Microsoft Building Footprints Download and Formats: Global Building Polygons for GIS

Microsoft Building Footprints is an open dataset of machine-detected building polygons derived from Bing Maps aerial and satellite imagery. Microsoft describes the current release as approximately 1.4 billion building detections from imagery acquired between 2014 and 2024, with additional 2025 imagery contributing new footprints and height estimates in later updates. The primary download workflow uses a Microsoft-maintained link index to retrieve country- and Bing quadkey-partitioned .csv.gz files whose contents are line-delimited GeoJSON. A separate Microsoft Planetary Computer catalog entry exposes a GeoParquet and Delta representation, but users should verify that its snapshot matches the release they need. The dataset is useful for settlement mapping, building-density analysis, humanitarian mapping, exposure modeling, and spatial joins. It is not a cadastral survey, address register, building inventory, or guarantee of current or complete building coverage.

Get Started Browse datasets

Finding every building in a region is a foundational geospatial task, but authoritative building data is often fragmented across cadastral agencies, local government portals, commercial providers, and inconsistent mapping projects. Microsoft Building Footprints offers a broad alternative: machine-detected building polygons derived from imagery and released for download under the CDLA Permissive 2.0 license. Its value is greatest when an analysis needs a consistent building layer across large areas and can tolerate the uncertainty and imagery vintage associated with automated extraction.

How EDK fits

For: GIS analysts, remote-sensing practitioners, humanitarian mapping teams, urban researchers, data engineers, cartographers, and developers building spatial applications that need broad building coverage.

Microsoft Building Footprints is a useful external building layer for understanding one important class of physical-world features: structures visible from aerial or satellite imagery. In an Earth-scale world model, footprints can act as searchable geographic anchors for buildings, settlements, land-use patterns, infrastructure relationships, and change analysis. The dataset becomes more useful when combined with roads, addresses, administrative boundaries, imagery, elevation, environmental layers, permits, or other location data. E2's relevance here is conceptual and analytical: a building footprint is not just a polygon to display, but a geographic object that can be found, compared, filtered, aggregated, measured, and related to surrounding features. Microsoft Building Footprints should remain an external input with its own provenance, coverage gaps, confidence values, imagery dates, and licensing conditions; it should not be treated as an E2-native dataset or as an authoritative source for every building.

•Use Microsoft Building Footprints as an external building layer when searching a region for structures and then relating those polygons to roads, waterways, administrative areas, imagery, public assets, environmental features, or other physical-world data.
•Compute building counts, footprint area, density, nearest-feature distances, settlement summaries, or spatial overlays by city, watershed, service area, or other geographic region, while retaining the source release and imagery-vintage metadata.
•Compare building detections with complementary location data such as roads, construction permits, land cover, surface water, or elevation to identify patterns that require further validation rather than assuming the automated footprints are complete or current.

Publisher

Microsoft / Bing Maps

Primary subject

Machine-detected building footprint polygons

Reported scale

Approximately 1.4 billion building detections in the current repository description

Imagery period

Primarily imagery acquired between 2014 and 2024, with later additions from imagery acquired through 2025

Primary bulk format

Gzip-compressed line-delimited GeoJSON files using a .csv.gz extension

Coordinate reference system

EPSG:4326

Dataset profile

license

CDLA Permissive 2.0, as stated in Microsoft's repository.

geography

Intended worldwide coverage, with substantial variation in completeness and quality by country, region, imagery source, and partition.

resolution

Vector building polygons; no single uniform ground resolution is specified for the complete dataset because source imagery and processing conditions vary.

update cadence

No fixed recurring schedule identified in the official documentation. Microsoft publishes dated repository updates and revises the download index as new or edited coverage becomes available.

temporal coverage

Imagery primarily acquired between 2014 and 2024 according to the current repository description, with later updates adding smaller groups derived from imagery acquired through 2025. Publication and imagery dates are not the same.

Formats

Gzip-compressed line-delimited GeoJSON stored in files with a .csv.gz extension.

GeoJSON after decompression and parsing.

GeoParquet and Delta representation through the Microsoft Planetary Computer catalog.

Coverage and download-index CSV/GeoJSON resources.

Converted GIS outputs such as GeoPackage, PostGIS tables, or partitioned Parquet, depending on the user's workflow.

Access methods

Download the official dataset-links.csv index and select partitions by location and Bing quadkey.

Download selected .csv.gz files from the URLs listed in the index.

Use Microsoft's Python or PowerShell examples to decompress and parse the line-delimited GeoJSON.

