the positive real number space. It has long
1986; Sanford et al. 1993). Research by Martín-Fernández et al. been recognized that many geochemical processes can be clearly
(2003) and Hron et al. (2010) provided various methods for finding described using element/oxide ratios that reflect the stoichiometric
replacement values for the censored data. For instance, the R package balances of minerals during formation (e.g. Pearce 1968).
‘zCompositions’ with the function (lrEM) (Palarea-Albaladejo & Geochemical data derived from mineral-based materials, when
Martín-Fernández 2008; Palarea-Albaladejo et al. 2014) can be used expressed in elemental or oxide form, are a proxy for mineralogy. If
to determine suitable replacement values for several of the elements. the mineralogy of a geochemical dataset is known, then the
Equally important is the distinction between missing values (i.e. no proportions of these elements can be used to calculate normative
data) and censored data. Missing values may not be censored values, mineral proportions (Caritat et al. 1994; Grunsky 2013).
requiring a decision on how they should be replaced, or if they should An essential part of the process discovery phase is a suitable
be used at all (Martín-Fernández et al. 2003). choice of coordinates to overcome the problem of closure. The
centred logratio (clr) transformation (Aitchison 1986) is a useful
transform for evaluating geochemical data. The principal compo-
Advanced analytics for process discovery nents (PCs) of clr-transformed data are orthonormal (i.e. statistically
independent) and can reflect linear processes associated with
Process discovery involves the use of unsupervised multivariate stoichiometric constraints. The PCs offer a significant advantage for
methods such as principal component analysis (PCA), independent subsequent process validation.
component analysis (ICA), multi-dimensional scaling (MDS), or
random forests (RFs), to name a few. Model-based process
discovery methods can also be used, such as model-based clustering
Advanced analytics for process validation
(MBC) or RFs. As described previously, statistical measures
applied to geochemical data typically reveal linear relationships, Process validation is the methodology used to verify that a
which may represent the stoichiometry of rock-forming minerals geochemical composition (response) reflects one or more processes.
and subsequent processes that modify mineral structures, including These processes can represent lithology, mineral systems, soil
hydrothermal alteration, weathering and water–rock interaction. development, ecosystem properties, climate or tectonic assem-
Physical processes such as gravitational sorting can effectively blages. Validation can take the form of an estimate of likelihood that
separate minerals according to the energy of the environment and a composition can be assigned membership to one of the identified
mineral/grain density. Mineral chemistry is governed by stoichi- processes. This is typically done through the assignment of a class
ometry and the relationships of the elements that make up minerals identifier or a measure of probability. The prediction of class
are easily described within the simplex, an n-dimensional membership can be done through techniques such as linear

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 Advances in geochemical data analytics for mineral exploration 221

Fig. 3. Mineral occurrences obtained from NUMIN (2017).

discriminant analysis (LDA), logistic regression (LR), neural A critical part of process validation is the selection of variables
networks (NN), support vector machines (SVMs), RFs or other that produce an effective classification. This requires the selection of
machine learning procedures. variables that maximize the differences between the various classes

Fig. 4. Sample quantile v. theoretical quantile plot showing the effect on Fig. 5. Screeplot of the eigenvalues and cumulative percentage of
the data distribution of imputation for Sb in lake sediments, Melville eigenvalues derived from a PCA applied to clr-transformed lake sediment
Peninsula. geochemical data, Melville Peninsula.

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 222 E. C. Grunsky & P. de Caritat

isometric logratio (ilr) yields the same results. Although the

covariance matrix of clr-transformed data is singular, classification
can still be undertaken using a generalized inverse and yields the
same classification results as the alr and ilr transforms. Analysis of
variance (ANOVA) applied to clr-transformed data enables the
recognition of the compositional variables (elements) that are most
effective at distinguishing between the classes. Choosing an
effective alr transform (choice of suitable denominator) or balances
for the ilr transform can be challenging and requires some
knowledge and insight about the nature of the processes being
investigated. ANOVA applied to the PCs derived from the clr
transform has been shown to be highly effective at discriminating
between the different classes (Grunsky et al. 2014 – Melville
Peninsula; Grunsky et al. 2017 – Australia). Because the dominant
PCs (PC1, PC2, …) commonly identify active processes, as
discussed above, and the lesser components (PCn, PCn-1, …, where
n is the number of variables) may reflect under-sampled processes
or noise, the use of the dominant components can be effectively
used for classification using only a few variables. Classification
results can be expressed as direct class assignment or posterior
Fig. 6. Principal component biplot of clr-transformed lake sediment probabilities (PPs) in the form of forced class allocation, or as class
geochemical data, Melville Peninsula. See text for an explanation of the typicality. Forced class allocation assigns a PP based on the shortest
element associations. Mahalanobis distance of a compositional observation from the
compositional centroid of each class. Class typicality measures the
Mahalanobis distance from each class and assigns a PP based on the
and minimizes the amount of overlap due to noise, unrecognized or F-distribution (Campbell 1984; Garrett 1990). This latter approach
under-sampled processes in the data. As stated previously, because can result in an observation having a zero PP for all classes,
geochemical data are compositional in nature, the variables that are indicating that its composition is not similar to any of the
selected for classification require transformation to logratio compositions defined by the class compositional centroids.
coordinates. For the purposes of classification, the choice of The application of a procedure such as LDA can make use of
coordinate system based on the additive logratio (alr) or the cross-validation procedures, whereby the classification of the data is

Fig. 7. Kriged map of PC2 derived from clr-transformed lake sediment geochemical data, Melville Peninsula, overlain with mineral occurrences.

