Integrating nonlinear ecosystem service responses into ecological security pattern construction using explainable machine learning: A case study of the northern Qinling Mountains foothills

Rapid urbanization and intensifying anthropogenic activities are fundamentally restructuring ecological processes in mountain–urban transitional zones, driving nonlinear ecosystem degradation and diminishing ecological resilience. Existing ecological security pattern (ESP) studies largely rely on static resistance assignments, making it difficult to capture the threshold responses of ecosystems to disturbances and limiting the realistic representation of ecological flows. Taking the northern foothills of the Qinling Mountains as a case study, this study integrates ecosystem service value (ESV) dynamic analysis with interpretable machine learning to construct a threshold-responsive ESP framework. To achieve this, the study analyzed ESV changes from 2000 to 2020, employed XGBoost-SHAP to identify nonlinear responses and critical thresholds of driving factors, translated these thresholds into ecological resistance rules, and constructed an ecological network using circuit theory. Network robustness was then tested under random failure and targeted attack scenarios. The results showed that: (1) ESV exhibited a V-shaped pattern (decline then recovery), with degradation concentrated in rapidly urbanizing areas; (2) precipitation, slope, and NDVI together accounted for 60.9% of the total mean absolute SHAP value across all factors; (3) the SHAP-based resistance surface better identified ecological transitional zones, bottlenecks, and critical corridors; (4) the threshold-responsive network exhibited higher structural stability under targeted attack than conventional methods. The framework has been validated in the mountain–urban transitional zone of the northern Qinling Mountains, and its core methodological workflow exhibits cross-regional transferability. Future research can further extend it to different geographical environments and ecosystem types, and explore adaptive optimization methods for LOWESS parameters to enhance the framework's universality.