A graph convolutional network model for evaluating potential congestion spots based on local urban built environments

Automatically identifying potential congestion spots in cities has significant practical implications for efficient urban development and management. It requires the ability to examine the relationships between urban built environment features and traffic congestion situations. This article presents a novel and effective approach for achieving the task based on a machine-learning technique and publicly available street-view imagery and point-of-interest (POI) data. The proposed multiple-graph-based convolutional network architecture can: (a) extract essential urban built environment features from street-view imagery and neighboring POIs; (b) model the spatial dependencies between traffic congestion on road networks via graph convolution; and (c) evaluate the risk level of road intersections to emerging congestion situations based on local built environment features. We apply the model to Wuhan in China, and predict the potential congestion spots across the city. The results confirm that the model prediction is highly consistent (about 85.5%) when compared to the ground-truth data based on traffic indices derived from a big taxi GPS trajectory dataset. This research enhances the understanding of traffic congestion situations under various geographic, societal, and economic contexts based on easily accessible road, street-view, and POI datasets at large spatiotemporal scales.