Optimization of Sensor Combinations for Simplified Estimation of Reference Crop Evapotranspiration Using Machine Learning and SHAP Interpretation

Accurately estimating reference crop evapotranspiration (ET0) is essential for agricultural water-resource management, yet the traditional Penman–Monteith (PM) method requires multiple meteorological variables and is difficult to apply in data-sparse regions. To explore more data-efficient alternatives, this study systematically evaluates several machine-learning (ML) models capable of capturing nonlinear relationships, using daily observations from 698 meteorological stations across China. In addition, we incorporate SHapley Additive exPlanation (SHAP), a game-theory-based interpretability approach, to quantify the contribution of input variables at both national and regional scales. The results show that the Random Forest (RF) model performs best (coefficient of determination, R2 = 0.957; mean absolute percentage error, MAPE = 9.214%), significantly outperforming multiple linear regression and approaching the accuracy of the PM method. SHAP analysis indicates that maximum temperature, sunshine duration, and month are the most influential factors nationwide. Geographic variables contribute less overall but become important in specific regions, such as Southwest China. The study also reveals pronounced spatial heterogeneity in the drivers of ET0, highlighting the necessity of regionalized interpretations. Furthermore, sensor-reduction experiments demonstrate that reasonable estimation accuracy can be maintained even without radiation or wind-speed observations, offering guidance for low-cost monitoring scenarios. Overall, this study provides transparent model comparisons for ML-based ET0 estimation, uncovers regional differences in controlling factors, and offers practical insights for designing meteorological monitoring strategies in data-limited environments.