Explainable Artificial Intelligence (XAI) framework using XGBoost and SHAP for assessing urban fire risk based on spatial distribution features

Urban fire risks intensify owing to rapid urbanization, increasing population density, and concentrated infrastructure. Traditional fire risk assessment models, such as Multi-scale Geographically Weighted Regression (MGWR), often overlook spatial heterogeneity and lack interpretability. This study applied an eXplainable Artificial Intelligence (XAI) framework integrating eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) for predictive modeling and for feature attribution, respectively, to assess the urban fire risk in Qingyang District in Chengdu, China. A key challenge was to address the spatial complexity of fire risks and ensure model interpretability for decision-makers. To overcome this challenge, we integrated spatial autocorrelation metrics and kernel density analysis to analyze the spatial patterns of key features. The framework outperformed MGWR in terms of prediction accuracy, with R² values of 0.998 for fire occurrence and 0.935 for fire loss, compared to the 0.995 and 0.856 obtained for MGWR, respectively. The mean squared error was 4.447 for fire occurrence and 0.00027 for fire loss, while 10.297 and 0.00055 were obtained for MGWR, respectively. Notably, the SHAP values proved to be effective in explaining the spatial heterogeneity of urban fire risk features. The main contributions of this study are as follows: (1) the integration of XGBoost and SHAP for interpretable urban fire risk prediction, (2) the application of spatial analysis to enhance model transparency, and (3) the development of a framework that provides actionable, spatially explicit insights for urban safety planning. This study provides a data-driven tool for policymakers that offers quantitative guidance for urban fire risk mitigation.