Building Footprint

Building footprints are based on the "Building" AI Layer. The Building layer is currently designed to detect non-habitable roofs only. It detects any roof of a permanent structure such as a house, unit, commercial building, garage, large garden shed or carport that is designed to keep the weather out. The current definition explicitly excludes the top of a building that is designed for regular use by people, such as a rooftop basketball court or car park. If a building is partly under construction, the completed part of the roof will be detected as building.

The vectorised Building outlines exported from AI Parcels take the Roof AI Layer and performs post-processing to impose regular shapes, such as straight edges and 90 degree corners. Where the building is determined to be of irregular shape, such as for round and highly complex roofs, the algorithm reverts to a natural roof contour. In Gen 3, the number of these "reversions" was drastically reduced to be very rare, and the vectorisation algorithm was enhanced to deal much more elegantly with:

  • Buildings with multiple axes (i.e. having one or more "bends" in them).
  • Angles that are not 90 degree corners.
  • Courtyards or other interior cutout spaces.,-97.6476033,22.00z,0d/V/20190607

Gen 1 Building AI Layer: Typical residential buildings, including successful capture of building under tree cover.,115.8980274,19.00z,0d/V/20191208

AI Parcel exports of Gen 1 Building Footprints. Note the high confidence buildings are shown with yellow fills, lower confidence buildings (such as garden sheds, and a poorly captured commercial building) in shades of green.

Gen 3 Building features pulled from the AI Feature API, and rendered in a GIS application with auto-generated labels from the metadata (confidence that each feature is a real building as a percentage, and area in metres squared) .


Fidelity Score

gen4-lightning_bolt-1.0 introduced the fidelity score on building footprint for the first time. Where the confidence is the probability that the building identified in the location actually exists, the fidelity score measures the quality of the vectorised building footprint polygon. One of the best uses for fidelity score is to programmatically identify buildings that require a different workflow or treatment due to CBD/downtown conditions (e.g. falling back to a previous digitised outline, or flagging for manual inspection). Using a proprietary machine learning model that compares the final outline to the prediction rasters, we can accurately determine whether the polygon matches exceptionally well (1.0) or very poorly (0.0), or somewhere between.

The list below provides guidance on how to interpret a fidelity score:

  • 0.9-1.0 - The vast majority of buildings have a fidelity score of 0.9 or greater, indicating an excellent match.
  • 0.7-0.9 - Fidelity scores between 0.7 and 0.9 will still be good enough for many applications, but may be degraded slightly due to challenging conditions such as extensive tree overhang, shadowing or architectural complexity.
  • 0.5-0.7 - Scores between 0.5 and 0.7 may be useful in situations where the shape is less critical, but should be handled with care.
  • 0.0-0.5 - The lowest scoring buildings often have low confidence as well, such as a heavily shadowed garden shed with overhanging branches. If the confidence is high but the fidelity is low, it is most likely a very large, complex commercial or industrial building that exhibits very complex architecture, or extensive parallax error (building lean due to height).

Fidelity scores in Omaha (left) and downtown New York (right). 0-1 scale mapped to red-yellow-green

Characteristics and Recommended Use

Commercial and Industrial Buildings

The evolution of Residential Building Footprints

The reason the pack was initially called "Residential Building Footprints" is that Gen 1 data performance in commercial and industrial buildings (C&I) such as warehouses and office buildings, and CBD (downtown) was reduced compared to the excellent quality with residential buildings. Approximately 80% of C&I buildings had good quality footprints, but the Gen 1 system found particularly large, featureless roofs challenging. This was resolved with the Gen 2 release, and now C&I buildings are detected with quality similar to their residential counterparts from Gen 2 onwards. From Gen 3 onwards, the improved vectorisation algorithm combined with ignoring parcel boundaries has made the complex world of large C&I buildings (where parcel boundaries are often wrong, out of date, or very large) on par with our residential performance. Combined with 3D mesh-related data (height and storeys) in the Building Characteristics AI Pack, the release of Gen 3 gave us the results we needed to remove the "residential" tag from this AI Pack.

Tall Skyscrapers and complex buildings at the heart of many cities are very challenging, due to building lean (parallax) and complex multi-level structures where the reasonable extent of a footprint is ambiguous even to a human looking at nadir imagery. CBD/Downtown areas should be checked for reasonable raster performance in MapBrowser to confirm whether they meet the quality needs of your particular use case. Please discuss with your account manager if downtown areas are critical to your use case, as we have a range of options to assist you. Using the DSM directly is an alternative that may be useful for situations such as RF propagation modelling amongst tall buildings. You can find out about exporting DSM from MapBrowser here: Export 3D.

The "confidence" value for each building is the best measure to determine whether the building is captured well, or whether the model was uncertain. Viewing the Building AI Layer in your area of interest can give a sense as to whether the building footprints are likely to be of good quality in an area, or direct calculations using the DSM are a better option for you.

Tree Overhang

The building footprints are particularly good at identifying corners and edges of buildings that are obscured by trees (whether or not leaves are present). The AI Layer does a good job in the majority of these situations (except when the tree overhang is so large that it is difficult for a human to estimate how far underneath the roof extends), and enhancements in the Building Footprint vectorisation process typically improve it further. We do not capture extreme overhang, where it is difficult for a human to estimate the extent of the building under the tree.

If the level of tree overhang is of interest, the actual shapes of overhanging trees (and their areas) are available in the Roof Characteristics AI Pack.

Vacant Lots

In the AI Parcel export CSV file, the "present" field denotes whether any buildings are in the parcel; any property parcel with building set to "N" indicates that it is a vacant lot. This is a good way to reliably identify vacant lots, or ones where the construction stage has not yet reached the adding of a roof. When there is an error on this flag (i.e. checking visually in MapBrowser you can see that the Y/N designation is incorrect), the incorrect cases are made up almost entirely of parcel boundary errors, where the building actually lies in the neighbouring property. 

Parcel Boundary Cuts

When an AI Parcel is processed, the property parcel boundary is used to identify objects within the parcel. Objects that protrude from the parcel (due to poor parcel alignment, or changed parcel boundaries) are cut to the boundary in Gen 1 and Gen 2. While this is often the desired outcome, it can result in unintended behaviour. From Gen 3 onwards, this is no longer an issue.

Shape / Object Quality - filtering out poor building outlines through confidence scores

The "confidence" attribute is the best way to filter out cases of potentially poor quality outlines. Almost all buildings have a confidence of greater than 90% (many are in fact at 100% confidence). Depending on how you need to use the data, you may choose to discard the couple of percent of buildings with confidence lower than 90%. This is by far the easiest way to filter out the commercial/industrial buildings that are captured poorly in Gen 1, or the other false positives that are mentioned below.

Roof classification typically include

House roofShed/Outbuilding roofCommercial roof

Partially obscured roofStadium roof

Roof classification typically exclude

Rooftop carparkAwningMobile home

Shade clothShipping containerBalcony/Terraces

Version History

For a precise changelog, refer to the AI Generations pages. This is a convenient summary to describe changes to this AI Pack over time:

Gen 1

"Residential Building Footprints", with great quality for residential, but lesser quality for C&I and downtown/CBD.

Gen 2

Major improvements in AI Layer raster for C&I buildings, and some improvement in downtown/CBD.

Gen 3

Renamed pack from "Residential Building Footprints" to Building Footprints". New vectorisation algorithm, and removal of parcel boundary related caveats and artefacts.

Gen 4

gen4-lightning_bolt-1.0 - Addition of "fidelity score" to building footprints to measure the quality of the vectorised outline.