code and data.zip

Segmenting urban street networks is critical in urban planning, transportation studies, and geographic information science. The existing methods, commonly relying on administrative boundaries or network connectivity, cannot fully capture the visual and morphological nuances inherent in street design, resulting in segments that diverge from human intuitive perceptions of neighborhoods and typical street morphological patterns. Spurred by this limitation, this study introduces a novel morphology-based segmentation method for urban street networks using Graph Neural Networks (GNNs). Specifically, the proposed approach represents urban street networks as a rook-graph, where nodes are street blocks and edges are adjacencies. Based on the street-block graph representation, nine morphological features of street blocks are extracted. Furthermore, a model architecture based on a weighted Graph Attention Network (weighted-GAT) with spatial-morphological constraints is designed. Finally, six typical morphological pattern samples are collected through OpenStreetMap. In the proposed method, the segmentation problem is converted to the classification of nodes into these categories using the weighted-GAT model. The results demonstrate that the developed approach outperforms current machine learning and unsupervised segmentation methods, achieving block-level classification accuracy of 0.91±0.01 on the Austin dataset. Quantitative comparisons against unsupervised baselines demonstrate overall superior segment coherence (normalized inequality = 0.0466; average nearest-neighbor distance = 0.0788). This study highlights the potential of weighted-GATs for urban street network segmentation, providing a more effective tool for urban analysis and planning by considering both morphological and spatial street features.