What is a Building Footprint? Data, Sources, Uses

A building footprint is the outline a building occupies on the ground: a polygon traced around its outer walls. That is the building footprint meaning in zoning, where planners measure it to enforce lot coverage rules. In GIS (geographic information systems), each building is one polygon with attributes such as height and floors. Insurers, utilities, and solar developers use building footprint data to locate each structure. Open datasets now map more than two billion buildings, so footprint analysis that once covered a city now covers the planet.

Key takeaways

• Footprint differs from floor area, which sums every floor, and from a roofprint, which includes the roof overhang.
• Most open footprints are traced from imagery by machine learning. A network labels building pixels, and a second step turns them into polygons with straight walls and right angles.
• Overture Maps buildings merges OpenStreetMap, Esri Community Maps, government data, and machine-learning footprints from Microsoft and Google into 2,533,842,612 buildings, each with a stable ID.
• Coverage is uneven. Each footprint shows the building on the date of the imagery or survey behind it.

What is a building footprint?

In zoning and real estate, a footprint is a measured area. Definitions vary by code. Law Insider collects examples such as "the horizontal area as seen in plan, measured from outside of all exterior walls and supporting columns." Some codes leave out decks, patios, and detached garages.

In mapping and GIS, a footprint is a polygon for each building. The polygon can carry attributes such as height, class, and source. Footprints are vector data, the polygon counterpart to the raster imagery they are often traced from.

Building footprint vs floor area vs roofprint

Measure	What it counts	Typical use
Building footprint	Ground area inside the outer walls, counted once	Lot coverage limits, site plans, exposure mapping
Floor area	The area of every floor added together	Density limits such as floor area ratio, leasing, valuation
Roofprint	The roof outline seen from above, including eaves	Footprints traced from satellite and aerial imagery

Does a building footprint include the overhang? A ground footprint does not: it stops at the exterior walls. A roofprint does, and most mapped footprints are roofprints, because imagery from above shows the roof. Overture's building schema defines the geometry as "its footprint or roofprint (if traced from aerial/satellite imagery)." Eaves make a roofprint slightly larger than the walls, so lot coverage computed from mapped footprints runs slightly high.

How building footprints are made

Survey and hand digitizing. City and county GIS teams draw footprints from survey plans, permits, and orthophotos. OpenStreetMap volunteers trace outlines over aerial imagery and tag them with building types. Hand-drawn outlines follow walls closely but cover only places someone mapped.

Lidar. Airborne lidar records the height of every laser return, so a building stands out as a block of points above the ground. Subtracting a bare-earth digital elevation model (the ground alone) from the surface of roofs and treetops gives building height. The block's edge gives an outline.

Machine learning on imagery. Continental datasets come from computer vision models on satellite and aerial imagery. Microsoft describes its pipeline in two stages: semantic segmentation, which labels building pixels with a deep neural network, and polygonization, which converts those pixels into polygons. The raw pixel outline is a staircase. Regularization replaces it with straight walls and right angles. It uses building priors, rules about expected building shape, which Microsoft says are "manually defined and automatically tuned" in its US release.

Footprint quality is scored by matching each predicted polygon to a reference polygon with intersection over union (IoU): the overlap area divided by the combined area of the two shapes. Precision is the share of predicted buildings that match a real one, and recall is the share of real buildings found. Microsoft reports 98.5% precision, 92.4% recall, and an IoU of 0.86 for its US footprints, and 94.4% precision with 70.9% recall for Africa in its global release.

What building footprint data contains

Attribute	Example fields
Geometry	Polygon or multipolygon outline
Identity	Stable ID, source record ID, release version
Height	Height, minimum height, number of floors
Classification	Subtype and class, such as residential, commercial, or house
Roof	Roof shape, material, color, and direction
Source	Dataset, license, update time, and confidence for machine-learning detections
Parts	Building parts for complex structures, linked to the parent building

Coverage of each field varies: 204,366,920 of 2,533,842,612 Overture buildings, about 8%, carry a height (query below). Microsoft trained a network to estimate height above ground from imagery and averages it within each polygon. Google Open Buildings includes a confidence score and a Plus Code for each building, with no information about building type or address.

