Data-driven attribution of evapotranspiration dynamics in the Heihe River Basin: Controlling factors from site measurements to regional satellite observations

A comprehensive understanding of evapotranspiration (ET) processes and their controlling factors is vital for sustaining agricultural production and effectively managing limited water resources. However, identifying ET drivers across scales remains uncertain, as satellite-based products assume pixel homogeneity, which conflicts with land heterogeneity. This limits the understanding of how key environmental factors vary across different spatial scales under uniform climatic conditions. In this study, scale-dependent ET dynamics within the Heihe River Basin were quantified by integrating decade-long in-situ flux measurements with multi-source satellite products using an interpretable ensemble machine learning framework (R2 = 0.87). At the site scale, ET variability was primarily driven by air temperature (Ta, 35%) and leaf area index (LAI, 28%), followed by vapor pressure deficit (VPD, 11%) and soil moisture (9%). Crucially for agricultural water management, our analysis identified specific physiological thresholds: VPD levels above 1.10 kPa were associated with reduced SHAP contributions, potentially reflecting stomatal regulation under high evaporative demand, while LAI acted as a central mediator, amplifying the effects of thermal energy on water flux. Regionally, climatic factors dominated ET patterns (61–82%), yet satellite products exhibited divergent sensitivities that impact basin-wide planning: GLEAM was overwhelmingly driven by Ta (58%), whereas PML v2 uniquely captured CO2 fertilization feedbacks (15%) absent in empirical models. These findings underscore that agricultural water allocation strategies must account for the non-linear interactions between vegetation structure and local climate, as well as the impacts of elevated CO2, to ensure food security and ecosystem sustainability under changing environmental conditions.