Deciphering the complex drivers of vegetation change in arid regions using an interpretable machine learning framework across multiple scales

Western China's endorheic basins are climate-sensitive, yet the non-linear coupling and scale-dependent drivers of their vegetation remain unclear. We applied the MS-XGSA (Multi-Scale XGBoost-SHAP Attribution framework), integrating machine learning and trend analysis, to examine NDVI dynamics across major basins and their riparian zones. Results reveal a hierarchical pattern where climate dominates at the basin scale, while topography and human activities act as localized modulators. As the spatial scale narrows to riparian buffers, the driving forces shift significantly toward hydrological dependence and concentrated anthropogenic interference. Vegetation exhibits pronounced non-linear responses to environmental stress, characterized by critical thermal and topographic thresholds that trigger abrupt browning, defined as a sudden and significant decline in vegetation greenness (NDVI), marking a rapid transition from stable growth to ecosystem degradation. Furthermore, the synergistic coupling of rising temperatures and declining precipitation creates compound risks that destabilize the ecosystem's water balance. A distinct threshold drift occurs between scales, with riparian zones displaying higher thermal tolerance. This quantifies their role as ecological refugia, where hydrological compensation and evaporative cooling mitigate macro-climatic pressure. This study characterizes the transition from climate-controlled regional processes to hydrologically-regulated local habitats, providing a scientific basis for identifying ecological tipping points and implementing scale-specific management to enhance arid ecosystem resilience.