Building footprint data sources

Dataset	Producer	Coverage	Method	License
Overture Maps buildings	Overture Maps Foundation	Global	Conflation of the sources below and others	Open Database License (ODbL)
Global ML Building Footprints	Microsoft	Global, 1.4 billion buildings	Machine learning (ML) on Bing Maps imagery from 2014 to 2024	Community Data License Agreement (CDLA) Permissive 2.0
US Building Footprints	Microsoft	50 states and the District of Columbia, 129,591,852 buildings	ML on Bing imagery	ODbL
Google Open Buildings v3	Google Research	1.8 billion detections over 58 million km² of Africa, South and Southeast Asia, Latin America, and the Caribbean	ML on 50 cm satellite imagery	CC BY 4.0 or ODbL
USA Structures	Federal Emergency Management Agency (FEMA) and Oak Ridge National Laboratory	US and territories, structures of 450 square feet and larger	ML on imagery, with occupancy type from authoritative sources	Publicly available
OpenStreetMap	Volunteers	Global, uneven	Hand digitizing	ODbL

Cities and counties publish their own, such as New York City's building footprints. USA Structures occupancy classes include single-family residential, restaurant, and hospital. VIDA's combined Google and Microsoft buildings are published as GeoParquet and FlatGeobuf, partitioned by country.

How Overture building footprints are conflated

Overture conflates its sources, merging records of the same building into one, in priority order. OpenStreetMap comes first "so that community knowledge and manual edits are always preferred." Esri Community Maps, Spain's Instituto Geográfico Nacional, and the City of Vancouver come next. Machine-learning roofprints from Microsoft, Google, and a dataset covering East Asian countries fill the gaps.

Two footprints match when their IoU exceeds 0.5. Matched buildings share one Global Entity Reference System (GERS) ID. A GERS ID is a stable identifier across releases. Overture also filters the inputs: it drops buildings in water, buildings of 900 m or taller, and machine-learning footprints of 10 m² or less. Because the release includes OpenStreetMap data, the buildings theme is licensed under ODbL. Overture's guide notes that machine-learning footprints have "lower footprint precision," most of all in the Global South.

This query counts every Overture building in the Havasu catalog by the dataset that supplied its geometry, and how many carry a height.

SELECT CASE WHEN GROUPING(get(FILTER(sources, s -> s.property = ''), 0).dataset) = 1 THEN 'All sources (total)'
            ELSE COALESCE(get(FILTER(sources, s -> s.property = ''), 0).dataset, '(none)') END AS geometry_source,
       COUNT(*) AS buildings,
       COUNT(height) AS with_height
FROM wherobots_open_data.overture_maps_foundation.buildings_building
GROUP BY ROLLUP(get(FILTER(sources, s -> s.property = ''), 0).dataset)
ORDER BY buildings DESC

Geometry source	Buildings	With height
All sources (total)	2,533,842,612	204,366,920
Google Open Buildings	898,130,959	96,506
OpenStreetMap	706,818,187	117,776,403
Microsoft ML Buildings	689,237,083	80,719,707
East Asia building dataset (Zenodo)	212,222,368	0
Esri Community Maps	15,119,688	4,948,247
Instituto Geográfico Nacional (Spain)	12,298,053	809,783
City of Vancouver	16,274	16,274

The three machine-learning sources supply 71.0% of the geometry. 8.1% of buildings carry a height, and Google's footprints almost never do. The full scan of 2.5 billion rows ran in 286 s.

Building footprint examples and uses

• Insurance exposure. Underwriters place each policy on its footprint. Ecopia gives the example of a policy priced from the center point of the parcel. That point can misplace the building relative to a flood zone or an eroding beach. Footprints joined to flood risk data are inputs to catastrophe modeling and probable maximum loss estimates.
• Rooftop solar. A rooftop solar technical potential study by Gagnon and colleagues (2016) at the National Renewable Energy Laboratory (NREL) used lidar for 128 US cities to find suitable roof planes. It put US rooftop potential at 1,118 gigawatts (GW), or 39% of national electricity sales, and found that only 26% of the total rooftop area of small buildings is suitable.
• Urban planning and zoning. Lot coverage, setbacks, and buildable area start from the footprint and the parcel it sits on.
• Disaster response. Analysts compare footprints against post-event imagery to count damaged buildings, a form of change detection.
• Telecom and location intelligence. Carriers count rooftops to plan fiber. A store or campus in POI (points of interest) data is a single point. Joined to a footprint, it gets the building's area.

