Google Open Buildings download and formats: footprints, CSV files, access, and GIS workflows

What Google Open Buildings contains

Google Open Buildings V3 contains approximately 1.8 billion detected building outlines across an inference area of roughly 58 million square kilometers. The detections were generated from high-resolution satellite imagery, described by Google as 50-centimeter imagery in the Earth Engine catalog. For each detected building, the polygon data includes a footprint geometry, the footprint area in square meters, a model confidence score ranging from 0.65 to 1.0, latitude and longitude for the centroid, and a full Plus Code for that centroid. (sites.research.google)

The dataset is intentionally limited in semantic scope. Google’s Earth Engine documentation states that it does not include the building type, street address, or other details beyond the geometry and associated attributes. A polygon therefore indicates a detected structure footprint, not whether the structure is residential, commercial, industrial, temporary, occupied, legally recorded, or connected to a particular address or parcel. (developers.google.com)

• •Polygon geometry is supplied as WKT POLYGON or MULTIPOLYGON text.
• •The confidence score is a model output, not a probability that every individual feature is correct.
• •Plus Codes identify the approximate centroid location and should not be treated as street addresses.
• •The dataset does not provide building height, use classification, ownership, parcel boundaries, or occupancy.

Geography, temporal coverage, and update cadence

Open Buildings V3 covers parts of Africa, South Asia, Southeast Asia, Latin America, and the Caribbean. Google reports the geographic extent as an inference area of approximately 58 million square kilometers, but coverage is not equivalent to complete, uniform coverage of every country or every populated place in those regions. Some areas were not processed, and sensitive areas may have been omitted. The official explorer is the appropriate way to check whether a particular area has detections. (sites.research.google)

The current V3 inference was carried out in May 2023. This is an inference date, not a statement that every image used for every building was captured in May 2023. Google explains that freshness depends on the availability of high-resolution source imagery and that some locations may rely on imagery several years old or may not have been available for processing. (sites.research.google)

• •Current V3 inference date: May 2023.
• •No fixed public refresh cadence is stated for the polygon release.
• •Geographic inclusion and data freshness vary by location.
• •Older V1 and V2 versions remain available through versioned cloud paths according to Google’s documentation.

Download options and file formats

The primary download format is compressed CSV. Polygon and point data are spatially sharded, with one CSV per S2 cell at level 4. Polygon files total approximately 178 GB across the release, with individual files reported as large as 7.8 GB. Point files total approximately 48 GB, with individual files reported as large as 2.1 GB. These sizes make a whole-dataset download a data-engineering project rather than a casual desktop GIS task. (sites.research.google)

Each polygon CSV contains text attributes and WKT geometry. A typical import process is to download only the S2 tiles intersecting the area of interest, decompress them, parse the WKT column, assign the appropriate geographic coordinate reference system, and write the result to GeoParquet, GeoPackage, FlatGeobuf, a spatial database, or another format suited to the next stage of analysis. The official page also provides tile geometry and URL metadata in GeoJSON, which can be used to plan selective downloads.

• •Spatially sharded CSV files, compressed for download.
• •WKT geometry in polygon files; centroid coordinates and Plus Codes in the attributes.
• •GeoJSON metadata for tile geometry and download URLs.
• •Google Earth Engine FeatureCollection: GOOGLE/Research/open-buildings/v3/polygons.
• •Colab-based regional download workflow and direct Google Cloud Storage access.

Suitable GIS and spatial-analysis workflows

A defensible local GIS workflow begins by defining the study area and selecting only the intersecting S2 level-4 tiles. After downloading and decompressing the relevant shards, inspect the CSV schema, parse WKT geometry, validate geometry types, and check for invalid or unusual polygons before loading the data into a spatial database or analysis format. Keep the original confidence, area, Plus Code, source version, and tile identifier so results remain traceable.

Confidence filtering should be treated as an analytical decision. The official score-threshold table supplies tile-specific estimates for reaching target precision levels. An analyst studying broad building prevalence may accept a lower threshold to preserve recall, while an operational map requiring fewer false detections may use a stricter, locally derived threshold. The threshold should be documented because changing it changes both the number and spatial distribution of retained buildings. (sites.research.google)

• •Use tile metadata to limit downloads to the area of interest.
• •Preserve version, inference date, tile, and confidence information in derived products.
• •Use the provided local threshold estimates instead of assuming one universal score cutoff.
• •Join footprints to roads, services, administrative boundaries, hydrology, imagery, or other reference layers only after checking coordinate systems and geometry quality.
• •Treat outputs as modeled observations and screen them against local evidence where accuracy matters.

License, strengths, and tradeoffs

The Google Research Open Buildings page states that the data is available under both the Creative Commons Attribution 4.0 license and the Open Data Commons Open Database License 1.0, allowing users to choose the license under which they use the data. The Google Earth Engine catalog lists CC BY 4.0 for the hosted V3 polygon collection. Users should review the current license notices, attribution requirements, and liability disclaimers before redistributing derived data or incorporating it into a commercial or public operational system. (sites.research.google)

• •Strength: continental-scale building footprints in regions where consistent building data may otherwise be difficult to obtain.
• •Strength: polygon, point, confidence, area, centroid, Plus Code, and tile-threshold information support several levels of GIS analysis.
• •Strength: multiple access paths, including regional downloads, cloud storage, Earth Engine, and an interactive explorer.
• •Tradeoff: coverage is geographically selective and not uniformly complete across the named regions.
• •Tradeoff: the release is based on imagery and inference carried out through May 2023, so it should not be assumed current everywhere.
• •Tradeoff: model errors, imagery offsets, omission of sensitive areas, and local variation in precision require validation and quality-aware filtering.
• •Tradeoff: the very large CSV shards require substantial storage, download bandwidth, parsing time, and spatial ETL infrastructure.
• •Tradeoff: there is no building type, address, parcel, ownership, occupancy, or authoritative legal-status information.