Integrating Explainable Machine Learning and Classification methods for Wildland Fire Danger Mapping

Abstract. In the context of contemporary wildland fire management, accurately understanding and predicting fire danger is fundamental to ensuring public safety and optimizing the allocation of suppression resources. Geospatial data combined with machine learning provides an effective framework for analyzing the complex interactions between weather patterns and site characteristics, thereby identifying conditions associated with elevated wildfire danger. Machine learning encompasses a range of algorithms developed to detect patterns in data. This study integrates a Random Forest model with Recursive Feature Elimination and Cross Validation (RFECV) and three classification methods to generate wildland fire danger maps. To address the opacity inherent in machine learning models, SHapley Additive exPlanations (SHAP) were utilized to quantify the marginal contribution of each variable to the predicted distance from fire. Given the critical role of classification in mapping outcomes, we evaluated the relative effectiveness of several classification methods, including Natural Breaks (NB), Geometrical Interval (GI), and Standard Deviation (STDI). The results indicate that the GI approach most effectively aligns high-risk classes with observed fire patterns. Assessing the sensitivity of wildland fire danger mapping across the three classification methods demonstrates both consistency and uncertainty in the resulting fire danger maps. Overall, the comparative results highlight the robustness of the proposed machine learning based framework for producing spatially explicit fire danger maps.