How to calculate a building footprint

For a rectangular building, multiply the length by the width of the outer walls. For any outline, compute the polygon's area. Footprints stored in longitude and latitude need a spheroidal area function, such as ST_AreaSpheroid in WherobotsDB, which returns square meters.

Lot coverage is total footprint area on a parcel divided by the parcel's area.

Building footprints in Wherobots

Wherobots hosts Overture buildings and building parts as Apache Iceberg tables in the Havasu catalog, along with combined Google and Microsoft Open Buildings. The Salem queries use one box around Salem, Oregon: longitude -123.10 to -122.96 and latitude 44.88 to 45.00, about 11 km by 13 km. Salem sits in Marion County, where Wherobots ran SAM3 (Segment Anything Model 3) on aerial imagery. SAM3 outlines every object that matches a text prompt such as "roofs," giving an independent check on Overture.

The walkthroughs read three tables:

Source	Format in Wherobots	Rows the queries read	Columns the queries read
Overture buildings, overture_maps_foundation.buildings_building	Apache Iceberg table in the Havasu catalog	2,533,842,612 worldwide; 66,200 in the Salem box; 25,352 in the Santa Monica box	sources, height, num_floors, class, geometry, bbox
SAM3 detections, rasterflow_output_samples.marion_county_sam3_vector	Vector table in the Havasu catalog, written by RasterFlow from a 133 GB mosaic of 2022 National Agriculture Imagery Program (NAIP) aerial imagery	102,374 roofs in the Salem box	geometry, label, bbox_score, bbox
Regrid parcel sample, partner_samples.regrid_parcels	Table in the Havasu catalog	9,651 Los Angeles County parcels in the Santa Monica box	parcelnumb, usedesc, county, geometry

The two SAM3 queries below join the roof polygons to the Overture footprints in WherobotsDB SQL.

Sources and attributes in Salem

Each Overture building has a sources array. The entry with an empty property path names the geometry's dataset, and an entry with the path /properties/height names the height's dataset. This query counts buildings, attribute coverage, and area by geometry source:

WITH b AS (
  SELECT id,
         get(FILTER(sources, s -> s.property = ''), 0).dataset AS geom_source,
         get(FILTER(sources, s -> s.property = '/properties/height'), 0).dataset AS height_source,
         height, num_floors, class,
         ST_AreaSpheroid(geometry) AS area_m2
  FROM wherobots_open_data.overture_maps_foundation.buildings_building
  WHERE bbox.xmin >= -123.10 AND bbox.xmax <= -122.96
    AND bbox.ymin >= 44.88 AND bbox.ymax <= 45.00
)
SELECT COALESCE(geom_source, 'All sources') AS geometry_source,
       COUNT(*) AS buildings,
       COUNT(height) AS with_height,
       COUNT_IF(height_source = 'Microsoft ML Buildings') AS height_from_microsoft,
       COUNT_IF(height_source = 'OpenStreetMap') AS height_from_osm,
       COUNT(num_floors) AS with_floors,
       COUNT(class) AS with_class,
       ROUND(PERCENTILE_APPROX(area_m2, 0.5), 1) AS median_m2,
       ROUND(SUM(area_m2) / 1e6, 3) AS total_km2
FROM b
GROUP BY ROLLUP(geom_source)
ORDER BY buildings DESC

Geometry source	Buildings	With height	Height from Microsoft	Height from OSM	With floors	With class	Median area	Total area
All sources	66,200	51,775	49,423	0	336	9,153	196.8 m²	17.263 km²
OpenStreetMap	63,176	49,444	49,423	0	336	9,153	199.1 m²	16.800 km²
Microsoft ML Buildings	3,024	2,331	0	0	0	0	117.1 m²	0.463 km²

In Salem, 95.4% of footprints are OpenStreetMap geometry, and Microsoft fills 3,024 gaps. Height coverage is 78.2%, far above the global 8.1%. Almost all of it is Microsoft's height estimates, which Overture attaches to the OpenStreetMap outlines it keeps. Number of floors is present on 336 buildings (0.5%) and class on 9,153 (13.8%). The query ran in 28 s.

