Using surface regolith geochemistry to map the major crustal blocks of the Australian continent

Abstract Multi-element near-surface geochemistry from the National Geochemical Survey of Australia has been evaluated in the context of mapping the exposed to deeply buried major crustal blocks of the Australian continent. The major crustal blocks, interpreted from geophysical and geological data, reflect distinct tectonic domains comprised of early Archean to recent Cenozoic igneous, metamorphic and sedimentary rock assemblages. The geochemical data have been treated as compositional data to uniquely describe and characterize the geochemistry of the regolith overlying the major crustal blocks across Australia according to the following workflow: imputation of missing/censored data, log-ratio transformation, multivariate statistical analysis, multivariate geospatial (minimum/maximum autocorrelation factor) analysis, and classification. Using cross validation techniques, the uniqueness of each major crustal block has been quantified. The ability to predict the membership of a surface regolith sample to one or more of the major crustal blocks is demonstrated. The predicted crustal block assignments define spatially coherent regions that coincide with the known crustal blocks. In some areas, inaccurate predictions are due to uncertainty in the initial crustal boundary definition or from surficial processes that mask the crustal block geochemical signature. In conclusion, the geochemical composition of the Australian surface regolith generally can be used to map the underlying crustal architecture, despite secondary modifications due to physical transport and chemical weathering effects. This methodology is however less effective where extensive and thick sedimentary basins such as the Eromanga and Eucla basins overlie crustal blocks.


Background
The composition of surface sediment or soil (s) varies spatially as a result of changes in climate (cl), organisms (o), relief (r), parent material (p) and time (t). This was expressed over 75 years ago by Jenny (1941) in the well-known expression: s = f (cl, o, r, p, t). Although this relationship has been understood conceptually for a long time, it is only recently that its quantitative prediction has become possible through (1) the development of large digital databases, (2) advances in analytical capabilities, and (3) application of multivariate and data mining statistical methods. The digital soil mapping community has been quick to show how this predictive capability can benefit knowledge discovery in pedology (e.g., McBratney et al., 2003). Applications of this new 'big data' approach in the geosciences are emerging. In a lake sediment study in Nunavut, Canada, Grunsky et al. (2014) concluded that multivariate statistical analysis of the geochemical data provided an objective basis for validating and potentially improving existing geological maps. Thus insights into parent material (p) of the sediments could be gained. Similarly, Drew et al. (2010) and Grunsky et al. (2013aGrunsky et al. ( , 2013b showed that soil geochemistry over the United States of America could accurately map climate zones (cl), ecosystems (~o), landforms (~r), and surface geology (~p) at the continental scale. All these variables are proxies for the soil formation controls listed above. Previous work by Caritat and Grunsky (2013) and Grunsky et al. (2014) demonstrated the usefulness of applying multivariate statistical methods such as principal component analysis to describe linear combinations of elements that ultimately are controlled by the stoichiometry of minerals and associated geochemical/geological processes.
The Australian continent preserves a diverse geological record spanning over 3 Ga, and is composed of different crustal blocks delimited by major crustal boundaries (e.g., Korsch and Doublier, 2016). In the present contribution we apply statistical analysis to the National Geochemical Survey of Australia (NGSA) dataset. We compare the geochemical composition of surface sediments collected on different crustal blocks, then calculate and map the posterior probabilities (PP) of all NGSA samples relative to those crustal blocks. The scientific question motivating this investigation can be stated as follows: Can we predict the location and extent (spatial distribution) of Australian crustal blocks based on surface regolith geochemistry using multivariate statistical and geospatial techniques? A positive answer to this question opens up an avenue for mapping deeper geological features from surface materials even with poor/absent outcrop, with implications for geological mapping as well as mineral exploration through cover. This effort aligns directly with two main themes of the current Australian government-industry-academia UNCOVER strategic initiative for 'providing the knowledge base and technology that will substantially increase the success rate of mineral exploration beneath post-mineralisation cover in Australia' (UNCOVER, 2016). The present contribution represents an extension of the earlier statistical analysis of NGSA data presented by Caritat and Grunsky (2013) and Mueller et al. (2014).

