EarthDefine Building Footprints 3D: Technical Specification

EarthDefine Building Footprints 3D is a nationwide dataset of US building footprints delivered as two geometry layers: building polygons, and address points carrying the attribute table. Footprints are extracted by EarthDefine's AI classification models from LiDAR and 30cm to 60cm color-infrared aerial imagery, then split using parcel data, addressed from multiple sources, and enriched with LiDAR-measured 3D attributes.

Our dataset distinguishes three units:

• Structure or str_UUID is one contiguous physical building mass, the thing you would trace walking around the outside walls.
• Building or bld_UUID is a structure split at parcel boundaries: a townhome row of eight units on eight parcels is one structure and eight buildings.
• Address or address is an address record inside a building; multiple points per building are normal, and every point falls spatially within its polygon.

How these three units combine

Not all buildings, structures, and address points map one to one. Below are six illustrated cases that describe how they are handled:

1. A single-family home that sits on one parcel: one parcel, one structure, one building, and carries one address.
2. A duplex or small multi-unit building that sits on one parcel: one parcel, one structure, one building with two units, with one address per unit.
3. A large house that sits across two parcels: two parcels, one structure, one building, and carries one address.
4. A townhome row across three parcels with a mix of two units that have one address each, and a third which has two addresses because it includes a rented basement unit: three parcels, still one structure, three buildings, but four addresses (one + one + two).
5. A house and a shed that sit on one parcel: one parcel, two structures, two buildings, each carrying the same address.
6. A house and a separate studio that sit on one parcel: one parcel, two structures, two buildings, two addresses.
7. A multi-storied building in a single parcel: one parcel, one structure, four buildings, each carrying a different address.
8. A large rural parcel with a house, a shed, and two industrial units: one parcel, four structures, four buildings, the house and shed share one address and the industrial units have two have distinct addresses (one + one + one).

This explains why the buildings count and the structures count (see Section 2) diverge, and why per-parcel aggregation needs to be split. The variation is the direct result of buildings not always mapping one-to-one with addresses even when parcels do.

To view a building you are familiar with, try the Buildings API - it will return the str_UUID, bld_UUID and address records for any address or coordinates.

Coverage: contiguous US, Puerto Rico, US Virgin Islands, Hawaii, Mariana Islands, American Samoa, and Alaska (partial), covering a total of 3227 out of the 3,244 counties.

Context: The buildings figure is higher than the contiguous structures figure because of the parcel splits described in Section 1. Both figures are provided to support appropriate comparisons with datasets that do not split structures at parcel boundaries.

Check coverage for a specific address or region here: EarthDefine Buildings API

Footprints derive from three source types:

• USDA NAIP* color-infrared imagery at 30cm to 60cm resolution
• State and local orthoimagery
• LiDAR point clouds

*U.S. Department of Agriculture National Agriculture Imagery Program

Every address point records the most recent source that detected its building (Source) and the date of that source (SourceDate), making the provenance queryable per record rather than stated only at the dataset level.

Production runs as a staged pipeline: imagery is classified by the extraction models; classified output is tiled and regularized into clean polygon geometry; LiDAR enriches regularized polygons with 3D measurements and contributes new and refined footprints where imagery alone underperforms; polygons are split at parcel boundaries; and buildings are addressed from multiple sources. Data is processed at the county level, with each county processed separately. Record counts are reconciled at every stage.

For details about how output is reviewed and corrected before it ships, and how those corrections feed future releases, please refer to Section 6.

We release quarterly. Each release carries model improvements plus any new LiDAR and state or local imagery that has arrived since the last one, so footprints, heights, and parcel splits keep improving between imagery refreshes. The source imagery itself is refreshed on a fixed two-year cycle per state, about half the states each year. In the most recent quarter we added 688,295 structures and extended measured heights to a further 8.66 million.

Identifiers persist across releases: str_UUID and bld_UUID are stable. In cases when the building shifts or changes significantly, new UUIDs are assigned.

A building is a permanent structure with a roof. That definition is applied consistently and decides most edge cases. Here are some common rules:

Captured as a building:
Houses, commercial buildings, and industrial buildings; parking garages; trailer homes and mobile homes sited in trailer parks; gazebos, covered baseball dugouts, and covered parking; glass sunrooms and pool houses; covered pool enclosures (usually attached to the main structure); greenhouses of clear glass or plastic; covered walkways connecting complexes of the same facility (hospitals, campuses); gas-station canopies; water towers; stadium seating with an alternate use beneath it (concessions, restrooms, locker rooms); solar panels serving a secondary structural use such as covered parking; silos and large cisterns; roofed structures at the end of docks; detached sheds and outbuildings, which are delivered as separate buildings.

