Microsoft Building Footprints Analysis Workflow: Satellite-Derived Building Polygons for GIS
Microsoft Building Footprints is a large, machine-generated collection of building footprint polygons derived from Bing Maps aerial and satellite imagery. The current GlobalMLBuildingFootprints project reports approximately 1.4 billion detected buildings from imagery captured between 2014 and 2024, with additional releases and edits documented through July 24, 2026. Data is distributed primarily as country- and quadkey-partitioned compressed files containing line-delimited GeoJSON, with related access through Microsoft’s Planetary Computer. This page focuses on how to obtain, normalize, validate, filter, join, and analyze the footprints in a defensible GIS workflow. The dataset is useful for settlement mapping, building-density analysis, infrastructure studies, humanitarian mapping, and creating building layers where authoritative local data is unavailable. It is not a cadastral register, address database, current building-permit inventory, or guaranteed complete representation of every structure.
Microsoft Building Footprints is best understood as a global building-detection layer rather than a conventional survey or continuously maintained map. Microsoft’s GlobalMLBuildingFootprints project uses computer-vision methods to detect building-shaped areas in Bing Maps imagery and polygonize those detections into downloadable geographic features. The result is a large, practical starting point for geographic analysis: analysts can estimate building density, identify settlement expansion, locate structures in data-poor regions, and combine building geometry with roads, waterways, administrative areas, imagery, population estimates, or environmental layers. The central challenge is that the records inherit the date, quality, geography, and visibility limits of the source imagery and machine-learning process.
How EDK fits
For: GIS analysts, remote-sensing practitioners, humanitarian mapping teams, urban researchers, geospatial developers, OpenStreetMap contributors, and data scientists working with large building inventories.
Building footprints are a direct representation of the built physical world: they turn structures that may be difficult to search consistently across countries into geographic objects that can be measured, compared, filtered, and joined to other place-based evidence. That makes Microsoft Building Footprints relevant to the broader world-model problem of finding and computing about physical features. In an E2-oriented workflow, the footprints could serve as one external evidence layer for questions such as which buildings lie near roads, waterways, schools, parks, infrastructure, or reported public-service issues; how building density changes across regions; or where detected structures overlap environmental and administrative zones. The dataset should remain distinguishable from authoritative addresses, local cadastral records, permits, or current imagery. Its value is as a broad, machine-generated building geometry layer that can be evaluated and combined with other geographic sources, not as a complete description of what each building is, who owns it, or whether it still exists.
- •Use building polygons as an external geometry layer for measuring building counts, footprint area, density, proximity, and spatial change across neighborhoods, cities, or broader regions.
- •Combine detected structures with roads, waterways, schools, parks, infrastructure, imagery, administrative boundaries, or public-service records to investigate how physical assets and human activity relate across place.
- •Use confidence values, imagery vintage, local reference data, and spatial validation to rank or qualify building detections before incorporating them into a broader searchable and computable representation of the physical world.
Dataset profile
Formats
Compressed line-delimited GeoJSON / GeoJSONL in files commonly ending .csv.gz
GeoJSON after decompression and conversion
GeoPackage or GeoParquet after analyst-managed ingestion
Planetary Computer catalog assets where available
Access methods
Official GlobalMLBuildingFootprints GitHub repository and release documentation
dataset-links.csv location table for country and quadkey download URLs
Microsoft Planetary Computer catalog and STAC-related access for the cataloged collection
Python, GeoPandas, Shapely, Dask, Spark, DuckDB, or other scripted processing tools
Quality and limits
Detection quality varies by geography, terrain, urban density, imagery source, and imagery date.
Microsoft reports evaluation metrics for sampled regional datasets, not a uniform accuracy guarantee for every location. (github.com)
False positives can include non-building objects or image artifacts, and false negatives can omit buildings.
Coverage gaps can result from imagery exclusion and low-probability detection filtering. (github.com)
Footprints do not provide universal addresses, ownership, occupancy, building use, legal status, construction date, or guaranteed current existence.
Height and confidence fields are not available consistently across all releases or all footprints.
Analysis workflows
Building counts and footprint-area aggregation by administrative unit, grid cell, neighborhood, or service area.
Settlement and urbanization screening by combining footprints with imagery vintages, roads, population surfaces, and land-cover data.
Humanitarian and development mapping in locations without accessible authoritative building layers.