The footprint area distribution

WITH b AS (
  SELECT get(FILTER(sources, s -> s.property = ''), 0).dataset AS geom_source,
         ST_AreaSpheroid(geometry) AS area_m2
  FROM wherobots_open_data.overture_maps_foundation.buildings_building
  WHERE bbox.xmin >= -123.10 AND bbox.xmax <= -122.96
    AND bbox.ymin >= 44.88 AND bbox.ymax <= 45.00
),
binned AS (
  SELECT geom_source, area_m2,
         CASE WHEN area_m2 < 20 THEN 1 WHEN area_m2 < 50 THEN 2 WHEN area_m2 < 100 THEN 3
              WHEN area_m2 < 200 THEN 4 WHEN area_m2 < 500 THEN 5 WHEN area_m2 < 1000 THEN 6
              WHEN area_m2 < 5000 THEN 7 ELSE 8 END AS bin
  FROM b
)
SELECT bin,
       ELEMENT_AT(ARRAY('<20','20-50','50-100','100-200','200-500','500-1,000','1,000-5,000','5,000+'), bin) AS area_m2_range,
       COUNT(*) AS buildings,
       COUNT_IF(geom_source = 'OpenStreetMap') AS osm,
       COUNT_IF(geom_source = 'Microsoft ML Buildings') AS microsoft,
       ROUND(SUM(area_m2) / 1e6, 3) AS total_km2
FROM binned
GROUP BY bin
ORDER BY bin

52,730 of the 66,200 footprints (79.7%) fall between 100 and 500 m², the size of a house or small commercial building. The 176 buildings of 5,000 m² or more are 0.3% of the count and 10.9% of the footprint area. Microsoft's gap-filling footprints skew small: 834 of its 3,024 are under 50 m², against 3,181 of OpenStreetMap's 63,176. The query ran in 14 s.

Checking Overture against SAM3 roof detections

Wherobots used RasterFlow, its tool for running vision models over imagery, to run SAM3 on aerial imagery of Marion County. The input was a 133 GB mosaic of U.S. Department of Agriculture (USDA) NAIP imagery. It wrote the detections to wherobots_open_data.rasterflow_output_samples.marion_county_sam3_vector. The table holds detections in eight labels, including roofs, solar panels, and shipping containers. The completeness check runs both ways: which Overture buildings have a detected roof, and which detected roofs have no Overture building.

For each Overture building in the Salem box, the query below keeps the highest IoU with any SAM3 roof. It then counts buildings that pass Overture's 0.5 match threshold. It is a spatial join on ST_Intersects. EXPLAIN FORMATTED shows WherobotsDB building an R-tree, a spatial index, over the roofs and copying it to every worker (a BroadcastIndexJoin). Each building is then tested only against nearby roofs.

WITH buildings AS (
  SELECT id, get(FILTER(sources, s -> s.property = ''), 0).dataset AS geom_source, geometry
  FROM wherobots_open_data.overture_maps_foundation.buildings_building
  WHERE bbox.xmin >= -123.10 AND bbox.xmax <= -122.96
    AND bbox.ymin >= 44.88 AND bbox.ymax <= 45.00
),
roofs AS (
  SELECT geometry
  FROM wherobots_open_data.rasterflow_output_samples.marion_county_sam3_vector
  WHERE label = 'roofs'
    AND bbox.xmin >= -123.10 AND bbox.xmax <= -122.96
    AND bbox.ymin >= 44.88 AND bbox.ymax <= 45.00
),
best AS (
  SELECT b.id, b.geom_source,
         MAX(ST_Area(ST_Intersection(b.geometry, r.geometry))
             / ST_Area(ST_Union(b.geometry, r.geometry))) AS best_iou,
         COUNT(r.geometry) AS roofs_touching
  FROM buildings b
  LEFT JOIN roofs r ON ST_Intersects(b.geometry, r.geometry)
  GROUP BY b.id, b.geom_source
)
SELECT COALESCE(geom_source, 'All sources') AS geometry_source,
       COUNT(*) AS buildings,
       COUNT_IF(roofs_touching > 0) AS any_roof_overlap,
       COUNT_IF(best_iou >= 0.5) AS matched_iou_50,
       COUNT_IF(roofs_touching >= 2) AS two_or_more_roofs,
       ROUND(PERCENTILE_APPROX(best_iou, 0.5), 3) AS median_best_iou
FROM best
GROUP BY ROLLUP(geom_source)
ORDER BY buildings DESC

