Classifying arbitrary urban morphotypes: a typology-aligned, open-world approach

Title: Classifying arbitrary urban morphotypes: a typology-aligned, open-world approach

pic-urban-morphotype-classification

Abstract

Analyzing urban morphology across different geographical contexts can reveal recurring patterns of urbanization and support the critical evaluation of urban design practices. Existing machine-learning approaches generally follow either supervised classification or unsupervised clustering. Supervised methods retain established morphological knowledge, but force all samples into predefined classes, whereas unsupervised methods can identify new patterns but often lack clear connections to existing urban-morphological concepts. To address this challenge, this study introduces generalized category discovery into urban morphology and establishes an open-world evaluation paradigm for jointly recognizing reference morphotypes and discovering categories withheld from training. Based on this paradigm, we propose AnyMor-VLM, a novel urban-morphotype-aligned, scalable framework that leverages vision-language models capable of identifying any geometry-based urban morphotypes. The AnyMor-VLM integrates techniques of language-guided representation learning, supervised learning, self-supervised contrastive learning, and cross-modal alignment learning. Comprehensive experiments show that the AnyMor-VLM achieves the strongest overall classification performance on the proposed benchmark and improves the recognition of established reference classes and novel categories withheld from training. Two proof-of-concept applications further illustrate the use of the urban-area delineation and the identification of informal-settlement-associated fabrics. These findings provide evidence that open-world techniques can bridge knowledge-driven classification and data-driven pattern discovery in computational urban morphology.