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 Advances in geochemical data analytics for mineral exploration 223

Fig. 8. Map of the residual values of Au ( ppb) estimated from a robust linear regression (Au ∼ PC1 + PC2 + PC3 + PC4 + PC5) of lake sediment
geochemical data, Melville Peninsula.

repeatedly run based on random partitioning of the data into a procedure. If a geospatial rendering of a posterior probability shows
number of equal-sized subsamples. One subsample is retained for no spatial coherence (i.e. no structure, or a significant amount of
validation and the remaining subsamples are used as training set. ‘noise’), then it is likely that the classification will be difficult to
This approach produces stable results and reduces the influence of interpret within a geological context. The most effective way to test
outliers (Aitchison 1986; Tolosana-Delgado 2006; Pawlowsky- this is through the generation and modelling of semi-variograms that
Glahn & Egozcue 2016). However, the subsequent derivation of describe the spatial continuity of a specific class based on PPs. If
maps displaying PPs, which are compositions in themselves, meaningful semi-variograms can be created, then geospatial maps
requires a suitable logratio transformation to deal with the non- of PPs can be generated through interpolation using the kriging
negativity and the constant sum constraint of compositional data. process. Maps of PPs may show low overall values but still be
Posterior probabilities are transformed using an alr transform spatially coherent. This is also reflected in the classification
followed by ordinary co-kriging after which a back transformation is accuracy matrix that indicates the extent of classification overlap
carried out for geospatial rendering. It is important to note that the between classes. Geospatial analysis methodology described by
alr transform cannot be used to estimate kriging variance (Aitchison Bivand et al. (2013) and the ‘gstat’ package (Pebesma 2004) in R
1986; Tolosana-Delgado 2006). Kriging variance can be estimated can be used to generate the geostatistical parameters and images of
by the calculation of the expected value and error variance the PCs and PPs from kriging.
covariance matrix by Gauss–Hermite integration (Pawlowsky-
Glahn & Olea 2004) after which a back transform can be applied.
Classification accuracies can be assessed through the generation Two case studies
of tables that show the accuracy and errors measured from the Melville Peninsula, Nunavut, Canada
estimated classes against the initial classes in the training sets used
for the classification. Process discovery and validation
The Melville Peninsula region, Nunavut, has been the focus of
geological mapping and lake sediment and till geochemical
Geospatial coherence
sampling for the past 40 years. The example presented here
The results from the classification of samples gathered in a highlights the value of multi-element geochemical data as an aid to
geochemical survey should bear a geospatial resemblance to the area regional geological mapping and exploration targeting for potential
sampled. The creation of maps is part of the process validation base and precious metal deposits through the evaluation of regional

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 224 E. C. Grunsky & P. de Caritat

Fig. 9. Map of the residual values of Cr ( ppm) estimated from a robust linear regression (Cr ∼ PC1 + PC2 + PC3 + PC4 + PC5) of lake sediment
geochemical data, Melville Peninsula.

geochemical survey data in the Melville Peninsula area (Figs 1 and sheet. The till on most of the plateau is immature, and the matrix
2). Recent work by Grunsky et al. (2014); Harris & Grunsky, (2015) tends to be sandy.
and Mueller & Grunsky, (2016) has evaluated the lake sediment and
till geochemistry in the context of predictive geological mapping The lake sediments are the result of reworking and sorting of the
and mineral resource potential. Figure 2 shows a generalized glacial till that developed during the retreat of the ice sheet.
geological map of the area and Figure 3 shows the principal mineral The lake sediment geochemical data used in the study of Grunsky
occurrences of the area. A study in the use of till geochemical data et al. (2014) have been published in the Geological Survey of
for predictive geologic mapping using multivariate spatial analysis Canada Open File 6269 (Day et al. 2009) based on earlier studies by
is summarized by Mueller & Grunsky, (2016) and is not discussed Hornbrook et al. (1978a, b). Details on the sampling methodology
here for the sake of brevity. and analytical protocols are documented in Open File 6269. Sample
The geology is comprised of poly-deformed and poly-metamor- pulps collected in the earlier field campaigns were re-analysed using
phosed Archean and Paleoproterozoic assemblages (Machado et al. aqua regia digestion and ICP-MS instrumentation. Pulps were also
2011, 2012; Corrigan et al. 2013; Grunsky et al. 2014). The area analysed using INAA. Where elements have been analysed using
was covered by the Laurentide ice sheet during the Foxe glaciation. both methods, the elements were evaluated in terms of detection
Sandy till covers much of the northern part of the Melville limit suitability and visual examination of the correlation of the
Peninsula. The central part of the area was covered by a cold-based element with each method. This included the evaluation of the
ice cap that preserved much of the pre-glacial landscape, which is degree of censoring. QA/QC protocols and reporting are provided in
composed of weathered regolith and boulder rubble with only local the reports by Day et al. (2009) and the data were considered
glacial transport (Dredge 2009; Tremblay & Paulen 2012). adequate for statistical processing. The R statistical package (R Core
According to Dredge (2009, p. 6): Team 2014) was used to process the data.
Following the protocols described above, the data were screened for
Glacially scoured lake basins and classic glacial erosion forms values reported < LLD. Data < LLD were imputed (estimated) using
are absent. Apart from a few scattered outcrops, the southern the function ‘impKNNa’ in the R package ‘robCompositions’ (Hron
plateau surface consists of weathered regolith, or bouldery et al. 2010). Figure 4 shows a quantile–quantile plot of imputed Sb
rubble that was glacially transported for short distances. values that minimizes bias in calculating statistical moments.
The main glacial landforms are distinctive subglacial and ice After adjusting the censored values to minimize statistical bias, a
marginal channels associated with wasting phases of the ice clr transform was applied to the data. These transformed values were

