Building footprint extraction at city scale

High-resolution optical and SAR imagery can delineate individual building outlines across entire cities. The accuracy ceiling depends on pixel size, roof geometry, and how honestly the method accounts for SAR distortion.

Sensors

Maxar WorldView-2 / WorldView-3: WorldView-3 delivers 0.31 m panchromatic and 1.24 m multispectral resolution, sufficient to resolve individual parapet edges and roof fixtures. WorldView-2 offers 0.46 m pan. Both support stereo collection for photogrammetric height derivation, with vertical accuracy typically within 1–3 m CE90 over urban areas.
Airbus Pléiades / Pléiades Neo: Pléiades Neo reaches 0.30 m pan resolution with daily revisit capability across most latitudes. Tri-stereo tasking produces dense point clouds that support building height estimation. Pléiades (first generation) provides 0.50 m pan with a heritage archive stretching back to 2012.
TanDEM-X: The TanDEM-X bistatic SAR pair operates at X-band (9.65 GHz) and has produced a global Digital Elevation Model at 12 m posting with relative height accuracy around 2 m. For urban areas, the WorldDEM product captures building volumes, enabling height-augmented footprint derivation where optical stereo is unavailable.
Capella Space SAR: Capella's X-band synthetic aperture radar delivers spotlight imagery at approximately 0.35 m resolution, with tasking possible through cloud cover and at night. SAR geometry introduces layover and foreshortening artefacts in dense urban scenes, but the same geometry encodes building height information that trained models can extract.
ESA Sentinel-2: At 10 m multispectral resolution, Sentinel-2 cannot resolve individual buildings in most cities. It is included here for context: it defines the lower boundary of what is usable for footprint work. Anything below roughly 1 m per pixel struggles to separate adjacent structures reliably.

What a rooftop actually tells you

A building footprint is a polygon. That sounds simple. In practice it encodes land tenure, tax liability, construction permit status, population density, and flood exposure all at once. Governments that lack current cadastral records, which is most of them, are working from field surveys that may be a decade old. Satellite-derived footprints can close that gap faster and more cheaply than any ground campaign.

The extraction process starts with the roof, not the foundation. Sub-metre optical imagery records the roof plane, parapets, and overhangs. Shadow length and direction, combined with the sun elevation angle at the time of acquisition, give a first-order building height estimate. For a 10 m tall structure imaged at 45-degree solar elevation, the shadow is roughly 10 m long. That geometry is deterministic, which is why shadow-based height estimation works even from a single-pass image, though accuracy degrades for clustered buildings where shadows overlap.

The physics of SAR: layover is not just noise

Synthetic aperture radar measures the round-trip travel time of microwave pulses. Tall structures return energy before the ground at their base does, because the top of the building is closer to the sensor in slant range. The result is layover: the upper part of the building appears displaced toward the sensor in the image. For a building 20 m tall imaged at a 30-degree incidence angle, the layover displacement is roughly 20 × cos(30°), or about 17 m. This makes naive footprint extraction from SAR imagery unreliable if the analyst treats it like an optical image.

Foreshortening is a related effect: slopes and façades facing the sensor are compressed in range, making them appear brighter and narrower than they are. Both artefacts are deterministic functions of incidence angle and building height, which means they are also information. Interferometric SAR techniques, and increasingly deep learning models trained on simulated SAR scenes, can invert these distortions to recover building height and footprint simultaneously. TanDEM-X exploits this principle at global scale. Capella's high-resolution spotlight mode makes the same physics available at the individual-building level.

The honest limit: SAR-derived footprints in very dense urban cores, where buildings are taller than the street width separating them, suffer from persistent shadow zones where the radar signal never reaches. No amount of processing recovers information from a geometric shadow. Multi-look acquisitions from different orbit directions partially fill these gaps.

Instance segmentation and the accuracy ceiling

Modern footprint extraction pipelines use convolutional neural networks, typically Mask R-CNN or derivative architectures, trained on labelled datasets such as SpaceNet or the INRIA Aerial Image Labelling benchmark. These models predict per-pixel building masks and then convert them to vector polygons. On well-lit, low-density residential areas imaged at 0.30–0.50 m resolution, intersection-over-union scores of 0.75–0.85 are achievable against manual reference data. That is good. It is not perfect.

The accuracy ceiling is set by three factors. First, imagery resolution: at 0.50 m per pixel, a 3 m parapet wall is only six pixels wide. Model predictions become unstable at that scale. Second, roof complexity: curved roofs, multiple levels, and rooftop plant rooms create ambiguous edges that confuse both edge-detection and segmentation approaches. Third, spectral contrast: flat concrete rooftops in arid cities often have near-identical reflectance to unpaved ground, making boundary detection genuinely difficult even for experienced human analysts. Multispectral bands help, but they do not eliminate the problem.

Post-processing matters as much as the model. Raw segmentation outputs produce ragged polygon edges that are cartographically unusable. Regularisation algorithms, which snap edges to dominant orientations and enforce right-angle corners where appropriate, convert probabilistic masks into cadastre-ready polygons. The trade-off is that regularisation can introduce systematic error on non-rectilinear buildings.

