Integrating Spatial Dependence into Machine Learning to Quantify the Impacts of 2D/3D Built Environment Features on Fire Risk

Clarifying the relationships between built environment characteristics and urban fire risk is important for developing effective fire prevention and planning strategies. However, spatial dependence and nonlinear relationships between the built environment and fire risk remain insufficiently understood. Accordingly, this study presents a geographically enhanced machine learning (GE-ML) framework that incorporates spatial adjacency into machine learning models through spatially weighted feature construction. Specifically, contiguity-based spatial weight matrices were used to derive spatially weighted features from 2D and 3D built environment variables. The Optimal Parameter-based Geographical Detector (OPGD) was applied to assess scale sensitivity and compare the explanatory power and interactions of the original and spatially weighted features. Six candidate models, including Ordinary Least Squares (OLS), K-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and eXtreme Gradient Boosting (XGBoost), were then evaluated under different feature configurations, followed by SHapley Additive exPlanations (SHAP) analysis of the selected model. Results show that: (1) spatial weighting generally increased the explanatory power of major built environment factors and their interactions, with Queen contiguity yielding higher q-values than Rook contiguity; (2) spatially weighted features improved predictive performance across different models, and GE-XGBoost achieved the highest R2 (0.7067) and lower residual spatial autocorrelation than GWR and GWRF; and (3) 2D and 3D built environment features accounted for 59.55% and 40.45% of the total SHAP importance, respectively, with Geo-TPD, Geo-BVD, Geo-PS, and Geo-LUI identified as the most important features. SHAP analysis further revealed nonlinear relationships and interactions between these features and predicted fire risk. These findings highlight the value of incorporating spatial adjacency information into fire risk modeling and support spatially differentiated fire risk management.