Identifying the driving mechanisms of ecosystem health in a typical ecologically fragile region: A study based on the XGBoost–SHAP model

• While ecologically improved overall, the Loess Plateau still faces localized degradation. • Afforestation suitability in the humid southeast, contrasting with the arid northwest. • Anthropogenic factors emerged as the primary driver of ecosystem health. • Key drivers showed contrasting spatial patterns, necessitating region-specific management. Accurately assessing ecosystem health is essential for effective environmental management, particularly in ecologically fragile regions like the Loess Plateau. However, existing studies face challenges in disentangling driving mechanisms, as traditional statistical methods cannot precisely quantify the independent contributions of multiple factors. To address these limitations, this study developed an ecosystem health index (EHI) based on the Vigor–Organization–Resilience–Service (VORS) theoretical framework and innovatively integrated spatial analysis, XGBoost–SHAP modeling, redundancy analysis (RDA), and geographically weighted regression (GWR) to systematically quantify the spatiotemporal dynamics and underlying drivers of EHI on the Loess Plateau from 2000 to 2020. The main findings were as follows: (1) EHI on the Loess Plateau increased steadily, and the optimization of its hierarchical structure demonstrates the effectiveness of restoration. However, significant spatial heterogeneity was observed. (2) The SHAP model and RDA jointly confirmed that anthropogenic factors were the primary drivers of EHI. Forest cover stands out as the most influential yet paradoxical factor, a finding that highlights the need for region-specific strategies. (3) The integrated VORS–XGBoost–SHAP approach offers a mechanistic and spatially explicit analysis of driving pathways. This finding necessitates a shift in management from blanket interventions to adaptive strategies that respect environmental constraints. Overall, this study provides a scientific basis for ecological management on the Loess Plateau, and introduces an interpretable machine learning–driven analytical framework that serves as a methodological paradigm for mechanistic analyses of ecosystem health in other vulnerable regions.