Opposite shifts in drought-season evapotranspiration controls across hydroclimatic regimes in China

Drought depletes water resources and can trigger substantial productivity losses and plant mortality. However, predicting drought impacts on water resources and ecosystem functioning remains difficult because evapotranspiration (ET) responses are highly uncertain. The sign and magnitude of drought-induced ET anomalies affect not only the water balance, but also land-atmosphere interactions and drought progression. Atmospheric drying (high vapor pressure deficit, VPD) can enhance ET, whereas soil moisture (SM) depletion suppresses soil evaporation and plant transpiration via stomatal regulation. Therefore, drought ET responses emerge from competing constraints imposed by atmospheric demand and moisture supply. Here we quantify how VPD and SM jointly control growing-season drought ET anomalies across hydroclimatic regimes in China using satellite remote sensing, physics-constrained machine learning ET estimations, and hydro-meteorological reanalysis data. ET is derived by coupling the Penman–Monteith framework with machine learning, yielding estimates that have been extensively validated and shown to perform robustly under data-limited conditions and during drought events. We then characterize the sign and magnitude of ET anomalies during drought by jointly considering meteorological, hydrological, and ecological drought metrics. Then, we disentangle the contribution of atmospheric demand and moisture supply constraints on ET anomalies based on the percentile binning method (assuming weak VPD and SM dependence in their short intervals), thereby distinguishing water demand-limited from water supply-limited regimes. The enhancement driven by atmospheric drying dominates in water demand-limited regions, while the suppression driven by soil moisture deficit prevails in water supply-limited regions, and both vary along dry-wet gradients. Finally, using an explainable machine learning approach (SHAP), we diagnose multiyear changes in these controls. We find regime-dependent trends with opposite signs: the positive VPD effect on drought ET anomalies declines in demand-limited regions, whereas the negative SM effect becomes less negative in supply-limited regions. These opposite-sign trends are primarily associated with evolving air-temperature and soil-moisture anomaly patterns, highlighting non-stationary drought controls on ET across China’s hydroclimatic regimes.