Predicting Wildfire Damage Using Machine Learning: Focusing on Meteorological and Environmental Variables

기계학습을 활용한 산불 피해 규모 예측: 기상 및 환경 변수를 중심으로

Wildfires have become increasingly frequent and severe in recent years, driven by rising global temperatures, prolonged droughts, and shifting precipitation patterns associated with climate change. These fires not only cause substantial ecological damage but also threaten human lives and infrastructure. As a result, the ability to accurately predict the scale of wildfire damage shortly after ignition is becoming a critical component of disaster preparedness and forest management. This study proposes a machine learning-based approach to predict the magnitude of wildfire damage using post-ignition environmental and geographic variables. The research utilizes wildfire incident data collected in South Korea between 2020 and 2024. Wildfires were classified into three categories-small, medium, and large-based on area burned and fire duration, following criteria adapted from national wildfire response manuals. To build predictive models, a diverse set of variables was used, including meteorological factors, drought indices, vegetation characteristics, and spatiotemporal information such as season and administrative region.Three classification algorithms-Random Forest, XGBoost, and Support Vector Machine (SVM) were applied. Due to the imbalance in class distribution, particularly the scarcity of large wildfire cases, data resampling techniques were employed to enhance model robustness. Among the models, XGBoost demonstrated the highest accuracy of 0.96 and achieved a recall of 0.89 for large wildfires, outperforming the other methods. These results suggest that combining real-time weather data with historical environmental information can help improve early predictions of the scale of the wildfire. The proposed model may assist in supporting faster response decisions and minimizing damage in high-risk areas.