Unveiling pedestrian injury risk factors through integration of urban contexts using multimodal deep learning

This study aimed to identify contributing risk factors for pedestrian injury by integrating socio-spatial and street-level contexts through multimodal deep learning to overcome the limitations of existing studies that only consider one type of data. To investigate how the two contexts assist in describing pedestrian injury risk, six multimodal deep learning models were established by varying the ratio integrating the two contexts. The developed model with the highest performance was interpreted by using two XAI methods: SHAP for socio-spatial context and Grad-CAM for street-level context. The results indicated that the street-level context mainly contributes to the pedestrian injury risk level, assisted by the socio-spatial context, which cannot be captured at the street-level. The three main contributing risk factors were identified through model interpretation: the fragmented sky view due to the locations of high-rise buildings, the placement of crosswalks in areas adjacent to public transits, and interregional sociodemographic disparities. This study provides insight into the use of integrating two different urban contexts to identify pedestrian injury risk factors, which are expected to support improvement strategies that enhance public health.