Not captured as a building:
RVs, campers and towed trailers anywhere on residential property, on roads, or at dealerships; boats and vehicles; shipping containers and semi-truck trailers; tents, canvas or mesh temporary structures; sporting-event seating with no alternate use beneath it, including small stands at school fields; standalone solar arrays; transmission towers and wind turbines; statues and monuments; uncovered docks, pools, trampolines, patios and porches; pool-adjacent cabanas and pergolas unless roofed; construction-material piles and pallets; mesh agricultural covers; road segments; industrial piping.

Geometry rules.
Uncovered courtyards and patios inside a building's boundary are excluded from the footprint, as are adjacent pavement and bare soil. Above-ground decks and balconies are included; ground-level patios are not; the distinction is made from elevation evidence (stairs, furniture, cast shadow). Multi-floor buildings are captured at their base. Where a building leans in the imagery, slight lean is resolved from the rooftop; intense lean is resolved toward the foundation. Buildings partially obscured by tree canopy or shadow are captured on the estimated complete extent of the building rather than only the visible portion, and complex multi-material roofs are resolved as one footprint using shadow and context rather than split by roof color.

There is no minimum-size threshold published for capture.

Our quality assurance runs at three levels.

• Pipeline stage gates. Record counts are reviewed with every new data integration to ensure a net positive data improvement is achieved. Large fluctuations in building counts are investigated when new LiDAR or imagery is introduced.
• Per-state release checks. Before a state ships, a named checklist verifies, among other checks: UUID consistency between point and polygon layers (no extras, no missing); no null address fields; no zero values in ground-elevation fields (LAG, HAG, mean elevation); story-count outlier review; source-date correctness against the LiDAR and imagery inventories; gap investigation over the resolved layer; dissolution of overlapping polygons along tile edges; and spot checks of residential addresses against independent map services.
• Human review. County-level classifier output is reviewed and signed off by a named human data analyst; every county is hand-verified before it ships. Where output needs correction, human data analysts edit annotations against source imagery under the written guideline of section 5, and corrections feed the next model release. The review work is itself audited: annotation output is re-reviewed in structured feedback rounds that classify errors into four categories (omissions, commissions, incomplete annotations, mislabeling), with worked examples issued back to reviewers each round.

Accuracy is measured at different levels relevant to your project:

Height accuracy. Heights are measured from LiDAR point clouds rather than estimated from imagery, and delivered as multiple values per building (maximum, mean, maximum-with-clutter) with ground elevations and roof slopes alongside. We do not publish a single product-wide height-accuracy figure; the published vertical specification of the underlying LiDAR collect is available for any area of interest.

Positional accuracy. Footprints derive from three sources, each with its own published horizontal specification, so we do not quote a single blended figure. The source specification for any area of interest is available on request.

Address confidence. Every address point carries Add_Conf, a low, medium or high confidence score based on the number of corroborating sources and the point's positional relationship to the building.

The polygon layer carries three fields (str_UUID, bld_UUID, Area_SqFt).

The point layer carries the full attribute table, addressing fields, Add_Conf, coordinates, 3D metrics (heights, ground elevation, roof slopes, stories, area, volume), Land Based Classification Standards (LBCS) land-use classification, a parcel identifier, source and source date, and a largest-structure-per-address flag, joined to the polygon layer on either UUID.

Coordinate system is geographic WGS84 (EPSG:4326), vertical datum NAVD88.

3D metrics require a 3D license and are computed from processed LiDAR point-cloud data; the LBCS land-use fields and parcel identifier (ll_uuid) derive from Regrid parcel data.

For the authoritative field-by-field reference, definitions, and types, please refer to the data dictionary below.

To see the fields populated, try Buildings API. It returns the full point record for any address, coordinates, or region of interest.

The data is available two ways: as a bulk delivery, or per lookup through the Buildings API. Bulk deliveries ship as standard GIS formats like ESRI Shapefiles and Geopackages. Depending on your project needs, we can deliver a joined version, only polygons with attributes (no points), repeating polygons for multiple addresses, etc.

The Buildings API returns the same records per lookup, queried by address, coordinate, or region of interest, with JSON responses. The response carries the point schema described in Section 7, so a single address lookup will show you the full attribute set on a building you already know, before you scope a bulk delivery.

Licensing: full commercial licensing, no share-alike obligations. The 3D license tier gates the 3D metrics; LBCS and parcel-identifier fields derive from Regrid parcel data.

Section 5 describes the capture rules used for the standard, nationwide product. Those rules can be adapted for a specific project: custom capture criteria, delivery formats, or coverage scoped to what a project actually needs.

Get in touch to talk through what's possible for your project. Or try the API on your own data first.