Developing Interpretable Pan Evaporation Forecasting Models for Wafra Agricultural Basin Based on Optimized Decision Tree Ensembles

The forecasting of pan evaporation has improved with the advancement of machine learning (ML) models. However, many existing approaches suffer from limited interpretability and uncertain transferability, limiting their practical value in water resources planning. This study developed a transparent and generalizable ML framework incorporating three decision tree-based models to estimate daily pan evaporation under hyperarid climatic conditions: the decision tree (DT), bagged tree (BT), and boosted regression tree (BRT). The DT model was optimized using a grid search algorithm to assess the value of hyperparameter tuning, whereas BT and BRT were implemented as unoptimized ensemble benchmarks. All models were trained using daily meteorological data from the Wafra station in Kuwait. The optimized DT model achieved the lowest mean absolute error (MAE) of 0.3975 mm/day compared with 0.4125 mm/day and 0.5734 mm/day for BT and BRT, respectively. Post hoc interpretability was performed using Shapley Additive Explanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and partial dependence plots (PDPs), revealing shortwave solar radiation and relative humidity as dominant drivers, with critical thresholds of >2,400 J/cm2 radiation, >6 m/s wind speed, and <25% humidity. The models were tested at two independent stations to assess spatial robustness: Kuwait International Airport (KIA) and Abdaly. At these stations, they maintained high predictive performance with only marginal degradation. The results demonstrate that the proposed approach offers a practical balance among predictive performance, interpretability, and spatial generalization, making it a viable tool for evaporation modeling in data-scarce, hyperarid environments.