Overture's 2.53 billion building footprints: keep the ID, treat the polygon as a draft

The Overture Maps Foundation's 2026-10-21 release lists 2,533,842,612 open building footprints and 4,486,107 building parts worldwide. The buildings count is identical to September's. That is not evidence of a dataset that holds its shape: the foundation re-promoted the September buildings data, so every per-source count shows 0.00% change because nothing new was published. Whether a building keeps the same identifier from one release to the next is a separate question, and only the registry changelog can answer it.

That distinction shapes how to use the data. The polygon is a draft that can be replaced by a better source at any release. The identifier, where the project commits to it, is the part worth building on.

Seven sources supply the footprints, and the polygon inherits its source's quality

The October table lists Google Open Buildings at 898,130,959 footprints, OpenStreetMap at 706,818,187 and Microsoft ML Buildings at 689,237,083. Those three add up to 2,294,186,229, about 2.29 billion. The other four rows total 239,656,383, about 9.5% of the headline figure: a dataset hosted on Zenodo and cited by DOI (212,222,368), Esri Community Maps (15,119,688), Spain's national mapping agency, the Instituto Geográfico Nacional (12,298,053), and the City of Vancouver (16,274). I did not read the Zenodo dataset's own documentation, so I cannot say how it was built. The seven rows sum exactly to the headline total.

When Overture first published the theme in December 2023, with 2,339,547,325 buildings, it described a precedence order with four named providers: OpenStreetMap first, then Esri Community Maps, then Google's higher-precision footprints, then Microsoft's machine-learning footprints, then Google's lower-precision ones. The Google tiers were split by a 90% precision confidence threshold that varies by S2 cell. The October table lists seven sources against those four, and I did not trace when the other three entered. I also looked for a current statement of the order in the October 2026 release notes and in the GERS documentation and found none. Treat the 2023 order as history, not as a description of today's data.

The practical consequence is that neighbouring buildings in one city can come from different sources with different geometric quality, and the polygon does not say which. Footprint area, a common proxy for floor area, then carries an error that depends on the source and cannot be seen in the shape.

A concrete test: for each city you value, draw a sample of a few hundred buildings per source, using the bridge files to find each building's source record. Measure their footprint area against a municipal or cadastral reference. Report the median and the 90th-percentile absolute percentage error separately for each source, and keep the sample out of any model training. A single pooled error figure would hide the source that is wrong.

A stable ID can turn a snapshot into a timeline, if the building is in the registry

Overture's answer to the matching problem is GERS, the Global Entity Reference System. A GERS ID is a 128-bit identifier that Overture says it keeps stable across releases. The project publishes a registry that records when each ID was first seen, last seen and last changed, plus bridge files that map GERS IDs to the record IDs in the source datasets.

The documentation sets limits. Its page states that presence in the registry is what makes an identifier a GERS ID. Features in the base theme and features of type building part carry deterministic UUIDs that look like GERS IDs but come from a single source, are not matched across sources, and carry no stability commitment. The registry currently covers buildings, divisions, division areas, division boundaries, places, road segments and connectors. For a pipeline, the rule follows: check each ID against the registry before treating it as a key, and mark anything missing as provisional.

Addresses show why the link needs care. The October release counts 475,263,892 addresses, up 0.23% from 474,186,531 in the previous release. One building often carries several addresses, and one address can span several buildings. Joining on address alone gives a many-to-many result that needs a confidence score.

Permit registers show the scale of the job. In Turin, our count of the city's public building-permit register shows between 1,815 and 2,135 applications filed in each month from March to July 2026, with 2,135 in July. We have not matched those applications to footprints here, so this is a measure of workload, not a matching result. Each application has to be resolved to a building once, whatever the register uses to name the place, and the result stored with the release in which it was matched. With a persistent ID, a second renovation of the same block adds a row to that building's history. Without one, the matching starts again every month.

Practical rules for the pipeline:

• Store the GERS ID with the release version in which you matched it.
• At each release, reconcile against the registry and changelog rather than re-matching everything.
• Match by overlap and address together, and keep the match score.

Open attributes are the weak point, and a European study measures it

Footprints say where a building is. Valuation needs height, age, type and floor count. A study published on 4 June 2026 in the ISPRS International Journal of Geo-Information, with a record in the European Commission's Joint Research Centre repository, compares six pan-European open building datasets across the 27 EU Member States: OpenStreetMap, EUBUCCO, Microsoft Global ML Building Footprints, Overture Maps, GHS-OBAT and the Digital Building Stock Model. It processes around 1.250 billion footprints and checks five attributes: height, typology, building age, floors and material.

The result is a trade-off. The remote-sensing products, GHS-OBAT and the Digital Building Stock Model, are the most complete for height, typology and age, but their values are coarse or aggregated. OpenStreetMap and Overture have richer schemas, but their completeness is low and uneven across countries and degrees of urbanisation. The authors also point out that high completeness can mask low information, because placeholder or aggregated values count as filled fields. They conclude that no single dataset is best on every criterion.

For a lender or asset manager, this means a fill rate is the wrong metric. A field that is 95% populated with a generic class tells a model almost nothing, so score informativeness: how often the value differs from the default and how often it matches a reference. For a developer reading a neighbourhood, open footprints are good for counting and locating buildings and weak for estimating age or height, where official records still do the work.

What to watch

Overture released on 22 July, 19 August, 23 September and 21 October 2026. If the monthly cadence holds, the November release will show whether the buildings theme changes again after October's re-promotion, and whether the per-source table gains or loses a provider.

The sharper test is the GERS changelog. If the number of changed or retired building IDs between releases stays small, the case for long-lived keys holds. If it rises, polygons are being replaced faster than identifiers can follow, and the case weakens.

For attribute quality, look for follow-up work on accuracy against official records, not only completeness. If open attributes prove accurate where they exist, the weak point shrinks. If they prove accurate only in a few well-mapped cities, open footprints remain a location layer.

For permit data such as Turin's, the useful check is a per-source error table for the matched buildings, published beside the match rate. A match rate without it says how many applications found a polygon, not whether the polygon was any good.