Long-Term Dynamics of Alpine Grassland Ecosystem Resilience in Northwestern Sichuan with Stagewise Shifts and Nonlinear Driver Responses

Alpine grassland ecosystems are increasingly exposed to global change and human activities, yet their long-term resilience dynamics, stagewise shifts, and nonlinear driver associations remain unclear in northwestern Sichuan, China. We estimated a grassland resilience index (RI) for 1985–2020 using a first-order autoregressive model within centered five-year moving windows. Inputs comprised monthly normalized difference vegetation index (NDVI), temperature, and precipitation data for 1983–2022. Exploratory change-pattern analysis, nested group cross-validation, Shapley Additive exPlanations (SHAP), and accumulated local effects (ALE) characterized temporal changes and predictor associations. Regional RI declined significantly under autocorrelation-adjusted inference. Exploratory classification assigned 54.01% of grassland to linear or candidate turning patterns; 75.56% of candidate turning areas showed increase-to-decrease patterns. Random forest achieved a mean outer-test R2 of 0.36 ± 0.02, decreasing to 0.24 ± 0.05 under 100 km spatial-block validation and 0.11 ± 0.02 without Year. Year accounted for 48.68% of full-model SHAP importance. Among the seven ecological predictors, temperature–energy factors contributed 51.96%, exceeding moisture-related factors (35.51%). However, this dominance was not robust to excluding Year (driver-rank Spearman ρ = −0.29). ALE revealed asymmetric temperature associations, whereas grazing responses varied across periods. These results indicate limited predictive power of ecological predictors alone and Year-sensitive driver attribution.