Global Urban Morphology Clustering

Goal

Do cities cluster into recognisable morphological types from street and building form alone, and how well does that form track continent or cultural sphere?

Methodology

Started from a global cities list (>10k population, capped at 2,000 per continent), classified by continent and cultural sphere (via a merged-countries boundary layer, with nearest-neighbour imputation for unclassified cities).

For each city, pulled an 800m-radius sample of buildings and drive-network streets from OSM (OSMnx), then derived 18 morphometrics with momepy: tessellation cell/building area ratio, street length and linearity, street profile width/deviation/openness, building perimeter-wall length, building adjacency, and nearest-neighbour distance - each as median and std across the sample. Cities with under 7% building coverage in the sample area were dropped as likely OSM data gaps. This step is one Overpass query per city per geometry type across ~10k candidate cities - run in batches per continent on Colab over several sessions, checkpointing every 50 cities.

6,577 cities had complete morphometrics after filtering. Standardised these, then:

• Fit a multinomial logistic regression per city’s 18-variable morphometric profile against continent and against cultural sphere, to see how separable those classes are on form alone.
• Ran PCA (18 components, first 12 explain ~97% of variance) and KMeans (k=3, chosen via elbow plot) on the components to get morphological clusters independent of geography.

Results

Morphometrics alone predict continent at 64.5% accuracy and cultural sphere at 57.2% (both in-sample, no held-out split - read as how separable the classes are on this feature set, not a generalisation claim). Europe and North America are the most distinct (74–79% recall); Oceania is barely separable from the rest (5% recall, small sample). By cultural sphere, Hindu and Southeast Asian spheres are entirely absorbed into other classes, while European, Western and Sub-Saharan African spheres hold together (65–78% recall).

KMeans on the PCA components settles on 3 clusters at a clear elbow. Mapped globally, the clusters track broad urban-development patterns more than strict continental boundaries: one cluster dominates North America, Australia and much of Europe (low building adjacency, wide streets); a second concentrates in South/Southeast Asia and parts of Africa (dense, high adjacency, narrow streets); a third is more mixed, prevalent across Latin America and Eastern Europe.