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 Advances in geochemical data analytics for mineral exploration 225

Fig. 10. Map of the residual values of Ni ( ppm) estimated from a robust linear regression (Ni ∼ PC1 + PC2 + PC3 + PC4 + PC5) of lake sediment
geochemical data, Melville Peninsula.

then used to carry out a PCA on the data. A useful tool that is derived rocks in the NW part of the map, and negative scores corresponding
from PCA is the screeplot, which is shown in Figure 5. The screeplot with the supracrustal assemblages in the Paleoproterozoic Penrhyn
shows the eigenvalues plotted in descending order. The figure Group in the southeastern part of the map. Thus, as an initial part of
indicates that eigenvalues decrease rapidly and that most of the process discovery, a PC biplot provides useful information on the
variation of the data is accounted for by the first five PCs. The geochemical nature and relationships of the data. Grunsky et al.
remaining PCs can be interpreted as under-sampled or random (2014) provide more detail on the use of PCA in this area. There are
processes. The five largest eigenvalues indicate that there is several other ways that processes can be discovered in geochemical
‘structure’ in the data that is controlled by mineral stoichiometry data as outlined previously in the process discovery section.
and hence, geological processes. The structure in the data can be
visualized using a PC biplot (Gabriel 1971). Figure 6 shows a biplot
of the first two PCs for the clr-transformed Melville Peninsula lake Mineral exploration targeting
sediment geochemistry. Three generalized features are evident in In the example provided here with Melville Peninsula lake sediment
this biplot, which accounts for 42% of the variability of the data. geochemical data, the underlying geology was tagged to the sample
First, the plot indicates the relative relationships of the elements sites, which are shown in the PC biplot of Figure 6. An analysis of
(loadings) that highlight the relative affinities of the sample sites and the number of lake sediment sites associated with specific
corresponding geological domains. Second, scores of the sample lithologies is summarized in Table 4 of Grunsky et al. (2014). In
sites associated with granitoid and gneissic rocks occur along the this case, eight dominant lithologies, derived from the revised
positive PC2 axis. Third, sample sites associated with the Prince geology of Machado et al. (2011, 2012), were tagged to the lake
Albert Group supracrustal rocks and locally associated granitoid sediment sample sites. As part of the process discovery phase, it is
rocks occur along the positive PC1 axis and the sites associated with reasonable to test the ability of the lake sediment geochemistry to
the Paleoproterozoic Penrhyn supracrustal rocks occur along the distinguish between the dominant lithologies. This can be done by
negative PC1, negative PC2 axes. Maps of the first and second PCs applying an ANOVA, in which the most significant PCs provide
are shown in Grunsky et al. (2014). The map of PC1 shows a maximum distinction between the lithologies. Grunsky et al. (2014)
clustering of positive values that correspond with a region of demonstrated that PCs derived from clr-transformed geochemical
granitoid material and rocks associated with the Prince Albert data provide an effective and efficient means to demonstrate
supracrustal and associated granitoid assemblages. The map of PC2 discrimination between lithologies. Since the PCs represent linear
(Fig. 7) shows positive values associated with granitic and gneissic combinations of elements that are mostly controlled by

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 226 E. C. Grunsky & P. de Caritat

Fig. 11. Map of the residual values of Zn ( ppm) estimated from a robust linear regression (Zn ∼ PC1 + PC2 + PC3 + PC4 + PC5) of lake sediment
geochemical data, Melville Peninsula.

stoichiometry, more geological information is contained in fewer on the assumption that for the sake of regression, the commodity
components; thus, this approach is more parsimonious and effective elements are independent. In fact, there is no easy way to create a
than using the elements. In the study by Grunsky et al. (2014), it was simple regression of an element in any of the logratio-transformed
found that the first six PCs accounted for most of the lithological spaces (alr, clr, ilr). Each transform presents different problems in
separation of the data. In contrast, almost all of the 44 elements were any analysis involving regression.
required to maximize differences between the lithologies. (4) High residual values derived from a regression of an element
PCA provides insight into processes controlled by mineral against the dominant PCs may reflect processes that are potentially
stoichiometry. Sampling strategies for large-scale geochemical associated with mineralization.
surveys are useful for highlighting dominant processes such as the (5) These high residuals when plotted on a map may highlight
underlying bedrock, but are seldom at a sufficient spatial sampling areas for mineral exploration follow-up.
density for detecting processes that have small spatial footprints,
such as veining or mineralization associated with an ore deposit. The screeplot of Figure 5 indicates that the first five PCs account for
The following assumptions are made when considering data from 65% of the overall variability in the data. The remaining 41 PCs
a geochemical survey: account for under-sampled or random patterns in the data. On this
basis, parametric linear modelling of specific commodity elements,
(1) The dominant PCs reflect linear combinations of elements Au, Cu, Ni, Cr and Zn can be applied on the first five PCs. The
and their relative relationships are controlled by mineral difference between the observed values of these elements and the
stoichiometry. Typically, rock-forming processes are highlighted estimated values derived from a linear regression define the
in the dominant principal components. These processes were residuals that may represent under-sampled processes that are
confirmed in the studies of Grunsky et al. (2014) and provide the possibly associated with mineralization.
confidence that the lesser components likely represent under- Mineral occurrence information was obtained from the Nunavut
sampled processes. Mineral Occurrences database (http://nugeo.ca/apps/showing/
(2) Under-sampled processes may represent processes such as showQuery.php). Only selected commodities are presented here to
alteration and mineralization associated with a specific deposit type. demonstrate the methodology of using residual values for resource
(3) A regression of a commodity of interest (e.g. Zn) against the prediction.
dominant PCs will reflect the association of that element with the Figures 8–11 show the residuals for linear models applied to the
dominant processes. The choice of using the raw values was based first five PCs derived from the clr-transformed lake sediment

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 Advances in geochemical data analytics for mineral exploration 227

Fig. 12. Map of the residual values of Cu ( ppm) estimated from a robust linear regression (Cu ∼ PC1 + PC2 + PC3) of lake sediment geochemical data,
Melville Peninsula.

geochemical data for untransformed values of Au, Ni, Cr and Zn. Figure 10 shows a map of residual Ni values, which are restricted
The function ‘rlm’ (robust fitting for linear models) from the MASS to the Penrhyn Group supracrustal volcanic and sedimentary rocks
library in the R statistical environment (R Core Team 2014; (Ps1/2, Ps3) within the south-central part of the map area. One Ni is
Venables & Ripley 2002) was used. Figure 12 shows the results of noted in the eastern part of the map area with a few isolated residual
modelling Cu against the first three PCs. The choice of three Ni values in the range of 150–500 ppm.
components is based on the observation that most of the variability High residual values of Zn occur throughout the Penrhyn Group
of Cu is contained within the first three components, and the supracrustal assemblage (Ps1/2, Ps3) in the southern part of the map
residuals generated based on linear modelling on the first three area, shown in Figure 11. These values are closely associated with
components may further highlight areas of potential mineralization. known Zn occurrences. A few isolated lower residual values of Zn
Figure 8 shows a map of the residual values of Au. Within the occur in the east-central part of the map area.
Penrhyn Group (Ps1/2, Ps3) of supracrustal rocks in the The high residual values of Cu (Fig. 12) occur within the western
southeastern part of the map area, four lake sediment sites show portion of the Penrhyn Group assemblages (Ps1/2, Ps3) in the
high values of residual Au and are associated with local Au southern part of the map area. Isolated high residual Cu values occur
occurrences. Other areas of high residual values are located in the within the K granitoid rocks in the central part of the map area and
northern part and Penrhyn Groups in the eastern part of the area. the Prince Albert Group supracrustal assemblages (APWmv,
One isolated elevated residual Au value is located in the western part APWs) in the east-central part of the area. High residual values
of the map area within Archean gneisses (APgn, Amgn). also occur in the northern part of the area, associated with K
Although there are no known Cr occurrences in the area, Figure 9 granites, mixed gneiss and sediments associated with the Prince
shows a map of residual Cr based on the first five PCs. Elevated Cr Albert Group.
values appear to be concentrated within the north-central part of the
map area possibly associated with the Prince Albert group of
volcanics (ultramafic, mafic and felsic), sediments and associated Thomson region, Australia
mafic gneissic rocks (APwmv, APWs, APgn). It is unknown if a Cr-
steel mill was used in the sample preparation process, which could Resolution of process discovery with regional geochemical
contaminate the sample material. Alternatively, gravitational effects survey data
on heavy minerals may have resulted in local accumulation of Cr- The southern Thomson Orogen region of northern New South
rich minerals. Wales and southern Queensland (Australia) was the subject of two

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 228 E. C. Grunsky & P. de Caritat

Fig. 13. Map of PC1 of the MMI extraction of coarse (<2 mm) surficial sediments from the southern Thomson survey study area in eastern Australia (a).
Values at the sampling sites are shown by coloured circles; kriging interpolation is shown as a semi-transparent continuous raster based on the same colour
scale (eight equi-percentile classes). First vertical derivative of the total magnetic intensity (TMI 1vd) surface (Nakamura 2015) is underlain (as shaded
relief ), as are catchment boundaries (as black/grey polygons). The green polygon between Brewarrina and Bourke shows the location of Thomson
Resources Ltd tenements EL7252/7253, enlarged in (b), which shows the location of exploration drill holes over a TMI image (http://www.
thomsonresources.com.au/projects/porphyry-copper-gold). Figure from Caritat et al. (2017) published with the permission of the Geological Society of
Australia. Map (b) reproduced with the permission of the Chief Executive Officer, Thomson Resources Ltd.

adjacent regional geochemical surveys (Caritat & Lech 2007; Main Thomson Orogen to the north and the Paleozoic Lachlan Fold
& Caritat 2016), whose results were combined. Here, the thick Belt to the SE (which includes the Macquarie Arc, a fertile province
sedimentary deposits of the Mesozoic Eromanga Basin cover the for magmatic arc-related mineralization such as porphyry Cu
inferred boundary between the largely unknown Paleozoic deposits), and the Paleoproterozoic Broken Hill Domain to the SW,

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 Advances in geochemical data analytics for mineral exploration 229

