An Interpretable Machine Learning Approach to Remote Sensing-Based Estimation of Hourly Agricultural Evapotranspiration in Drylands

Obtaining evapotranspiration (ET) estimates at high spatiotemporal resolution is a fundamental prerequisite for clarifying the patterns and controlling factors of agricultural water consumption in drylands. However, most existing ET products are provided at daily or coarser spatial–temporal scales, which limits the ability to capture short-term variations in crop water use. This study developed a novel hourly 10-m ET estimation method that combines remote sensing with machine learning techniques. The approach was evaluated using agricultural sites in two arid regions: the Heihe River Basin in China and the Bekaa Valley in Lebanon. By integrating hourly eddy covariance measurements, Sentinel-2 reflectance data, and ERA5-Land reanalysis meteorological variables, we constructed an XGBoost-based modeling framework for hourly ET estimation, and incorporated the SHapley Additive exPlanations (SHAP) method for model interpretability analysis. Results demonstrated that the model achieved strong performance across all sites (R2 = 0.86–0.91, RMSE = 0.04–0.05 mm·h−1). Additional metrics, including the Nash–Sutcliffe efficiency coefficient (NSE) and percent bias (PBIAS), further confirmed the model’s robustness. Interpreting the model with SHAP highlighted net radiation (Rn), 2-m temperature (t2m), and near-infrared reflectance of vegetation (NIRv) as the dominant factors controlling hourly ET variations. Significant interaction effects, such as Rn × NIRv and Rn × t2m, were also identified, revealing the modulation mechanism of energy, vegetation status and temperature coupling on hourly ET. The study offers a practical workflow and an interpretable framework for generating high-resolution ET maps, thereby supporting regional water accounting and land–atmosphere interaction research.