Is There a Dataset of Building Facade Materials by City? Where the Labels Live

Ranked by how directly they answer “what are the buildings in this city made of”:

The trade-off sits in grain. BMAT is the only source naming a material for a building you can point to on a map; the other three describe aggregates — measured tonnages, census counts or country-level archetype averages.

Each of the 73 cities ships as its own GeoPackage of building footprint polygons, and every polygon carries a facade material from nine classes: brick, concrete, glass, metal, stone, stucco, tile, wood and other. Labels were inferred from 147 million multi-temporal street-view images collected between 2007 and 2025 by a fine-tuned vision-language model reporting an F1-score of 0.91. A materials field keeps each building's label history, so a facade recorded as brick and later as concrete shows the change instead of overwriting it.

Geometry and four attributes — subtype, class, names and sources — come from Overture Maps, so rows arrive with context beyond the bare label. Beside every GeoPackage sit two PNG files: a material-distribution visualisation and a spatial coverage and completeness map for that city.

Scale dictates handling. The package spans 219 files and about 20.6 GB, of which the 73 GeoPackages account for roughly 20.2 GB — Tokyo_material.gpkg alone is about 1.25 GB. Version 2 was posted on figshare on 2026-06-09 (first online 2026-03-10) and the record is static, so nothing newer is committed. Classifier-training workflows on top of these labels are mapped on the data scientists use cases page.

If the question is how much material sits in a city's buildings rather than what each facade looks like, Building Material Stocks and Embodied Carbon — Two Chinese Urban Agglomerations answers it for 38 Chinese cities across 2000–2020: 26 xlsx tables totalling 132,931,068 records tracking 11 materials including lime, cement, steel and glass. City-level sheets hold 8,778 records each for total material stock and for total embodied carbon emissions, and each per-material gridded table holds 5,538,063 records at 500x500 m resolution.

Japan is the standout official count. The Tabulation on Construction Material of Dwellings from the October 2023 Housing and Land Survey publishes pre-tabulated dwelling counts by construction material across Major Metropolitan Areas and Metropolitan Areas — 38 database tables plus 36 downloadable files in that filtered view, out of a related listing of 177 datasets reaching prefecture and municipal level. The survey repeats roughly every five years, and all 38 tables were refreshed on e-Stat on 2025-01-29.

Sanity-check the output against the Chinese dataset's gridded embodied carbon — the only city-scale measured benchmark in the slice. Every layer in that chain is accessible on demand and none requires an integration key.

Browse every cataloged record on the construction-materials data hub.