A local exploration company also using soil MMI geochemistry

had coincidentally and independently developed an empirical (not
statistically derived) vector to mineralization consisting of an
enrichment in Sr, Ca and Au concomitant with a depletion in REEs
to successfully site drill holes targeting Cu–Au mineralization in the
region (J. Macauley, pers. comm. 2013). Thus, it was postulated that
the regional map of PC1 from the Thomson regional geochemical
survey could serve as a first order mineral potential map for
porphyry Cu–Au mineralization (Main & Caritat 2016). This could
potentially be extended to other Cu–Au mineralization types as
well. The implications of these findings are that several areas within
the southern Thomson region were identified to potentially be of
interest for exploration purposes at the regional scale: an area
between Brewarrina and Bourke and its eastward continuation
towards Walgett and spreading north and south; four regions in the
centre of the study area (one c. 150 km NW of Cobar, one c. 50–
150 km west of Bourke, one c. 40 km NE of Hungerford, and one c.
40 km NW of Eulo); a region c. 80 km east of White Cliffs; around
Wilcannia; and in the northwestern part of the area stretching from
the Koonenberry Belt to Tibooburra.
A test for this concept emerged when another local exploration
company reported some historical as well as new drill hole
intersections of sub-economic Pb–Zn–Cu–Au mineralization at
Warraweena, between Brewarrina and Bourke in northern New
South Wales (tenements EL7252/7253) (http://www.
thomsonresources.com.au/projects/porphyry-copper-gold). Burton
et al. (2008) attributed the mineralization at Warraweena to
porphyry-type deposits related to volcanic arcs. The boreholes
on tenements EL7252/7253 were sited not based on
geochemistry but on regional and detailed airborne magnetic
geophysical surveys (Fig. 13b). These tenements happened to be in
the centre of a circular anomalous area of high positive PC1 values
in the regional MMI survey. This was taken as providing supporting
evidence for the prospectivity model developed above (Main &
Caritat 2016).

Fig. 14. Graphical representation of the scaled and ordered eigenvalues of

PC1 for 30 elements analysed after MMI extraction in both (a) the Upscaling from regional to continental scale
regional (southern Thomson) and (b) the continental (NGSA) scale
Next, Caritat et al. (2017) tested whether the continental scale
studies. Figure from Caritat et al. (2017) published with the permission of
the Geological Society of Australia.
National Geochemical Survey of Australia (NGSA) data, which
also included MMI data, would yield similar PCA results, and if
so, if the patterns would be potentially useful for prioritizing
mineral exploration in certain regions. The number of NGSA sites
(which is well known for a multitude of mineral deposits but most was c. 1200 and the area was >6 million km2, yielding an average
significantly its world-class base metal endowment) (Fig. 13a). density of one site per 5200 km2. The PCA conducted on the same
One hundred and twenty top outlet sediments (0–25 cm), similar suite of 30 elements and following the same procedure
to floodplain or overbank sediments but with a potential (imputation, clr transformation) yielded as PC1 (53.3% of
contribution of aeolian material in places, were sampled by the variance) the association of Ca–Sr–Mg–Cu–Au–Mo at the positive
regional surveys over a combined area of 185 000 km2 (yielding an end, and of REEs-Th at the negative end (Fig. 14b). This PC
average site density of 1 sample per 1540 km2) and analysed by composition was noted to be strikingly similar to that obtained for
ICP-MS for multi-element geochemistry after weak extraction using the smaller southern Thomson regional survey (Caritat et al.
the Mobile Metal Ion® (MMI) procedure (Mann 2010; SGS 2017), 2017). The resulting national scale patterns of PC1 are shown in
among other analyses. MMI extraction was designed to target Figure 15. The implications of this research (Caritat et al. 2017)
loosely bound elements adsorbed on the surfaces of soil particles, were that (1) the dominant element associations determined by
oxyhydroxide coatings and organic matter; thereby potentially multivariate statistics at the regional scale are robust enough to
representing elements that moved up in the regolith post also emerge at much lower density continental scale surveys and
mineralization (Mann 2010). PCA carried out after imputation of vice-versa; and (2) several provinces potentially prospective for
censored values and clr transformation of the data for 30 elements Cu–Au-base metal mineralization could be determined by the map
indicated that the variance explained by PC1 was 49.3% (Main & of PC1 from the NGSA dataset (Pilbara-Capricorn-northern
Caritat 2016). The composition of PC1 was dominated by positive Yilgarn, Albany-Fraser, western Eucla, eastern Eucla, western
loadings of Ca, Sr, Cu, Mg, Au and Mo, and negative loadings of Amadeus, Victoria, Adelaide, Eromanga, southern Georgina-Isa,
the rare earth elements (REEs) and Th (Fig. 14a). This composition Curnamona-western Murray and southern Surat geological
was interpreted to dominantly reflect regolith processes (calcrete regions). This was the first study (to our knowledge) of upscaling
formation: Ca, Sr, Mg; clay minerals: REEs; and accumulation of a methodology for mineral prospectivity analysis based on the
resistate minerals: REEs) with a potential overprint of mineraliza- statistical analysis of surface geochemical surveys from the
tion expressed by the Cu, Au and Mo components. regional scale to the continental scale.

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 230 E. C. Grunsky & P. de Caritat

Fig. 15. Map of PC1 of the MMI extraction of coarse (<2 mm) surficial sediments from the NGSA project study area. Values at the sampling sites are
shown by coloured circles; kriging interpolation is shown as a continuous raster based on the same colour scale (eight equi-percentile classes). Geological
regions (Blake & Kilgour 1998) are overlain (as grey polygons); those discussed in the text are labelled. The southern Thomson project area (Fig. 13a) is
shown by the black polygon. Figure from Caritat et al. (2017) published with the permission of the Geological Society of Australia.

