From buildings to blocks: Graph contrastive learning for integrated urban architectural morphology classification

Urban block classification is fundamental for understanding urban morphology and guiding renewal strategies, yet traditional methods based on morphological indicators fall short in capturing complex spatial relationships among buildings within blocks. Adopting an urban architecture perspective that treats buildings and blocks as integrated systems, this study proposes GCL-AE (Graph Contrastive Learning with Autoencoder), an unsupervised framework that bridges architectural and urban scales. The framework employs graph attention networks to model inter-building relationships as fully-connected graphs, where attention mechanisms adaptively identify dominant spatial interactions. Through contrastive learning, the model learns discriminative structural embeddings without manual labeling, which are then fused with block morphological indicators through an autoencoder-based strategy. Applied to Nanjing's urban fabric, the method successfully identified 13 distinct morphological types, demonstrating superior performance in distinguishing structurally different but morphologically similar blocks. The discovery that approximately 30% of morphologically relevant information resides in inter-building relationships provides empirical validation for the urban architecture paradigm. Comparative experiments confirm that incorporating spatial relationships significantly enhances classification accuracy. The framework provides data-driven tools for differentiated renewal strategies based on integrated structural-morphological patterns rather than uniform standards.