Abstract

The multi-element aqua regia National Geochemical Survey of Australia (NGSA) database is used to demonstrate an improved method for quantifying the degree of geochemical similarity (DOGS2) between soil samples. The improvements introduced here address issues relating to compositional data (closure, relative scale). After removing the elements with excessive censored (below detection) values, the rank-based Spearman correlation coefficient (rs) between samples is calculated for the remaining 51 elements. Each element is given equal weight through the rank-based correlation. The rs values for pairs of samples of known similar origin (e.g. granitoid-derived) are significantly positive, whereas they are significantly negative for pairs of samples of known dissimilar origin (e.g. granitoid- v. greenstone-derived). Maps of rs for all samples in the database against various reference samples are used to obtain correlation maps for lithological derivations. Likewise, the distribution of soils having a geochemical fingerprint similar to established mineralized provinces can be mapped, providing a simple, first order mineral prospectivity tool. Sensitivity of results to the removal of up to a dozen elements from the correlation indicates the method to be extremely robust. The new method is compliant with contemporary compositional data analysis principles and is applicable to various digestion methods.

Multi-element databases, often containing in excess of 50 elements, are a common product of modern soil geochemistry programs, primarily due to the quality and quantity of data produced by modern instrumentation, in particular inductively coupled plasma-mass spectrometry (ICP-MS). Presentation of informative statistical analysis derived from such datasets can be challenging and is commonly limited to one or two elements of interest, either because they are sought-after commodities or pathfinders in an exploration program or are potential contaminants in an environmental impact assessment. The alternative approach is to consider a geochemical composition as a multi-dimensional ‘whole’ and treat the data in a multivariate way. Correlation analysis can thus be used to define elements with common geochemical behaviour. Recently, principal component analysis (PCA) has been used to objectively ‘discover’ suites of elements with common characteristics (e.g. Caritat & Grunsky 2013; Zhang et al. 2014a), which can then guide follow-up interpretation. Grunsky (2010) provided a comprehensive discussion of multivariate data analysis techniques, including PCA and cluster analysis methods. The systematic and objective examination of databases for pattern recognition or inference of lithology and geology commonly requires advanced statistical and/or coding skills. The widely varying concentrations of elements in a multi-element compositional database and their interdependencies present special problems for statistics-based methods of analysis (Aitchison 1986).

Determining quantitatively how geological samples are similar or not based on major, trace element or isotopic geochemistry has applications in many fields. These include: sediment provenance, archaeology, agriculture, environmental investigations, geological mapping, digital soil mapping, resource evaluation, geochemical exploration, and forensic geochemistry (e.g. McBratney et al. 2003; Welte & von Eynatten 2004; Keegan et al. 2008; Feng et al. 2011; Bowen & Caven 2013; Frei & Frei 2013; Reid et al. 2013; Zhang et al. 2014b; Mann et al. 2015, 2016; Blake et al. 2016; Sylvester et al. 2017).

Mann et al. (2016) used the multi-element Mobile Metal Ion® (MMI - Mann et al. 1998; Mann 2010) dataset from the National Geochemical Survey of Australia (Caritat & Cooper 2011a) to develop the concept of degree of geochemical similarity (DOGS). This straightforward concept was introduced to statistically assess the similarity of two rock, sediment or soil samples to one another, using all the statistically significant elements, i.e. those with all or most results reported above the detection limit, in the database. By choosing a single reference sample and calculating the Pearson correlation coefficients of all the other samples relative to it, Mann et al. (2016) plotted maps making use of the exploratory data analysis symbology (Tukey 1977). However, those elements with concentrations far removed from the mean (e.g. MMI Ca and MMI Au) and thus likely to be outliers are systematically heavily weighted in this parametric statistical method. This can lead to unrealistically high Pearson correlation coefficients. In order to overcome this problem, we use rank-based Spearman rather than concentration-based Pearson correlation coefficients in the current paper to obtain geochemical information about sample pairs and produce maps of this improved degree of geochemical similarity (dubbed DOGS2 to distinguish it from the earlier method). Also, we use data from the NGSA aqua regia (AR) digestion database, rather than from the MMI database as Mann et al. (2016) did, to test if the method applies to stronger digestion data. Scheib's (2013) review of the NGSA AR data in relation to Western Australia included some statistical correlation based on ranking of log-transformed data.

Aqua regia (HCl:HNO3 3:1 molar mixture) digestion of samples is commonly used in exploration and environmental geochemistry programs. For instance, AR digestion of soil and sediment samples has been recommended in geochemical exploration for mineral deposits (Church et al. 1987; Rubeska et al. 1987; Heberlein 2010). In Europe the legislation requires soils, sediments and sludges to be analysed after AR digestion, e.g. for environmental impact assessment or remediation (REACH 2008). This digestion method is also recommended by the United States Environmental Protection Agency and the International Standards Organization for trace element analysis of soils (e.g. see Tighe et al. 2004; Gaudino et al. 2007; USEPA 2015; ISO 2016). Aqua regia dissolves most sulphates, sulphides, oxides and carbonates, but only partially attacks silicates, which are common rock-forming and therefore soil-forming minerals (Gaudino et al. 2007). Thus, AR is not a total digestion method and should instead be considered a strong partial digestion method (Taraškevičius et al. 2013). It has been argued that this is a useful characteristic of the digestion, especially for trace metals and metalloids, as it avoids unnecessary ‘dilution’ by silicate matrix elements (Kisser 2005).

