Computational modeling study shows that spatially weighted machine learning improves urban fire risk prediction, highlighting the significant role of 2D and 3D built environment features.
Key Points
To develop a geographically enhanced machine learning framework that integrates spatial adjacency to quantify the nonlinear effects of 2D and 3D built environment factors on urban fire risk.
Constructed spatially weighted 2D and 3D built environment features using Queen and Rook contiguity-based spatial weight matrices.
Assessed scale sensitivity and spatial interactions using the Optimal Parameter-based Geographical Detector (OPGD) across six algorithms: OLS, KNN, MLP, Random Forest, LightGBM, and XGBoost.
Evaluated model performance against GWR and GWRF, and quantified feature contributions and nonlinear interactions using SHapley Additive exPlanations (SHAP).
Spatial weighting enhanced the explanatory power of major built environment features, with Queen contiguity generating higher q-values than Rook contiguity.
The geographically enhanced XGBoost (GE-XGBoost) model achieved the highest predictive performance with an R² of 0.7067 and lower residual spatial autocorrelation compared to GWR and GWRF.
Two-dimensional and three-dimensional features accounted for 59.55% and 40.45% of total SHAP importance, respectively, with Geo-TPD, Geo-BVD, Geo-PS, and Geo-LUI emerging as the most critical predictors.