Predicting Anthropogenic Wildfire Occurrence Using Explainable Machine Learning Models: A Nationwide Case Study of South Korea

Anthropogenic wildfires account for the majority of wildfire ignitions in human-dominated landscapes, yet their spatial drivers remain insufficiently understood at national scales. This study aims to identify key factors influencing anthropogenic wildfire occurrence and to develop a robust and interpretable prediction framework using nationwide data from South Korea. Wildfire occurrence records from 2011–2021 were integrated with daily meteorological, environmental, and socio-economic variables at a 1 km grid resolution. A stacking ensemble model combining Random Forest, XGBoost, LightGBM, Extra Trees, and logistic regression was implemented to improve predictive robustness under rare-event conditions. Model performance was evaluated using ROC–AUC, PR–AUC, and threshold-optimized F1-scores, and variable contributions were interpreted using feature importance and SHAP analyses. The ensemble model achieved a PR–AUC of 0.934 and an ROC–AUC of 0.941. Relative humidity and maximum temperature were identified as influential meteorological variables, while human-accessibility-related variables, particularly distance to roads and agricultural land, showed consistently high contributions to spatial ignition probability. These findings indicate that anthropogenic wildfire occurrence is shaped by interactions between fire-weather conditions and spatial patterns of human accessibility. The proposed framework provides a scalable approach for understanding anthropogenic wildfire mechanisms and supporting prevention strategies in forested landscapes.