Machine-learning wildfire susceptibility mapping with SHAP-based explainability in Türkiye’s fire-prone regions

Wildfire management in Türkiye’s Mediterranean and Aegean regions, characterized by frequent and often extreme fires, requires spatial assessments of wildfire susceptibility to identify and mitigate risks. Our study employs machine-learning (ML) models to map wildfire susceptibility and uses explainable AI (XAI) to provide transparent interpretation of the predictive drivers and their interactions. We retrieved fire ignitions from Visible and Infrared Radiometer Suite (VIIRS) imagery for the period between 2013 and 2022, along with 21 fire conditioning factors encompassing climatic, topographic, anthropogenic, and vegetative variables. We used Gradient Boosting Machine (GBM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) as ML classifiers for fire susceptibility. GBM outperformed the other techniques with an area under the curve (AUC) of 0.84. We used SHapley Additive exPlanations (SHAP) to quantify the contributions of conditioning factors. Among the key predictors were temperature, leaf area index and solar radiation. A pairwise interaction analysis further underlined the interconnection of climatic variables, like solar radiation and wind speed, in exacerbating fire risks. Spatiotemporal analysis revealed August and July as peak fire months, with Hatay, Antalya, İzmir, and Mugla as the most heavily affected regions. The susceptibility map indicates that 13.9% of the study area is under very high fire risk. This study advances the integration of XAI with ML for wildfire susceptibility mapping and introduces a framework that could benefit decision-making and predictive power for wildfire management.