Diagnostic Analysis of Crosswalk Safety Hazards in Pedestrian Environments: A SHAP-Enhanced Machine Learning Approach With Street-View Imagery

This study investigated the influence of built environment features on crosswalk safety in dense urban settings, with a focus on visual streetscape characteristics extracted from street-view imagery using both semantic and instance segmentation. It used data from 36,750 crosswalks in Seoul, South Korea, to rigorously evaluate multiple machine learning algorithms for predicting pedestrian crash risk. Among the models assessed, the Random forest (RF) demonstrated the highest precision, aligning with the objective of enhancing pedestrian safety through accurate risk identification. The RF model enhanced by SHapley Additive exPlanations (SHAP) achieved strong predictive performance (precision: 0.91), and SHAP analysis identified visual features, particularly sky openness ratio, building coverage, and sidewalk ratio, as influential factors affecting crash risk. A lower sky openness ratio combined with a higher building ratio was associated with increased crash likelihood, whereas greater sidewalk coverage and the presence of traffic control measures, including traffic lights and crosswalk time indicators, mitigated risk. Interaction effects further highlighted the complexity of urban safety, showing that combinations of streetscape and infrastructural elements can amplify or reduce hazards. These results highlight the importance of combining visual and structural data for thorough risk assessment and further the use of interpretable machine learning in urban safety research. The findings imply that to address particular combinations of built environment elements that increase the risk of crosswalk crashes, policy and planning initiatives should concentrate on context-sensitive interventions, particularly by placing bus stops strategically, maintaining tree canopies for visibility, clearing visual clutter, and improving pedestrian infrastructure.