Methods

Sample collection and preparation

In the National Geochemical Survey of Australia (NGSA) project (www.gsa.gov.au/ngsa), the lowest point of each catchment, as determined by digital elevation and hydrological modelling, was the target site for sample collection, whether it was near the catchment boundary or, in the case of an internally draining catchment, toward its centre (Lech et al. 2007). The NGSA samples are similar to floodplain sediments where alluvial processes dominate, but can also be strongly influenced by aeolian processes in many parts of arid and semi-arid Australia (e.g. Gawler Region; see Caritat et al. 2008). At each target site a surface (0–10 cm deep) ‘Top Outlet Sediment’ or TOS sample, and a deeper (on average c. 60–80 cm deep) ‘Bottom Outlet Sediment’ or BOS sample were collected. To the extent that these sediments are biologically active (e.g. presence of roots) and pedogenised (e.g. presence of soil horizons), they can also be referred to as soils (e.g. Caritat et al. 2011).

All samples were prepared in a central laboratory (Geoscience Australia, Canberra). The samples were oven dried at 40°C, homogenised and riffle split into an archive sample for future investigations and an analytical sample for immediate analysis. The latter was further riffle split into a bulk subsample, a dry sieved <2-mm (US 10-mesh) grain-size fraction subsample and a dry sieved <75-μm (∼US 200-mesh) grain-size fraction subsample (Caritat et al. 2009). Each of these subsamples was further split into aliquots of specific mass/volume as per analytical requirements.

Sample analysis and quality control/quality assessment

The analytical program for the NGSA project was extensive (Caritat et al. 2010; Caritat & Cooper 2016), with the relevant analysis for the present study being the determination of AR-soluble concentrations of 60 elements (Ag, Al, As, Au, B, Ba, Be, Bi, Ca, Cd, Ce, Co, Cr, Cs, Cu, Dy, Er, Eu, Fe, Ga, Gd, Ge, Hf, Hg, Ho, In, K, La, Li, Lu, Mg, Mn, Mo, Na, Nb, Nd, Ni, Pb, Pr, Rb, Re, Sb, Sc, Se, Sm, Sn, Sr, Ta, Tb, Te, Th, Tl, Tm, U, V, W, Y, Yb, Zn, Zr) using ICP-MS in an external commercial laboratory (Actlabs, Perth). For Au determination, a 25.00 ± 1.00 g aliquot of the sample, and for multi-element analysis (all elements above excluding Au), a 0.50 ± 0.02 g aliquot of the sample were digested in a hot AR solution to leach the acid-soluble components. Once the sample had cooled to room temperature, the solution was diluted, capped and homogenised. The sample was then allowed to settle in the dark before being diluted further with 18 MΩ/cm water. All details are given in Caritat et al. (2010).

Quality control measures of the NGSA included:

  • a field sampling manual (Lech et al. 2007) describing in detail all standard operating procedures to ensure homogenous field practices;

  • the use of the same field equipment and consumables provided centrally to all field parties to avoid random contamination due to variable quality of tools and storage bags;

  • the use of gloves during sample collection to minimize contamination;

  • the double labelling of all samples to minimize sample mix-up;

  • the collection of composite samples at each site to address the natural heterogeneity of soil and sediment, such as the ‘nugget’ effect (e.g. Ingamells 1981);

  • the collection of field duplicates at 10% of the sites to assess sample collection, preparation and analysis uncertainty;

  • the randomization of sample numbers including the field duplicates to avoid spurious spatial anomalies due to instrument drift or memory effects;

  • the insertion of blind laboratory duplicates to assess sample preparation and analysis uncertainty; and

  • the insertion of blind internal project standards, exchanged project standards (e.g. Reimann et al. 2012), and certified reference materials (CRMs) at regular intervals in samples submitted to the laboratory to assess accuracy and instrument drift.

Quality assessment results are detailed in Caritat & Cooper (2011b). In summary: elements with >50% censored values (i.e. below the lower limit of detection, LLD) were B, Ge, Hf, Lu, Re, Ta, and W; those with >50% relative standard deviation (RSD) on laboratory duplicates were B (note: numerous censored values), Na, Re (note: numerous censored values), and Se; those with >20% negative or positive bias (i.e. <80% or >120% of the certified values of the CRMs) were Mn; and those with >70% RSD on field duplicates were Ag, Au, Cd, Re (note: numerous censored values), and Te.

Data analysis

All the NGSA data and metadata are open access and free to download from the project website (www.ga.gov.au/ngsa). Here we focus on the fine fraction (<75 μm) of the NGSA TOS samples digested by AR and analysed by ICP-MS, a subset of the NGSA database that hitherto has been little studied. Elemental values below the LLD were replaced in the final database with values set to half LLD. Nine elements (B, Ge, Hf, In, Lu, Re, Ta, Tm, W) with >40% of values <LLD were removed from the database, leaving a 51 element suite for this study (Ag, Al, As, Au, Ba, Be, Bi, Ca, Cd, Ce, Co, Cr, Cs, Cu, Dy, Er, Eu, Fe, Ga, Gd, Hg, Ho, K, La, Li, Mg, Mn, Mo, Na, Nb, Nd, Ni, Pb, Pr, Rb, Sb, Sc, Se, Sm, Sn, Sr, Tb, Te, Th, Tl, U, V, Y, Yb, Zn, Zr). Note that of these, only three elements (Hg, Nb, Te) have between 20% and 40% of values <LLD; all the others have <20% of values <LLD. Haslauer et al. (2017; their Fig. 1) have shown that this imputation method (replacing <LLD values by half LLD) does not bias outcomes when up to 40% of results are <LLD. Whilst two elements, Al and Sr, had reported upper limits of detection (ULDs), no sample exceeded the ULD for Al and only 12 samples returned values above the ULD for Sr (1000 mg/kg or ppm); these were assigned a value of twice ULD. Compositional data issues such as closure and relative scale (Aitchison 1986) were addressed in the method presented here by using rank (non-parametric) statistics. Rank transformation delivers data that are not summing up to a fixed value (regardless of whether all components have been analysed or not), are not expressed in relative units (dimensionless), and are not subject to distortion by outliers and therefore are not skewed. Mainstream spreadsheet calculation software Microsoft Office/Excel (products.office.com) and open access geographical information system software QGIS (www.qgis.com) were used for data, spatial and graphical analysis.

