Predicting Irrigated Rice Soil–Water Conditions Using Multispectral Remote Sensing and Machine Learning in Semi-Arid Australia

Detecting soil–water conditions ranging from dry to fully ponded in rice fields using solely multispectral remote sensing is crucial for irrigation water management practices focused on water savings in Semi-Arid Australia. To this end, this research employed the Minimum Redundancy Maximum Relevance (mRMR) algorithm to identify a set of multispectral remote sensing indices for use with Machine Learning (ML) to predict three soil–water conditions in irrigated rice: “Flooded”, “Saturated”, and “Dry”. Two models were developed: Model 1, using the most frequently used remote sensing indices in the literature; Model 2, including multispectral variables selected by the mRMR algorithm. Model 2 achieved the highest performance, with an accuracy of 0.64 and a kappa of 0.46, and ROC-AUC values of 0.87, 0.62, and 0.83 for “Flooded”, “Saturated”, and “Dry”, respectively. All models exhibit high confusion rates between “Flooded” and “Saturated” conditions, suggesting that multispectral remote sensing doesn’t provide sufficient information to distinguish these soil–water conditions. The SHAP analysis revealed that vegetation-sensitive indices encoding information on crop biomass, plant moisture, and senescence status were the primary drivers of soil–water condition prediction.