Urban street network morphology classification through street-block based graph neural networks and multi-model fusion

Precise categorization of urban street network patterns is essential for urban planning and morphology analysis. Current classification methods typically rely on a single model type, which causes them to struggle in considering topological, geometric, visual, and global features simultaneously, leading to suboptimal results. To address this, we propose a novel fusion model that integrates three submodels: our proposed street-block graph neural network (SBGNet), a convolutional neural network (CNN) using ResNet-34, and a multi-layer perceptron (MLP). SBGNet represents urban street networks as graphs, with nodes as street blocks and geometric features as node attributes, while spatial arrangements serve as edge features. Topological and geometric features are extracted from this street-block representation. The CNN extracts visual features from monochrome images of the street network, and the MLP computes global descriptors using OSMnx and NetworkX to extract global features. Features from SBGNet, CNN, and MLP are concatenated and processed through a downstream MLP for classification. Experiments on a dataset of 6,700 samples from 12 major U.S. cities demonstrate that our fusion model significantly outperforms traditional methods, achieving a PR AUC of 0.88 ± 0.01, with improvements of 5% to 33% over baseline models, validating its effectiveness in classifying complex urban street networks.