The National Geochemical Survey of Australia
The National Geochemical Survey of Australia (NGSA) project, a cooperation between the Federal and State/Northern Territory (NT) geoscience agencies in Australia, was part of the 5-year Onshore Energy Security Program managed at Geoscience Australia between 2006(Johnson, 2006. The NGSA was initiated to bridge a significant knowledge gap in Australia about the composition of surface regolith at the continental scale by providing internally consistent pre-competitive (i.e., government sponsored, regional and publicly available) data and knowledge to aid exploration for energy and mineral resources (Caritat and Cooper, 2016). The NGSA delivered a continental geochemical atlas , a series of reports, and a public-domain geochemical dataset (www. ga.gov.au/ngsa).
Previously existing geochemical data in Australia were perceived to have limited spatial coverage, poor internal comparability and/or inadequate extent of reported elemental compositions (Caritat et al., 2008a). During a series of pilot projects that preceded the NGSA a variety of sampling media, depths and grain-size fractions were tested at different sampling densities (Caritat et al., 2005(Caritat et al., , 2008bCaritat, 2007a, 2007b). It was found that regolith samples collected on floodplains or similar lowlying depositional landforms near the outlet of large catchments, or 'catchment outlet sediments', provided a ubiquitous sampling medium suitable for low-density geochemical mapping in Australia. These materials usually are a fine-grained mixture of detrital material originating from the main rock and soil types found within a catchment (Ottesen et al., 1989;Bølviken et al., 2004). In places, a component of eolian material is inevitably present in these settings in Australia. The effect of quartz-rich sand undoubtedly will interfere with the statistical analysis of geochemical data from surficial materials, however, the origin of the quartz-rich sand is likely the result of weathering from nearby crustal blocks. Thus, the eolian effect is expected to 'smear' the boundaries of the blocks. Pedogenic processes have occurred to some extent in those landforms and to those materials over time, and the sample material collected can be referred to as soil (topsoil and subsoil) developed on (allochthonous) fluvial ± eolian sediment as the parent material (Caritat et al., , 2012. Many geochemical surveys have been carried out in Australia by Federal and State/NT geoscience agencies, industry, and/or academic institutions (e.g., Cox and Curtis, 1977;Ryall and Nicholas, 1979;Grunsky, 1990Grunsky, , 1991Bain and Draper, 1997;Cohen et al., 1999;Davy et al., 1999;Morris et al., 2003;Cornelius et al., 2008). Before the NGSA, none was continental in scope mainly because a 'universal' sampling medium had not been recognised and also because of the perception that a relatively high sampling density was required, which would have been prohibitively expensive (Walker, 1978;L. Wyborn, pers. comm., 2003). Thus, the NGSA  is the first local attempt at delivering a uniform, internally consistent geochemical atlas and database using a common sampling medium, harmonised sampling protocols, and state-of-the-art multi-element analytical equipment for an area of 6,172,000 km 2 in extent covering~81% of the continent (Fig. 1). By necessity, the sampling density of NGSA was very low, on average 1 site/5200 km 2 , similar to, for instance, the European FOREGS project (stream sediments and soils collected at 1 site/4700 km 2 ; Salminen, 2005;De Vos and Tarvainen, 2006) and the 'Shacklette dataset' collected over the conterminous United States of America (soil and other surficial materials collected at 1 site/5925 km 2 ; Shacklette and Boerngen, 1984;Gustavsson et al., 2001). Recent applications of the NGSA data were reviewed in Caritat and Cooper (2016).

Major crustal blocks of Australia
We classify the surface geochemistry dataset according to the major crustal blocks (MCBs) of Australia. The MCBs were obtained from polygonising and simplifying the major crustal boundaries of Doublier (2015, 2016), which were defined to better  understand the crustal architecture and geodynamic evolution of Australia's key geological provinces and basins. For the generation of the major crustal boundaries dataset, the main crustal breaks interpreted from regional, deep seismic reflection lines collected across the continent since 1980 were used as starting points. Next, those major crustal boundaries were mapped in plan view away from the seismic profiles, using geological (e.g., outcrop maps, drill hole information, geochronology, isotope data) and geophysical (e.g., gravity, aeromagnetic, magnetotelluric) data. While some boundaries are exposed (e.g., Yilgarn Craton), others are covered by younger sediments/regolith and occur at greater depth. In this case the position of the boundary in map view represents the vertical surface projection of the position interpreted at depth in the deep seismic reflection data. Because a boundary between crustal blocks is not necessarily vertical and can be oblique, the surface projection of a contact between blocks can be located at some horizontal distance from the oblique upward continuation of the boundary, especially under deep sedimentary basins such as the Eromanga Basin in northeastern Australia. It is important to note that most deep crustal materials are crystalline basement rocks, except in some cases where metamorphosed sedimentary rocks may exist at quite some depth, e.g., under the Thomson Orogen in northeastern Australia.
Where crustal blocks were too small relative to the density of NGSA sampling points, they were merged either together or with an adjacent, larger block in the most geologically sensible fashion for the present analysis (M. Doublier, pers. comm., 2015). A map of the resulting 30 MCBs is shown in Fig. 1a. Table 1 shows the number of NGSA samples included in each MCB. Of the 30 MCBs, eight (MCBs 9, 17, 25-30) did not have a sufficient number of samples to warrant further statistical processing and were consequently dropped from the dataset. Supplementary material Table 1 shows the summary statistics of the chemical composition for each of the 22 MCBs that were retained for statistical evaluation and prediction. Fig. 1b shows the distribution of surface lithology (Raymond, 2012) and of the geological regions (Blake and Kilgour, 1998) over Australia, which will be referred to in the Results and discussion below.