Spearman DOGS2 method

A Spearman rank-based correlation is obtained by calculating a standard Pearson correlation on ranks, rather than on raw concentration values. Accordingly, the Spearman DOGS2 method proposed here consisted of six steps:

  1. Remove elements with more than 40% of values <LLD and, for the remaining elements, replace censored values by half LLD.

  2. Rank each element using the RANK.AVG function in Excel: 
    =RANK.AVG(Samplei,Range(1:N),1),
    (1)
    where N, the number of samples, is 1055, i ranges from 1 to N, and the final 1 is a flag indicating that the ranking is to be performed in ascending order (increasing from low to high concentrations).
  3. Select a suitable reference sample.

  4. Calculate the Spearman correlation coefficient rs v. the reference sample for all elements (here 51) using the CORREL function of Excel on the ranked values.

  5. Plot selected pairs of samples on an XY scatter diagram for pairwise comparison (thus 51 points).

  6. Map the rs values using 10 Jenks natural breaks (see below) classification in QGIS, with the topmost class representing the samples most similar to the reference sample.

The Spearman coefficient is applied here because it circumvents issues inherent to compositional data, the ranks of concentrations being no longer subject to the closure effect or in proportional units. Additionally, ranked data have a similar standardized range for each element, removing the heavily weighted influence of elements with concentrations far removed from the mean concentration (e.g. Au and Ca). This range is typically from 1 to 1055 in the present dataset although, because the RANK.AVG function assigns an average rank between the previous and the next rank when ties (samples with the same concentration) occur, this can vary somewhat. Thus, if the ordered lowest concentrations for an element are, say, 1, 1.2, 1.2, 2.7 mg/kg, the corresponding ranks will be 1, 2.5, 2.5, 4. For the present dataset with 51 variables, an rs of 0.4 is significant at p < 0.002 and an rs of 0.5 is significant at p < 0.0001.

Application of DOGS2 to key NGSA catchment sites

The average size of the NGSA catchments is c. 5200 km2, an area that in most cases will include more than one lithology. In order to demonstrate the capabilities of the DOGS2 methodology, NGSA catchments likely to be representative of different but dominant underlying geochemical characteristics ultimately derived from bedrock have been chosen as reference sites; in each case the TOS sample represents the integrated geochemistry of the catchment. The aim of comparisons in this case is to exemplify and validate the outcomes of DOGS2 catchment comparisons, not to delineate specific lithologies; the latter is an aim which would be more accurately realised using residual soils over known and identified geological sites as reference samples. Reference catchment sites were chosen after consulting 1:250 000 scale geology maps, the Australian mineral deposits and occurrences databases OZMIN and MINLOC (Sexton 2011), and, the element concentration values within the NGSA AR database. The result is that we have identified a small number of ‘acid felsic’, ‘sediment’ and ‘mafic/ultramafic’ dominated catchments as well as some hosting ‘typical’ gold, copper, and iron deposits and occurrences with which we illustrate the scope and potential of the DOGS2 technique. Their locations are shown on Figure 1.

Selecting which DOGS2 results to highlight

The results of DOGS2 analysis can be spatially mapped in various ways (1) using Exploratory Data Analysis classifications (Mann et al. 2016); (2) by applying an arbitrary r cut-off value, e.g. 0.5, to highlight higher values deemed ‘closely related’; or (3) by identifying the highest class of samples using ‘statistical breaks’ within the range of rs values; this highest class denotes closest relationship to the reference sample. In the present case rs values obtained by the Spearman correlation approach described above are classified into 10 classes using Jenks natural breaks optimization (Jenks 1967), related to Fisher's discriminant analysis, as the classification method. The method iteratively uses different breaks in the dataset to determine which set of breaks has the smallest in-class variance. This visualization provides an effective method for interrogating the database for samples with closest affinity (say, top one or two Jenks classes) to the chosen reference sample. Maps are plotted here using QGIS software in Lambert conformal projection.

Results

Results from the AR analysis of the NGSA TOS fine fraction samples are summarized in Table 1 for 51 elements. Basic concentration and rank statistics are given and clearly illustrate that all elements regardless of their general abundance in these samples (ranging from <0.0001 mg/kg for Au to 400 000 mg/kg for Ca) play an equal role once rank values are used (all within 1–1055 range). Using these ranks for subsequent data analysis circumvents compositional data limitations of closure and relative scale. The compositions of sample pairs are independent and thus correlation analysis and linear regression are statistically valid techniques.

Application of DOGS2 to catchments of known provenance

The XY scatter plots of the rankings of any two samples provide a useful insight into the Spearman correlation and the geochemistry. Sample #2007191559 is a sample from a sub-catchment of the main Swan-Avon drainage system near Beverley on the western side of the Yilgarn Craton (see Fig. 1). The 1:250 000 scale geological sheet for Corrigin (Muhling & Thom 1985a) indicates that the main rock types for this catchment are adamellite, gneiss and migmatite (i.e. acid felsics) with minor (<5%) greenstone enclaves. The rankings in Figure 2 compare this sample with sample #2007190602 from near Merredin, also from the Swan-Avon system, some 140 km to the NE; the 1:250 000 geological sheet for Kellerberrin (Muhling & Thom 1985b) indicates that the main rock types for this catchment are adamellite (Kellerberrin batholith), adamellite with xenoliths of gneiss (i.e. acid felsics), and again with less than 5% greenstone enclaves.

