Answers are grounded in this paper and cite their sources.
Building-use information underpins urban building energy modeling, disaster risk assessment, and evidence-based planning, yet semantic labels are missing for most buildings in open databases; the most widely available proxies are OpenStreetMap (OSM) building-type tags, which largely denote type or physical form rather than independently observed current use. Machine learning models that infer such building-type tags from footprint morphology promise scalable label enrichment, but it remains unclear how well such models travel between cities. This exploratory study benchmarks the cross-city transferability of morphology-based OSM building-type-tag inference using an interpretable, openly documented, and computationally lightweight machine learning pipeline; inference of actual current use would require external validation against authoritative records and is treated here as a downstream hypothesis only. From OpenStreetMap extracts of five European cities (Berlin, Amsterdam, Vienna, Barcelona, and Budapest), 3.22 million building footprints were processed and 197,747 labeled buildings were sampled, each described by 22 footprint- and context-level morphometric indicators. Gradient boosting classifiers achieved within-city macro-F1 of 0.831–0.914 (binary residential/non-residential) and 0.595–0.699 (four-class) under spatially blocked cross-validation. On the class-enriched benchmark samples, pairwise transfer retained 85.7% of within-city performance on average, with a minimum of 61.5%; when the same transferred predictions are reweighted to each target’s observed tagged-subset prevalence, mean transferred macro-F1 falls to 0.621 and non-residential precision to 0.09–0.50, and source rankings can reorder, so the class-enriched matrix does not by itself support operational source selection. Transfer degradation was positively associated with the Wasserstein distance between the morphological feature distributions of city pairs (Spearman ρ = 0.66 descriptively; the association is dominated by Amsterdam and largely disappears when Amsterdam pairs are excluded, ρ = 0.13). SHAP analysis indicated that building-size heterogeneity was a recurring predictive signal across the five models, whereas adjacency- and density-related signals varied more across cities. Learning-curve analyses on the class-enriched benchmark indicate that a pooled multi-city model is able to rival locally trained models on comparable class-enriched samples when the target city is morphologically similar to the source pool, but small local samples quickly dominate otherwise; these fractions refer to the class-enriched benchmark samples and do not translate directly into city-level annotation budgets. The proposed pipeline runs end-to-end on a standard computer with exclusively free and open data and software, providing an openly documented benchmark for AI-assisted urban analytics and hypothesis-generating evidence on when morphology-based tag models may be reused across cities.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: