Spatial heterogeneity effect of built environment on traffic safety using geographically weighted atrous convolutions neural network.

The built environment exerts a significant influence on the frequency and severity of traffic accidents. Spatially uniform assumptions on the impacts of built environment factors commonly employed in existing research may lead to inconsistent and contradictory findings. While some studies have investigated spatial heterogeneity using geographically weighted regression models (GWR), these approaches frequently neglect critical aspects including the road network distance between built environment features and the non-linear decay of influence relationships. To address these methodological limitations, this study develops a geographically weighted atrous convolutional neural network regression model (GACNNWR) to more accurately capture the spatial heterogeneity in the impact of built environment factors on traffic safety. Based on empirical data of traffic accidents and built environment from Jinan City, our results demonstrate that the GACNNWR model outperforms traditional analytical methods such as GWR model. Intersection density and bus stop density are identified as having a more substantial impact on traffic accidents compared to population density, land use mix, and destination accessibility. Additionally, population density is shown to exert a bidirectional influence on traffic accidents, while the spatial variability in the effects of land use mix is relatively pronounced. These findings provide important implications for the design of context-sensitive built environments and the formulation of localized traffic safety management strategies aimed at mitigating crash risks.