These are the soil attribute products of the Tasmanian Soil Attribute Grids. There are 8 soil attribute products available from the TERN Soil Facility. Each soil attribute product is a collection of 6 depth slices. Each depth raster has an upper and lower uncertainty limit raster associated with it. The depths provided are 0-5cm, 5-15cm, 15-30cm, 30-60cm, 60-100cm & 100-200cm, consistent with the Specifications of the GlobalSoilMap.

Attributes: pH - Water (pHw); Electical Conductivity dS/m (ECD); Clay % (CLY); Sand % (SND); Silt % (SLT); Bulk Density - Whole Earth Mg/m3 (BDw); Organic Carbon % (SOC); Coarse Fragments >2mm (CFG).

These products were developed using datasets held by the Tasmanian Department of Primary Industries Parks Water & Environment (DPIPWE) Soils Database. The mapping was made by using spatial modelling and digital soil mapping (DSM) techniques to produce a fine resolution 3 arc-second grid of soil attribute values and their uncertainties, across all of Tasmania.

Note: Previous versions of this collection contained a Depth layer. This has been removed as the units do not comply with Global Soil Map specifications. Lineage: The soil attribute maps are generated using spatial modelling and digital soil mapping techniques.

Soil inventory:

Tasmanian soil site data originates from the DPIPWE soils database, a compilation of various historical soil surveys undertaken by DPIPWE, CSIRO, Forestry Tasmania and the University of Tasmania. This database contains morphological and laboratory data for all the soil sites.

Data Modelling :

A raster stack of all covariates was generated and the target variable (each soil property and depth) individually intersected with the covariate values to provide the calibration and validation data. All modelling was undertaken in ‘R’ (R Development Core Team 2012), using Regression tree (RT), specifically the Cubist R package (Kuhn, Weston et al. 2012; Kuhn, Weston et al. 2013; Quinlan 2005). The RT approach is a popular modelling approach for many disciplines (Breiman, Friedman et al. 1984), and has been widely used with DSM (Grunwald 2009; Kidd, Malone et al. 2014; McKenzie and Ryan 1999). Cubist develops the regression trees by first applying a data mining-approach to partition the calibration and explanatory covariate values into a set of structured ‘classifier’ data. The tree structure is developed by repeatedly partitioning the data into linear models until no significant measure of difference in the calibration data is determined (McBratney, Mendonça Santos et al. 2003). A series of covariate-based rules (conditions) is developed, and the linear model corresponding to the covariate conditions is applied to produce the final modelled surface. For this modelling exercise, the number of rules was set within the model controls to let the Cubist algorithm decide upon the optimum number of rules to generate.

Uncertainty Leave-one-out-cross-validation (LOOCV) was applied to the Cubist model to generate rule-based uncertainties, using only those covariates forming the conditional partitioning of that rule, following Malone et al (2014). The LOOCV, applied to an individual Cubist model for each rule, effectively produced a mean value for each RT partition, with the upper and lower 5 and 95% quantiles of the prediction variation providing the lower and upper prediction uncertainty values respectively, at the 90% Prediction Interval (PI). A 10-fold cross validation was used to run this process 10 times across all data to produce mean modelling diagnostics and validations, and reduce modelling bias due to sensitivity to training data variance.