Modeling Multi-Temporal Fire Niches for Wildfire Susceptibility Mapping Using Active Fire Data

Rising global wildfires threaten semi-arid ecosystems, necessitating advanced predictive mapping for mitigation. This study employs a comparative machine learning approach to develop seasonal wildfire susceptibility maps, capturing dynamic fire drivers for effective management in semi-arid regions. This study specifically investigated seasonal effects on wildfire patterns by evaluating four machine learning techniques (ANN, GLM, XGBoost, and MaxEnt) across distinct temporal scales. A comprehensive fire occurrence database was established using 12 years of active fire data. Twelve predictors, spanning four categories: meteorological parameters, satellite-derived variables, topographic features, and anthropogenic indicators were developed to capture seasonal variations in fire drivers. The results indicated that MaxNet and XGBoost consistently outperformed other models in multi-season wildfire risk prediction, achieving remarkable accuracy. Land cover emerged as the most critical wildfire predictor (> 60% influence). Models highlighted human activity as the dominant driver in warm seasons, while climatic and vegetation factors dominated in colder periods. The central and northern high-elevation regions (> 3000 m) are consistently identified as the highest-risk areas across models and seasons, especially during the warm season. Our findings highlight the crucial role of seasonal dynamics in wildfire risk. Accurate wildfire risk mapping relies on advanced ensemble models that precisely identify hotspots while balancing sensitivity, specificity, and spatial interpretability to guide mitigation strategies. This study provides practical tools for seasonally-targeted wildfire management.