This study used two soil datasets, referred to as the calibration and validation datasets. The calibration dataset was used to build the spatial prediction model, and the validation dataset was used to test the prediction quality of the calibrated model.

The calibration dataset data were generated as part of the NGSA project (https://www.ga.gov.au/about/projects/resources/national-geochemical-survey, last access: 5 May 2023), a collaborative project between Geoscience Australia and the Australian states and Northern Territory between 2007–2011, which aimed to document the soil geochemical concentration levels and patterns across Australia. Details on the project, analysis, sampling methods, and the measurement of other parameters can be found in de Caritat and Cooper (2011b, 2015) and de Caritat (2022).

The NGSA collected samples at 1315 sites (including field duplicates) at or near the outlet of large catchments with a total area coverage of 6.17×106 km2 and an average sampling density of one site for every 5200 km2 (de Caritat and Cooper, 2011b). The target sampling medium was floodplain sediments away from river channels, though in various places in Australia, aeolian modification of floodplain sediments can be important; thus, the medium was called “catchment outlet sediment” rather than floodplain sediment. These geomorphological entities are typically vegetated and biologically active (plants, worms, ants, etc.), thereby making the collected materials true soils (e.g. SSSA, 2022), albeit soils all developed on transported alluvium parent material. Due to limitations to access, samples from some parts of South Australia and Western Australia could not be obtained.

Samples were collected from two depths, namely “top outlet sediment” (TOS) from 0–10 cm depth, and “bottom outlet sediment” (BOS) from, on average, ∼60–80 cm depth. All of the samples were air-dried, homogenised, and dry sieved to <2 mm and <75 µm prior to various analyses for 60 plus elements (see de Caritat et al., 2009, 2010, for a full description of the NGSA sample preparation and analytical methods, respectively).

In this contribution, we use Li concentrations after aqua regia digestion, as the NGSA did not report total Li. A 0.50 ± 0.02 g aliquot of sample (<2 mm) was digested in aqua regia (1.8 mL of HCl + 0.6 mL of HNO3) at 90 ± 3 ∘C for 2 h to leach acid-soluble components. Once the sample had cooled to room temperature, 17.5 mL of diluent was added, and the sample was inverted 10 times to homogenise the content. The sample was further diluted 50 times prior to analysis, using inductively coupled plasma mass spectrometry (ICP-MS) in a commercial laboratory (de Caritat et al., 2010). For the remainder of the paper, any reference to Li concentrations is understood to mean aqua-regia-extractable Li unless otherwise noted. Any Li measurements that fell below the detection limit (0.1 mgkg-1) were replaced with half the detection limit (0.05 mgkg-1). A detailed quality assessment of the NGSA data is given in de Caritat and Cooper (2011a), where a relative analytical precision (repeat analysis of TILL-1 Certified Reference Materials (CRM)) of 12 % and a relative overall precision (based on field duplicates) of 39 % were reported. The distribution of sampling sites and Li concentration levels for both TOS and BOS are shown in Fig. 1.

As an independent validation dataset, we used the geochemical dataset from the Northern Australia Geochemical Survey (NAGS) project (Main et al., 2019). This dataset contains 773 observations located in the Tennant Creek–Mount Isa region in the Northern Territory and Queensland, with an approximate sampling density of one sample every 500 km2 and collection in 2017. The distribution of these samples is also shown in Fig. 1. These samples were collected, prepared, and analysed following the NGSA protocols (de Caritat and Cooper, 2011b), albeit at a higher sampling density. However, only TOS samples were collected in NAGS. Furthermore, these NAGS samples were collected at a different time and analysed in a different laboratory compared to the NGSA dataset. To address the analytical variation that could potentially arise, a levelling method was applied using the TILL-1 CRM standards (Main and Champion, 2022). First, the subset of the NGSA dataset that covers the spatial area of the NAGS dataset was extracted. Then a Kolmogorov–Smirnov test was used to verify if the samples from the two datasets (subset of the NGSA and NAGS) were similar. A correction factor to relate the two datasets based on the TILL-1 CRM standards was then calculated and applied as a multiplier to the NAGS dataset to level its data to the NGSA dataset.

