This AI Pack can be purchased as an add-on to the Residential Building Footprints pack. It adds AI Layers and AI Parcel outputs related to:
- The dominant material of the roof (shingle, tile, metal or other)
- The presence of various structural shapes on the roof (hip, gable, flat and turret presence)
- Trees overhanging the building.
3D Roof Characteristics
In Gen 3 data, a new attribute was introduced - Dominant Roof Pitch. This is calculated from our 3D data using a similar approach to the 3D attributes in AI Pack: Building Characteristics. See the description there for when and how the 3D related attributes are made available.
Definition - Dominant Roof Pitch
Measured in degrees from horizontal, this is the pitch of the largest area of the roof. It is calculated by assessing the surface normals of the roof area, and identifying the most prevalent surface normal. Where multiple pitches have been used in a roof, this will NOT provide an average - it will be (approximately) the most commonly appearing pitch by area.
Roof Material includes a group of three AI Layers:
It also includes a "dominant roof material" category in AI Parcel exports (both as a column in the spreadsheet and as an attribute associated with a building footprint polygon in the GeoPackage file), which can take the value of "Tile", "Metal", "Shingle" or "Other".
The Dominant roof material is the material taking up the largest horizontal area of the roof (as long as it covers at least 50% of the total area). So a mostly tile roof with a metal extension would be denoted as tile. If there is no clear dominant roof material, it is classified as "Other".
Roof tiles are defined by their structure and thickness: clearly laid out lines with shadows on each row, and ridge caps. As such, this includes any tile materials (primarily terracotta or concrete), regardless of colour.
|Terracotta tiles||Concrete tiles|
Roof comprised of metal sheets, typically zinc, copper and steel alloys, regardless of whether they are smooth, corrugated, or with distinct ridges. As can be seen in the statistical performance figures, metal is the most challenging roof type to detect correctly. Some metal is clear (with corrugated ridges). However, particularly in the US, metal roofs are often flat and relatively uniform in colour, and can easily be confused with other types of roof that are similarly featureless or contain ridges that may look like corrugations.
|Corrugated metal roofs|
Roof comprised of a patchwork of shingles. Roof shingles are a roof covering consisting of individual overlapping elements. These elements are typically flat, rectangular shapes laid in courses from the bottom edge of the roof up, with each successive course overlapping the joints below. Visually, they have less distinctive straight ridge lines and clear rows compared to tile roofs . We do not distinguish between wood, metal, laminated asphalt, or any of the numerous shingle materials nor colours.
Any roof material not described by the materials defined above. As we add more roof materials in future releases, "Other Roof" will reduce in scope to exclude them.
|Concrete roofs||Asphalt/Fibreglass roofs||Very poor condition roofs with|
Characteristics and Recommended Use
The dominant roof material is provided to avoid confusion in the very common situation where a shingle or tile roof has a small metal extension. In some situations (due to lighting conditions for example) it can be very difficult to tell the difference between tile and shingle, or between metal and other fairly featureless roof types. This shows up as less bright areas in the AI Layer viewer. The dominant roof material simplifies this information by making the best choice of overall roof material given the information available.
Tiles are much less common in certain parts of the US, which means that the majority of roofs identified as tile are lower confidence. This is because a certain percentage of shingle roofs will always be mistaken for tile, and an absence of actual tile roofs means that these errors form a much larger percentage of those detected as tile.
Shingles are almost never used in Australia whereas they are the most common roof type in most parts of the US. This means that the majority of roofs in Australia identified as shingle are lower confidence. This is because a certain percentage of tile roofs will always be mistaken for shingle, and an absence of actual shingle roofs means that these errors form a much larger percentage of those detected as shingle.
The practical performance of the metal roof attribute varies with region, depending on how prevalent it is, and how common other roof materials are.
The lower number of "Other" roof material in Australia means that less of the "Other" predictions are confident (as Australia is almost exclusively tile, metal and concrete roofs).
The set of roof shape attributes each detects whether there is at least one structural element of the roof with that definition.
All parts of the above roof are hip
A roof, or part of a roof, that is made up of three sloping triangles (more than three are considered a partial turret).
At least one hip element on the roof (either a whole hip end, or a small "mini hip" joint as part of a more complex roof).
|At least one gable element on the roof (similarly, a large end of a building, or a small gabled roof feature). A roof, or part of a roof, that has inverted "v" shaped ends.|
|At least one flat panel on the roof (including a small covered balcony extension, as long as the roof is permanent, not fabric awning).|
The centre house has a partial turret
|At least one round, pointed turret (made of connected triangles pointing to a common centre). The definition includes both full round ones, and a half turret feature that is comprised only of a few triangles that do not form a complete circle.|
|A dutch gable roof looks mostly like a hip end, but with a small gable-like section on top. they are easily discernable from a top-down view, as per the picture on the left (which has two gable ends).|
Roof drawings courtesy: https://unitecontractors.com/roofing-contractors-boston-services/
Characteristics and Recommended Use
The Roof Shape data differs from that of Dominant Roof Material, in that there is no sense of an overarching dominant shape. For example, it is difficult to determine the dominant shape of a rectangular building with one hip end, one gable end, and a turret on the side. By the same token, no reasonable area estimate can be provided. Viewing the AI Layers you will be able to identify a blob around the key part of the feature, e.g. the corner where three panels join for a hip. This blob is carried over and attached to the relevant building when an AI Parcel is exported.
The main reason flat roofs have a low confidence is that a "shed" roof (flat section at a slight angle greater than a few degrees) is marked as incorrect. The other reason is fabric awnings protruding over a balcony are almost impossible to distinguish from a top-down aerial image, and the model is more cautious at predicting the presence of flat roofs. If you don't require this distinction, you can safely include lower confidence flat roofs.
The primary reason for turret roofs having low confidence is that they are rare, small features. In practical use, high confidence turrets tend to be the clearest, sharpest examples of full round turrets. The lower confidence turrets are the partial ones that can look like a cross between a turret and fancy hip (with e.g. 4x triangles meeting to cover a building end or corner, rather than three).
A "Tree Overhang" AI Layer is available for viewing in MapBrowser . Polygons of tree overhang, as well as area estimates, are available in an AI Parcel export.
Any case where there is some tree (leaf on or off) protruding over the top of a Building Footprint.
Characteristics and Recommended Use
When presented as a "Y" or "N" flag in the CSV file, Tree Overhang may not be directly useful. It detects any amount of overlap at all, such that more than half of the properties in our coverage area are listed as having tree overhang. Where the overhang is used to identify trees that may need trimming, or calculating property risk, it is recommended you use the area estimate to set a meaningful threshold (e.g. only look at buildings with >10m2 of overhang). In more advanced use cases, you could use the area directly to rank properties from highest to lowest amount of overhang, or as a continuous variable to use in a statistical model. It is also possible to consider the shape and position of the tree overhang sections on the building, obtained from the geospatial file of exported AI Parcels.
Note that in Gen 1 and Gen 2 , the tree overhang is calculated on Building Footprints before enhancement algorithms are applied. This means that the tree overhang polygons do not always push quite to the edge of the sharp corners introduced by the enhanced Building Footprint polygons.
The confidence values for Tree Overhang are particularly low, because the edges of trees that overhang a roof often fade out gradually between high confidence and very low confidence. The area estimate of Tree Overhang is recommended as a better means of adjusting whether mild, moderate or extreme overhang should be included for your particular application.
For a precise changelog, refer to the AI Generations pages. This is a convenient summary to describe changes to this AI Pack over time:
Available as described above.
3D Attributes (Dominant Roof Pitch), and Dutch Gable added.