Motivation for this study
Caritat and Grunsky (2013) applied principal component analysis (PCA) to the four combined sample types of the NGSA dataset ) after a centred log-ratio (clr) transformation was applied (see Supplementary material). The resulting principal components enabled the description and interpretation of a number of geological and geochemical processes, and formed the basis for extending the NGSA data analysis into the prediction of spatially and geochemically defined regions, as developed here. For the present study, MCBs interpreted from geological and geophysical datasets Doublier, 2015, 2016) were used as a framework to validate any geochemically distinct blocks. Of particular significance in the present study is the ability to identify and predict these distinct crustal blocks from the near-surface soil geochemistry. Since their formation these crustal blocks have been subjected to various physical and chemical processes (e.g., uplift, weathering, erosion, alluvial, colluvial and eolian deposition, and locally glaciation), implying that their primary geochemical signature has been in part obscured/modified by secondary processes.
The use of the centred log-ratio transformation ensures that the data are represented within the real number space along with the relative relationships of the data. The compromise with this approach is the possibility of not achieving multivariate normality in the transformed data. However, we are not able to achieve perfection in the mathematical/statistical treatment of the data, as seen in the summary statistics of Supplementary Table 2, which contains summary statistics for the minimum/maximum autocorrelation factors (MAF) for each of the major crustal blocks. In this study, we intend to show that the definition of the MCBs is controlled by the relative relationships of the soil geochemistry. Non-normality may have an influence on this but our conclusion is that the influence is not significant. Our experiments with robust versus non-robust methods indicated that the results of the classification do not change by any significant amount. Our conclusions are the same regardless of the use of robust versus non-robust methods.

Methods
Details on the sample collection, preparation and analysis, as well as the methods of data evaluation, analysis and classification are provided as Supplementary material. The workflow developed from the methods employed in this study is graphically represented in Fig. 2. In the following, 'TOS' refers to 'top outlet sediment' (0-10 cm depth), 'BOS' to 'bottom outlet sediment' (~60-80 cm depth), 'c/g' to 'coarse-grained' (b2 mm grain-size fraction), and 'f/g' to 'finegrained' (b75 μm grain-size fraction), as detailed in the Supplementary material.

Results and discussion
3.1. Process discovery using minimum/maximum autocorrelation factor (MAF) analysis MAF analysis does not offer the equivalent 'process discovery' of patterns or trends of multivariate data that can be observed in PCA biplots. Based on the previous work by Caritat and Grunsky (2013) and preliminary research for this paper, PCA biplots of the NGSA data coded by MCB identifiers demonstrated similar linear trends and patterns that exhibit distinct differences and overlaps amongst the MCBs. However, MAF analysis also includes the spatial context of the multi-element geochemistry, which will provide a spatially coherent pattern of trends associated with the MCBs, forming the basis for the following discussion on process validation.

