Year-round daily wildfire prediction and key factor analysis using machine learning: a case study of Gangwon State, South Korea

Abstract

Under climate change and human’s dominant influence, wildfires have been increasing in frequency and scale, highlighting the demand of effective wildfire prediction and response. While prior research often maps high-risk areas, a few studies predict wildfire occurrences at specific dates and locations. This study aims to develop a year-round daily wildfire prediction model using machine learning, targeting Gangwon State in the Republic of Korea, and examine major influencing factors. We integrate meteorological elements (e.g., temperature, humidity, precipitation), forest-related variables (e.g., coniferous forest ratio, forest growing stock volume), and socioeconomic indicators (e.g., agricultural and cemetery land ratios) to identify salient predictors. We compare multiple algorithms, including Logistic Regression, XGBoost, and Random Forest, and use SHAP (SHapley Additive exPlanations) to enhance interpretability. The Extra Tree model achieves the highest AUC (0.839), and Random Forest demonstrates the best recall (0.828). SHAP results confirm that meteorological factors—especially relative humidity, precipitation, and temperature—are crucial, with forest- and socioeconomic variables also showing consistent effects. Applying a machine learning–based approach to daily wildfire prediction, integrating climate, environmental, and anthropogenic factors nationwide, and refining the temporal and spatial resolution of input data helps to advance wildfire prevention and response strategies in practice.