Despite there being some elements with different rankings in the two samples (e.g. Ho, Yb, Zn, Cd) a high positive Spearman correlation coefficient (rs = 0.75) confirms that these samples are geochemically very similar. Moreover, the linear regression on the ranks has a slope close to unity (0.88) and a small intercept (84.9). It is concluded that this geochemical fingerprint is diagnostic of ‘acid felsic’ geochemistry.

The ranking diagram for the TOS from the Beverley acid felsic catchment v. the TOS selected from a mafic/ultramafic dominated catchment 100 km east of Kalgoorlie, is very different (Fig. 3); the 1:250 000 geological sheet for Kurnalpi (Williams & Doepel 1971) suggests that the main rock units for the latter (#2007190484) are mainly from the Mulgabbie (basic) and to a lesser extent Gindalbie and Gundockerta (turbidites, clastics, acid volcanics) Formations.

In this case the Spearman coefficient of correlation is negative (rs = −0.55). The linear regression on the ranks has a negative slope (−0.81) and a large intercept (986). This is driven by elements such as the REEs, U and Th, which are elevated in the granitoid-derived soil and depleted in the greenstone-derived soil, and elements such as Au, Co, Cu, V, Se, Sr and Ca, which have the opposite behaviour.

The ranking diagram between two soil samples from two predominantly mafic/ultramafic catchments is shown in Figure 4. In this case the Kalgoorlie greenstone (#2007190484) sample is compared to one further south in the Yilgarn Craton (#2007190142) – from a catchment near lake Cowan 100 km east of Widgiemooltha. The Widgiemooltha 1:250 000 geological sheet (Griffin & Hickman 1988) indicates that the main rock types in this catchment are metamorphosed Archaean mafic and ultramafic rocks with minor amounts of mafic/ultramafic material from Proterozoic dykes. The Erayinia granitoid complex is marginal to the NE.

The high positive Spearman correlation coefficient (rs = 0.66) between these two samples together with the positive slope (0.72) and low intercept (160) of the linear regression suggest a high degree of geochemical similarity. This is evident on the distribution of elemental values on the XY scatter plot (Fig. 4) where elements such as Cr (labelled), Ni, V, Sr, Se, Sc (unlabelled) in the top right of the diagram are all ranked relatively highly in both samples. The REEs (within the ellipse in Fig. 4) are ranked relatively low in both of these samples. It appears that the Spearman correlation coefficient of sample pairs has the potential to provide a meaningful measure of the degree of geochemical similarity between them.

Selection and use of reference samples

The selection of reference samples and the lithological and geochemical characteristics used to select them are shown in Table 2; their locations are shown in Figure 1.

Selection of the first four reference samples in Table 2 is facilitated by the fact that in general catchments in the Yilgarn Craton are granite-dominated towards the west, and greenstone-dominated towards the east, although it is likely that all catchments comprise multiple/mixed lithologies. Column 4 of this table shows the concentrations of elements Ca, Ce, Fe and Ni, which represent diagnostic characteristics. The various lithologies as identified here appear to be represented by elevated concentrations of various elements in the NGSA samples: acid felsics (Ce), mafics/ultramafics (Ni), marine carbonates (Ca), and banded iron formation (Fe). The concentrations of Cu in catchment samples #2007190065 and #2007191289, which will be investigated in some detail below, are both elevated and similar. Likewise values for Au in the Coolgardie and Ballarat catchment samples are similar and higher than in other reference samples and were used in the selection of these catchments as reference samples representative of catchments dominant in gold prospective lithologies (‘auriferous geology’).

As with selection of the reference catchments, ‘identified samples’ (those highlighted in red or orange on the accompanying maps) with closest affinity to the reference samples identified by the DOGS2 process are subject to the same constraints imposed by the NGSA sampling method (e.g. low sampling density, large catchments and ensuing likely mixed lithological input, variable distance of sample site to bedrock outcrop/subcrop, weathering history, etc.). Only a selected number of ‘identified samples’ with closest affinity to the reference sample (as determined by the Jenks breaks, see Methods section) are plotted in each case for clarity.

From each reference sample a Spearman correlation coefficient for very other sample in the NGSA database can be obtained. The correlation coefficients can be plotted as a single independent variable located on a map by the sample coordinates (for further detail please refer to the Methods section).

Degree of geochemical similarity map v. Beverley acid felsic catchment

The map with Beverley acid felsic catchment as the reference is shown in Figure 5. This map highlights the top Jenks class many of which are located in the Yilgarn Craton; Cassidy et al. (2006) discuss some of the reasons why more of the Yilgarn Craton catchments are not in this top class. In addition, samples within the same class are located immediately SE of the Yilgarn Craton in the Proterozoic Albany–Fraser geological region and to the NW of the Craton where rivers draining the northern part of the Yilgarn Craton reach the Indian Ocean. The quartz monzonites and adamellites of the Albany-Fraser geological region are thought to be reworked Archean granitic material (www.ga.gov.au/provexplorer/province). The high DOGS2 of a soil sample here supports this contention in at least one catchment. Other samples with a similar high degree of similarity (rs > 0.58) occur along northern Australian coastlines and are related to acid felsics within catchments on Cape York Peninsula and in Arnhem Land. A group of top Jenks class samples are shown in central Australia, and one is also evident in western Victoria in the vicinity of the Grampian Mountains (Victoria Valley granites), and another in southern South Australia near a granite quarry.

