Explainable Artificial Intelligence to Unveil Intrinsic Characteristics of Conditioning Factors Governing Forest Fire Susceptibility

Abstract. Forest fires are typically triggered by natural factors or human negligence and accidents, spreading across vast areas and causing extensive damage to vegetation, wildlife, and ecosystems. Machine learning algorithms have recently become important tools for their efficiency in generating high-quality wildfire susceptibility maps in the literature. Despite their success in achieving promising thematic accuracies, they are typically criticized for their black box structure and their limited ability to interpret the resulting susceptibility maps. This study aims to address these limitations by exploring the inherent characteristics of geospatial covariates controlling the wildfire phenomena with local and global underlying factors of wildfire phenomena with the application of explainable artificial intelligence (XAI). For this purpose, three ensemble machine learning algorithms, including random forest, XGBoost, and NGBoost, were initially inputted with 11 conditioning factors to produce wildfire susceptibility maps. The internal mechanisms of these models were then interpreted using global and local XAI techniques. The results showed that the NGBoost had the highest predictive performance with an overall accuracy of 81.42%, and outperformed the other algorithms by approximately 5% to 8%. The global explainability analysis with the SHAP technique revealed that topographical parameters, such as elevation and valley depth, were the most influential factors in wildfire susceptibility. On the other hand, local analyses conducted with the LIME technique for three randomly selected instances highlighted the significant influence of parameters such as elevation, wind speed, and valley depth on individual wildfire cases.