Enhancing Soil Water Prediction in Arid Climates Using Multipredictor Machine-Learning Models and SHAP-Based Interpretability

Efficient water management in irrigated watersheds requires the timely and precise application of water to optimize crop yields and sustain resources. Although machine-learning (ML) models have shown promise in capturing hydrological dynamics, their forecasting performance is often limited by the quality and comprehensiveness of the input predictors. This study presents a computational ML framework for predicting soil moisture variability in hyperarid environments using historical data (2007–2014) from three sites in Kuwait. Soil moisture records were processed using standard statistical methods to remove long-term trends. The study tested three machine-learning models: support vector machines (SVMs), Gaussian process regression (GPR), and bagged trees ensemble (BTE). To enhance the transparency and practical applicability of the results, SHapley Additive Explanations (SHAP) were employed as an interpretability tool to highlight which environmental factors most influence soil moisture. This approach not only enabled the achievement of high predictive accuracy but also provided clear insights for sustainable water management. Among the models, BTE achieved overall best performance, with R2 up to 0.987, root-mean-square error (RMSE) as low as 0.008 g/cm3, and mean absolute error (MAE) as low as 0.002 g/cm3 across stations and depths. This study effectively integrated data-driven models into smart irrigation systems to provide recommendations for enhancing water resources security in water-scarce regions. Furthermore, it provides preliminary knowledge for policymakers in other arid regions to develop sustainable water conservation measures and strengthen existing capacities to respond to water shortage scenarios. The proposed framework presents a practical and effective tool for regulating irrigation systems and guiding evidence-based water management strategies in hyperarid environments.