Degree of geochemical similarity map v. Kalgoorlie mafic/ultramafic

Figure 6 shows the map obtained when the Kalgoorlie mafic/ultramafic dominated catchment is chosen as the reference sample for DOGS2 analysis. Proximal samples with DOGS2 in the top Jenks class with respect to that reference sample are limited to the eastern goldfields portion of the Yilgarn Craton. Only in this part of the Yilgarn Craton are catchments dominated by mafic/ultramafic lithology. Three top Jenks class samples with high rs relative to the Kalgoorlie mafic/ultramafic catchment sample also occur within the western Eucla geological region near the boundary to the Albany-Fraser Belt. Recent Au discoveries in this belt, such as the 6 million ounce Tropicana deposit (Doyle et al. 2013), occur in felsic granulite thought to be the reworked margin of the Craton. A few catchments with a high degree of similarity (rs > 0.52) to the Kalgoorlie mafic/ultramafic dominated catchment also occur in the Pilbara geological region. This similarity is probably based on geochemistry resulting from a combination of predominantly mafic/ultramafic bedrock lithology and regolith generated in arid to semi-arid terrain. Two catchments also with high similarity to Kalgoorlie greenstone catchment occur in near coastal environments, on the Exmouth Gulf in Western Australia and on the Nullarbor Coast in South Australia; these are more likely to be marine carbonates (dolomites?) than locally derived greenstone, although the possibility of transport of mafic/ultramafic material from the Pilbara and Yilgarn Cratons cannot be excluded. Looking at the map of Australia highlighted with the highest Jenks class as shown in Figure 6, it would appear that catchments with highest similarity to the Kalgoorlie mafic/ultramafic dominated catchment are proximal to known mafic/ultramafic outcrop/subcrop (Yilgarn, Pilbara, northeastern New South Wales, central eastern Queensland, Tasmania, and Victoria).

Degree of geochemical similarity map v. Nullarbor limestone

Both the NGSA MMI Ca map of Mann et al. (2012) and the modelling of Wilford et al. (2015), among other evidence, indicate that carbonate terrain is common in the arid and semi-arid areas of Australia. However, when a carbonate soil sample from the Nullarbor Plain (#2007190159) is chosen as the reference, and the other NGSA AR samples correlated with it, the distribution of the most highly correlated samples is more limited (Fig. 7). There is, however, more than one type of carbonate (high Ca) terrain in Australia and it is likely that multi-element geochemistry distinguishes them. All except four of the samples with the highest DOGS2 relative to the Nullarbor Plain reference (Fig. 7) are within the Eucla geological region (Blake & Kilgour 1998). The Nullarbor Plain in the Eucla geological region may be the most extensive Miocene carbonate deposit described to date (O'Connell et al. 2012). The four other samples are either just outside the eastern boundary of the Eucla region (just inside the Gawler geological region), or in the St Vincent and Pirie basins within the Adelaide geological region, and the Murray Basin to the east of the Eucla region. The first of these probably represents a mechanically transported geochemical signature. The three others are in basins of more recent (Cenozoic) marine sediments that have progressively been overlain in part by fluvial-lacustrine sediments (www.minerals.dmitre.sa.gov.au). Thus, it appears that sample #2007190159 is a specific reference for marine limestone. For much of the remainder of inland Australia soil samples have a (much) lower degree of similarity to Nullarbor limestone sediments. Clearly a large number of secondary or regolith-derived calcretes in Australia are distinct from these primary marine sediments when multi-element geochemistry is considered.

The XY scatter plot of the rankings for the Nullarbor marine limestone sample and a regolith calcrete from the Gascoyne River catchment located in the NW of the Yilgarn Craton (see Fig. 1) in Western Australia is shown in Figure 8. With an rs < 0.01 the marine and regolith carbonate samples are distinguished (i.e. shown to have close to no similarity whatsoever) by this technique. Calcium is ranked slightly more highly than Mg in the marine limestone (ranks of 1000 v. 947). While both samples have similar high rankings of Fe (c. 800), the ranking for Mg (1054) is higher than that for Ca (925.5) in the regolith calcrete. Thus, pedogenic carbonates have lower Ca/Mg ratios (ranks of 925.5/1054) and Ca/Sr (925.5/1041) ratios than marine carbonates (ranks of 1000/947 and 1000/922.5, respectively). The calcrete sample also has significantly higher ranks for U, V, Mo, Co, As, and Au (all >850) than its marine counterpart (all <460). Valley calcretes are known hosts for the U mineral carnotite (K2(UO2)2V2O8.3H2O) with the U being derived from the weathering of granite (Mann & Deutscher 1978). Regolith carbonates, or pedogenic calcretes, are also known to be effective sampling media for exploration for gold deposits, with their Au content being the most effective vector to mineralization (Reith et al. 2011; Lintern 2015). Conversely, relative to the calcrete, the marine limestone is clearly more enriched in REEs (ranks >450 v. <270) and especially chalcophile elements Pb, Cd, Sn, and Tl (ranks >840 v. <350). It should be noted that whilst REEs are more highly ranked in marine limestone than the pedogenic carbonate the rankings are moderate in comparison to those from acid felsic catchments, and maybe indicate that the pedogenic carbonate is associated with mafic/ultramafic source rocks very low in REEs.

Degree of geochemical similarity map v. Mt Isa/Cloncurry mineralization

A catchment sample site some 150 km to the north of Mt Isa and Cloncurry on the Leichardt River provides a suitable reference site for the mineralized Mt Isa/Cloncurry district. This is a well-known and historic Pb-Zn and Cu mining area (Pirajno 2010). There are also many other major and minor ore deposits and mineral occurrences in the region, including Cu-Pb-Zn (Mt Isa), Cu-Au (Ernest Henry), U, Ag-Pb-Zn (Cannington), and Au (Tick Hill). The NGSA reference sample (#2007190065) contains 50.3 mg/kg Cu (rank 1024.5). The sample site is at the confluence of Gunpowder Creek and the Leichardt River, which is prone to episodic flooding; the sample probably represents a composite geochemical fingerprint for geology hosting a number of deposit types and styles, but because of the Gunpowder Creek's proximity, it is likely to be predominantly that of clastic-dominated siliciclastic-carbonate sediment hosted massive sulphide deposit type for which Mt Isa is a classic example. The rs map for all NGSA AR samples against this sample is shown in Figure 9.

