A relation-aware heterogeneous graph neural network for the hierarchical regionalization of urban morphology

Urban morphology provides a material context for how cities are perceived, used and planned. Indicator-based classifications have made urban-form analysis increasingly scalable, but most represent blocks either as independent observations or primarily through spatial proximity. They therefore struggle to capture both locally coherent patterns and morphologically similar fabrics that recur in distant parts of a city. This study develops a relation-aware graph workflow that distinguishes spatial adjacency from global morphological similarity and organizes street blocks into an interpretable multi-scale hierarchy. Applied to 4364 blocks in Nanjing using nine interpretable morphological features, the workflow first separates broad differences in development intensity and then progressively differentiates urban fabrics by block grain, building height, internal building morphometrics and vegetation-built-surface context. The resulting hierarchy retains recurring fabrics across the city while making their local spatial organization legible. Downstream analyses show that plan-defined functional areas contain different mixtures of the same fabrics, while morphology-POI-profile correspondence scores and road-context patterns vary across areas and hierarchical levels. These findings show that urban fabric is neither a single-scale partition nor a proxy for land use or transport context. The workflow provides planners and designers with an auditable basis for comparing physical character across scales and for identifying where morphological diagnosis should be combined with functional and infrastructural evidence.