A total of 19 environmental covariates (Table 1) characterising the factors of climate, parent material, soil, and topography, which contribute to soil formation, were considered in this study.

The first factor is climate. Water (humidity) and temperature affect the rate of mineral weathering and thus soil formation. Hence, we included precipitation, evaporation, and temperature data (Harwood, 2019), along with the topographic wetness index (TWI) data (Gallant and Austin, 2012b), informing about the relative wetness within a landscape. In short, the TWI was derived from the partial contributing area product, which was computed from a hydrologically enforced digital elevation model, and from the percent slope product, which was computed from the smoothed digital elevation model (DEM-S; Gallant and Austin, 2012b).

The second factor is parent material (i.e. degree of weathering and mineralogical composition), including gamma-ray radiometric and total magnetic intensity. Gamma-ray radiometric surveys provide estimates for the concentrations of gamma-ray-emitting radio-elements like potassium (K), uranium (U), and thorium (Th) at/near the soil surface. The gamma-ray radiometric data were measured from airborne surveys throughout most of Australia (Poudjom Djomani et al., 2019). In this study, we used a complete gamma-ray survey grid where gaps in the airborne coverage were filled in using covariate machine learning (Wilford and Kroll, 2020). Gamma-ray radiometric data have been found to be a useful covariate in identifying surface processes such as sediment transport and weathering (Wilford, 2012; Wilford et al., 1997) and detecting radioactive mineral deposits and occurrences (Alhumimidi et al., 2021; Dickson et al., 1996; Dickson and Scott, 1997; Wilford et al., 2009). Total magnetic intensity (TMI), which measures variations in the Earth's magnetic field intensity caused by the contrasting content of various rock-forming minerals in the crust (Poudjom Djomani et al., 2019), could also potentially identify geological features and processes.

The third factor is the soil itself, particularly the relevant physical soil properties. As previous studies, e.g. by Kabata-Pendias (1995) and Robinson et al. (2018), highlighted the high correlation between Li and clay content of soil, soil texture was used as a covariate. The soil texture spatial information (sand and clay contents) was derived from Malone and Searle (2021), which contained updated information on soil texture across Australia derived using a digital soil mapping approach. The sand and clay fractions were developed by integrating field morphological (n=180498) and laboratory measurements of soil texture fractions (n=17367) from the Soil and Landscape Grid of Australia (SLGA). The SLGA is based on a comprehensive compilation of soil attributes across Australia, including the NGSA dataset. These sand and clay content maps (Malone and Searle, 2021) were for specific depth intervals (0–5, 5–15, 15–30, 30–60, 60–100, and 100–200 cm). They were converted to the depths corresponding to the NGSA Li measurement (0–10 and ∼60–80 cm) using the mass-preserving spline function, described in Bishop et al. (1999) and modified by Malone et al. (2009). Soil reflectance in the visible, near-infrared (NIR), and shortwave–infrared (SWIR) spectra captured by remote sensing images provides information on soil composition. However, the unprocessed images consist of a mixture of soil, bedrock, vegetation, and clouds. By removing the influence of seasonal vegetation, Roberts et al. (2019) were able to document the “barest” state of soil, so critical in mapping the physical characteristics of soil and rock. This was done by combining Landsat 5, 7, and 8 observations of the past 30 years to remove the contamination by seasonal vegetation, cloud cover, shadows, detector saturation and pixel saturation. The model used to develop the Barest Earth product was validated using the NGSA spectral archive (Lau et al., 2016).

Finally, topography is represented by elevation and slope. These factors also play an important role, as they affect how water is added to and/or lost from soil. The elevation was derived from the DEM-S which was obtained from the 1 arcsec resolution Shuttle Radar Topography Mission (SRTM) data acquired by NASA in February 2000 (Gallant et al., 2011). The slope covariate was also calculated from DEM-S using the finite difference method (Wilson and Gallant, 2000). The different spacing in the E–W and N–S directions due to the geographic projection of the data was accounted for by using the actual spacing in metres of the grid points calculated from the latitude.

All covariates were reprojected to EPSG: 3577 (GDA94 datum; Australian Albers equal area projection) and resampled to a common spatial resolution of 3 km prior to any analysis. All the environmental covariates used are shown in Table 1.