Concluding remarks references cited in this paper provide further detail on the use
and application of other methods.
This paper outlines a systematic approach by which regional and The examples presented in this contribution are based on weak or
exploration scale geochemical survey data can be evaluated for the partial digestions in which many of the silicate minerals are not
purposes of discovering geological/geochemical processes that decomposed. Despite this limitation, the results presented here
define underlying geology and features associated with mineral demonstrate that both the underlying lithologies and processes
systems. As a contribution to Exploration ‘17, this paper related to alteration and mineralization can be detected using
summarizes advances in the use of statistical and geospatial multivariate statistical methods.
methods since the last review of geochemical methods at Geological mapping and mineral exploration programs usually
Exploration ‘07 (Grunsky 2010). The concept of process have access to a wealth of digital data that, when used effectively,
discovery facilitates the construction of geological process provide an enormous amount of information from which mineral
models that assist in identifying the dominant processes, which, exploration and mapping models can be constructed. In the authors’
in turn, assist in unmasking and ‘discovering’ under-sampled or view, the use of high quality multivariate and geospatial methods
‘rare-event’ processes associated with mineral deposits. The through the application of modern, advanced data analytics, at
methodology of applying the consecutive steps of process detailed and regional scales, is the next step forward in finding the
discovery followed by process validation provides a systematic, greenfields mineral provinces of the future.
transparent, defensible and reproducible way for extracting useful
information from a range of data that represent geological/
geochemical processes. There are many methods available for Acknowledgements This manuscript is a minor modification of the
manuscript that was published in the Proceedings of Exploration ‘17 (Grunsky &
both process discovery and process validation. In this paper we Caritat 2017). We gratefully acknowledge Chris Nind for permitting the reprint of
have highlighted only a few of these methods. Numerous these proceedings. We also would like to acknowledge the support of Charles

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 Advances in geochemical data analytics for mineral exploration 231