Height augmentation and volumetric outputs

A flat footprint polygon is useful. A footprint with a height attribute is considerably more so. It enables gross floor area estimation, population density modelling, and structural vulnerability assessment for earthquake or flood risk. Three methods exist for adding height to a footprint layer.

Photogrammetric stereo from WorldView-3 or Pléiades Neo produces dense point clouds with 1–3 m vertical accuracy over open rooftops. The method requires two or more images of the same area from different angles, collected within the same pass or on consecutive orbits. TanDEM-X WorldDEM provides pre-computed height values at 12 m posting globally, which is sufficient for block-level height statistics but too coarse to assign heights to individual buildings in dense cities. Shadow-based estimation, as described earlier, works from a single image but accumulates error in shadow-overlap zones and is sensitive to errors in the recorded sun angle metadata.

None of these methods reliably measures the height of a building's interior floors. Gross floor area derived from satellite data is always an estimate, and a cautious analyst should treat it as accurate to within 15–25 percent rather than as a substitute for a building permit record.

Where the method works and where it does not

Footprint extraction performs best in cities with low-to-medium building density, consistent roof materials, and reliable cloud-free imagery. Sub-Saharan African secondary cities, South and Southeast Asian peri-urban zones, and Pacific island capitals are contexts where the gap between administrative records and actual built extent is largest, and where satellite-derived footprints offer the most practical value. Satellize's crop-estimation work in Tonga demonstrates the same underlying principle in a different domain: administrative boundaries and ground truth diverge, and imagery resolves the disagreement.

The method struggles in three environments. Tropical cities with persistent cloud cover require SAR as the primary sensor, which introduces the distortion problems described above. Cities with very tall, closely packed buildings, think central Hong Kong or Manhattan, produce layover and shadow zones that leave significant blind spots. And informal settlements with highly irregular, overlapping structures and mixed materials push segmentation models toward their worst-case performance, because training datasets are dominated by formal construction. Separate model fine-tuning on local imagery is usually necessary before production-scale extraction in those contexts.

Cloud cover is not a permanent obstacle. SAR acquisitions are unaffected by cloud, and optical imagery can be composited from multiple passes to build cloud-free mosaics. But compositing introduces temporal inconsistency: a mosaic assembled over three months may contain buildings that were demolished and others that were started after the first acquisition. For rapidly changing cities, that matters.

Typical figures

Best optical resolution	0.30 m (WorldView-3 pan, Pléiades Neo pan)
Best SAR resolution	~0.35 m spotlight (Capella Space X-band)
Minimum detectable building footprint	Approximately 25–50 m² at 0.50 m imagery; smaller structures become unreliable below this threshold
Height accuracy (stereo photogrammetry)	1–3 m CE90 over open rooftops; degrades to 3–6 m in dense urban cores
Height accuracy (TanDEM-X WorldDEM)	~2 m relative vertical accuracy at 12 m posting
Typical footprint extraction accuracy (IoU)	0.75–0.85 on formal residential areas at sub-metre resolution; lower on informal or industrial areas
Revisit (commercial tasking)	1–3 days for WorldView/Pléiades depending on latitude and priority; near-daily for Pléiades Neo
Archive depth	WorldView-2 from 2009, Pléiades from 2012; TanDEM-X global DEM from 2010–2015 acquisition campaign
Delivery formats	GeoJSON, Shapefile, GeoPackage, PostGIS-compatible; height attributes as separate field or 3D CityGML
Cloud cover constraint	Optical unusable above ~20% cloud fraction per scene; SAR unaffected

Analytics Satellize can run

Building footprint vector layer	Instance segmentation (Mask R-CNN class of models) applied to sub-metre optical imagery, followed by polygon regularisation	GIS layer (GeoJSON or GeoPackage) with per-building unique ID, area, and perimeter attributes
Height-augmented footprint layer	Shadow-length estimation from single-image sun geometry, or dense point cloud from stereo WorldView-3 / Pléiades Neo collection	GIS layer with building height and estimated gross floor area fields; optional 3D CityGML export
Change detection: new construction and demolition	Polygon differencing between footprint layers from two acquisition dates; confirmed by pixel-level change detection on co-registered imagery pairs	Change report with added, removed, and modified building polygons; timestamped GIS layer
SAR-derived urban volume model	Layover and foreshortening inversion using incidence-angle geometry; optionally fused with TanDEM-X WorldDEM height values	Raster building height model (GeoTIFF) at 2–5 m posting, clipped to city extent
Footprint-to-parcel matching report	Spatial join of extracted footprints against client-supplied cadastral parcel boundaries; discrepancy flagging by area and centroid offset	Tabular report and GIS layer flagging parcels with unregistered structures or significant footprint-area mismatches
Building density and coverage ratio statistics	Zonal statistics computed over administrative or planning zones using footprint layer and imagery-derived impervious surface mask	Summary table and choropleth-ready GIS layer by zone

Sources

Who does the work

We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.