Abstract
Increasing water scarcity and climate variability have intensified the need for precise agricultural irrigation management. Accurate estimation of crop coefficients (Kc) is critical for determining crop water requirements, especially in arid and semi-arid regions. However, conventional methods for estimating Kc often rely on generalized plant characteristics, which may not account for local climatic variations. In this study, we address this challenge by predicting the daily crop coefficient for soybean using four machine learning models: Extreme Gradient Boosting (XGBoost), Extra Tree (ET), Random Forest (RF), and CatBoost. These models were trained on meteorological data from Suhaj Governorate, Egypt, spanning 1979-2014. Additionally, SHapley Additive exPlanations (SHAP), Sobol sensitivity analysis, and Local Interpretable Model-agnostic Explanations (LIME) were applied to evaluate model interpretability and consistency with physical processes. Among the models evaluated, the ET model achieved the highest accuracy, with r = 0.96, NSE = 0.93, RMSE = 0.05, and MAE = 0.02. XGBoost and RF also performed well, each obtaining r = 0.96, NSE = 0.92, RMSE = 0.06, and MAE = 0.02. In comparison, CatBoost demonstrated slightly lower accuracy, with r = 0.95, NSE = 0.91, RMSE = 0.06, and MAE = 0.02. SHAP and Sobol analyses consistently identified the antecedent crop coefficient [[Formula: see text]] and solar radiation (Sin) as the most influential variables. LIME results revealed localized variations in predictions, reflecting dynamic crop-climate interactions. This study underscores the importance of integrating interpretable machine learning models to enhance both predictive accuracy and reliability while maintaining alignment with critical physical processes. The proposed framework offers a robust tool for improving daily Kc estimation, thereby supporting more sustainable irrigation practices and climate-resilient agriculture.