Predicting wildfires triggered by human-caused ignition: A spatial framework integrating AI models

Human-caused ignitions—including agricultural residue burning, land clearing, and various open-fire practices—are now a major contributor to wildfire occurrence in South Korea, accounting for more than 40% of recent events and concentrating along rapidly expanding wildland–urban interface (WUI) zones. To improve national-scale understanding of these ignition processes, this study develops an integrated wildfire risk assessment framework that combines Random Forest (RF), Generalized Additive Models (GAM), and Geographically Weighted Regression (GWR). Using 3703 georeferenced human-caused wildfire records (2001–2025) and high-resolution climatic, land-use, vegetation, and demographic datasets, we identify key ignition drivers, quantify nonlinear environmental thresholds, and map spatial heterogeneity in human-driven wildfire susceptibility. RF achieved the strongest predictive performance (AUC = 0.829), while GAM revealed sharp increases in ignition probability below 15% relative humidity, around 15 m/s wind speeds, and at intermediate NDVI levels (0.2–0.4). GWR showed substantial regional variability in these effects, particularly within peri-urban landscapes where human activity and fuel continuity intersect. High-risk areas (probability ≥0.70) accounted for 16.96% of the national territory and were concentrated in northern Gyeonggi-do, eastern Gangwon-do, and parts of Chungcheong and Jeolla provinces. By integrating the complementary strengths of RF, GAM, and GWR, this study provides operational ignition thresholds and high-resolution risk maps that support evidence-based land-use planning, targeted burning restrictions, and climate-adaptive fire management. The framework offers a transferable approach for regions facing similar challenges associated with human-caused wildfire ignitions and rapidly evolving socio-ecological landscapes. • A national-scale assessment of human-caused wildfire ignition risk in South Korea using an integrated RF–GAM–GWR framework. • RF achieved the highest predictive performance (AUC = 0.829), while GAM identified nonlinear climate–vegetation thresholds and GWR captured pronounced spatial heterogeneity. • Key ignition drivers spanned human proximity, forest structure, topography, and climate, including building proximity, slope degree, low humidity (