Irrigation water use estimation based on knowledge-guided interpretable machine learning

To accurately estimate field-scale irrigation water use, this study developed a knowledge-guided machine learning model (KGML-irri) that integrates the Budyko water-energy balance framework with the XGBoost algorithm. The KGML-irri model combines the Budyko-Fu parameter with meteorological, soil, vegetation, and management factors to estimate field-scale irrigation water use, while SHapley Additive exPlanations (SHAP) were used to quantify the contributions of individual input features. The results showed that KGML-irri substantially outperformed purely data-driven machine learning models on the test set (R² = 0.83 vs. 0.49—0.55). The Budyko-Fu parameter was accurately predicted, and irrigation water use estimates maintained consistently high accuracy across the study years. The Budyko-Fu parameter contributed most strongly to irrigation water use estimation, with evapotranspiration and soil moisture emerging the primary factors controlling its variability. These findings demonstrate that embedding the Budyko water-energy balance framework within machine learning can substantially improve the accuracy of field-scale irrigation water use estimates and provide a robust methodological basis for improved agricultural water resources management at regional scales.