Abstract: Objectives: Arid and semi-arid resource-based urban agglomerations are located in areas where ecologically vulnerable regions overlap with intensive resource development. Identification of grassland ecosystem vulnerability and application of explainable machine learning to analyze the importance and response characteristics of assessment factors can support regional ecological understanding, restoration planning, and green transition. Methods: Grassland cover trajectories in the Hohhot-Baotou-Ordos-Yulin urban agglomeration were identified using the China land cover dataset and net primary productivity (NPP) data from 2001 to 2025. Sensitivity, adaptability, and vulnerability were quantified, and a tabular prior-data fitted network (TabPFN) with Shapley additive explanations (SHAP) was used to analyze the contributions, nonlinear responses, and interactions of climate, socioeconomic, surface ecological, and topographic factors. Results: Grassland cover was generally stable, with stable grassland accounting for 67.70%. Gained grassland was concentrated in the Mu Us Sandy Land and the surrounding restoration areas, whereas lost and fluctuating grassland occurred mainly near urban margins, mining areas, and cropland edges. Mean annual NPP was higher in the east and south and lower in the west and northwest, with a regional mean value of 174.188 gC·m^-2·a^-1. Sen-MK and Sen-Hurst analysis showed that the significant and persistent NPP increases accounted for 90.43% and 94.10%, respectively. Vulnerability showed that the marked spatial heterogeneity, and severe vulnerability had the largest share of 21.15%, mainly in the Ordos Plateau, margins of the Mu Us Sandy Land, and local mining areas. TabPFN performed the best, with a mean five-fold cross-validation R² of 0.827 5. Mean normalized difference vegetation index (NDVI), mean precipitation, and NDVI standard deviation were important assessment factors, and their importance differed among grassland trajectories. Conclusions: Nonlinear threshold and interaction analysis indicated that vegetation state, water availability, thermal conditions, and topography were closely associated with the spatial differentiation of grassland vulnerability. A zonal governance scheme integrating protection, restoration, consolidation, and regulation was developed to support vulnerability identification and ecological restoration in arid and semi-arid resource-based urban agglomerations.