The correlation map (Fig. 9) shows a number of samples in different areas of Australia with a high geochemical affinity (top two Jenks classes, or rs > 0.39) relative to the Mt Isa/Cloncurry catchment reference sample. Some of these areas host known historic or operating Cu mines (e.g. Mt Cuthbert and Mt Gordon in Queensland; Myrtle and Redbank in the Northern Territory; Cadia and Cobar in New South Wales).

In Western Australia, 27 samples with a DOGS2 in the top two Jenks classes relative to the Mt Isa reference occur in catchments stretching from the Albany-Fraser geological region (1) in the SE, through the Yilgarn (13), to the Capricorn (6), Pilbara (2) and northernmost Pinjarra (5) geological regions to the NW. Some of these anomalies, again, host known (current or historic) Cu mines or prospects in their catchment areas (e.g. Sherlock Bay, Ashburton Copper). The five anomalies in the Pinjarra geological region are likely due to transported sediment and their geochemical signatures from the Capricorn and/or Pilbara geological regions. The sample in the Albany-Fraser region, which contains 45 mg/kg AR Cu (rank 1006), is, however, likely to be reflecting a real mineral potential for Cu deposits, as there are known prospects (Mammoth, Fraser Range Target, Simons Hill, Alpha) within the sample's catchment.

Other well-known Cu districts, however, do not show up within the two highest rs classes on the map of correlation to the Mt Isa/Cloncurry catchment sample, despite the fact that NGSA samples have elevated Cu values in these regions. Examples include the Gawler and Curnamona geological regions (South Australia and western New South Wales), which are known to host IOCG deposits, e.g. Olympic Dam, Prominent Hill and Carapateena (Pirajno 2010). This is not too surprising as these deposits are buried by more than 200 m of unrelated cover.

A second Cu reference sample was chosen from the Curnamona region. Sample #2007191289 with 36.8 mg/kg Cu (rank 962) and 1.1 µg/kg (ppb) Au (rank 673) was taken from the drainage entering the north of Lake Frome, near the South Australia-New South Wales border. Samples that fall in the two upper Jenks classes after calculation of Spearman correlation relative to the Curnamona reference #2007191289 are shown on Figure 9 with red and orange squares. They form a relatively coherent grouping in South Australia, which extends across the Gawler and Curnamona regions into western New South Wales and the SE of the Northern Territory; it coincides with much of the area shown on Figure 9 with known Cu mineralization likely to be of the IOCG type. One ‘outlier’ is from the southern Thomson region of New South Wales, an area which was studied by MMI as a pilot project to the NGSA (Mann et al. 2009) and also the subject of a PCA study of the MMI NGSA data (Caritat et al. 2017), whilst two samples further north in the Eromanga region of central Queensland are probably ‘false positives’. There is one sample from the Cu mineralized Mt Isa/Cloncurry district of this type, and two samples from the coastal Pilbara of Western Australia which overlap samples of the Mt Isa type; the latter are from outlet sediments of the Sherlock and De Grey Rivers, catchments which host VMS style Cu-Zn-Pb deposits.

The element ranking diagram for sample #2007190065 (Mt Isa type) v. #2007191289 (Curnamona type) (not shown) indicates that Se, Mo, Bi, Mn, Gd and Co have appreciably higher rankings in the former, whilst Na, Te, Sn, Sr, Al and Ca have higher rankings in the latter reference.

Degree of geochemical similarity v. Roy Hill banded iron formation (BIF)

The Pilbara of Western Australia is well known for the economic iron deposits hosted by banded iron formation (BIF) (Blake & Meakins 1993). Figure 10 shows the distribution of catchment samples in the NGSA program when referenced to a catchment sample #2007190917, located on the Fortescue River near the Roy Hill iron mine.

Samples with rs > 0.56 relative to the Roy Hill reference sample also occur further downstream on the Fortescue River, where the Bight, Murrays Hill, Fig Tree and Fingers iron mines are found, and around Newman some 100 km to the south, close to Mt Newman iron ore mine. The series of samples with top class DOGS2 relative to the Roy Hill reference that stretches SE from the Pilbara to the Yilgarn regions have several iron ore mines/deposits in their catchments or close proximity (e.g. Hardey, Rocklea, Robinson Range, Beyondie, Wiluna West Iron and Goongarrie South Iron). Of the two anomalous samples in South Australia, one is close to the Iron Knob/Monarch Fe mine) and the easternmost one is in a catchment 70 km east of Burra where the Razorback Iron Project is currently being evaluated for magnetite potential (https://www.magnetitemines.com/razorback-iron-project).

The sample in the Jenks upper class relative to the Roy Hill reference in the Northern Territory lies in the Proterozoic belt which extends south from Tennant Creek.

Degree of geochemical similarity map v. Coolgardie and Ballarat (gold) references

Catchment outlet sample #2007190717 is located 12 km east of Coolgardie near a small salt lake, Brown Lake. The sample contains 43.7 µg/kg Au after AR digestion (i.e. 437 times LLD or rank 1050) in the <75-µm soil fraction. The dot map shows a number of catchment outlet samples in the auriferous Eastern Goldfields of Western Australia within top DOGS2 class (rs > 0.53) relative to sample #2007190717 (Fig. 11). On the Yilgarn Craton these extend from east of Widgiemooltha in the south to Wiluna in the north. Further north of this, another sample, #2007190167 is on the Gascoyne River just to the north of the Yilgarn Craton margin and west of the Marymia gold mining district which includes the Marymia and Plutonic mines. Even further north are a number of samples with rs > 0.53 relative to #20070717 in the Pilbara Craton; not all of these, including some on the Fortescue River, are in the vicinity of known gold mining areas. These warrant follow-up sampling, as they were also observed as anomalous in a previous publication (Mann et al. 2013). In South Australia several samples within the top DOGS2 class relative to #2007190717 occur in the auriferous Gawler Craton, and further east another anomalous sample occurs 20 km to the north–NW of Tamworth in New South Wales. It is of note that many Victorian, New South Wales and Queensland auriferous locations do not have catchment outlet samples which relate closely to the Coolgardie reference.

