Unraveling soil moisture dynamics with dual‐scale interpretable machine learning: Cover cropping and irrigation insights in semi‐arid agriculture

Abstract Soil moisture prediction remains a major challenge in arid and semi‐arid agroecosystems, where highly variable rainfall, intermittent irrigation, and heterogeneous soil properties create strong nonlinearities in soil‐water dynamics. Reliable forecasts and interpretable models are needed for optimizing irrigation scheduling and improving water‐use efficiency under such conditions. The purpose was development of a novel, dual‐resolution framework that integrates machine learning (ML) and deep learning (DL) approaches with explainable artificial intelligence to evaluate soil moisture dynamics under different cover crop treatments. A 2‐year field experiment was conducted in Clovis, NM, where soil moisture, temperature, and meteorological variables were continuously monitored across five treatments: fallow, pea, oat, pea–oat mixture, and a six‐species mixture. ML models (random forest, light gradient boosting machine LGBM, and eXtreme gradient boosting) and DL models, such as long short‐term memory and Transformer models, were trained on both daily and 5‐min datasets. SHapley Additive exPlanations (SHAP) indicates that features from the three most recent days had the strongest influence on daily predictions, informing a 72‐h lookback for short‐term modeling. LGBM achieved the highest accuracy, while Transformer reduced temporal lag in high‐frequency predictions. SHAP analysis further identified treatment‐specific sensitivities and scale‐dependent relationships between environmental variables and soil moisture content, with historical soil moisture and total water input dominating daily predictions, and solar radiation, temperature, and recent water events prevailing at finer scales. This framework improves both predictive accuracy and interpretability of soil water dynamics, supporting data‐driven irrigation and crop management in arid and semi‐arid regions.