The correlation matrix of the Li concentrations to all the other element concentrations and environmental covariates was generated using Pearson's correlation method. Strong correlation was defined as >0.5, moderate was defined as 0.35 to 0.5, and weak correlation was defined as <0.35. Note that this classification was generated to facilitate interpretation of this dataset only and is not implied to be a general rule.

Here, we used the machine learning model “cubist” to relate soil observations to the environmental covariates. Cubist is a tree-based regression algorithm based on the M5 theory (Quinlan, 1993). This algorithm creates partitions of data with similar spectral characteristics and creates one or more rules for each partition. If the partition rules are satisfied, then the linear regression of that partition is used to create the prediction (Eq. 1). Each rule can be defined as follows: 1If [condition is true], then [regression], else [apply next rule]. The cubist model has two tuning parameters: committees (number of sequential models included in the ensemble) and neighbours (number of training instances that are used to adjust the model-based prediction). A comprehensive combination of committees (5, 10, 20, 30, 40, 50) and neighbours (0, 1, 5, 9) was tested to tune the cubist model. To obtain the best estimates of optimum parameters, a 10-fold cross-validation approach was utilised. Based on the optimum parameters, 50 bootstrap models (“sampling with replacement”) were trained. The flowchart of the process is shown in Fig. 2.

The performance of the prediction models was then evaluated using both an internal evaluation and the external, independent validation dataset. An internal evaluation of the model was conducted using “out of bag” samples, which were not used during the development of the bootstrap models. The NAGS dataset was used to evaluate the performance on the independent dataset (top depth only). The following metrics, briefly explained below, were used: adjusted coefficient of determination (Radj2), Lin's concordance correlation coefficient (LCCC) (Lin, 1989), root mean square error (RMSE), bias, and ratio of performance to interquartile distance (RPIQ). Radj2 is a measure of the linear association between observed and predicted values; LCCC measures the agreement between the observed and predicted values in relation to the 1:1 line while accounting for the magnitude of the differences; RMSE is a measure of the differences between the observed and predicted values; bias is the measure of the difference between the mean of the observed and the mean of the predicted values; and RPIQ is a measure of performance that takes into account the distribution of the values and can be calculated as a fraction of the interquartile range of the observed values (Q3-Q1) and the RMSE (RPIQ=(Q3-Q1)/RMSE) (Bellon-Maurel et al., 2010).

Variable importance analysis was also conducted to evaluate the contributions of each covariate to the Li prediction. The relative variable importance is measured as the percentage of times the environmental covariate is used either as conditions for a rule or as predictors (usages) within the linear regression model when certain conditions are met. These bootstrap models were then used to generate output maps with the same extent and resolution. The final map output was derived based on the mean prediction of the bootstrap models; similarly, the standard deviation map was obtained based on the standard deviation of the prediction from the bootstrap models.

All the data analytics, modelling, and mapping procedures in this study were conducted in the R statistical open-source software (R Core Team, 2021). Besides the base R functionality, the R packages used in this study included “cubist” (Kuhn and Quinlan, 2021) for fitting cubist models, “caret” (Kuhn, 2022) for tuning the hyperparameter of the cubist model, and “raster” (Hijmans, 2021) for handling raster layers and generating soil map predictions. All soil maps were produced in ArcMap version 10.8 (ESRI, 2019) using the Albers equal area projection (EPSG:3577).

The distribution of 1315 aqua-regia-soluble Li concentration values (NGSA dataset; de Caritat and Cooper, 2011b) was positively skewed (Fig. 3) with concentrations ranging from 0.05–67.4 and 0.05–56 mgkg-1 for TOS and BOS, respectively. Only limited observations above 20 mgkg-1 of Li concentrations were found in this study for both TOS (n=76) and BOS (n=95). The mean concentration of TOS (7.6 mgkg-1) was slightly lower than that of BOS (8.8 mgkg-1). These concentrations were lower than those observed for the mean aqua-regia-soluble Li concentrations in European soil at 11.3 mgkg-1 (Négrel et al., 2019), as well as those found in upper continental crust (both in loess and shales) at 35 mgkg-1 total Li (Teng et al., 2004). A soil geochemical survey in the USA shows total soil Li concentrations with a range of <1–300 mgkg-1 (median 20 mgkg-1) for soils from 0–5 cm and a range of <1–280 mgkg-1 (median 24 mgkg-1) for soil samples from the C horizon (Smith et al., 2019). Similarly, total Li concentrations of up to 400 mgkg-1 have been reported in China (Liu et al., 2020). These latter Li concentrations, measured using an extraction of four acids, were considerably higher than the aqua regia extraction data from the NGSA dataset.