Beaudry and Ken Witherly that enabled us to make this contribution to Egozcue, J.J., Pawlowsky-Glahn, V., Mateu-Figueras, G. & Barcelo-Vidal, C. 2003.
Exploration ‘17. The predictive mapping studies undertaken over the Melville Isometric logratio transformations for compositional data analysis. Mathematical
Peninsula were supported by the Geo-mapping for Energy and Minerals Program Geology, 35, 279–300, https://doi.org/10.1023/A:1023818214614
(GEM) of Canada (http://www.nrcan.gc.ca/earth-sciences/resources/federal- Gabriel, K.R. 1971. The biplot graphic display of matrices with application to
programs/geomapping-energy-minerals/18215). Collaboration with the geo- principal component analysis. Biometrika, 58, 453–467, https://doi.org/10.
science agencies of all States and the Northern Territory was essential to the 1093/biomet/58.3.453
success of the NGSA project and is gratefully acknowledged. We thank all the Garrett, R.G. 1990. A robust multivariate allocation procedure with applications
land owners and custodians, both nationally and within the Thomson region, for to geochemical data. In: Agterberg, F.P. & Bonham-Carter, G.F. (eds)
granting access to field sites for the purposes of sampling, and the laboratory staff Proceedings of Colloquium on Statistical Applications in the Earth Sciences.
for assistance with preparing and analysing the samples. We are grateful to our Geological Survey of Canada Paper, Ottawa, Ontario, 89–9, 309–318.
internal, Exploration ‘17 and GEEA reviewers Roger Skirrow, Steve Cook, Bob Geological Survey of Northern Ireland (GSNI) 2007. Tellus Project Overview,
Garrett, Natalie Caciagli and the late Peter Winterburn for their helpful comments http://www.bgs.ac.uk/gsni/tellus/index.html
for improving the manuscript. PdC publishes with permission of the Chief Grunsky, E.C. 2010. The interpretation of geochemical survey data.
Executive Officer, Geoscience Australia. Geochemistry, Exploration, Environment Analysis, 10, 27–74, https://doi.
org/10.1144/1467-7873/09-210
Grunsky, E.C. 2013. Predicting Archean volcanogenic massive sulfide deposit
Funding The National Geochemical Survey of Australia project was potential from lithogeochemistry: application to the Abitibi Greenstone Belt.
supported by Commonwealth funding through the Onshore Energy Security Geochemistry: Exploration, Environment, Analysis, 13, 317–336, https://doi.
Program, and Geoscience Australia appropriation (http://www.ga.gov. au/ngsa). org/10.1144/geochem2012-176
Grunsky, E.C. & Caritat, P.de 2017. Advances in the use of geochemical data for
Scientific editing by Scott Wood mineral exploration. Exploration 2017, Proceedings, October, 2017, Toronto.
Grunsky, E.C., Corrigan, D., Mueller, U. & Bonham-Carter, G.-F. 2012.
Predictive geologic mapping using lake sediment geochemistry in the Melville
Peninsula. Geological Survey of Canada, Open File, 7171, 1 sheet.
References Grunsky, E.C., Mueller, U.A. & Corrigan, D. 2014. A study of the lake sediment
Aitchison, J. 1986. The Statistical Analysis of Compositional Data. Chapman and geochemistry of the Melville Peninsula using multivariate methods:
Hall, New York, 416. applications for predictive geological mapping. Journal of Geochemical
Aitchison, J. 1999. Logratios and natural laws in compositional data analysis. Exploration, 141, 15–41, https://doi.org/10.1016/j.gexplo.2013.07.013
Mathematical Geology, 31, 563–580, https://doi.org/10.1023/A:1007568008032 Grunsky, E.C., Caritat, P.de & Mueller, U.A. 2017. Using surface regolith
Bivand, R., Pebesma, E. & Gomez-Rubio, V. 2013. Applied Spatial Data geochemistry to map the major crustal blocks of the Australian continent.
Analysis with R. 2nd edn. Springer, New York, 405. Gondwana Research, 46, 227–239, https://doi.org/10.1016/j.gr.2017.02.011
Blake, D. & Kilgour, B. 1998. Geological Regions of Australia 1:5,000,000 scale Harris, J.R. & Grunsky, E.C. 2015. Predictive lithological mapping of Canada’s
(Data Set). Geoscience Australia, Canberra, https://ecat.ga.gov.au/ North using Random Forest classification applied to geophysical and
geonetwork/srv/eng/catalog.search;jsessionid=3BE16A8B235E117FC9F544 geochemical data. Computers & Geosciences, 80, 9–25, https://doi.org/10.
8B361B954E#/metadata/32366 1016/j.cageo.2015.03.013
Buccianti, A. & Grunsky, E.C. 2014. Compositional data analysis in Hornbrook, E.H.W. & Lynch, J.J. 1978a. National geochemical reconnaissance
geochemistry: are we sure to see what really occurs during natural processes? release NGR 33A-1977, regional lake sediment and water geochemical
Journal of Geochemical Exploration, 141, 1–5, https://doi.org/10.1016/j. reconnaissance data, Melville Peninsula, Northwest Territories, Geological
gexplo.2014.03.022 Survey of Canada, Open File 522A, 1978.
Buccianti, A., Mateu-Figueras, G. & Pawlowsky-Glahn, V. (eds) 2006. Hornbrook, E.H.W. & Lynch, J.J. 1978b. National geochemical reconnaissance
Compositional Data Analysis in the Geosciences: From Theory to Practice. release NGR 33B-1977, regional lake sediment and water geochemical
Geological Society, London, Special Publications, 264, 212, https://doi.org/ reconnaissance data, Melville Peninsula, Northwest Territories, Geological
10.1144/GSL.SP.2006.264 Survey of Canada, Open File 522B, 1978; 14 sheets.
Burton, G.R., Dadd, K.A. & Vickery, N.M. 2008. Volcanic arc-type rocks Hron, K., Templ, M. & Filzmoser, P. 2010. Imputation of missing values for
beneath cover 35 km to the northeast of Bourke. Quarterly Notes of the compositional data using classical and robust methods. Computational Statistics
Geological Survey of New South Wales, 127, 1–23, https://digsopen.minerals. and Data Analysis, 54, 3095–3107, https://doi.org/10.1016/j.csda.2009.11.023
nsw.gov.au/digsopen/ Machado, G., Houlé, M.G., Richan, L., Rigg, J. & Corrigan, D. 2011. Geology,
Campbell, N.A. 1984. Some aspects of allocation and discrimination. In: southern part of Prince Albert Greenstone Belt, Prince Albert Hills, Melville
Howells, W.W. & van Vark, G.N. (eds) Multivariate Statistical Methods in Peninsula, Nunavut. Geological Survey of Canada, Canadian Geoscience
Physical Anthropology. D. Reidel, Dordrecht, 177–192. Map, 44 ( preliminary), 1 sheet, 1 CD-ROM.
Campbell, N.A. 1986. A General Introduction to a Suite of Multivariate Machado, G., Corrigan, D., Nadeau, L. & Rigg, J. 2012. Geology, northern part
Programs. CSIRO Division of Mathematics and Statistics, unpublished report. of Prince Albert Greenstone Belt, Prince Albert Hills, Melville Peninsula,
Caritat, P.de & Cooper, M. 2011. National Geochemical Survey of Australia: The Nunavut. Geological Survey of Canada, Canadian Geoscience Map, 81
Geochemical Atlas of Australia. Geoscience Australia Record, 2011/20, 557 p ( preliminary), scale 1:25 000.
(2 Volumes), http://www.ga.gov.au/metadata-gateway/metadata/record/gcat_ Main, P. & Caritat, P.de 2016. Geochemical Survey of the Southern Thomson
71973 Orogen, Southwestern Queensland and Northwestern New South Wales – The
Caritat, P.de & Grunsky, E.C. 2013. Defining element associations and inferring Chemical Composition of Surface and Near-Surface Catchment Outlet
geological processes from total element concentrations in Australian Sediments. Geoscience Australia Record, 2016/11, 136. Available at http://
catchment outlet sediments: multivariate analysis of continental-scale www.ga.gov.au/metadata-gateway/metadata/record/83230
geochemical data. Applied Geochemistry, 33, 104–126, https://doi.org/10. Mann, A.W. 2010. Strong versus weak digestions: ligand-based soil extraction
1016/j.apgeochem.2013.02.005 geochemistry. Geochemistry: Exploration, Environment, Analysis, 10, 17–26,
Caritat, P.de & Lech, M.E. 2007. Thomson Region Geochemical Survey, https://doi.org/10.1144/1467-7873/09-216
Northwestern New South Wales. Cooperative Research Centre for Landscape Martín-Fernández, J.A., Barceló-Vidal, C. & Pawlowsky-Glahn, V. 2003.
Environments and Mineral Exploration Open File (Report, 145), 1021 p+CD- Dealing with zeros and missing values in compositional data sets using
ROM. Available at http://crcleme.org.au/Pubs/OFRSindex.html nonparametric imputation. Mathematical Geology, 35, 253–278, https://doi.
Caritat, P.de, Bloch, J. & Hutcheon, I. 1994. LPNORM: a linear programming org/10.1023/A:1023866030544
normative analysis code. Computers & Geosciences, 20, 313–347, https://doi. Mueller, U.A. & Grunsky, E.C. 2016. Multivariate spatial analysis of lake sediment
org/10.1016/0098-3004(94)90045-0 geochemical data; Melville Peninsula, Nunavut, Canada. Applied Geochemistry,
Caritat, P.de, Main, P.T., Grunsky, E.C. & Mann, A. 2017. Recognition of 75, 247–262, https://doi.org/10.1016/j.apgeochem.2016.02.007
geochemical footprints of mineral systems in the regolith at regional to Nakamura, A. 2015. Magnetic Anomaly Map of Australia (Data Set). Geoscience
continental scales. Australian Journal of Earth Sciences, 64, 1033–1043, Australia, Canberra. Available at http://www.geoscience.gov.au/gadds
https://doi.org/10.1080/08120099.2017.1259184 NUMIN 2017. Nunavut Minerals, http://www.nunavutgeoscience.ca/pages/en/
Chung, C.F. 1985. Statistical treatment of geochemical data with observations numin.html
below the detection limit. Geological Survey of Canada, Current Research, Palarea-Albaladejo, J. & Martín-Fernández, J.A. 2008. A modified EM alr-
Part B, Paper 85-1B: 141–150. algorithm for replacing rounded zeros in compositional data sets. Computers
Corrigan, D., Nadeau, L. et al. 2013. Overview of the GEM Multiple Metals – & Geosciences, 34, 902–917, https://doi.org/10.1016/j.cageo.2007.09.015
Melville Peninsula project, central Melville Peninsula, Nunavut. Geological Palarea-Albaladejo, J., Martín-Fernández, J.A. & Buccianti, A. 2014.
Survey of Canada, Current Research, 2013–19, 17. Compositional methods for estimating elemental concentrations below the
Day, S.J.A., McCurdy, M.W., Friske, P.W.B., McNeil, R.J., Hornbrook, E.H.W. limit of detection in practice using R. Journal of Geochemical Exploration,
& Lynch, J.J. 2009. Regional lake sediment and water geochemical data, 141, 71–77, https://doi.org/10.1016/j.gexplo.2013.09.003
Melville Peninsula, Nunavut ( parts of NTS 046N, O, P, 047A and B). Pawlowsky-Glahn, V. & Buccianti, A. (eds) 2011. Compositional Data Analysis,
Geological Survey of Canada, Open File, 6269, 12. Theory and Applications. Wiley & Sons, New York, 378.
Dredge, L.A. 2009. Till geochemistry, and results of lithological, carbonate and Pawlowsky-Glahn, V. & Egozcue, J.J. 2016. Spatial analysis of compositional
textural analyses, northern Melville Peninsula, Nunavut (NTS 47A,B,C,D). data: A historical review. Journal of Geochemical Exploration, 164, 28–32,
Geological Survey of Canada, Open File, 6285, 11. https://doi.org/10.1016/j.gexplo.2015.12.010

