Sensors
- Airbus Pléiades Neo tri-stereo: 30 cm panchromatic resolution; three along-track stereo angles acquired in a single pass, yielding dense point clouds and DSMs with 1–2 m vertical RMSE under good contrast conditions. The tri-stereo geometry reduces occlusion shadows compared with a simple stereo pair.
- Maxar WorldView-3 stereo: 31 cm panchromatic; agile off-nadir tasking allows same-day stereo pairs. Vertical accuracy for well-textured urban surfaces is typically 1–3 m RMSE. Archive depth back to 2014 supports multi-epoch change analysis.
- TanDEM-X: X-band SAR interferometry; 12 m posting global DEM with 2 m relative vertical accuracy over urban areas. Unaffected by cloud and independent of sun angle, making it the preferred city-scale height baseline. Penetrates neither rooftop cladding nor vegetation canopy.
- ICESat-2 ATL08: Photon-counting lidar; 17 m along-track ground footprint at 100 m posting. Published vertical accuracy for terrain is better than 10 cm, but urban building heights require careful filtering of the ATL08 canopy-height product. Useful for calibrating and validating DSM-derived heights at city scale.
What a rooftop elevation actually tells you
A building footprint records where a structure sits. A digital surface model records how high its roof sits above the ground. The difference between those two numbers, divided by an assumed floor-to-floor height, gives an estimated storey count. In practice, floor-to-floor heights vary: residential construction in much of sub-Saharan Africa runs around 2.8–3.0 m per storey, while commercial office stock in East Asian cities often reaches 3.5–4.0 m. Any floor-count estimate carries that uncertainty unless local ground-truth or planning records constrain it.
The practical value is not precision down to the last floor. It is the ability to rank a city's building stock by approximate height class, identify where density is concentrating, and detect change between two acquisition epochs without sending a single surveyor into the field. That is a different kind of answer from what footprint mapping alone can provide.
Tri-stereo geometry and why it beats a simple pair
A stereo pair produces a surface model by matching corresponding pixels across two images taken from different angles. Occlusion is the persistent problem: tall buildings block the view of adjacent shorter structures, leaving voids in the point cloud. Tri-stereo acquisition, as offered by Pléiades Neo in a single pass, adds a third angle that partially fills those voids and improves matching confidence in dense urban fabric.
The resulting DSM typically achieves 1–2 m vertical RMSE against lidar reference data over well-textured rooftops. Smooth, low-contrast surfaces such as white membrane roofing or flat concrete degrade matching quality and widen that error band. Vertical walls are almost always poorly reconstructed; the height information comes from the roof surface, not from the façade.
WorldView-3 stereo pairs offer comparable resolution but require two separate tasking passes unless the archive already holds a suitable historic image at the right off-nadir angle. That archive depth, stretching back to 2014 in many cities, is its principal advantage for change detection work.
Multi-epoch differencing: finding the floors that were added
The most operationally useful product is not an absolute height map but a height-change map. Subtract a DSM from 2019 from one from 2024 and any positive residual above roughly 2.5–3 m, co-located with an existing building footprint, is a candidate new floor. Below that threshold the signal drowns in the combined vertical noise of both models.
Systematic errors matter more than random noise here. If the two DSMs were generated from imagery with different sun angles or different stereo base-to-height ratios, biases can masquerade as real change. Co-registration to a stable reference, typically TanDEM-X or a lidar survey where one exists, is not optional. It is the step that separates credible change detection from artefact mapping.
TanDEM-X provides a consistent city-scale baseline free of cloud and illumination effects, though its 12 m posting means individual buildings narrower than roughly 20–25 m are poorly resolved. For dense informal settlements with small irregular structures, the optical stereo models are more informative despite their cloud sensitivity.
Flat roofs, pitched roofs and the height ambiguity problem
A flat-roofed building presents its actual floor level to the sensor. The DSM height is the roof-deck elevation and the storey-count estimate is as reliable as the model's vertical accuracy allows. A pitched roof is different. The DSM records the ridge, not the eaves. Depending on roof pitch, the ridge can sit 1.5–4 m above the top occupied floor. A two-storey house with a steep pitched roof can return a DSM height consistent with a three-storey flat-roofed building.
Resolving this requires either a roof-type classification layer derived from the imagery itself, or local knowledge of predominant construction typologies. In cities where pitched roofs are confined to a recognisable residential typology and flat roofs to commercial or multi-family stock, a simple morphological rule applied to the footprint shape and spectral signature can reduce the ambiguity substantially. Where typologies are mixed at fine grain, the ambiguity is honest and should be reported as a confidence interval rather than suppressed.
