Putting regional digital soil mapping into practice in Tropical Northern Australia

- Digital soil mapping for 155,000 km2 of Northern Australia is described.
- Rule-based data mining and environmental correlation are combined.
- Prediction of three target variable data types is described: numeric, binary, and categorical.
- Three data assessment methods are described.
- Digital soil mapping worked under challenging data and operational circumstances.

Tropical Northern Australia is a vast region dominated by extensive cattle grazing. Australia is seeking opportunities to intensify land use in this region through irrigation. We describe a digital soil mapping (DSM) approach for the Flinders and Gilbert catchments (combined area, 155,000 km2) in north Queensland to supply soil property (i.e. target variable) maps for an eventual crop-specific irrigation suitability assessment.We applied a statistically-based survey design to identify new soil sampling sites. This data was merged with legacy soil data to produce a combined dataset of 1951 soil sites used in our DSM approach to map 16 soil target variables. Our mapping relied on the RuleFit3 analytical toolset and environmental correlation employing 13 predictors from terrain analysis, mapping and remote sensing. We present prediction and evaluation for three categories of target variables, namely: numeric, binary and categorical using the examples of surface pH (H2O) (numeric), rocky/non-rocky (binary) and permeability (categorical).Prediction quality was evaluated using internal cross-validation, independent validation, and non-parametric bootstrapping. Under internal cross-validation the models achieved R-squared of 0.49 for surface pH (H2O) (numerical), 76% classification accuracy for rocky/non-rocky (binary) and 56% classification accuracy for soil permeability (categorical). Under independent validation the target variables achieved a R-squared of 0.67 for surface pH (H2O), and accuracy of 93% for rocky/non-rocky accuracy of 63% for soil permeability. Non-parametric bootstrapping conducted on the surface pH (H2O) estimation showed the most reliable predictions to be in the pH range of 6-7.Despite the practical pressures imposed on the project (i.e. short project duration, and study area remoteness, size and difficult access), our DSM approach delivered the soil data within required specifications for the crop-specific irrigation suitability assessment. The approach gives a working framework of other soil mapping exercises sharing similar practical (i.e. environmental, logistical, data) constraints.