Diagnosing and correcting openet satellite evapotranspiration across us climate zones: Error decomposition, explainable AI, and uncertainty quantification

Satellite evapotranspiration (ET) from the OpenET platform supports water resource management across the United States but exhibits biases that vary by climate. While machine learning (ML) can improve satellite ET accuracy, prior studies report aggregate metrics without explaining how corrections work or providing operational uncertainty quantification. This study applies an XGBoost correction framework integrating OpenET satellite outputs with ground meteorological observations from 22 AmeriFlux cropland eddy covariance sites across six Köppen climate zones. The framework achieved 41.9% RMSE improvement over OpenET, while a configuration excluding all ground meteorological inputs still achieved 28.5% improvement. Error decomposition revealed that the dominant correction mechanism shifts seasonally: summer corrections were driven primarily by bias removal (61.3% of improvement), reflecting systematic overestimation during peak ET, while winter corrections were driven by variance reduction (96.8%), addressing satellite noise from cloud contamination. Across climates, humid zones improved primarily through variance reduction (68–81%), while water limited zones, including BSk and dry-summer Csa sites, benefited more from bias correction (40–52%). SHAP analysis stratified by climate yielded patterns consistent with these mechanisms. Leave-one-site-out cross-validation indicated strong transferability in humid climates (Dwb R² = 0.82) but limited generalization in BSk sites (R² = −0.30), while leave-one-cluster-out validation provided a more conservative spatial-transferability estimate by accounting for colocated site clusters. Conformal prediction provided empirical uncertainty quantification: constant-width intervals achieved 92.6% overall coverage, while climate-adaptive intervals produced environment specific uncertainty bounds ranging from ±0.81 to ±2.13 mm/day across climates. This framework provides a pathway toward operational ET correction with uncertainty bounds and diagnostic insights for satellite algorithm improvement.