Depth to sodicity constraint mapping of the Murray-Darling Basin, Australia

Abstract

The Murray-Darling Basin is a highly significant agricultural region in Australia, contributing AUD $24 billion to the Australian economy each year through food and fibre production. While the region provides favourable conditions for agriculture, it faces a multitude of challenges, including climate change, drought, and soil degradation. The occurrence of excess sodium ions in the soil, referred to as ‘sodicity’, is a form of soil degradation which is also a significant constraint to production. Maps of sodicity exist in the Murray-Darling Basin, but only cover portions of the region and most are only measures of surface Exchangeable Sodium Percentage (ESP), rather than at depth. Therefore, the current study sought to produce 3D maps of ESP at 10 cm depth increments to 100 cm, and maps of depth to ESP constraints across the Murray-Darling Basin. Random Forests served as the model algorithm, and our results provide known measures of accuracy and precision for a transparent and reliably assessed product. We aimed to produce a more parsimonious model, avoiding the inclusion of a superfluous number of covariates, through a simplified recursive feature elimination procedure, which resulted in a reduction of 20 to 9 covariates for modelling. Following model tuning, independent testing showed the 3D model was of a moderate quality (LCCC = 0.66, r2 = 0.49, RMSE = 5.99 %, Bias = 0.14 %). Spatial cross-validation demonstrated that the model did not predict so well at under-represented locations (LCCC = 0.45, r2 = 0.29, RMSE = 7.19 %, Bias = 0.77), producing a substantially lower fit compared to the independent test fit. Area of Applicability analysis supported this notion, identifying that the model was only applicable to small pockets across the region and suggesting we were extrapolating too far beyond the feature space of the training points. Although there is room for improvement, providing transparency via measures of accuracy, precision, and applicability, provides end users with a starting point for further analyses and exploration.