Process validation of MAF using analysis of variance (ANOVA)
A critical step in process validation is the identification of variables that assist in uniquely defining the classes assigned to the data through the application of an ANOVA. ANOVA can be carried out on the clr transformed concentrations or the MAFs derived from the clr transformed data. The results of the ANOVA based on the clr transformed elements for the TOS c/g and BOS f/g are shown in Fig. 3a The plots of F-values for the elements describe the "relative" discriminating power for the variables (elements). Maximum discriminating power is obtained by using all of the elements. However, this can create degrees-of-freedom problems if there are too few observations. Thus, as a rule-of-thumb, a reduced set of elements are selected based on the identification of "inflection points" along the F-value trend. In Fig. 3a, Y-Er-Yb-Dy-Gd-Th-Cr-Na-Ho-Tb could be considered the "best" suite of elements.
For both media there is a gradual decay of F values for the elements. For TOS c/g the ten best discriminators are the elements Y, Er, Yb, Dy, Gd, Th, Cr, Na, Ho and Tb whereas the variables Au, F, As, LOI, Be, Al, Rb and Mg are poorest for group separation.
For BOS f/g eight of the ten best discriminators are the same as for TOS c/g albeit the discriminatory power differs. Similarly, five of the elements (As, Au, Be, F and Rb) are poor at group separation for both media.
Figs. 4a,b show the ordered F-values based on the MAFs for the TOS c/g and BOS f/g sample types. These plots are significantly different from the ordered plots of F-values for the variables of Figs. 3a,b. The charts that accompany Figs. 4a,b show that MAF1, followed by MAFs 1, 2, 5, 4, 3 and 6 have high F-values relative to the other MAFs for the TOS c/g sample type. The remaining MAF F-values are quite low and the graph shows a gradual decay in their values. For the BOS f/g the chart shows that MAFs 1, 2, 5, 3, 4 and 6 account for most of the discrimination between the MCBs. We found that the grain-size based F-value patterns (c/g vs f/g) have a greater contrast and discrimination ability compared to the depth based ones (TOS vs BOS), a result echoing the PCA based findings of Caritat and Grunsky (2013). For the sake of brevity, the remainder of the discussion of the results therefore will focus on the TOS c/g and BOS f/g classifications.