The red squares in Figure 11 show NGSA samples with top Jenks class similarity to another type of gold deposit, for which the NGSA sample #2007190042 near Ballarat, in Victoria, was chosen to be the reference. This sample contains 18.4 µg/kg Au after AR digestion (rank 1039) in the <75-µm soil fraction. The Jenks upper class cohort v. this reference produces samples in most areas in the eastern states shown as auriferous, up to and including the Palmer River goldfield on the Cape York Peninsula. The ranking diagram for the Coolgardie reference (#2007190717) v. Ballarat reference (#2007190042) shows that the Coolgardie reference has very high rankings for many elements such as Cd, Cu, Zn, Te, Sc, Ga, Al relative to the Ballarat reference, which in turn has higher rankings for elements such as Nb, Ce, Nd, Sm, Gd, Pb and Hg. It is clear that these reference samples pick out a different set of NGSA samples, showing that not only is the high Au content important but the complete geochemical fingerprint allows targeting of very specific conditions, such as comparing Archaean greenstone v. Eastern Australian (younger sediment-related) style Au mineralization. This approach, which considers significant Au value and other associated elements, could be important for outlining areas prospective for Au in covered terrain.

Discussion

It is fundamental to this method that Spearman correlation coefficients differ appreciably in soils derived from different lithologies. For this to occur the concentrations subsequent to AR digestion and ICP-MS analysis of many if not most elements have to vary sufficiently between samples. It is also important that in digestions of similar soils, the ranking of elements digested remains similar. In Table 3 the rankings of some key elements in soils from catchments with different dominant lithologies are presented. Whilst the similarity of individual element rankings from catchments with similar lithologies (top six rows) evidently is conducive to the high DOGS2 between respective pairs, it is instructive to also examine the geochemical makeup of samples representing mineralized areas (bottom five rows).

Felsic igneous bedrock, represented by the granite-derived soil samples in Table 3, have their rankings dominated by REEs Ce, La and Yb, followed by Al, K, Mg and Ag; Ba, Fe and Cu have the lowest ranks. Mafic lithologies (greenstone-derived samples) have their rankings dominated by Ca, Ni, Mg, Fe, Cu and Au; Ag, Ce and Yb have the lowest ranks. Carbonate-derived soils have different geochemical signatures depending on their origin: marine limestone has high ranks for Ca, K, Mg and Al, with low Au, La and Ce; regolith carbonate has high ranks for Mg, Ni, Cu and Ca, with low La, Ce and Yb. The change in relative ranking order between Ca and Mg between marine (Ca > Mg) and pedogenic (Mg > Ca) carbonates noted here is in accord with the findings of Wolff et al. (2017), indicating that practical geochemical indicators derived from raw geochemical concentrations can be preserved when converting to rank statistics The differences highlighted in Table 3 allow for a practical differentiation of marine v. pedogenic carbonates: the latter will likely have higher Mg, Ni, Cu and Au and lower Yb, La, Ce and Ag rankings (and concentrations) than their marine counterparts. Thus elevated Mg/REE, Ni/REE, Mg/Ca, Cu/Ag or Au/Ag rank ratios in carbonate-derived soils will tend to indicate a pedogenic (calcrete) origin.

The Mt Isa type Cu mineralization style is indicated by a geochemical fingerprint with elevated Cu, Fe, La and Ce ranks, and depressed Al, Ca and Mg ranks. The Curnamona style of Cu mineralization, while still having an elevated Cu rank, also has high rankings for Ca, K and Fe, and low ranks for Yb and Ag. Banded Iron Formation mineralization typical of the Pilbara region is diagnosed by soils with raised Cu, Fe, Ni and Ag ranks accompanied by low Ce, Au, Ca and La ranks. It appears that soils from auriferous regions in general have high Au and Ni ranks, and intermediate to low Ba, Ca and Yb. ‘Archean Yilgarn’ style Au mineralization differs from ‘Paleozoic Eastern Australian’ style Au mineralization by its higher Cu, Al, Ca and K ranks and its lower La, Ce, Au and Ag ranks.

Sensitivity analysis shows that this outcome is relatively independent of values of Au and Cu in the samples; in column 3 of Table 4 the rs values for a number of reference samples, including the Coolgardie (#2007190717) and Ballarat samples (#2007190042) with Au and Cu removed from the Spearman rs calculation are compared to those in column 2 where all 51 elements are included. The changes in all cases are minimal – commodity elements Au and Cu have only a minor effect on the Spearman calculation. This extends to elements such as Mg and Ni, which it might be thought have an important effect on comparison of catchments with greenstone lithology (#2007190484 v. #2007190142) and Ca and Fe, the major elements in many aqua regia digests. The sensitivity analysis suggests the method is extremely robust and relies on the contribution from a large number of elements to generate similarities (or differences).

It is only when the 12 rare earth elements (lanthanides) are removed, in addition to Au, Cu, Mg, Ni, Ca and Fe that some reference comparisons (column 6 of Table 4) show differences in rs of >0.1 compared to the values in column 2. The Spearman rs shown in bold in Table 4 between a predominantly granitoid catchment (#2007191559) and a greenstone dominated one (#2007190484) decreases i.e. the samples appear more similar with removal of the REEs indicating their importance in distinguishing acid felsic from mafic/ultramafic lithologies. Removal of the REEs however decreases the high rs between granitoid samples #2007191559 and #2007190602 illustrating that elements other than the REEs are also involved in the differences between these two granitoid dominated catchments. Similar conclusions can be drawn in the case of the marine carbonate and pedogenic calcrete comparison (#2007190159 v. #2007190881) and the overall geology of the two Cu prospective catchments (#2007190065 v. #2007191289).

