What Is a Building Footprint? The Definition, the Data Sources, and How to Get One

A building footprint is the outline of a building where it meets the ground, recorded as a polygon on a horizontal plane. Trace round the base of the walls, flatten that shape onto a map, and you have one.

Almost none of the footprints you can download are that. They were traced from a photograph taken from above, so they follow the roof. The rest of this page is about that gap, where the open data lives, and what to do when it does not cover you.

The definition, and what sits just outside it

A footprint is a horizontal projection. It carries no height, no floors and no interior. Two buildings with the same footprint can be a bungalow and a tower.

Four things get called a footprint and are not one. Each of them has cost me an afternoon at some point.

Footprint or roofprint

Aerial and satellite cameras see the roof. A model or a human tracing that image draws the roof edge, and the roof edge is not the wall base.

Three things pull them apart. Eaves and overhangs push the roof outward, usually by 30 to 80 cm on a house, which is a few percent of a small building's area. Lean, the sideways displacement of anything tall in an off-nadir image, moves a whole tower sideways by metres. And a car port, a veranda or a canopy has a roof with no walls under it at all.

Overture writes this into its own schema. Its buildings guide says the geometry is "expected to be the most outer footprint, roofprint if traced from satellite/aerial imagery". Google says the same thing about tall buildings in its FAQ: the model "is trained to detect building rooftop rather than base", so on high-rises the polygon sits where the roof appears, not where the building stands.

The error compounds when you multiply by thousands of buildings. Solar potential wants the roof, so a roofprint is the right shape. Ground sealing, flood exposure and setback compliance want the base, and a roofprint quietly overstates every one of them.

Footprint or parcel

A parcel, or cadastral boundary, is a legal object. It is the land somebody owns. The building sits inside it, usually covering a fraction of it. Parcel data comes from a land registry and changes when ownership changes. Footprints come from imagery or survey and change when the building does. Joining the two is a common and useful operation. Treating one as the other gives you a garden counted as a roof.

Footprint or floor area

Gross floor area is the footprint times the number of storeys, roughly, and it is what tax and planning rules usually care about. The building envelope is the 3D skin: footprint plus height plus roof shape. A footprint alone tells you neither. If you need floor area, you need a storey count or a height, and only some datasets carry one.

One building, or several

Terraced housing is the case that breaks the most pipelines. A row of eight houses shares walls, so from above it is one continuous roof with no visible seam. Detectors return one polygon. Cadastres return eight. Neither is wrong, they answer different questions, and you have to decide which one your analysis needs before you count anything.

Courtyards are holes, so a building around a courtyard is a polygon with an interior ring, and any tool that drops interior rings inflates its area. A church with a nave and a separate tower may arrive as a multipart geometry. And the canopies again: a petrol station forecourt roof is a roof over nothing.

What people use footprints for

Solar potential, by roof area and orientation. Flood and wildfire exposure, and the insurance pricing behind it. Population estimation where the census is old or absent. Property tax assessment and detection of unpermitted extensions. Telecoms and network planning, where a building is both an obstacle and a customer. 3D city models, extruded from footprint plus height. Damage assessment after an earthquake or a storm, by comparing two dates.

All of them need the same thing, one polygon per building, with an area you can trust.

Where to get building footprint data

Five sources cover most needs. Everything in this table was read from the publisher's own page on 23 August 2026.

Read the licences before you pick. ODbL is a share-alike licence, so a derived database goes back out on the same terms, and that clause has killed more than one client deliverable. CDLA Permissive 2.0 and CC BY 4.0 only ask for credit. Google publishes under both CC BY 4.0 and ODbL and lets you choose, which it did on purpose so that the OpenStreetMap community and everybody else could each use the data without arguing about compatibility.

The three global sources disagree about the same ground, and the disagreement is the useful part.

If you only want a file for one area, our free building footprint downloader reads all three global sources for a box you draw and hands you GeoJSON with the attribution inside it. No account, and the box caps at 25 km².

When open data is not enough

Four situations, and I hit all four in a normal year.

The area is not covered. Google stops at the edge of the Global South. Microsoft goes where Bing imagery went. OpenStreetMap goes where somebody bothered.

The data is older than the question. Google Open Buildings ran its inference in May 2023. A subdivision finished in 2025 is not in it, and no amount of downloading changes that.

The buildings are new, or gone. Damage assessment and construction monitoring both need two dates, and open data gives you one.

You need a shape nobody publishes. Roof planes for a solar study, or buildings out of your own 5 cm drone flight, where a 30 cm global model has nothing useful to say.

Then you extract your own.

Four ways to get footprints for your own area

Three of the four run inside . Same questions for each: what installs, whether it needs a GPU or an account, whether your imagery leaves your machine, what goes in, what comes out, the cost for real work, and one limit.

The open datasets, downloaded. Install nothing. No GPU, no account. Nothing leaves your machine except the bounding box you ask about. In goes a box, out comes GeoJSON with the licence and the attribution written into the file. Free. The limit: it can only return what the publishers already mapped, so a new estate stays missing. Our downloader.

Our AI Segmentation. QGIS plugin manager, then an account. No GPU, no weights download, no Python environment. Automatic and the cloud save in Semi-Auto send the imagery inside the zone you draw to our servers in Europe. The small Semi-Auto model runs locally and keeps the imagery on your machine. In goes a zone over any raster or basemap plus a word such as building. Out comes a styled GeoPackage, one polygon per object, each with a class, a confidence score and an area. Free for 3 km² of automatic detection a month, then €39 a month for 300 km². The limit: it returns more than you asked for, so vents, skylights and courtyard sheds come back as correctly outlined polygons that are not buildings, and you filter them out by confidence and area. Product page.

Deepness. Plugin manager, then it installs OpenCV and ONNX Runtime on first start. Around twenty minutes when it works. No account. Everything runs on your machine, CPU for small rasters, your own CUDA setup for speed. In goes any raster layer plus an ONNX model file you supply. Out comes a vector or raster layer. Free, Apache 2.0. The limit: it ships no detection ability of its own, and a model trained on 25 cm imagery will read your 5 cm drone flight as noise without warning you. Repository and documentation.

Mapflow, by Geoalert. Plugin manager, then sign in. No GPU. Runs on their servers, so the imagery leaves your machine on every run. In goes an area you draw over one of their imagery sources. Out come vector layers: buildings with optional height, roads, fields, forest. Buildings cost 10 credits per km² at $0.10 a credit, rounded up to whole square kilometres, so a 10 km² town is about $10. The free plan gives 250 credits over at most 25 km², with export blocked. The limit: terraced housing merges into one polygon, which their own screenshots show. Plugin documentation.

For the wider comparison, including the local SAM-based options, see the AI plugin roundup and the QGIS AI hub.

Try it free in QGIS, no card needed

Check the layer before you trust it

Five checks, in the order I run them. They take about fifteen minutes and they have saved me from shipping a wrong number more than once.

The next step, if you now want the polygons rather than the definition, is the how-to: detect building footprints in QGIS automatically, which covers the filtering step that decides whether the layer is usable.