Explanatory Models for Regional Fire Risk Using Nonlinear and Spatiotemporal Information

This study examined changes in the goodness of fit of regional fire risk explanatory models associated with the incorporation of nonlinear, spatial, and temporal information across 229 local administrative districts in South Korea from 2019 to 2023. Seven models were analyzed using six fire-risk indicators based on fire incidents, casualties, and property damage normalized by population and area as dependent variables and 12 regional characteristics as independent variables. The results demonstrated that incorporating nonlinear, spatial, and temporal variables consistently improved the goodness of fit, with nonlinear and spatial information yielding greater gains than temporal information. The Random Forest spatiotemporal linear regression model, which incorporated all three types of information, exhibited the highest goodness of fit across all six indicators, with a mean coefficient of determination of 0.756. Furthermore, models incorporating spatial information also showed an overall decrease in the residual Moran's I, indicating reduced spatial dependence in the residuals. These findings underscore that jointly considering nonlinear relationships and spatial heterogeneity is critical for explaining regional fire risk and that integrating nonlinear, spatial, and temporal information can further improve model performance.