Geometry source	Buildings	Overlap a SAM3 roof	Match at IoU 0.5	Touch 2+ roofs	Median best IoU
All sources	66,200	58,669	49,697	13,383	0.674
OpenStreetMap	63,176	56,497	48,164	12,969	0.676
Microsoft ML Buildings	3,024	2,172	1,533	414	0.609

88.6% of Overture buildings overlap at least one SAM3 roof, and 75.1% match one at IoU 0.5. Matching is weaker for Microsoft geometry (50.7%) than for OpenStreetMap geometry (76.2%). Microsoft's footprints are smaller (median 117.1 m² against 199.1 m²), and a small outline loses more IoU for the same edge error. 13,383 buildings touch two or more roofs, where SAM3 returned more than one roof polygon over a single building. The query ran in 167 s.

The reverse check counts SAM3 roofs with no Overture building under them:

WITH buildings AS (
  SELECT geometry
  FROM wherobots_open_data.overture_maps_foundation.buildings_building
  WHERE bbox.xmin >= -123.10 AND bbox.xmax <= -122.96
    AND bbox.ymin >= 44.88 AND bbox.ymax <= 45.00
),
roofs AS (
  SELECT geometry, bbox_score, ST_AreaSpheroid(geometry) AS area_m2
  FROM wherobots_open_data.rasterflow_output_samples.marion_county_sam3_vector
  WHERE label = 'roofs'
    AND bbox.xmin >= -123.10 AND bbox.xmax <= -122.96
    AND bbox.ymin >= 44.88 AND bbox.ymax <= 45.00
),
unmatched AS (
  SELECT r.* FROM roofs r
  LEFT ANTI JOIN buildings b ON ST_Intersects(r.geometry, b.geometry)
)
SELECT 'all roofs' AS roof_set, COUNT(*) AS roofs, ROUND(SUM(area_m2) / 1e6, 3) AS total_km2,
       ROUND(PERCENTILE_APPROX(area_m2, 0.5), 1) AS median_m2, ROUND(AVG(bbox_score), 3) AS mean_score,
       COUNT_IF(area_m2 >= 50) AS roofs_50m2_plus
FROM roofs
UNION ALL
SELECT 'no Overture overlap', COUNT(*), ROUND(SUM(area_m2) / 1e6, 3),
       ROUND(PERCENTILE_APPROX(area_m2, 0.5), 1), ROUND(AVG(bbox_score), 3),
       COUNT_IF(area_m2 >= 50)
FROM unmatched

SAM3 found 102,374 roofs in the box with 15.287 km² of area, against Overture's 66,200 buildings and 17.263 km². 18,010 roofs touch no Overture building, but they total only 0.369 km² and their median area is 0.2 m², so most are slivers. 1,648 of them are 50 m² or larger. These are the candidates for buildings missing from Overture or built after its sources were captured. The query ran in 136 s.

Neither dataset is ground truth. The comparison lists the buildings to review.

Lot coverage from footprints and parcels

This query computes lot coverage for parcels in part of Santa Monica, California. It joins Overture footprints to the Regrid parcel sample in Wherobots open data. For each parcel it clips every footprint to the parcel boundary and merges overlapping footprints before measuring area. A building that crosses a property line counts on both sides, and overlaps count once. The left join keeps parcels with no mapped building, at zero coverage.

WITH parcels AS (
  SELECT parcelnumb, usedesc, geometry,
         ST_AreaSpheroid(geometry) AS parcel_m2
  FROM wherobots_open_data.partner_samples.regrid_parcels
  WHERE county = 'los-angeles'
    AND ST_Intersects(geometry, ST_PolygonFromEnvelope(-118.50, 34.00, -118.47, 34.03))
),
buildings AS (
  SELECT id, geometry
  FROM wherobots_open_data.overture_maps_foundation.buildings_building
  WHERE bbox.xmax >= -118.51 AND bbox.xmin <= -118.46
    AND bbox.ymax >= 33.99 AND bbox.ymin <= 34.04
),
coverage AS (
  SELECT p.parcelnumb, p.usedesc, p.parcel_m2,
         COALESCE(ST_AreaSpheroid(ST_Union_Aggr(ST_Intersection(p.geometry, b.geometry))), 0) AS footprint_m2
  FROM parcels p
  LEFT JOIN buildings b ON ST_Intersects(p.geometry, b.geometry)
  GROUP BY p.parcelnumb, p.usedesc, p.parcel_m2
)
SELECT usedesc,
       COUNT(*) AS parcels,
       ROUND(PERCENTILE_APPROX(footprint_m2 / parcel_m2, 0.5), 2) AS median_lot_coverage
FROM coverage
WHERE parcel_m2 > 0
GROUP BY usedesc
HAVING COUNT(*) >= 50
ORDER BY parcels DESC

Median lot coverage is 0.59 for the 5,477 single-family parcels and for the 1,301 parcels with five or more apartments. It is 0.74 to 0.79 for the three commercial store and retail categories and 0.02 for commercial parking lots. The query ran in 80 s.