Based on the data collected by the NGSA project, the highest concentrations of Li for both TOS and BOS were found in northernmost Queensland (Cape York Peninsula), as shown in Fig. 1 and Table 2. Other regions that have elevated concentrations of Li were located in the Goldfields–Esperance region (Table 2) in Western Australia, which has been recognised as one of the most resource-rich areas on the planet (Champion, 2019), and the region around the Victoria–New South Wales border (Fig. 1). Some of the findings correlate well with the existing Li mine sites in Australia (red triangles in Fig. 1). The largest deposit of Li found in Australia is the Greenbushes deposit, south of Perth. Other regions include Mount Marion and Earl Grey in the Yilgarn Craton and Pilgangoora in the Pilbara Craton (Champion, 2019; see Table 1). In July 2019, Strategic Metals Australia (SMA) found a new Li exploration target near Cairns, in the Georgetown province of north Queensland (Gluyas, 2019). However, this discovery has not been updated in the data collected by Geoscience Australia because considerable work such as drilling, modelling, resource calculation, and feasibility studies are needed to bring the discovery to the feasibility stage.

The final cubist model was tuned with 20 committees and 9 neighbours, which resulted in the lowest RMSE compared to the other combinations of hyperparameters, indicating an optimised cubist model.

From the cubist model, we can infer the relative importance of the covariates by calculating the percentage of times a covariate is being used in the model. The variables used by the cubist model can be further split in terms of “importance as conditions within rule” and “frequency of usage as predictors in models”.

For Li prediction in TOS, the variables clay, PTA, TRA, and EPA are of higher importance in the conditions than other variables (Fig. 6). This implies that the model separates out prediction values based on climate covariates along with clay content. However, within the regression models, the top five variables most frequently used in the regression were the Landsat band 2, band 6, band 1, band 3, and gamma radiometric total dose. The first three Landsat bands (red, green, and blue) and band 6 (SWIR2) have been commonly used to predict soil properties, delineate geological boundaries, and differentiate between vegetation zones (Khorram et al., 2012), while the gamma radiometric dose discriminated the various soil types and their mineral make-up. The next set of covariates were annual precipitation and clay and sand contents, indicating they have lower importance as predictors. As indicated in the correlation analysis, slope was not significant.

For the BOS model, the TRA variable had the highest importance in the conditions of the model (Fig. 6) for Li predictions, separating high and low values. EPA, clay content, and PTA also affect model conditions. Overall, parameters that were more frequently used as predictors in the BOS model were similar to those for TOS, i.e. the top five are gamma radiometric dose and Landsat bands 2, 1, 6, and 3. In the BOS model, however, there was a higher importance of the clay content (sixth most used) compared to the TOS model (ninth). The usage of slope covariate as predictor is similarly low (last) for both TOS and BOS.

The cubist model led to the generation of spatial predictions of aqua-regia-soluble Li concentration in alluvium-derived soils across Australia at two depths (Fig. 7). So far, there are only five known Li mines in Australia (mostly in Western Australia), all of which are located within areas that were predicted to have a higher background concentrations of soil Li, especially for the BOS model (>∼8 mgkg-1) (Fig. 8).

In Australia, the largest producer of spodumene is the Greenbushes Li operation, located approximately 250 km south-southeast of Perth. In the most recent public report, the company reported combined measured and indicated resources of 118.4×106 t (Mt) of ore at 2.4 % Li2O containing proved and probable reserves of 61.5 Mt at 2.8 % Li2O (Champion, 2019). Other locations explored for Li include Mount Cattlin and Mount Marion in the Goldfields–Esperance region and Pilgangoora of East Pilbara. In a recent review (Champion, 2019), these projects' reports estimated Li resources ranging from 11.8 to 71.3 Mt at 1.01 % to 1.37 % Li2O. The predicted soil Li concentrations at the known Li mine sites range from 4.5 to 7.3 mgkg-1 for TOS and from 7.1 to 12.6 mgkg-1 for BOS. The highest TOS and BOS concentrations of Li proximal to a known mine site are for the Mount Marion deposit in Western Australia.

