Arid zones are vital yet fragile components of terrestrial ecosystems. However, the nonlinear response of ecosystem resilience to extreme climate events and anthropogenic disturbances remains insufficiently understood, and the thresholds of key influencing factors are still unclear. We innovatively developed a state-function-interaction resilience framework tailored to arid zones and applied it to a 31-year assessment of the Loess Plateau, China. Through a methodological innovation that couples eXtreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP) and Geographically and Temporally Weighted Regression (GTWR), we jointly quantified nonlinear effects, detected thresholds, and captured spatiotemporal nonstationarity in the impacts of extreme climate events and human activities on ecosystem resilience. Our findings revealed significant increases in ecosystem resilience on the Loess Plateau (1990-2020), with the impervious surface ratio, number of heavy rainfall days, and cropland coverage emerging as the dominant influencing factors. Threshold analysis revealed that when impervious surface cover exceeds 20%, cropland coverage surpasses 40%, or when the annual number of heavy rainfall days exceeds 6, these factors have threshold effects on resilience. Notably, we found that moderate extreme precipitation can increase the resilience of arid ecosystems. These findings underscore the importance of identifying nonlinear thresholds of influencing factors, providing a basis for adaptive land use planning and climate adaptation strategies. This study offers transferable methodologies and decision-support frameworks for ecological management in arid systems facing similar coupled stress factors.