At state scale, one Wherobots project scored every building in Colorado, 2,771,126 Overture buildings, for catastrophe risk. In the Marion County SAM3 study, county-wide roof area from SAM3 and Overture agreed within 7%.

Language models were trained on text, documents, databases, and the internet, so they can explain a footprint and cannot count the buildings inside a flood zone. The Model Context Protocol (MCP) is a standard way to connect an AI agent to tools and data. An agent connected to these tables through the Wherobots MCP server can answer that question with a spatial query.

Building footprint research and what changes at scale

From pixels to buildings

Mnih and Hinton applied neural networks to aerial image labeling in Learning to detect roads in high-resolution aerial images (ECCV 2010). Mnih's thesis, Machine Learning for Aerial Image Labeling (University of Toronto, 2013), extended the approach to buildings. The thesis also released the Massachusetts Buildings dataset: 151 aerial images of the Boston area at 1 m, labeled from OpenStreetMap footprints.

Ronneberger, Fischer, and Brox's U-Net (MICCAI 2015) and He, Gkioxari, Dollár, and Girshick's Mask R-CNN (ICCV 2017) came from outside remote sensing and became standard. U-Net labels every pixel. Mask R-CNN gives each object its own mask. Maggiori, Tarabalka, Charpiat, and Alliez asked whether semantic labeling methods generalize to any city with the Inria dataset (IGARSS 2017), which trains on some cities and tests on others. Van Etten, Lindenbaum, and Bacastow described SpaceNet (2018), a corpus of labeled satellite imagery on AWS with public prize competitions. Its first two challenges were building footprint extraction, and the second covered Las Vegas, Paris, Shanghai, and Khartoum.

Microsoft's 2018 release of US footprints used a neural network (ResNet34 with RefineNet upsampling layers), followed by a polygonization step. Sirko and colleagues described Google's approach in Continental-scale building detection from high resolution satellite imagery (2021). A U-Net was trained on 100,000 images with 1.75 million hand-labeled buildings across Africa. Training used mixup (blending pairs of images) and self-training (reusing the model's own labels on unlabeled images), and produced the first 516 million Open Buildings footprints. Zhu and colleagues' GlobalBuildingAtlas (Earth System Science Data, 2025) derived 2.75 billion building polygons and heights from PlanetScope imagery. Shi and colleagues mapped 281 million buildings in East Asia (Journal of Remote Sensing, 2024), the East Asia source in Overture's table above. Promptable models are the latest step: Kirillov and colleagues' Segment Anything (2023), then Carion and colleagues' SAM 3 (2025), which segments every object named in a text prompt.

Polygon regularization

A segmentation mask is a raster, and a GIS needs a polygon with a handful of vertices. The classic tool is Douglas and Peucker's line simplification algorithm (Cartographica, 1973), which drops vertices that sit within a tolerance of the line between their neighbors. It shortens a staircase but leaves it slanted, so footprint pipelines add building priors: dominant orientations, right angles, and parallel walls.

Recent work learns the polygon directly. Li, Wegner, and Lucchi's PolyMapper (ICCV 2019) predicts building outlines and road networks as vertex sequences. Girard, Smirnov, Solomon, and Tarabalka's frame field learning (CVPR 2021) adds a frame field output that encodes edge directions and corners, which guides polygon extraction. Zhao, Persello, and Stein's building outline delineation (ISPRS Journal, 2021) predicts key vertices and connects them in sequence into regularized outlines.

Completeness bias in footprint data

A footprint dataset is complete where someone mapped or where a model was trained well, and both are uneven. Haklay's comparison of OpenStreetMap with Ordnance Survey data (Environment and Planning B, 2010) found that OpenStreetMap covered the most deprived areas of England less well than wealthier ones. Hecht, Kunze, and Hahmann measured the completeness of OpenStreetMap building footprints over space and time (ISPRS IJGI, 2013). Herfort and colleagues estimated completeness for 13,189 urban areas worldwide (Nature Communications, 2023) and found that OpenStreetMap building data is distributed unequally.