From height model to gross floor area
Gross floor area (GFA) is the figure that planners, tax authorities and property markets actually use. Deriving it from satellite data requires three inputs: a building footprint (covered by the sibling page on footprint extraction), an estimated storey count, and an assumption about whether each floor is fully enclosed. The product of footprint area and storey count gives an upper bound on GFA; setbacks, terraces and mechanical floors reduce it.
At city scale, the aggregate GFA estimate is more reliable than any individual building figure. Errors in storey count tend to be distributed roughly symmetrically across a large stock, so the portfolio-level sum is less sensitive to individual misclassifications than a parcel-by-parcel assessment. That aggregate figure is what matters for housing supply analysis, infrastructure loading calculations and property tax base estimation. It is also the figure that is hardest to obtain any other way without a full cadastral survey.
Satellize runs this workflow as part of its analytics service, combining commercial stereo tasking with open-source footprint layers and client-supplied cadastral data where available. The Tonga crop-estimation programme demonstrated the same multi-source integration logic in a different domain; the height-change pipeline applies the same epoch-differencing architecture to urban vertical stock.
Honest limits and what to do about them
Cloud cover is the operational constraint that never goes away for optical stereo. A city in a humid tropical climate may have fewer than 60 cloud-free days per year, and stereo acquisition requires simultaneous clear conditions across the full scene. SAR-based methods avoid this but trade spatial resolution and point-cloud density for weather independence.
The 1–2 m vertical RMSE of tri-stereo optical DSMs translates directly into storey-count uncertainty. At a 3 m floor-to-floor height assumption, a 2 m error is two-thirds of a storey. For buildings below four storeys, that is a large relative error. The method is most reliable for mid-rise and high-rise stock, precisely the stock where upward densification is most consequential for planning. For low-rise informal settlements, the uncertainty band often exceeds the signal.
ICESat-2 ATL08 can provide independent spot-check validation of DSM heights along its ground tracks, which repeat every 91 days. The coverage is not dense enough to substitute for a full DSM, but it is dense enough to detect systematic bias in a stereo product and to calibrate floor-count estimates before they enter a planning database.
Typical figures
| Best available spatial resolution (optical stereo) | 30 cm (Pléiades Neo), 31 cm (WorldView-3) |
| DSM vertical accuracy (optical tri-stereo, urban) | 1–2 m RMSE typical; degrades to 3–5 m on smooth or low-contrast rooftops |
| DSM vertical accuracy (TanDEM-X, urban) | 2 m relative vertical error; 12 m posting limits resolution of narrow structures |
| Minimum detectable height change (epoch differencing) | ~2.5–3 m above combined model noise floor; approximately one storey |
| Revisit / tasking latency (commercial optical stereo) | 1–3 days for priority tasking; archive delivery within 24 hours where coverage exists |
| Cloud sensitivity | Optical stereo requires cloud-free acquisition; TanDEM-X unaffected by cloud |
| Archive depth | WorldView-3 from 2014; Pléiades from 2012; TanDEM-X global baseline 2010–2015 |
| ICESat-2 ATL08 ground track repeat | 91-day exact repeat; 17 m footprint, 100 m along-track posting |
| Typical area coverage per stereo collect | Up to 100 km² per Pléiades Neo tri-stereo strip; scalable with mosaicking |
| Delivery formats | GeoTIFF DSM, GeoPackage or Shapefile with height and storey-count attributes, CSV summary by planning zone |
Analytics Satellize can run
| Building height map | Semi-global matching (SGM) or patch-based stereo on tri-stereo imagery; DSM normalised against DTM derived from ground-classified points | GeoTIFF raster of roof heights above ground, clipped to building footprint polygons |
| Storey-count classification layer | Roof height divided by locally calibrated floor-to-floor height assumption; classified into 1–2, 3–5, 6–10, 10+ storey bins with confidence flag | Polygon GIS layer with storey-count attribute and per-building uncertainty band |
| Height-change detection (epoch differencing) | Co-registered DSM subtraction between two acquisition dates; change thresholded at 2.5 m minimum; false-positive suppression using footprint intersection | Change polygon layer flagging structures with statistically significant upward growth, with estimated floors added |
| Gross floor area estimate by planning zone | Footprint area multiplied by storey count; aggregated to cadastral or planning-zone boundaries; roof-type correction applied where classification data permits | Tabular report and choropleth layer of estimated GFA per zone, with uncertainty range |
| DSM accuracy validation against ICESat-2 | ATL08 terrain and canopy height photons intersected with DSM; bias and RMSE computed along ground tracks; calibration offset applied if systematic bias detected | Validation report with per-track residuals and adjusted DSM where bias correction is warranted |
| Upward densification trend summary | Multi-epoch height-change layers stacked across three or more acquisition dates; per-building height trajectory fitted; areas of sustained vertical growth identified | Time-series chart per district and ranked list of highest-growth parcels for planning authority review |
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.