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user
 232 E. C. Grunsky & P. de Caritat

Pawlowsky-Glahn, V. & Olea, R.A. 2004. Geostatistical Analysis of Interpretation of the GEMAS Data Set. Geologische Jahrbuch, B102, 528 p, 1
Compositional Data. Studies in Mathematical Geology. Oxford University DVD.
Press, New York, 7, 181. Sanford, R.F., Pierson, C.T. & Crovelli, R.A. 1993. An objective replacement
Pearce, T.H. 1968. A contribution to the theory of variation diagrams. method for censored geochemical data. Mathematical Geology, 25, 59–80,
Contributions to Mineralogy and Petrology, 19, 142–157, https://doi.org/10. https://doi.org/10.1007/BF00890676
1007/BF00635485 SGS 2017. Mobile Metal Ions (MMI), http://www.sgs.com/en/mining/
Pebesma, E.J. 2004. Multivariable geostatistics in S: the gstat package. Computers exploration-services/geochemistry/mobile-metal-ions-mmi
& Geosciences, 30, 683–691, https://doi.org/10.1016/j.cageo.2004.03.012 Smith, D.B., Cannon, W.F. & Woodruff, L.G. 2011. A national-scale geochemical
R Core Team 2014. The R Project for Statistical Computing. R Foundation for and mineralogical survey of soils of the conterminous United States. Applied
Statistical Computing, Vienna, Austria. Available at http://www.R-project.org/ Geochemistry, 26, 250–255, https://doi.org/10.1016/j.apgeochem.2011.03.116
Reimann, C., Matschullat, J., Birke, M. & Salminen, R. 2009. Arsenic Tolosana-Delgado, R. 2006. Geostatistics for constrained variables: positive
distribution in the environment: the effects of scale. Applied Geochemistry, data, compositions and probabilities. Applications to environmental hazard
24, 1147–1167, https://doi.org/10.1016/j.apgeochem.2009.03.013 monitoring. PhD thesis, University of Girona, 215.
Reimann, C., Matschullat, J., Birke, M. & Salminen, R. 2010. Antimony in the Tremblay, T. & Paulen, R. 2012. Glacial geomorphology and till geochemistry of
environment: lessons from geochemical mapping. Applied Geochemistry, 25, central Melville Peninsula, Nunavut. Geological Survey of Canada, Open File,
175–198, https://doi.org/10.1016/j.apgeochem.2009.11.011 7115.
Reimann, C., Birke, M., Demetriades, A., Filzmoser, P. & O’Connor, P. (eds) Venables, W.N. & Ripley, B.D. 2002. Modern Applied Statistics with S. 4th edn.
2014. Chemistry of Europe’s Agricultural Soils, Part A: Methodology and Springer, Berlin, 495.

Downloaded from http://pubs.geoscienceworld.org/geea/article-pdf/20/2/217/5017032/geochem2019-031.pdf

by Australian National University user