Characterizing urban form archetypes through landscape and morphology parameterization and dimensionality reduction

Cities exhibit inherent heterogeneity due to their multidimensional evolution and therefore create complex relationships between their characteristics and spatial forms. Approaches to understand such relationships often employ domain-driven classification schemes with predetermined categories, which may risk oversimplifying the intra-urban complexity and constraining the transferability to cities from different regions. To overcome these challenges, this study develops a data-driven framework with strategic use of unsupervised and supervised methods to extract shared latent structures across cities and reveal the mechanism that drives the cluster formation. The framework utilizes dimensionality reduction on urban descriptor parameters measuring the composition, configuration, and morphology of the landscape, derived from remote sensing datasets. Latent representations of the parameters are segmented based on proximity in the embedded space to identify structural similarities and produce distinct classes of urban forms. Explainability analysis not only shows the effect of latent representation dimensionality in capturing broad urban forms and fine local variations but also reveals a continuous gradient with overlapping transitions where emergent typologies exist and reflect patterns of composition, cohesion, fragmentation, and morphological differences. Overall, this framework enables contextual understanding to emerge from measurable properties without prior assumptions.