Linear discriminant analysis (LDA)
LDA and classification of the MCBs were applied to the MAFs of the NGSA samples. Supplementary Table 3 shows a matrix of classification accuracies from the LDA for the four sample types and the 22 classes. The diagonal along each matrix gives the proportion of correctly classified samples for each MCB, based on the training sets derived from cross validation. These classification accuracies range from 0.04 (MCB 19 TOS c/g) to 0.94 (MCB 14 TOS c/g). The overlap (confusion) of the classification is shown in the off-diagonal elements. The overall accuracies of the classification, which are defined as the sum of the diagonal of the classification accuracy matrix, divided by the row sums of the classification accuracy matrix, are shown in Table 2 with values of: 0.55 (TOS c/g), 0.59 (TOS f/g), 0.54 (BOS c/g) and 0.55 (BOS f/g). These averages are weighted according to the number of sites within each MCB. These nearly identical accuracies suggest that there is little difference between the four media with respect to their ability to correctly classify the MCBs. For some of the MCBs the predictive accuracies are different between the media and this is discussed below.
In the following we are going to discuss maps of PPs for selected MCBs. Although the PPs are compositional in nature, the application of a log-ratio transformation is problematic because of the existence of 0 probabilities and collinearity issues so that compositional cokriging cannot be used for the generation of maps. Thus, the PP maps were interpolated using soft indicator co-kriging. The following MCBs were selected for closer scrutiny (listed from west to  15,18,1,3,19,11,2,6, and 10, because they cover a representative range of larger to smaller blocks, and situations with greater to lesser degrees of accuracy in the predictions. For brevity, only three maps of the co-kriged MCB predictions are shown, namely for MCBs 4, 2 and 6. The results for all MCBs are available in the Supplementary material. Fig. 5 shows maps of the co-kriged PP for MCB 4 based on a nonrobust LDA. Only PP values N 0.2 are shown. Regardless of sample type, the predictive accuracy of MCB 4 is high especially in the southwestern part of the block. High PPs (warm colours) are marginally more extensive for the f/g than for the c/g sample types. PPs N 0.2 only occur within and near the border of MCB 4, indicating that its composition is geochemically quite unique based on the statistical methods applied in this study. Fig. 6 shows the PPs for MCB 2. This block shows PPs as high as 0.9 (c/g fractions) and overlaps with MCBs 1 (BOS c/g), 5, 6, 7, 10, 11, 13, 19, and 20. This is also reflected in Supplementary material Table 3. The highest probabilities (N0.6), however, are contained within the boundary of MCB 2. The smearing into the adjacent crustal blocks likely reflects the fact that much of the surface material for the sediments atop this crustal block consists of the Mesozoic Great Artesian (Eromanga) Basin, which contains clastic material from the surrounding highlands, particularly to the southwest (MCB 13), northwest (MCB 19) and east (MCBs 10, 5). Conversely, younger (mostly Cenozoic) sedimentary basins to the south (Murray-Darling Basin; MCB 6), west (Lake Eyre Basin; MCB 20), and north (Carpentaria/ Karumba Basin; MCB 11), partly contain recycled clastic material from the Great Artesian Basin, explaining the compositional overlap in these directions. Thus, over the area of MCB 2, very little basement material proper crops out (e.g., the Anakie Inlier) and thus the surface geochemistry in this case more likely reflects the overlying basins' geochemical signature rather than the true basement signature. Note the similarity, in places, between the PP contours and the Eromanga geological region (see Fig. 1b). For the BOS c/g sample type there is minor confusion with the central region of MCB 1, which is likely attributable to similar mineralogy.
In Fig. 7, the PPs of MCB 6 reach a maximum of 0.6, which is low relative to many of the other MCBs. There is some overlap observed with MCBs 2, 5, 11, 13, and 24, which is also indicated by the confusion values in Supplementary material Table 3. Most of these overlaps, which are the result of recycled clastic material and recent erosional/depositional processes have been discussed above. Table 2 also shows the diagonal elements of the four matrices coloured according to the accuracy of prediction. They are also sorted according to the predictive accuracy of the TOS c/g sample type. The table shows that the MCBs with the highest classification accuracies are: MCBs 14, 1, 2, 4, 8, 3, 15 and 18, whilst those MCBs with the lowest classification accuracies are MCBs 19, 22, 13, 16, 20, 24, 23 and 11. This is also geographically expressed in Figs. 8a and 9a mapping the accuracies for the TOS c/g and BOS f/g sample types for each of the MCBs. Both Figs. 8a and 9a show similar features, namely that the crustal blocks with the highest prediction accuracies are MCBs 14, 1, 2, 4, 8 and 3 regardless of sample type.
Measures of entropy (i.e., randomness; see Supplementary material), φ (local classification uncertainty), and 'most likely' MCB assignment for the TOS c/g and BOS f/g sample types are mapped spatially in Figs. 8b-d and 9b-d respectively.
In Fig. 8b, the lowest values of entropy (0.5 and below) occur mostly within MCBs 1,2,4,7,14,15,16 and 18. The highest entropies (1 and above) occur mainly in parts of MCBs 2,5,6,8,11,13,16,22,23 and 28. Fig. 9b shows the values of entropy for the BOS f/g sample media. The patterns of high and low measures of entropy are similar to the TOS c/g patterns of Fig. 8b. In both maps, the entropy values indicate greater variability in the northeast and southeast regions as well as in a central region and along some of the margins of the Yilgarn craton in the west. Conversely, the lowest entropy is displayed in the central north and south of the continent, as well as in the southwest and central east. The cause of this lower entropy (decreased disorder) may be the greater compositional stability or homogeneity of the (1) Eromanga, (2) Eucla, (3) Yilgarn and (4) Pilbara geological regions (see Fig. 1b), which are dominated, respectively (and in a simplified way) by (1) clastic sediments, (2) clastic and carbonate sediments, (3) granite and greenstone bedrock, and (4) felsic and mafic intrusive bedrock with sedimentary rocks and banded iron formations. MCBs 1 and 7 in northern central Australia also have low entropy in all four sample types, but the composition of surface materials and lithology of the basement is not as clearly homogeneous as for the other regions mentioned.
Local uncertainties (φ) for the TOS c/g and BOS f/g ( Fig. 8c and 9c) are quite similar. Low uncertainties (0.4 and below) are associated with MCBs 1,2,3,4,7,8,9,14,15,16 and 18. Elevated values of φ (0.8 and above) representing greater uncertainty are associated with MCBs 6/13/22, 2/11, 16 and 19/20. Fig. 9b (BOS f/g) shows patterns nearly identical to those in Fig. 8b, with larger areas of low entropy in MCBs 1, 4, 8 and 14. As for TOS c/g low values φ (0.4 and below) are associated with MCBs 1,2,3,4,7,8,9,14,16 and 18. The overall accuracy for sample type TOS c/g (0.47) is nearly equal to that of the BOS f/g sample type (0.48). Elevated values of φ (0.8 and above) are associated with MCBs 5/6/13/22, 2/11, 16, 18 and 19/20. The spatial patterns of φ (uncertainty) and of increased entropy (randomness) in Figs. 8 and 9 are very similar. The preliminary interpretation is suggested to be that increased uncertainty relates to a greater heterogeneity/mixture of the lithological source material in the above crustal blocks/regions. Regardless of interpretation, it is clear that greater uncertainty and disorder exist as spatially coherent regions for all four sample types. Figs. 8d and 9d show maps of the 'most likely' MCB based on the compositional similarity of a sample to a MCB from the PPs described above. Most of the cells match the MCB in which they are located, (e.g., MCBs 1, 2, 3, etc.). However, there are a few exceptions. MCBs 19 and 20 appear to be in large parts compositionally similar to MCB 2 for all four sample types. Here, transport of clastic sediments across these MCB boundaries in the Eromanga geological region (see Fig. 1b) can explain the compositional similarity. The southeastern part of MCB 6 appears to have a composition similar to that of MCB 24 to varying degrees in the four sample types. These regions being far removed from one another (eastern Victoria and northeastern Queensland), the compositional similarity cannot be explained by fluvial or eolian transport, but must instead be caused by coincidental similarity in the basement rocks (source materials for the sediments collected).