Proximity and accessibility analysis relating buildings to roads, waterways, schools, clinics, parks, utilities, hazards, or public-service records.
Quality-controlled enrichment using confidence thresholds, geometry validation, local reference data, and manual inspection.
Change analysis using separate dated releases or imagery vintages, with careful control for coverage and model-version differences.
Value inside EDK
What Microsoft Building Footprints contains
The dataset contains polygon geometries representing building-shaped features detected from aerial or satellite imagery associated with Bing Maps. Microsoft describes the extraction process as two stages: semantic segmentation identifies pixels likely to represent buildings, and polygonization converts groups of detected pixels into vector footprints. The current global project reports about 1.4 billion buildings derived from imagery captured between 2014 and 2024, although the exact count and geographic composition can change as releases are added or revised. (github.com)
Records may contain more than geometry. Microsoft documents footprint confidence scores between 0 and 1 for releases that include them, where the score summarizes the model’s building-pixel probabilities within the polygon. Some releases also contain estimated building heights in meters; structures without height estimates may use -1 as a placeholder. Height availability is therefore a subset property, not a universal attribute of every footprint. (github.com)
- •Building footprint polygons derived from machine-learning detection of imagery.
- •Optional confidence scores in newer or updated releases.
- •Estimated building heights for selected portions of the coverage.
- •Country and quadkey-based distribution rather than one small worldwide download.
- •EPSG:4326 geographic coordinates. (github.com)
Geography, imagery vintage, and update cadence
Microsoft describes the project as worldwide, but coverage is not uniform. The repository’s coverage maps and release notes show additions and edits arriving by geography, imagery source, and processing batch. Imagery has included sources such as Maxar, Airbus, IGN France, Vexcel, and other Bing Maps imagery. The relevant date for a footprint is therefore the vintage of the underlying image or processing release, not necessarily the date on which a user downloaded the file. (github.com)
There is no fixed recurring update cadence stated in the project documentation. Instead, Microsoft posts discrete additions and edits. The official repository lists updates through July 24, 2026, including new footprints and height estimates derived from 2025 imagery, but that changelog should not be interpreted as a promise of monthly, quarterly, or annual refreshes for every location. Analysts should record the download date, release path, source file, and any available imagery or release metadata. (github.com)
- •Broad international coverage, but not guaranteed wall-to-wall completeness.
- •Imagery dates vary by place and source.
- •Updates are irregular and release-based rather than governed by a published global schedule.
- •Missing tiles, older imagery, and local quality differences must be expected.
Formats and access methods
The primary public distribution uses compressed files with a .csv.gz extension whose contents are line-delimited GeoJSON, also called GeoJSONL or newline-delimited GeoJSON. The extension can be misleading: the file is not a conventional comma-separated table. Microsoft’s example script reads the compressed resource as JSON lines, converts each geometry into a GeoPandas geometry, assigns EPSG:4326, and writes GeoJSON for use in GIS software. (github.com)
The normal download workflow is to obtain dataset-links.csv, select the country or geographic partition of interest, download only the relevant compressed files, decompress them, and stream or batch-process the records. For large areas, use a spatially aware pipeline rather than attempting to open every file directly in a desktop GIS. Python with GeoPandas and Shapely is suitable for smaller subsets; Dask, Spark, DuckDB, or other streaming and parallel tools are more appropriate for very large partitions. Microsoft’s repository explicitly recommends parallel processing or memory-efficient segmentation for large files. (github.com)
- •Compressed line-delimited GeoJSON in .csv.gz files.
- •GeoJSON output for ordinary GIS applications after conversion.
- •Country and quadkey partitioning through dataset-links.csv.
- •Planetary Computer catalog and cloud access for supported collection assets.
- •Streaming and parallel processing recommended for large regions.
A defensible Microsoft Building Footprints analysis workflow
Start by defining the geographic question and the required vintage. If the goal is a current building count, do not treat the newest repository update as proof that every footprint is current; inspect the imagery dates and the release history for the specific area. If the goal is historical change, preserve separate versions and avoid mixing footprints generated from materially different imagery vintages without documenting the difference.
Next, use the link table to select only the relevant country and quadkey partitions. Decompress a small sample first and inspect the schema, geometry validity, coordinate reference system, confidence field, height field, and any release-specific attributes. Convert the sample to a GeoPackage or GeoParquet test layer, then check for invalid polygons, duplicate features, self-intersections, implausible areas, and unexpected detections such as vehicles, containers, bare ground, vegetation, or other image artifacts. Microsoft’s own documentation warns that quality varies and that the data should not be blindly imported into OpenStreetMap. (github.com)
- •Define geography, purpose, and imagery vintage before downloading.