Machine-learning footprints fill those gaps unevenly too, as Microsoft's US and Africa recall figures show. Gonzales of Oak Ridge National Laboratory matched Microsoft and Google footprints building by building in urban and rural Ethiopia (GIScience 2023). Gonzales notes that no open dataset has complete global coverage, so analysts must combine overlapping sources or choose between them.

Temporal currency of footprints

A footprint records a building at the date of the imagery it was traced from. Microsoft's US footprints mix 2019 and 2020 imagery for 73,250,745 buildings with older imagery that averages about 2012 for the rest. Overture's conflation ranks sources by priority, so an OpenStreetMap outline last edited years ago outranks a newer machine-learning detection of the same building. New subdivisions, demolitions, and rebuilt homes after a wildfire all appear late.

Van Etten and Hogan's SpaceNet 7 Multi-Temporal Urban Development Challenge (NeurIPS 2020 competition) tracked individual buildings across 24 monthly Planet mosaics at 4 m. Sirko and colleagues' building and road detection from Sentinel-2 (2023) produced the Open Buildings 2.5D Temporal dataset. It holds yearly building presence, counts, and heights from 2016 to 2023 at 4 m effective resolution. Yearly rasters are more current than traced outlines but show less of each building's shape.

What scale changes

A city of 66,200 buildings fits on a laptop. Two billion buildings joined to parcels, hazard rasters, and a second detection set do not. That join needs spatial partitioning, splitting the data by region across machines, and a spatial index, so each building is compared only with nearby candidates. Yu, Wu, and Sarwat introduced that design for Spark in GeoSpark (ACM SIGSPATIAL 2015). Yu, Zhang, and Sarwat described its spatial partitioning, indexing, and query processing in Spatial data management in Apache Spark: the GeoSpark perspective and beyond (GeoInformatica, 2019). GeoSpark became Apache Sedona, the engine under WherobotsDB. That is why the Salem join and the 2,533,842,612-row source count run as plain SQL.

Conclusions

The Overture, Salem, and SAM3 results show what a building footprint in open data is: an outline from one image date, often traced by a model, with few attributes.

1. Treat a mapped footprint as a roofprint unless the source says otherwise. 71.0% of Overture geometry comes from machine-learning sources traced from imagery, so lot coverage computed from it runs slightly high.
2. A footprint's attributes are sparser than its outline. Height is on 8.1% of Overture buildings worldwide and 78.2% in Salem, where floors (0.5%) and class (13.8%) are rarer. An analysis that needs them needs a second source.
3. A footprint dataset can miss buildings, so a building count in a flood zone depends on the dataset as well as the place. One IoU join against independent detections checks completeness. 75.1% of Salem's Overture buildings matched a SAM3 roof at IoU 0.5, and 1,648 unmatched roofs of 50 m² or more are candidates for missing buildings.
4. Machine-learning footprints matched less well in Salem. Microsoft footprints matched at 50.7% against 76.2% for OpenStreetMap, with median areas of 117.1 m² and 199.1 m². Review small buildings before trusting a 0.5 threshold.
5. A footprint shows a building on its imagery date, so record that date. Microsoft's US imagery dates from 2019 and 2020 or averages about 2012, and the SAM3 detections are from 2022. A disagreement can be new construction, a demolition, or a model miss.

Related terms

• Parcel data: land records under each footprint
• POI data: businesses and services inside buildings
• Spatial join: matching footprints to parcels, tracts, and hazards
• Lidar data: elevation points that measure building height
• Catastrophe modeling: estimating losses across insured buildings
• Change detection: finding new and damaged buildings between image dates
• Satellite imagery analysis: computer vision behind machine-learning footprints

Read more from Wherobots

• How well does SAM3 detect building footprints?
• How to score every building in a state for catastrophe risk
• El Niño 2026, atmospheric rivers, and California's burn scars: one SQL engine, two data models
• Tutorial: How to use GERS IDs in Wherobots for geospatial data integration

Query Overture buildings, parcels, and hazard layers with a Wherobots free trial at cloud.wherobots.com.