Comparison of the outcomes from this study with those from the previous DOGS (Mann et al. 2016) study is difficult to do in detail because (a) the analytical technique is different (AR v. MMI), (b) the DOGS2 technique in this study utilises Spearman correlation (ranking) of analytical data rather than a Pearson correlation of log-transformed data, and (c) the method of assessment of ‘upper class’ uses Jenks breaks rather than upper two percentiles. Notwithstanding this, the outcomes share similarities; many of the identified (‘upper class’) samples from the present study are within the same geological regions that contained the upper two percentiles of MMI samples (Mann et al. 2016). One important difference however is that the Spearman r (rs) values are in all cases lower than the equivalent Pearson r values; this is almost certainly due to the ranking procedure. In some cases the Spearman ranking coefficient is negative (e.g. granite v. greenstone). This increased spread of correlation coefficient values is an advantage in sample comparisons, making the technique more sensitive.

We tested the results of a clr transformation compared to our DOGS2 rank-based method. The two approaches give similar results, with mostly the same NGSA samples being identified as the most similar to a reference sample (e.g. #2007191559, Fig. 5), and the resulting maps are consequently also highly correlated. The rank-based approach however appears to be much more discriminating, giving a range of Spearman correlation coefficients (rs) for the NGSA dataset of −0.8 to +0.9, whereas the correlations of all samples against the same reference sample based on clr-transformed data only varies from +0.73 to +0.98. In other words, catchments with a lithology that is opposite (e.g. mafic) to the reference sample (e.g. felsic) have a clear (and expected) negative rs in the DOGS2 method, yet are still shown to have a high positive correlation coefficient with the clr method (0.73, significant at p << 0.01 for this dataset). The resulting maps show very much the same regions being highlighted as similar to the reference, but the contrast is much better with DOGS2.

Alternative and perhaps more conventional methods of measuring ‘closeness’ of geochemical data are through the Euclidean or Aitchison distances (with distance decreasing as similarity increases). Preliminary comparison with the DOGS2 method suggests that for most reference samples tested the results are generally compatible and yield similar spatial patterns on maps; in a few cases, however, results appear poorly correlated between DOGS2 and the inverse of distance. Whilst it is not known which method is necessarily correct and further investigation is required to understand implications, the DOGS2 method provides a simple, robust and tested (herein and in Mann et al. 2016) method of quantifying geochemical similarity in large datasets.

Conclusions

Aqua regia (AR) data from the National Geochemical Survey of Australia (NGSA) database have been used to assess the degree of geochemical similarity (DOGS2) of catchments based on catchment outlet soil samples and Spearman (rank-based) correlation coefficients (rs); reference samples from catchments with a dominant geochemistry have been used to produce correlation-based maps that show the distribution of the DOGS2 of all catchments in the NGSA AR dataset relative to each reference sample. Reference catchment sites dominant in acid felsic, mafic/ultramafic and carbonate (both marine and calcrete) lithologies on the one hand, and banded iron formation, copper and gold mineralization on the other hand, have been chosen to demonstrate the method and to provide diagnostic geochemical information. It is interesting that in each case some of the samples identified to have highest similarities to the reference are located other than proximal to the reference, i.e. Australia-wide, in other geological provinces.

Based on the NGSA AR sample set, proposed criteria for a meaningfully high DOGS2 between pairs of samples are: rs > 0.4 (significant at p < 0.002), linear regression slopes >0.7 and intercepts <200. Sample pairs that show no particular relationship to one another may have rsc. 0, linear regression slopes c. 0 and variable intercepts. Strongly antithetic relationships, such as felsic- v. mafic-derived soils, will have negative rs, negative linear regression slope values and variable intercepts.

Utilization of all significant (not overly censored) elements in the analysed multi-element suite ensures that the method is objective by taking the full compositional fingerprint into account. This comparative multi-element method provides information for geochemical interpretation complementary to that provided by single element interrogation of databases. Preliminary applications are suggested to be: differentiation of soils derived from various source lithologies, diagnostic distinction of a marine v. terrestrial origin of carbonate in soil, and a first order mineral prospectivity tool. The approach presented here is an improvement on a previous development of DOGS, which ensures the analysis is appropriate for compositional data. The new method of determining the degree of geochemical similarity, DOGS2, could find wide application in the fields of sediment provenance, archaeology, agriculture, environmental investigations, geological mapping, digital soil mapping, resource evaluation, geochemical exploration, and forensic geochemistry.

Acknowledgements

NGSA was led and managed by Geoscience Australia and carried out in collaboration with the geological surveys of every State and the Northern Territory under National Geoscience Agreements. The authors acknowledge and thank all landowners for granting access to the sampling sites and all those who took part in sample collection. The sample preparation and analysis team at Geoscience Australia and analytical staff at Actlab's Perth laboratories are all thanked for their contributions. Constructive reviews of earlier versions of this manuscript were provided by Andrew McPherson and David Champion (Geoscience Australia), for which we are thankful. We are grateful also to Graeme Bonham-Carter for taking the time to scrutinize the DOGS2 method and provide feedback on how it compares to the measure of ‘distance’ between compositional vectors. Paul Morris and an anonymous journal reviewer are thanked for their helpful advice and comments on the original manuscript. PdC publishes with permission from the Chief Executive Officer, Geoscience Australia.

Scientific editing by Gwendy Hall

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/)