Buildings

Buildings Default model

Building roof polygons over a suburban area

Example result: one polygon per building roof.

Cost

10 credits per km² + imagery

Example: 10 km² on Mapbox zoom 18 ≈ 130 credits. Your own imagery costs nothing.

Finds building roofs on high-resolution imagery and returns one polygon per building. Objects under 20 m² are dropped to keep the layer clean. Building types, heights and cleaner shapes are options you switch on when you start a run.

What you get

A polygon layer of building roofs.

Works best with

Aerial or satellite imagery at 0.6–0.3 m per pixel (zoom 18).

Usage examples

Refresh a municipal building layer before a field survey
Count structures in a settlement for planning or emergency response
Find parcels with new or missing buildings

Version: 2026-07-06
Geo domain: Global
Method: Segmentation
GSD / zoom: 0.6–0.3 m / z18

Scenarios with this model

Each scenario is the Buildings model with one option switched on. Choose options in the run form of Mapflow Web or the QGIS plugin.

Map every building

Default run

Roof outline of every building in your area, as editable polygons.

10 cr/km² + imagery · zoom 18

Map every building

Building polygons coloured by building type

Sort buildings by type

+ Building types

Every building polygon carries a type, ready to filter and style.

10 cr/km² + imagery · zoom 18

Sort buildings by type

Building footprints extruded by estimated height

Building heights and 3D footprints

+ Heights (beta)

A height for each building, with the outline moved from roof to ground.

20 cr/km² + imagery · zoom 18

Building heights and 3D footprints

Sort buildings by type

Buildings + Building types

Cost

10 credits per km² + imagery

Example: 10 km² on Mapbox zoom 18 ≈ 130 credits. Your own imagery costs nothing.

Runs the Buildings model with classification switched on, so each roof polygon gets a building class. Classification adds no extra cost.

What you get

Building polygons with a class attribute.

Works best with

Aerial or satellite imagery at 0.6–0.3 m per pixel (zoom 18).

Usage examples

Compare residential and non-residential stock by district
Pull industrial and commercial sites into their own layer
Feed zoning and land-use analysis with fresh data

The full list of types is in the building classes reference.

Building heights and 3D footprints

Buildings + Heights (beta)

Cost

20 credits per km² + imagery

Example: 10 km² on Mapbox zoom 18 ≈ 230 credits. Your own imagery costs nothing.

Adds a height model that reads shadow length and visible walls. The result is a 3D footprint: the building contour projected to ground level instead of the roof outline.

What you get

Footprint polygons with a height attribute, ready for the QGIS 3D map view.

Works best with

0.6–0.3 m imagery (zoom 18) with visible shadows. Helps most on oblique imagery, where roofs look shifted.

Usage examples

Extrude buildings in the QGIS 3D map view
Estimate floor area and building volume
Screen areas for shade and line-of-sight studies

The model does not see footprints directly, because walls hide them in the image. Like a cartographer, it moves the roof outline down to the bottom of the wall.

All options

Building types: Assigns every building one of the main classes (see reference). No extra cost.
Regularization: Corrects irregular contours. Shapes are replaced with rectangles, circles or polygons with 90° angles that fit the original shape. The result is more map-friendly, but some of the original mask accuracy can be lost.
Simplification: Simplifies the building shape while staying close to the original mask, including curved and complex shapes. Fewer right angles than regularization.
Merge with OpenStreetMap: Where OpenStreetMap coverage is good, each building that has a matching OSM object (Jaccard index) is replaced with the OSM polygon, and rotated to align with the nearest OSM road. The result is then no longer based on the image, so buildings can be shifted from their actual position.
Height estimation (beta): A regression model estimates height from visual cues such as shadow length and visible walls, and returns 3D footprints: the contour projected to ground level instead of the roof outline. Most useful on oblique imagery, where roofs appear shifted. +10 credits per km².

Note

Buildings smaller than 20 m² are removed from the result to avoid clutter.

Benchmarks

Buildings v.2026-07-06 (Global, segmentation, 0.3 m / z19) was evaluated against manually annotated ground truth. Metrics are area-based: IoU is the intersection-over-union of the predicted and ground-truth building masks, and F1, precision and recall are computed on the overlapping mask area.

Aerial imagery

AOI (location)	Predicted features	IoU	F1	Precision	Recall
United States — Fort Myers	65	0.907	0.951	0.983	0.921
United States — Phoenix	47	0.903	0.949	0.944	0.953
Canada — Rigaud	103	0.870	0.930	0.934	0.927
Australia — Adelaide	92	0.868	0.929	0.914	0.945
New Zealand — Lower Hutt	106	0.867	0.929	0.956	0.903
New Zealand — Wellington	76	0.838	0.912	0.900	0.924
United Kingdom — London	68	0.813	0.897	0.873	0.922
Côte d’Ivoire — Bangolo	102	0.663	0.797	0.844	0.755
South Africa — Worcester	156	0.592	0.744	0.867	0.651
Aerial (mean of 9 AOIs)	815	0.813	0.893	0.913	0.878

Area-based IoU / F1 / Precision / Recall measured against ground-truth building masks; evaluation run 2026-06-13.

Satellite imagery

AOI (location)	Predicted features	IoU	F1	Precision	Recall
Russia — Ufa	426	0.821	0.902	0.897	0.907
Saudi Arabia — Riyadh	225	0.817	0.899	0.920	0.879
India — Bangalore	508	0.779	0.876	0.859	0.894
India — Thane	429	0.768	0.869	0.889	0.849
United Arab Emirates — Abu Dhabi	142	0.751	0.858	0.828	0.890
Satellite (mean of 5 AOIs)	1730	0.787	0.881	0.879	0.884

Area-based IoU / F1 / Precision / Recall measured against ground-truth building masks; evaluation run 2026-06-16.