Detection and clustering of urban form types with machine learning: insights into Thessaloniki's urban planning and evolution

Abstract Advances in Machine learning open new frontiers in the systematic analysis of urban form. The study presents a scalable and interpretable framework that derives an urban-form typology by performing unsupervised clustering of 17 multi-scale morphological indicators encoded at the cadastral plot scale. The method adds positional information with the Getis-Ord Gi* spatial autocorrelation metric to encourage spatially homogeneous clusters. The study employs a combination of UMAP for non-linear dimensionality reduction and BIRCH for scalable clustering. Caveats of using the plot as a spatial unit are mitigated via filtering, tessellation and buffering. Applied to the metropolitan area of Thessaloniki, Greece, the framework identifies 14 urban form types organized into five families with similar characteristics. The resulting typology reveals, in a Conzenian fashion, patterns of urban development rooted in the city’s modern history. Results are validated quantitatively with performance metrics and qualitatively using aerial imagery and established knowledge of Thessaloniki’s planning and evolution.