Microsoft Building Footprints: Global Building Polygons for GIS, Mapping, and Spatial Analysis
Microsoft Building Footprints is an open geospatial dataset of machine-detected building footprints derived from Bing Maps aerial and satellite imagery. The current GlobalMLBuildingFootprints repository describes approximately 1.4 billion detected buildings from imagery acquired between 2014 and 2024, with later refreshes adding footprints and estimated heights from imagery captured through 2025. The data is distributed primarily as country- and Bing quadkey-partitioned line-delimited GeoJSON compressed as .csv.gz files, with a downloadable link index. A separate Microsoft Planetary Computer catalog entry provides an older, spatially partitioned GeoParquet representation in a Delta table. The dataset is useful for settlement mapping, building-density measurement, humanitarian mapping, urban analysis, exposure modeling, and creating base layers where local building data is unavailable. It is not a cadastral database, address register, building inventory, or guaranteed current representation of every structure.
Building footprints are one of the most useful ways to turn imagery into an analyzable representation of the built environment. They let GIS users count structures, estimate built-up area, measure proximity to roads or hazards, identify settlement patterns, and connect the physical shape of buildings to demographic, environmental, infrastructure, and public-service data. Microsoft Building Footprints is designed for that broad geospatial role. It contains machine-detected building polygons created from Bing Maps imagery using semantic segmentation followed by polygonization. The project has expanded through multiple regional releases and now presents a worldwide distribution with more than a billion detected structures. Some releases also include estimated building height and confidence attributes, although coverage of those attributes is not universal. The dataset is particularly valuable when a consistent building layer is needed across many countries and authoritative local data is unavailable or difficult to obtain. Its main challenge is that it represents model detections derived from imagery rather than legal or administrative records. Analysts must therefore treat coverage, positional accuracy, completeness, imagery date, confidence, and licensing as explicit parts of the workflow.
How EDK fits
For: GIS analysts, remote-sensing practitioners, humanitarian mapping teams, urban researchers, infrastructure and risk analysts, open-mapping contributors, data engineers, and developers building location-based or Earth-observation applications.
Microsoft Building Footprints is a direct example of converting observations of the physical world into computable geographic objects. A building polygon can become a unit for spatial search, proximity analysis, aggregation, exposure measurement, change comparison, or joins with roads, waterways, imagery, hazards, population estimates, permits, public services, and infrastructure. For a world-model perspective, the important distinction is between detecting a structure and understanding its broader place in the world. Microsoft provides a building geometry and, for some records, estimated height or confidence. Other questions—such as the building’s address, use, ownership, occupancy, construction status, relationship to a road, or exposure to a flood zone—require additional geographic context. E2’s relevance is therefore strongest when building footprints are treated as one physical feature layer among many, enabling buildings to be found, related to surrounding features, and used as units in geographic computation rather than treated as a complete description of a property.
- •Use building polygons as geographic units for measuring proximity to roads, waterways, public assets, environmental features, hazards, or infrastructure, while preserving the distinction between detected geometry and authoritative local records.
- •Combine building footprints with other physical-world layers to rank or compare areas by building density, built-up area, apparent structure height, access to services, or exposure to environmental conditions.
- •Search for and aggregate structures within a city, region, corridor, or custom boundary, then compare the resulting physical pattern with imagery, land cover, transportation, demographic, or public-data layers.
Dataset profile
Formats
Line-delimited GeoJSON in .csv.gz files
GeoJSON coverage and height-coverage indexes
GeoParquet assets in a Delta table through Microsoft Planetary Computer
STAC metadata and collection/item JSON
GeoPackage, FlatGeobuf, spatial database, or other derived formats after conversion
Access methods
Download selected partitions using the project’s dataset-links.csv index
Query the Microsoft Planetary Computer STAC API by region, quadkey, geometry, or catalog metadata
Read the Planetary Computer Delta/GeoParquet representation with compatible Python, GeoPandas, PyArrow, or DuckDB workflows
Process large compressed partitions using streaming scripts, Dask, Spark, or a spatial database
Use desktop GIS after decompressing and converting selected line-delimited GeoJSON files
Quality and limits
Model detections may omit buildings, include false positives, merge or split structures incorrectly, or misalign polygon edges.
Quality varies by geography, urban density, terrain, imagery source, and imagery date.
Some areas are missing because imagery was excluded or detection probability was low.
Height estimates are available only for part of the footprint collection and are model estimates rather than surveyed measurements.
Confidence values describe footprint detection confidence and do not guarantee completeness, height accuracy, positional accuracy, or legal validity.
The repository and Planetary Computer catalog describe different snapshots, counts, processing dates, and licenses.
The dataset does not provide a universal address, parcel, ownership, occupancy, building-use, permit, or construction-date layer.
