Grassland degradation in semi-arid regions is largely driven by complex climate-human interactions, but most assessments rely on a single remote-sensing indicator, limiting the detection of structural and functional shifts. To overcome this constraint, we established a comprehensive framework that combines multiple indicators and sensor datasets. This approach integrates fractional vegetation cover and net primary productivity spanning 2001-2022, along with aboveground biomass data for 2001-2020, to systematically assess long-term grassland dynamics in the Ili region of northwestern China. Degradation classifications were validated using Landsat 8 and Sentinel-2 imagery with texture-enhanced Random Forest models. Among the three indicators, AGB exhibited the highest reliability (overall accuracy approximate to 0.95; user's accuracy >0.70) and strongest temporal responsiveness to extreme events. Trend analysis revealed weak but consistent long-term declines across all indicators, with severe degradation peaking in 2008 (70.8%) and 2014 (65.9%), while partial recovery occurred in 2002 and 2016. Spatially, mountainous areas showed greater resilience, whereas lowland plains and agropastoral ecotones were more vulnerable. Using an XGBoost-SHAP framework, we quantified nonlinear driver contributions and identified temperature as the dominant long-term control, followed by anthropogenic factors, which had moderate influence (population 1.28, livestock 1.24, and gross domestic product 0.45). Integrating biomass-based structural metrics with greenness and productivity indicators improves the detection of degradation dynamics and clarifies climate-human interactions in fragile grassland ecosystems.