Temporal instability in pedestrian injury severity: evidence from school and non-school zones using street-level imagery

Built environment features derived from Google Street View (GSV) images offer valuable insights into urban safety, yet their role in pedestrian injury outcomes within school zones remains underexplored. This study analyses four years (2018-2021) of police-reported pedestrian-vehicle crash data from Seoul, South Korea, combined with GSV-based built environment features extracted using deep learning, to examine the annual temporal instability in injury severity between school and non-school zones. A random parameters multinomial logit model with heterogeneity in means was applied to capture unobserved variability, and marginal effects were estimated to assess factor-specific impacts across time and locations. Results indicate that higher proportions of buildings and traffic lights are associated with lower fatality probabilities. Significant temporal instability was observed, with factors such as drivers aged 56-65, evening periods, male drivers, and lower proportions of road and vegetation in school zones increasing the likelihood of severe or fatal injuries, particularly in recent years. Out-of-sample prediction simulations revealed notable differences in factor effects between school and non-school zones, highlighting the effectiveness of targeted safety interventions in school zones. Overall, the findings emphasize the need for dynamic assessments of built environment factors to support evidence-based pedestrian safety policies.