Analysis workflows
Define the study area and identify the relevant region and Bing quadkey partitions.
Download only the required partitions, decompress the .csv.gz files, and inspect the actual fields and geometry encoding.
Validate and repair geometries where necessary, then reproject to a suitable local or equal-area CRS for measurement.
Calculate footprint counts, total built area, density, size distributions, or estimated height summaries by grid, boundary, road corridor, watershed, or other analysis unit.
Join footprints with roads, public assets, land cover, imagery, surface-water layers, terrain, hazards, population, or infrastructure datasets.
Filter or stratify by confidence and height availability where those attributes exist, while retaining missing-value flags.
Visually inspect representative samples and compare with current imagery or authoritative local data before operational or regulatory use.
For change analysis, compare explicit releases or imagery vintages and document whether differences reflect construction, edits, detection changes, or updated source imagery.
For OpenStreetMap or humanitarian mapping, use the footprints as review or tasking inputs and follow the relevant community import and validation procedures.
Value inside EDK
What Microsoft Building Footprints contains
The dataset contains polygon geometries representing buildings detected in Bing Maps imagery. Microsoft describes a two-stage extraction process: a neural-network semantic-segmentation step identifies pixels likely to be buildings, and a polygonization step converts those detections into building-shaped polygons. The result is a large vector layer rather than imagery itself.
The repository’s current introduction reports 1.4 billion detected buildings from Bing Maps imagery between 2014 and 2024, including imagery associated with Maxar, Airbus, and IGN France. The project history also records later additions from Vexcel and Maxar imagery captured through 2025. Because the repository has accumulated several releases, counts depend on the distribution date and on whether a user is referring to the current combined release or an older catalog snapshot.
- •Primary feature: building footprint polygon.
- •Possible attributes: estimated height and footprint confidence, depending on release and region.
- •Coordinate reference system documented by the repository: EPSG:4326.
- •Source process: machine-learning segmentation followed by polygonization.
- •The dataset does not inherently provide addresses, building names, ownership, occupancy, use, legal parcel boundaries, or construction dates.
Geography, imagery vintage, and update cadence
The project is intended for worldwide use, but coverage is not uniform. The repository publishes coverage maps and partitions data by region and Bing Maps level-9 quadkey. Microsoft also notes that some locations are missing because imagery was excluded when it was dated before 2014 or had a low probability of detection. The project describes that probability as loosely related to proximity to roads and population centers, which means apparent gaps may occur in rural, remote, or otherwise difficult areas.
Imagery vintage varies by place. The older repository documentation identifies Bing Maps imagery from 2014 through 2021 for the original global distribution, while the current release history records regional updates based on imagery captured as late as 2025. A dataset download date is therefore not the same thing as the observation date of every building. For time-sensitive analysis, inspect the release notes, region, tile, and any available imagery metadata rather than assigning one global date to the entire layer.
- •Global intent with region-dependent completeness.
- •Imagery dates vary substantially across countries and tiles.
- •Updates are irregular rather than daily, monthly, or quarterly on a guaranteed schedule.
- •Record the release date, region, tile, and imagery vintage used in reproducible work.
Formats and access methods
The primary project distribution is organized through a dataset-links.csv index. The downloadable files are partitioned by country and Bing Maps level-9 quadkey and use a .csv.gz extension even though their contents are line-delimited GeoJSON. Microsoft provides a conversion script showing how to decompress the files and make them easier to open in GIS software. The partitioning is important: downloading a whole-country or whole-world collection can be impractical, while selecting only the quadkeys intersecting a study area can make the workflow manageable.
For large-scale processing, the repository specifically points to memory-efficient and parallel approaches such as Spark and Dask. In a desktop GIS, users commonly decompress selected partitions, give them a .geojsonl or .jsonl-style extension where appropriate, validate the geometries, and convert them to GeoPackage, FlatGeobuf, a spatial database, or another format suited to the project.
- •Direct distribution: compressed line-delimited GeoJSON in .csv.gz files.
- •Index: dataset-links.csv with region and quadkey download references.
- •Cloud-native catalog route: GeoParquet in a Delta table through Microsoft Planetary Computer.
- •Programmatic discovery: Planetary Computer STAC API and collection metadata.
- •Large-file processing: use spatial partitioning, streaming, Dask, Spark, or a spatial database.
Suitable GIS and spatial-analysis workflows
A practical workflow begins by defining the study area and selecting only the relevant regions or quadkeys. After downloading the partitions, decompress the files, inspect the schema, confirm the coordinate reference system, and validate polygon geometry. If the analysis involves area, distance, or density, reproject the data into an appropriate local or equal-area coordinate system before calculating measurements; EPSG:4326 coordinates should not be used directly for meter-based area or distance calculations.