- •Select partitions through the official link table.
- •Inspect schemas and validate geometries on a small sample.
- •Filter by confidence only with geographic validation and sensitivity testing.
- •Project data before calculating area or distance.
- •Join to roads, boundaries, imagery, infrastructure, environmental layers, or other place data.
- •Preserve source release, download date, filters, and validation results.
License, strengths, and tradeoffs
The current global project states that the data is licensed under the Community Data License Agreement Permissive 2.0. Users should read the license and the repository’s current notices before redistribution or incorporation into a product. The older United States-specific repository used a different Open Data Commons Open Database License, so licensing should be checked against the exact release rather than inferred from the dataset’s name. (github.com)
Microsoft Building Footprints is particularly strong when an analysis needs broad building geometry across many countries, a repeatable download structure, and a starting layer for areas without accessible local building data. Its machine-generated scale can make regional or global screening feasible where manual digitizing would be too slow. The tradeoff is that model-derived coverage is not equivalent to a legal or operational inventory. It may omit structures, include false positives, simplify boundaries, or represent buildings from imagery that is several years old.
- •Strength: very large geographic reach and billions of detected structures. (github.com)
- •Strength: machine-readable vector geometry suitable for GIS, spatial joins, and scalable analysis.
- •Strength: optional confidence and height attributes can support ranking and filtering in selected areas. (github.com)
- •Tradeoff: uneven completeness, imagery vintages, and accuracy across locations.
- •Tradeoff: no universal addresses, building names, ownership, occupancy, land use, construction date, or legal status.
- •Tradeoff: large partitioned files require engineering work for efficient processing.
- •Tradeoff: direct-download and Planetary Computer representations may differ by release and should be version-checked.
Questions teams ask
Is Microsoft Building Footprints a current building inventory?
No. It is a machine-generated footprint dataset derived from imagery with varying capture dates. Microsoft documents release updates, but does not publish a fixed global refresh schedule or guarantee that every footprint represents a currently standing building. Use recent local data or imagery when currency matters. (github.com)
What format are the Microsoft building footprint downloads?
The main distribution uses compressed files with a .csv.gz extension containing line-delimited GeoJSON records. They can be streamed or decompressed and converted to GeoJSON, GeoPackage, GeoParquet, or another GIS-friendly format. (github.com)
Does every footprint include a height and confidence score?
No. Confidence scores are documented for newer or updated releases, while estimated heights are available only for selected coverage. Microsoft states that structures without height estimates may contain -1, and older records may use -1 for confidence as a placeholder. (github.com)
Can the data be imported directly into OpenStreetMap?
Microsoft advises against blindly importing the data. Local quality should be checked, and any import should follow OpenStreetMap’s import community guidelines and local consultation practices. (github.com)
What is the best way to analyze a large country or city?
Use dataset-links.csv to select only the relevant partitions, process the GeoJSONL records in a streaming or parallel workflow, validate geometries, reproject before area calculations, and write the result to a spatial format such as GeoParquet or GeoPackage. Desktop GIS is useful for inspection and smaller extracts, but very large areas usually require batch or distributed processing. (github.com)
Keep exploring
google open buildings analysis workflow
A related machine-generated building-footprint workflow with different geographic coverage, attributes, confidence handling, and temporal products.
google open buildings download and formats
Useful comparison for file formats, spatial shards, access methods, and practical GIS ingestion of large building datasets.
landsat collection 2 analysis workflow
Provides imagery and quality-mask context for validating or comparing building detections against satellite observations.
Sources
Microsoft GlobalMLBuildingFootprints repository
Primary project documentation for coverage, release history, license, formats, quality, coordinate system, and limitations.
Open sourceMicrosoft GIS-friendly conversion example
Primary example showing how to read the compressed line-delimited GeoJSON files and convert them for GIS use.
Open sourceMicrosoft Planetary Computer catalog
Official Microsoft catalog entry showing the Microsoft Building Footprints collection and catalog context.
Open sourceMicrosoft USBuildingFootprints repository
Primary documentation for the earlier United States-specific release and its distinct licensing and vintage details.
Open source