Use the Microsoft Planetary Computer catalog and associated STAC/cloud data workflows for the cataloged GeoParquet representation.

Process large partitions incrementally with streaming, GDAL, GeoPandas, Dask, Spark, DuckDB, or a spatial database.

Quality and limits

Machine-detected polygons are not guaranteed to represent every physical structure.

Microsoft documents missing areas caused by imagery filtering, tile exclusions, and low detection probability.

Model performance can vary by geography and urban form; the repository specifically cautions that dense urban areas are not perfect.

Imagery acquisition dates vary, so the data is not a single-date global inventory.

Confidence scores, where present, describe footprint detection and not necessarily height accuracy.

Height attributes are incomplete and may use placeholder values when unavailable.

Important results should be checked against authoritative local data, current imagery, or field observations.

The bulk-download release and Planetary Computer snapshot may have different update histories.

Analysis workflows

Coverage-first download: inspect the coverage resource, identify intersecting partitions, preserve source metadata, and retrieve only the required files.

GIS normalization: decompress the .csv.gz files, parse line-delimited GeoJSON, validate geometries, assign EPSG:4326, and write to GeoPackage, PostGIS, or another analysis-ready format.

Urban metrics: calculate building counts, footprint area, density, size distributions, and spatial concentration by administrative or custom analysis zones.

Spatial joins: relate footprints to roads, addresses, administrative boundaries, waterways, hazards, land cover, elevation, permits, or public assets.

Exposure screening: intersect building polygons with flood, wildfire, earthquake, environmental, or infrastructure-related geographic layers, then validate high-consequence results.

Change screening: compare releases or footprint data with newer imagery and local records, interpreting differences as candidates for review rather than confirmed construction or demolition.

Cloud-scale analysis: use the Planetary Computer's GeoParquet/Delta representation or convert bulk partitions to partitioned Parquet for columnar and distributed processing.

Quality assessment: stratify results by geography, imagery vintage, confidence score, urban density, and availability of height attributes before reporting regional statistics.

Value inside EDK

What Microsoft Building Footprints contains

Microsoft Building Footprints contains polygon geometries representing buildings detected in aerial and satellite imagery. Microsoft describes the footprints as machine-generated from Bing Maps imagery, including imagery from providers such as Maxar, Airbus, IGN France, and Vexcel. The current repository introduction reports approximately 1.4 billion detections from imagery acquired between 2014 and 2024. The repository also records later additions, including new footprints and height estimates derived from imagery acquired in 2025. (github.com)

A typical record is a GeoJSON feature with a building geometry and associated attributes. Depending on the release and source imagery, attributes may include a confidence value and an estimated height. Microsoft notes that confidence scores describe the footprint detection, not the accuracy of the height estimate. Height is also not available for every building; structures without height estimates may use a placeholder value. (github.com)

•Machine-detected building polygons derived from imagery.
•Global distribution with uneven geographic coverage and varying imagery dates.
•Some releases include detection confidence and estimated height attributes.
•The dataset is a building-footprint layer, not a cadastral or address database.

Geography, imagery vintage, and update cadence

The dataset is intended to provide worldwide coverage, but “worldwide” does not mean that every place has complete, uniform, or equally recent data. Microsoft's documentation says imagery older than 2014 was excluded from processing and explains that filtering and tile exclusions can produce squares or other areas with missing data. The repository also warns that the model is imperfect, particularly in dense urban areas, while recall may be better in rural settings. (github.com)

Imagery vintage is heterogeneous. The repository's update history identifies different additions by geography, imagery provider, and acquisition period. For example, the current README records broad historical coverage from 2014–2024 and smaller 2025-derived updates published in 2026. This is not the same as a single global capture date, and the dataset should not be described as a continuously current building inventory. (github.com)

•Geographic completeness varies; inspect the published coverage layer before downloading.
•Imagery dates differ across countries, regions, and individual partitions.
•No fixed global update cadence is identified in the official documentation.
•A release date or repository update date is not necessarily the imagery acquisition date.