Building-density analysis is a natural use. Analysts can calculate footprint area, count polygons, summarize total built area, or normalize those measures by administrative boundaries, grid cells, parcels, watersheds, census geographies, or custom planning areas. The polygons can also be joined to roads to estimate access, to surface-water layers for screening potential exposure, or to elevation and land-cover data for settlement and environmental studies.
- •Clip by region or quadkey before loading data into desktop GIS.
- •Validate geometries and reproject before computing area, distance, or density.
- •Join footprints to boundaries, roads, hazards, imagery, population, or infrastructure layers.
- •Use confidence and height attributes selectively and document missing values.
- •Visually inspect samples and compare with local authoritative data before operational use.
License, strengths, and tradeoffs
The current GlobalMLBuildingFootprints repository states that its data is licensed under CDLA Permissive 2.0. The Microsoft Planetary Computer collection metadata identifies its cataloged GeoParquet/Delta representation as ODbL-1.0. These are different distributions or snapshots, so licensing should be checked at the exact source and version used. A project that redistributes derived or combined data should obtain legal advice where the licensing implications are material.
The dataset’s central strength is scale: it offers a common building-footprint concept across a very large geographic extent. It is also comparatively practical for computation because the data is partitioned spatially and can be processed without loading the entire global collection into memory. Its weaknesses are equally important: completeness and shape accuracy vary by location, imagery dates are heterogeneous, some regions have gaps, and model detections are not authoritative property records.
- •Strength: very large geographic coverage suitable for cross-region analysis.
- •Strength: polygon geometries can be combined with standard vector and raster GIS layers.
- •Strength: quadkey partitioning supports selective downloads and distributed processing.
- •Tradeoff: machine-detected features can contain omissions, false positives, overlaps, or boundary errors.
- •Tradeoff: imagery vintage and update timing vary by region; the layer is not uniformly current.
- •Tradeoff: access routes may represent different snapshots and licenses, including CDLA Permissive 2.0 and ODbL-1.0.
Questions teams ask
Is Microsoft Building Footprints a complete global building database?
No. It is a large machine-detected footprint dataset, not a guaranteed complete inventory. Microsoft documents regional variation, missing areas, false positives, imperfect recall, and imagery-dependent coverage. It should be treated as an analytical base layer and validated for the geography and decision at hand.
What is the difference between the GitHub download and Microsoft Planetary Computer access?
The GitHub-linked distribution uses compressed line-delimited GeoJSON partitions indexed by region and quadkey, while the Planetary Computer catalog provides a spatially partitioned GeoParquet representation in a Delta table and exposes STAC metadata. They may represent different snapshots. The current repository identifies CDLA Permissive 2.0, while the Planetary Computer collection metadata identifies ODbL-1.0 for its cataloged representation.
Does every footprint include height and confidence?
No. Height and confidence coverage is not universal. Microsoft states that structures without estimated height use -1, and older footprints may also use -1 for confidence. These attributes should be checked for availability by release, region, and record rather than assumed to exist everywhere.
Can Microsoft Building Footprints be imported directly into OpenStreetMap?
Microsoft advises against blind imports or overwriting existing community work. The project says quality varies by location and recommends local inspection and compliance with OpenStreetMap import guidelines. A footprint can be useful as a mapping aid, but it is not automatically an authoritative edit.
What should I record for a reproducible analysis?
Record the exact access route, repository or catalog URL, download or processing date, region and quadkey partitions, dataset version if available, imagery or processing vintage, filtering rules, coordinate transformation, geometry-repair steps, and license. This is especially important because the project receives irregular regional updates and different catalog snapshots may not match.
Keep exploring
google open buildings
A closely related global building-footprint dataset that provides a useful comparison of machine-detected structure polygons, coverage, confidence, and intended GIS use.
google open buildings analysis workflow
A workflow-focused companion for filtering, validating, joining, and interpreting large machine-generated building datasets in spatial analysis.
Sources
Microsoft GlobalMLBuildingFootprints repository and README
Primary project documentation for coverage, release history, formats, extraction method, quality metrics, coordinate reference system, limitations, and current repository licensing.
Open sourceMicrosoft Planetary Computer STAC collection: Microsoft Building Footprints
Authoritative catalog metadata for the GeoParquet and Delta-table access route, spatial partitioning, STAC fields, processing date, extent, and catalog-specific license.
Open sourceMicrosoft Planetary Computer dataset documentation
Microsoft’s human-readable overview for the Planetary Computer representation and its programmatic access context.
Open sourceCommunity Data License Agreement: Permissive 2.0
The official license text referenced by the current GlobalMLBuildingFootprints repository.
Open source