Artificial intelligence-driven forest fire land susceptibility mapping: A remote sensing and machine learning approach

Forest fires pose a significant threat to ecological stability, particularly in climatically vulnerable and topographically complex regions like Khyber Pakhtunkhwa, Pakistan. This study aims to identify the key environmental and anthropogenic drivers of fire occurrence in the Hazara Division and to evaluate the predictive performance of machine learning models for spatial susceptibility mapping. We hypothesize that climatic variability, vegetation conditions, and anthropogenic accessibility collectively influence forest fire distribution patterns, while elevation serves as an indirect spatial indicator of these environmental gradients. This research aims to comparatively evaluate the predictive performance of four machine learning classifiers: Random Forest (RF), Support Vector Machine (SVM), XGBoost, and Gradient Boosting Machine (GBM) for forest fire susceptibility mapping. A dataset of 2100 georeferenced points (1050 fire and 1050 non-fire) was used, and 34 predictors were initially considered, compiled from four thematic domains: topographic (elevation, slope, landforms, etc.), vegetation based Normalized Difference Vegetation Index (NDVI), Net Primary Productivity (NPP), Leaf area index (LAI), etc., climatic (Rainfall, Temperature maximum (Tmax), Temperature minimum (Tmin), Soil moisture (SM)), and anthropogenic (Distance to roads, settlements, waterways). After multicollinearity analysis using the Variance Inflation Factor (VIF < 5), 13 key predictors were retained. Receiver Operating Characteristic (ROC) analysis revealed that among the evaluated classifiers, RF achieved the highest predictive performance (AUC = 0.90), followed by SVM (AUC = 0.87), XGBoost (AUC = 0.86), and GBM (AUC = 0.85). RF and XGBoost demonstrated comparatively stronger responses to topographic and vegetation indices, while SVM showed a stronger response to climatic parameters. Across all models, elevation, anthropogenic factor, and temperature consistently ranked among the top predictors, validating the importance of terrain and human proximity. Spatial results indicate that lower-elevation peri-urban forest margins in southern Hazara exhibit higher susceptibility. While the modelling framework demonstrates strong predictive capability, uncertainties may arise from satellite-based fire detection accuracy and differences in spatial resolution among environmental predictors. The resulting susceptibility maps provide a practical decision-support tool for targeted mitigation and improved forest management planning.