Global Divergence in Trend Covariability Between Surface Evapotranspiration and Water Availability: A Scenario‐Conditioned Machine Learning Attribution

Evapotranspiration (ET) and precipitation (P) jointly shape terrestrial water availability, defined here as P − ET and regulate land‐atmosphere water and energy exchanges. However, their long‐term coupled changes remain insufficiently understood. We examined global land patterns of trends in ET and P − ET during the period of 1982–2020 using GLEAM v3.8a and ERA5‐Land ET together with MSWEP v2.8 precipitation. For the annually selected driest and wettest months, random‐forest models were trained on 1982–2010 samples and evaluated using a temporal holdout for 2011–2020, separately for four combinations of ET and P − ET trend directions. Both ET products showed widespread ET increases, whereas P − ET trends were spatially heterogeneous and frequently differed in sign from ET trends. Joint consideration of mean P − ET, (P − ET)/ET and their long‐term trends showed that some regions with positive mean water availability experienced a diminishing precipitation surplus relative to actual ET. Moreover, both RF and SHapley Additive exPlanations (SHAP) analyses showed that predictive contributions varied among precipitation conditions and trend scenarios. Temperature‐related information was prominent for ET in selected scenarios; antecedent vegetation state was more informative under the driest‐month condition where ET and P − ET trends differed in sign and contemporaneous precipitation dominated wettest‐month P − ET predictions. These results identify regional variability in water availability and show that the associated predictive relationships are condition‐ and scenario‐dependent rather than governed by a single universal factor.