Spatial Prediction of Forest Fire Occurrence Integrating Human Proximity: A Machine Learning Approach for Korea’s Eastern Coast
Jeman Lee, Sujung Ahn, Sangjun Im
Forest fire occurrence prediction remains challenging despite advances in operational fire danger rating systems. In South Korea, the Korea Forest Fire Danger Rating Index (KFDRI) incorporates meteorological conditions, terrain (elevation, aspect), and forest type to assess regional fire danger. While KFDRI successfully assesses environmental fire danger at the pixel level, it does not explicitly account for human activity patterns that create substantial occurrence variability among locations with similar environmental conditions. This limitation is critical in human-dominated landscapes where where the main source of fire occurrence is anthropogenic. This study developed a Random Forest (RF) model to predict forest fire occurrence probability and propose management priorities during the forest fire prevention season (November–May) along the eastern coast of Korea, explicitly integrating human proximity variables (distance to agricultural areas and roads) with topographical (elevation, slope, aspect), surface fuel load, and meteorological variables (SMAP soil moisture, cumulative precipitation). Using forest fire occurrence records (1112 fire occurrence records) and background samples from 2015 to 2024, the model was trained with monthly stratified sampling and 10-fold cross-validation. The model achieved stable classification performance, with an overall F1-score of 0.515 and accuracy of 0.733. According to the SHAP (SHapley Additive exPlanations) analysis, distance to agricultural areas, elevation, slope, aspect, 5-day cumulative precipitation, and forest type were the most influential predictors. In particular, occurrence probability tended to increase in areas close to agricultural land (<180 m), at low elevations (≤200 m), on moderately steep slopes (≥8°), on south- and west-facing aspects, and under dried conditions. These results emphasize that fire occurrence risk is primarily structured by human proximity within areas of similar environmental danger. We propose an operational integration in which the RF model provides a 30 m “where-to-focus” occurrence layer that is used alongside KFDRI’s daily danger rating to prioritize prevention and patrol efforts.