Machine learning approaches for spatial modeling of agricultural droughts in the south-east region of Queensland Australia.

A quantitative understanding of the hydro-environmental factors that influence the occurrence of agricultural drought events would enable more strategic climate change adaptation and drought management plans. Practical drought hazard mapping remains challenging due to possible exclusion of the most pertinent drought drivers, and to the use of inadequate predictive models that cannot describe drought adequately. This research aims to develop new approaches to map agricultural drought hazard with state-of-the-art machine learning models, including classification and regression trees (CART), boosted regression trees (BRT), random forests (RF), multivariate adaptive regression splines (MARS), flexible discriminant analysis (FDA) and support vector machines (SVM). Hydro-environmental datasets were used to calculate the relative departure of soil moisture (RDSM) for eight severe droughts for drought-prone southeast Queensland, Australia, over the period 1994-2013. RDSM was then used to generate an agricultural drought inventory map. Eight hydro-environmental factors were used as potential predictors of drought. The goodness-of-fit and predictive performance of all models were evaluated using different threshold-dependent and threshold-independent methods, including the true skill statistic (TSS), Efficiency (E), F-score, and the area under the receiver operating characteristic curve (AUC-ROC). The RF model (AUC-ROC = 97.7%, TSS = 0.873, E = 0.929, F-score = 0.898) yielded the highest accuracy, while the FDA model (with AUC-ROC = 73.9%, TSS = 0.424, E = 0.719, F-score = 0.512) showed the worst performance. The plant available water holding capacity (PAWC), mean annual precipitation, and clay content were the most important variables to be used for predicting the agricultural drought. About 21.2% of the area is in high or very high drought risk classes, and therefore, warrant drought and environmental protection policies. Importantly, the models do not require data on the precipitation anomaly for any given drought year; the spatial patterns in AGH were consistent for all drought events, despite very different spatial patterns in precipitation anomaly among events. Such machine-learning approaches are able to construct an overall risk map, thus assisting in the adoption of a robust drought contingency planning measure not only for this area, but also, in other regions where drought presents a pressing challenge, including its influence on key practical dimensions of social, environmental and economic sustainability.