Wildfire Susceptibility Mapping in Karabuk Province, Türkiye Using Machine Learning Algorithms

Abstract. Wildfires are among the most devastating natural hazards, increasingly intensified by climate change and anthropogenic pressures. Accurate susceptibility mapping is essential for disaster preparedness, risk mitigation, and sustainable land management. This study investigates the performance of five boosting-based machine learning algorithms—Gradient Boosting Machine (GBM), Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CatBoost), and Extreme Gradient Boosting (XGBoost)—in wildfire susceptibility assessment. Thirteen conditioning factors representing topographic, vegetational, climatic, and anthropogenic drivers were integrated into the models after preprocessing and multicollinearity checks. The results show that XGBoost, CatBoost, and LightGBM significantly outperformed GBM and AdaBoost, with XGBoost achieving the highest predictive accuracy (94.5%), AUC (0.939), and Kappa index (0.890). Feature importance analysis revealed that land cover, NDVI and temperature were the most significant factors, followed by slope, wind speed and proximity to human settlements. The susceptibility maps produced by the best-performing models provided spatially consistent and interpretable patterns, successfully delineating high-risk areas. This research confirms the effectiveness of advanced ensemble learning techniques, particularly XGBoost, in improving the accuracy and interpretability of wildfire susceptibility mapping. The findings provide actionable insights for forest management, land-use planning, and the development of early warning systems, contributing to more resilient strategies against escalating wildfire threats in a changing climate.