Seasonal Forest Fire Susceptibility Mapping in South Korea Using Explainable Machine Learning

다국어 초록 (Multilingual Abstract)

Seasonal forest fire dynamics in monsoon-influenced temperate regions, such as South Korea, reflect complex interactions among climate variability, vegetation stress, topography, and human activities. Despite increasing national attention to forest fire hazards, seasonal-level susceptibility mapping remains limited, and the interpretability of models is rarely addressed. This study aimed to map the risk of fire using multivariable data, including topographic, climatic, biophysical, and anthropogenic factors from 2019 to 2023 and validate the results using 2024 data, utilizing near-real-time fire data from the FIRMS (Fire Information for Resource Management System). Three ML (Machine Learning) models, RF (Random Forest), XGBoost (eXtreme Gradient Boosting), and LGBM (Light Gradient Boosting Model), were compared to evaluate prediction accuracy and interpretability using SHAP (Shapley Additive Explanations) and PDP (Partial Dependence Plots). Based on the final accuracy results, the XGBoost model achieved the highest accuracy compared to the RF and LGBM models. The top feature importance based on SHAP results revealed that spring was characterized by land surface temperature, summer and fall by the NDR (Normalized Burn Ratio), and winter by population density. Notably, NBR and population density consistently ranked among the top five most influential features across all seasons. The relative importance of predictors varied by season, with thermal conditions being the most influential in spring, vegetation indices in summer and autumn, and human-related factors in winter. Overall, this study demonstrates that utilizing machine learning and interpretable models enhances our understanding of how the dynamics of key driver factors vary in response to forest fires across different seasons, which can inform specific forest fire management and mitigation strategies tailored to each season.