Stage-dependent dynamics of Ebinur Lake under hydrological restructuring: An uncertainty-aware explainable machine-learning diagnosis

Terminal lakes in arid inland basins are highly sensitive to coupled climatic variability, hydrological restructuring, and human water use. However, their long-term surface-area responses are often difficult to interpret because multiple drivers are strongly correlated and may act nonlinearly. This study investigated the stage-dependent lake surface area (LSA) dynamics of Ebinur Lake, a typical closed terminal lake in northwestern China, during 1990–2023. An annual effective LSA dataset was derived from Landsat surface reflectance imagery, and multi-source climatic, hydrological, groundwater, and land-use data were integrated using an uncertainty-aware PCA–RF–SHAP diagnostic workflow. The results showed that Ebinur Lake experienced pronounced interannual variability and four stages of surface-area evolution: fluctuating expansion during 1990–2003, contraction during 2004–2015, short-term recovery during 2016–2019, and rapid contraction during 2020–2023. Although the basin showed a warming–wetting climatic tendency, this signal did not translate into sustained lake recovery. Instead, enhanced potential evapotranspiration, cropland expansion, reduced inflow connectivity, declining runoff, and weakened groundwater support suggest that effective recharge to the terminal lake was increasingly constrained under basin-scale hydrological restructuring. PCA extracted two composite drivers explaining 80.75% of the total variance. PC1 represented integrated basin-scale water-resource pressure, whereas PC2 reflected precipitation–runoff variability. Model validation showed limited out-of-sample predictive skill, so SHAP results were interpreted as exploratory model-based associations rather than causal evidence. These findings suggest that the long-term shrinkage of Ebinur Lake is associated with cumulative basin-scale water-resource pressure rather than climatic variability alone. The proposed uncertainty-aware diagnostic workflow provides an interpretable ecological-informatics perspective for understanding terminal lake responses under coupled environmental pressures.