Conclusions
The results presented here identify two important features. First, the MCBs as defined through a combination of geophysical data and geological information are, for the most part, geochemically distinct. Despite long histories of weathering and deposition, the geochemical character of near-surface sediment atop these blocks persists and is by-and-large statistically unique, except where thick (several km) and widespread (100,000s km 2 ) younger sedimentary sequences cover the basement (e.g., Eromanga Basin over the Thomson basement, MCB 2). Second, the use of MAF analysis, which represents linear relationships of the data along with a measure of geospatial proximity, defines a metric from which the prediction of MCBs can be carried out with slightly greater prediction accuracy than PCA (Mueller and Grunsky, 2016). Use of MAFs is effective at describing linear combinations of elements that define distinct geochemical compositions.
The accuracy of prediction for the MCBs is highly variable. This variability is the result of compositional overlap between the MCBs. Those blocks with relatively low PPs show overlap with other crustal blocks. The off-diagonal elements of Supplementary material Table 3 show which blocks overlap with each other. The overlap of PPs can be due to a primary similarity of compositions between the MCBs or it can be due to compositional mixing through surficial processes. In most cases, the latter prevails in Australia (long periods for surface weathering and transport) especially with the sampling medium analysed here (fluvial sediment locally with eolian influence). Application of this method to outcrop rocks and drilled, fresher bedrock in areas of cover would likely reflect to a greater degree the inherent mineralogical/geochemical dissimilarities between crustal blocks.
Compositional overlap occurs in both contiguous blocks (overlap with adjacent MCBs) and, less commonly, in a spatially discontinuous form; the latter represents a similarity of compositions based on similar geochemistry (mineralogy). Thus, compositional overlap that is not geospatially associated can be interpreted as similar mineralogy and/or lithologies within these MCBs or the cover material (e.g., Cenozoic sedimentary or eolian cover). Compositional overlap between adjoining MCBs may be due to a combination of similar geochemistry of the source lithology and modification and mixing of the original compositions through mechanical and/or hydromorphic transport processes.
In areas where the PP is low (e.g., MCB 19), the overlap with other blocks (MCBs 1 and 2) suggests that surficial processes may mask the unique character of the block, particularly its potential to host base metal mineralisation.
The findings from this study demonstrate that the MCBs that form continental Australia are largely geochemically distinct and that despite long histories of uplift, burial, metamorphism, alteration through hydrothermal and groundwater interaction, with subsequent weathering, erosion, transport and deposition, these compositions persist throughout the geological history of the continent. The implications of this for mineral exploration, for instance, include the prospect of being able to geochemically recognise different parts of the architecture of Australia and use regionally meaningful geochemical background values against which anomalous samples, perhaps reflecting mineral deposits, can be identified.

Acknowledgements
The National Geochemical Survey of Australia project would not have been possible without Commonwealth funding through the Onshore Energy Security Program, and Geoscience Australia appropriation. Collaboration with the geoscience agencies of all States and the Northern Territory was essential to the success of the NGSA project and is gratefully acknowledged. We also thank all the landowners for granting access to their properties for the purposes of sampling, and the laboratory staff for assistance with preparing and analysing the samples. We thank Michael Doublier and David Huston (Geoscience Australia) for their constructive internal reviews of the original manuscript, which was significantly improved as a result. Two anonymous journal referees and Editor Franco Pirajno are gratefully acknowledged for their efforts in critically reviewing and further enhancing this paper. PdC publishes with permission from the Chief Executive Officer, Geoscience Australia.