Dataset

Open 3D city models published as CityParquet on object storage, listed from their STAC catalogue.

Open 3D city models, converted to CityParquet and served over HTTPS from Cloudflare R2. Every package is public and needs no credentials: DuckDB, GDAL or any Parquet reader can query it in place, fetching only the columns and row groups a query touches.

The list below is not maintained by hand. Every package carries its own STAC Item (metadata.json, with the city3d extension), the Items are gathered into one collection per source, and the collections into one catalogue:

https://cityparquet.open3d.city/data/
  catalog.json                 -- the catalogue: one child link per collection
  plateau/
    collection.json            -- one Item per municipality
    chiyoda-ku/
      metadata.json            -- the package's STAC Item
      building.parquet         -- … one object table per CityGML module
      transportation.parquet
      …
  3dbag/
    collection.json            -- a collection of one package
    metadata.json
    building.parquet
  aura-zae/
    collection.json
    …
  montreal-terrain/            -- and montreal-2009/, montreal-2016/, montreal-2020/
    collection.json
    …

The page reads that catalogue in your browser and renders what it finds: the city objects, Levels of Detail, city object types and extent of each package, straight from its Item. A collection added to the catalogue appears here without a change to this page.

Querying a package

A package is a directory. Point a reader at one of its tables:

INSTALL httpfs; LOAD httpfs;

SELECT object_type, count(*)
FROM read_parquet('https://cityparquet.open3d.city/data/plateau/chiyoda-ku/building.parquet')
GROUP BY ALL;

Or at the whole package, one table per module, with the cityjson extension — see the DuckDB tutorials.

The catalogue itself is plain STAC, so any STAC client can browse it, for example STAC Browser.

What the conversion keeps

Each collection is converted from its publisher’s own files. A package converted from the City3D STAC catalogue links its source archive (rel=via) and the published STAC Item it derives from (rel=derived_from); the 3DBAG package, converted from the 3DBAG tiles directly, carries neither. The licence and provider shown for a collection are the source’s.

• PLATEAU (Japan, CityGML 2.0). The 62 municipalities of Tokyo and the city of Yokohama, one package each. Coordinates stay in the source CRS, JGD2011 (EPSG), stored longitude-first as GeoParquet requires. The uro Application Domain Extension (PLATEAU’s own attributes), the urf urban-planning features and the dem relief features are not encoded.
• 3DBAG (the Netherlands, CityJSON). Every building in the country as one package of 21.5 million rows, in Amersfoort / RD New + NAP (EPSG).
• Auvergne-Rhône-Alpes economic activity zones (France, CityGML 2.0). LoD3 buildings of the region’s economic activity zones, one package per source file, in RGF93 / CC46 (EPSG). The source covers the zones only, not whole municipalities. One file, 2024_BC1_Vichy_ZAE_2021_LoD3, declares only a vertical CRS (NGF-IGN69 height, EPSG) over Lambert-93 coordinates; its package keeps the declared CRS, so it has no geographic extent and is absent from the map.
• Montréal (Canada), four collections from the Ville de Montréal, all in NAD83(CSRS) / MTM zone 8 (EPSG). The 2015 terrain model as CityGML 2.0 TIN reliefs, one package per borough or municipality; LoD2 buildings with textures as captured in 2016 (six boroughs) and in 2020 (three boroughs), CityGML 2.0, one package per borough; and the LoD2 buildings of the 2009 tile VM05, converted from its CityJSON edition.

Where a source has no LoD0 footprint, the converter derives one from the object’s ground surfaces — its GroundSurface faces, or failing those its downward-facing ones — and leaves LoD0 empty when it finds no acceptable ground. The LoD0 column is plain GeoParquet, so a GIS can read those footprints as ordinary 2D layers.