Wildfire susceptibility mapping using three machine learning algorithms: the case study of the temperate region of Iran

Fire has been present in the Earth Planet for 400 million years and was a key tool for hunter-gatherers and farmers. However, fire also risks properties and causes casualties when out of control. It is necessary to determine the fire hazards to achieve sustainable management. Three novel models consisting of XGBoost, CatBoost, and LightBoost were implemented to prepare a probability distribution map showing fire susceptibility in Astara City, northern Iran. The case study (42,600 ha) is part of the Hyrcanian temperate forests and is one of the most sensitive cities to wildfire in this region. A SHapley Additive exPlanations (SHAP) model was used to investigate the contribution of different factors to the fire simulation in the applied algorithms. To assess the performance of the applied algorithms, the accuracy indexes consist of sensitivity, specificity, PPV, NPV, and AUC. Among the algorithms, the LightGBM model has the highest performance with accuracy indexes of 0.9712, 0.963, 0.9765, 0.966, and 0.98, respectively. The SHAP degree chart showed that between the contributed factors, the distance from road factor has the highest effect on wildfire, with a negative contribution to the simulation up to −0.25. According to the LightGBM model, very low, low, moderate, high, and very high susceptibility classes cover 51 %, 13 %, 8 %, 14 %, and 15% of the study area, respectively. The findings of this research can be used to develop strategic and operational plans for controlling and managing wildfires effectively in the temperate region of Iran and other forest areas of the world.