Although most Li exploration to date has been conducted in Western Australia, our map indicates that other regions in Australia are potentially anomalous in Li (Fig. 7). These areas are located for instance within the central western region of Queensland and visually correspond to areas of widespread black cracking (smectite-rich) soils or Vertosols (Isbell, 2021). An elevated concentration of Li was also observed over parts of the Eucla Basin, which has a widespread distribution of Fe-oxide-rich regolith with carbonate accumulations (Johnson, 2015; Wilford et al., 2015). The sources of carbonate include weathered Proterozoic and Palaeozoic carbonate bedrock, vast marine sediments that extend across the low-lying and offshore areas associated with Cenozoic sedimentary basins, and abundant widespread pedogenic carbonates (Johnson, 2015). This is in line with de Vos et al.'s (2006) observations, where higher Li concentrations of up to 56 mgkg-1 were identified in calcareous soil (high carbonate accumulation) in comparison to those of organic soil (1.3 mgkg-1). The Fe in the Fe-oxides and oxyhydroxides that help retain Li may be released from oxidation of primary minerals during weathering (Kabata-Pendias, 2010). The ultimate origin of Li within these clay-, iron-, and carbonate-rich soils remains to be established in the case of Australia. Other regions of potential interest occurring in different soil types are located in southern New South Wales and parts of Victoria.

We further explored the correlation of Li concentration against soil orders (Searle, 2021). Figure 9 shows the range of Li concentrations across various soil types identified at the sampling locations. The Li concentration tended to be slightly higher in Vertosols, Calcarosols, and Dermosols. These observations indicate Li accumulated in a more uniform soil profile with less differentiation between top and subsoils. In addition, clay soils (Vertosols) and soils with high CaCO3 (Calcarosols) appeared to have larger Li concentrations. These observations supported the anomalous Li predictions in various parts of Australia mentioned earlier.

The highest predicted values on the Li digital soil maps are 28 and 22 mgkg-1 in TOS and BOS, respectively. Although a higher Li concentration was expected to be observed in the deeper layer, the model used in this study was not able to support such predictions yet. This is most likely because the covariates used within the model represent observations from TOS instead of BOS. The variance of covariates within BOS was not obtained, hence yielding lower-accuracy predictions.

While we have successfully modelled soil Li distribution in Australia and validated it using an independent sample dataset, we recognise that there are limitations to this study's approach. (1) The NGSA data used apply to catchment outlet sediment representing the local accumulation of mainly detrital minerals. Therefore, strictly speaking, the predictions developed herein apply only to similar alluvial soils. (2) The NGSA data were measured using an aqua regia digestion that only extracts a portion of the total Li found in soil. The results could potentially be improved if total Li was measured. Most of the observations collected had a relatively low concentration; having more representative samples at higher concentrations might improve the prediction accuracy. (3) Despite the large amount and spread of data, the NGSA does not cover the whole of Australia. Notably, there is a data gap in parts of Western Australia and South Australia. However, no more extensive geochemical dataset than the NGSA exists in Australia. (4) The environmental covariates used in the study were selected based on our understanding of relevant soil-forming processes. (5) There is also limited information on how the covariates vary with depth except for the soil texture (sand and clay content) data. The inclusion of more environmental covariates related to depth and soil mineralogical information may improve the predictive capability of these machine learning models. Note that quantitative mineralogical data are currently being acquired on the NGSA samples, both as X-ray diffraction data on whole sediment samples and clay fractions (de Caritat and Troitzsch, 2021) and as automated mineralogy using energy dispersive spectrometry on heavy mineral fractions (de Caritat et al., 2022a, b, c).

The final product was only validated in one area within Australia (Tennant Creek–Mount Isa region in the Northern Territory and Queensland). Despite our predictions of elevated soil Li in parts of Queensland, New South Wales, and Victoria, ground-truthing is required to confirm them, and further work is necessary to determine the origin of the contained Li.