Spatially non-stationary wildfire risk formation driven by nonlinear ignition–driver interactions in a mountainous socio-ecological system

In coupled human–environment systems, ignition heterogeneity and spatial non-stationarity are central to precision wildfire prevention. Using a DPS–CHANS framework and a spatially explicit Geo-XGBoost model, this study examines wildfire evolution and nonlinear interactions in Liangshan Prefecture (2004–2024). Human-caused ignitions accounted for 95.7% of identified fires, with agricultural burning (158 cases) and smoking (62 cases) being the dominant production and non-production sources, respectively. Wildfire occurrence showed strong spatio-temporal concentration, although high ignition-frequency zones did not always correspond to large burned areas. Agricultural burning generated the largest cumulative burned area (4440 ha), whereas powerline-related fires were less frequent but more severe per event. Elevationally, human-caused fires clustered at 1500–2500 m, while lightning-caused fires peaked at 2516.5 m. Nonlinear interaction analysis revealed ignition-specific and spatially non-stationary risk formation. Agricultural fire risk increased when VPD exceeded 1.02 kPa, with the strongest interaction observed for VPD × relative humidity (0.389). Lightning fires were concentrated in high-mountain canyons and primarily controlled by temperature and lightning frequency (VPD × vegetation interaction = 0.153). These findings demonstrate that wildfire risk in mountainous regions is an ignition-specific product of coupled meteorological, ecological, and human processes. Our results advocate for a transition toward process-based, ignition-specific, and regionally differentiated management strategies.