Reducing uncertainty of evapotranspiration products across China under extreme climatic conditions using flux-tower observations and machine learning

Accurate estimation of terrestrial evapotranspiration (ET) is critical for understanding land–atmosphere interactions and supporting sustainable water resource management under climate change. However, large uncertainties persist in China due to complex surface heterogeneity, sparse ground observations, and frequent extreme events. This study presents the first comprehensive evaluation of evapotranspiration products over China using observations from 64 eddy covariance flux towers, assessing multiple daily and monthly datasets across diverse climate zones, land surface types, and five representative extreme climatic conditions, including high temperature (Temp), high vapor pressure deficit (VPD), high precipitation (Pre), high wind speed (WS), and drought. An explainable machine learning framework based on XGBoost was further developed to reduce ET uncertainties and identify dominant controlling factors, with independent validation conducted using additional flux sites. The results show that: 1) Under overall conditions, daily ET products exhibit larger uncertainties than monthly products, with correlation coefficients (r) generally ranging from 0.37 to 0.59 for daily estimates and from 0.61 to 0.85 for monthly estimates. Under extreme climatic conditions, daily ET accuracy declines sharply, with the mean correlation decreasing from 0.634 to 0.332, showing the strongest degradation under high VPD and relatively weaker sensitivity under high WS. Monthly ET products are less affected by extremes but still show notable performance deterioration during drought conditions. 2) In typical dense flux observations regions, none of the ET products capture spatial heterogeneity effectively, and estimation accuracy declines noticeably. 3) The XGBoost model significantly enhances ET estimation at both daily (r = 0.908) and monthly (r = 0.931) scales, particularly under extreme climatic conditions. ET products, solar radiation (Rs), and VPD are the primary contributors associated with higher estimated ET in the model. Independent validation across 12 sites confirmed robustness of XGBoost, demonstrating strong performance even under extreme climatic conditions. Overall, this study recommends integrating flux-tower-informed machine-learning fusion with existing ET products as an effective pathway to reduce ET uncertainty and enhance the reliability of regional water-cycle assessments and climate-impact analyses in complex climatic regions.