Grassland ecosystems in arid regions face mounting stress from intensified climate variability and anthropogenic disturbance. Despite the predictive capabilities of machine learning models, their lack of interpretability challenges the transparency of resilience drivers. This study integrates temporal autocorrelation (TAC) metrics with explainable machine learning (ML) to assess grassland resilience dynamics in an arid ecosystem from 2001 to 2023. Results reveal spatial divergence, with reduced resilience in radiation-dominated arid zones and stronger recovery in hydrothermally stable areas. The model identifies temperature variability and vegetation activity as dominant contributors to resilience trends, exhibiting marked heterogeneity across grassland types. By quantifying both structural and dynamic aspects of resilience, this framework enhances interpretability and diagnostic precision, offering a practical tool for identifying degradation risks and ecological tipping points. Findings support region-specific adaptation strategies and provide a robust foundation for sustainable grassland governance in the face of accelerating global change.