Driving mechanisms and threshold identification of landscape ecological risk: A nonlinear perspective from the Qilian Mountains, China

• Exploring spatiotemporal drivers of landscape ecological risk and establishing a nonlinear constraint-based evaluation framework. • Using Geo-Detector, XGBoost–SHAP, constraint lines, and spline regression to identify key thresholds in landscape ecological risk. • Linking threshold regulation to ecosystem management, offering innovative solutions for vegetation restoration and spatial planning. Landscape Ecological Risk Assessment is a core component of spatial governance and regional ecological protection, as well as a fundamental task in ecosystem management. This study uses the Qilian Mountains ecosystem as a case study, innovatively integrating the Geo-Detector model, XGBoost-SHAP model, and constraint line method to explore the spatiotemporal dynamics and driving mechanisms of Landscape Ecological Risk (LER) from a nonlinear perspective between 2000 and 2023. The main findings are as follows: (1) Temporal Evolution Characteristics: The annual variation rate of the Landscape Ecological Risk Index (LERI) was 0.0011 yr −1 ( R 2 = 0.0861, p = 0.1641), showing weak fluctuations. The proportion of Extremely Low-Ecological Risk Areas and Low Ecological Risk Areas remained stable within the range of 50.56 % to 64.07 %, and the ecological security pattern remained relatively stable. The area of Extremely High Ecological Risk Areas decreased significantly, with an annual reduction rate of −0.0791 × 10 4 km 2 yr −1 ( R 2 = 0.5655, p < 0.001), indicating continuous improvement in regional ecological quality. (2) Driving Mechanism Analysis: The Geo-Detector model showed that the primary driving factors, ranked by explanatory power, were Grazing Intensity (GI) (Q = 0.2472), Land Surface Temperature (LST) (Q = 0.2145), Elevation (Q = 0.1605), Annual Precipitation (Q = 0.1546), Downward Shortwave Radiation (DSR) (Q = 0.1032), and Annual Mean Temperature (Q = 0.0942), with a total explanatory power of 80.83 %. The XGBoost-SHAP model identified the top six significant factors as GI (SHAP = 0.0918), Specific Humidity (SH) (SHAP = 0.0454), Annual Precipitation (SHAP = 0.0452), DSR (SHAP = 0.0344), Wind Speed (WS) (SHAP = 0.0259), and Elevation (SHAP = 0.0251), with a total contribution rate of 87.46 %. Interaction analysis revealed that the nonlinear synergistic effect between GI and climate factors was the most significant, particularly the interactions between GI and Annual Precipitation (Q = 0.434) and GI and Elevation (Q = 0.419). (3) Threshold Response Characteristics: Elevation exhibited a concave-downward constraint effect ( R 2 = 0.7867), with a critical threshold at 4200 m. Beyond this threshold, the constraint intensity on LER increased. A significant threshold inflection point for DSR was found at 2502 W/m 2 . Climate constraint thresholds revealed that when Annual Precipitation < 200 mm, Mean Temperature < -6°C, and Specific Humidity < 2.8068 g/kg, the constraint effect on landscape risk was enhanced. Grazing Intensity exhibited a dual-threshold response: 3.35 SU/ha was the critical point for rapid increases in landscape risk, while 14.36 SU/ha marked the threshold for abrupt ecological stability loss. Beyond this threshold, the fragmentation of landscape structure sharply increased, significantly raising the risk of ecological collapse. The nonlinear constraint mechanism model of “driving factors − Landscape Ecological Risk” proposed in this study overcomes the limitations of traditional threshold determination methods and provides an accurate and quantitative tool for mountain ecosystem restoration and spatial planning. The findings offer significant practical value for balancing regional ecological protection with sustainable development.