Download options and file formats

The primary bulk-download workflow uses Microsoft's dataset-links.csv index. The index identifies partitions by location and Bing quadkey and provides URLs for the compressed data files. The files use a .csv.gz extension, but Microsoft explains that their contents are line-delimited GeoJSON rather than ordinary tabular CSV. After decompression, each line can be parsed as a GeoJSON feature. (github.com)

This partitioning is important for practical GIS work. Instead of downloading one monolithic world file, users can select the countries or quadkey tiles that intersect their area of interest. Microsoft's example Python script reads the link index, filters by location, loads the line-delimited JSON, converts geometry objects into a GeoPandas layer, assigns EPSG:4326, and writes GeoJSON output for use in common GIS software. The repository also provides a PowerShell example for downloading and decompressing selected partitions. (github.com)

•Bulk index: dataset-links.csv.
•Primary files: gzip-compressed line-delimited GeoJSON with a .csv.gz extension.
•Coordinate reference system documented by Microsoft: EPSG:4326. (github.com)
•Cloud representation: GeoParquet and Delta through the Microsoft Planetary Computer catalog.
•Common conversion targets: GeoJSON, GeoPackage, PostGIS, Parquet, or other spatial database formats.

Suitable GIS and spatial-analysis workflows

A defensible workflow begins with coverage inspection. Download the coverage layer or query the link index, identify the partitions intersecting the study area, and preserve the source URLs and retrieval date. Decompress only the required files, parse them as line-delimited JSON, validate geometry, and write them to a format suitable for the next operation. For large areas, stream or process partitions incrementally rather than loading the entire dataset into memory.

For mapping and urban analysis, useful operations include counting detected structures by administrative unit, calculating total footprint area, estimating building density, identifying settlement clusters, and comparing footprints with roads, rivers, surface water, land cover, or elevation. For exposure screening, footprints can be intersected with hazard zones or service areas. For humanitarian or development mapping, they can provide a preliminary structure layer where local building data is unavailable. In each case, report the imagery or release vintage and avoid presenting machine detections as verified ground truth.

•Inspect coverage and select partitions before downloading.
•Retain source URLs, release information, imagery-vintage notes, and retrieval dates.
•Validate geometries and normalize the data into a local spatial format.
•Use confidence values where available, but define thresholds for the specific geography and application.
•Cross-check important results against authoritative local data, imagery, or field observations.

License, strengths, and tradeoffs

Microsoft states that the data is available under the CDLA Permissive 2.0 license. Users should read the license and preserve required notices when redistributing or incorporating the data into another product. The license does not turn the machine-generated footprints into authoritative records, and it does not remove the need to evaluate source quality, attribution, privacy, or downstream data-governance requirements. (github.com)

•Strength: exceptionally broad scale for a building-footprint dataset, making cross-region and international analysis more feasible.
•Strength: partitioned downloads support selective retrieval instead of requiring a single world-scale file.
•Strength: machine-readable GeoJSONL and cloud-oriented GeoParquet options fit both conventional GIS and analytical data-engineering workflows.
•Tradeoff: imagery vintages and model quality vary substantially by geography.
•Tradeoff: gaps, false positives, missed structures, and dense-urban errors can affect counts and area estimates.
•Tradeoff: the data does not provide authoritative addresses, ownership, occupancy, legal building status, or a guaranteed current inventory.
•Tradeoff: the bulk file extension is misleading; users must parse compressed line-delimited GeoJSON rather than ordinary CSV.

Questions teams ask

Is Microsoft Building Footprints a global dataset?

It is designed as a worldwide building-footprint dataset, and Microsoft reports approximately 1.4 billion detections. Coverage is not uniform, however. Some areas are missing because of imagery filtering, tile exclusions, processing conditions, or limited detection quality. Inspect the official coverage information before treating a region as complete. (github.com)

What format are the Microsoft Building Footprints downloads?

The principal downloads are .csv.gz files whose contents are line-delimited GeoJSON. After decompression, each line can be parsed as a GeoJSON feature. Microsoft provides Python and PowerShell examples showing how to download selected partitions, decompress them, assign EPSG:4326, and convert them for common GIS tools. (github.com)

Does the dataset include building heights?

Some releases include estimated building heights, but height coverage is incomplete. Microsoft documents height estimates separately from footprint confidence and notes that structures without height estimates may use a placeholder value. Do not assume every footprint has a reliable height attribute. (github.com)

How often is Microsoft Building Footprints updated?

Microsoft does not state a fixed recurring global update schedule in the cited repository documentation. Instead, it publishes dated update notices and revises the download index when new or edited coverage is released. The current README lists an update dated July 24, 2026, based on 2025 imagery. (github.com)

Can Microsoft Building Footprints be used as an address or cadastral database?

No. The dataset contains imagery-derived building detections, not authoritative addresses, parcel boundaries, ownership records, occupancy information, permits, or legal building classifications. It can support geographic screening and analysis, but important decisions should use authoritative local sources and independent validation.

Keep exploring