Building height and density estimation from satellite data
Stereo photogrammetry, SAR tomography and shadow analysis let analysts derive building heights and urban density from satellite data, without a single site visit. Accuracy varies by sensor, urban geometry and cloud cover, and the honest figures matter for planning and property decisions.
Sensors
- Airbus Pléiades Neo: 30 cm panchromatic resolution. Tri-stereo acquisition mode collects three overlapping passes in a single overfly, enabling dense point-cloud generation and digital surface models at 50 cm posting. Revisit is theoretically daily at mid-latitudes with the four-satellite constellation.
- Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral. Stereo and tri-stereo tasking supports DSM generation at around 50 cm posting. Eight shortwave-infrared bands add material discrimination but do not improve height accuracy directly.
- TanDEM-X: X-band SAR interferometric pair operated by DLR. The global TanDEM-X DEM has 12 m posting and absolute vertical accuracy better than 10 m (LE90), with relative accuracy around 2 m over flat terrain. Tomographic processing of multi-baseline stacks can resolve individual building layers in research settings, though this requires dedicated acquisitions not available in standard catalogue.
- Copernicus DEM GLO-30: Derived from TanDEM-X data, freely available at 30 m posting. Useful as a bare-earth reference for differencing against commercial DSMs, though it captures some above-ground features and is not a true DTM. Vertical accuracy varies by terrain slope and land-cover type.
What a floating roof gives away
A digital surface model captures the top of everything: rooftops, tree canopies, bridge decks. A digital terrain model captures the bare ground beneath. Subtract one from the other and you have a normalised DSM, a layer in which every non-zero pixel represents something standing above the earth. Over a city, that layer is a height map of the built environment.
The subtraction sounds simple. The difficulty is that the two models rarely come from the same sensor at the same moment. Pléiades Neo tri-stereo acquisitions produce a DSM at 50 cm posting from a single overfly, but the bare-earth reference is typically the Copernicus DEM GLO-30 at 30 m, or a separately acquired LiDAR DTM where one exists. Misregistration between the two surfaces, even at sub-pixel level, introduces systematic height errors that compound in dense areas. Analysts routinely apply ground-control point correction and co-registration before differencing.
Stereo photogrammetry: the geometry of precision and its limits
Tri-stereo acquisition gives three convergent views of the same scene. Photogrammetric matching finds conjugate points across the image triplet and triangulates their 3-D position. With Pléiades Neo at 30 cm resolution, independent validation studies published in Remote Sensing journals report height RMSE values in the range of 0.5 to 1.5 m over open, well-contrasted rooftops. That is good enough to distinguish a two-storey terrace from a four-storey mansion block.
Urban canyons are a different problem. When buildings exceed roughly three to four times the street width, the lower portions of facades and the ground between them become invisible to any off-nadir stereo sensor. Matching algorithms have nothing to work with in those occluded zones and either interpolate across the gap or produce artefacts. In practice, height accuracy in dense city-centre blocks with narrow streets degrades to RMSE values of 3 to 5 m or worse, and footprint edges become uncertain by a metre or more in plan. This is not a failure of processing; it is geometry. Buyers of height data for dense urban cores should expect and budget for this uncertainty.
Cloud cover is the other hard constraint. Optical stereo requires clear sky across all three passes of a tri-stereo acquisition. In persistently cloudy climates, obtaining a clean tri-stereo collect over a specific city block can take weeks or months, even with daily revisit capability.
Shadow length as a low-cost cross-check
Every high-resolution optical image contains a second height signal: shadows. The length of a building's shadow is a function of the solar elevation angle at acquisition time and the building's height. Both the solar geometry and the image timestamp are recorded in the metadata, so the calculation is deterministic.
Shadow-based height estimation is coarser than stereo photogrammetry, with published accuracy typically in the range of 1 to 3 m RMSE for isolated buildings with clean shadow edges. It fails when shadows overlap between adjacent buildings, when the ground is sloped, or when the shadow falls on a surface that is itself elevated. Its value is as a rapid, archive-exploitable cross-check: historical imagery going back decades can be mined for shadow lengths to reconstruct how a building stock has changed over time, even where no stereo pair was ever acquired.
SAR tomography and what it adds
Synthetic aperture radar penetrates cloud and operates at night. TanDEM-X interferometry underpins the Copernicus DEM and provides a global height layer, but its 12 m posting cannot resolve individual buildings in dense urban areas. SAR tomography, which synthesises a 3-D image from a stack of acquisitions at different incidence angles, can in principle resolve individual floors of a building. In practice, this requires many coherent acquisitions over the same target, careful phase unwrapping, and significant processing expertise. It remains primarily a research technique for urban applications, not a routine commercial product.
For property intelligence purposes, the most practical SAR contribution is as a cloud-independent complement to optical stereo: use TanDEM-X or Copernicus DEM GLO-30 to establish a coarse height baseline across a city, then task Pléiades Neo or WorldView-3 stereo for the specific parcels where 50 cm posting matters.
From height map to density metric
Raw building height is an input, not a product. Property analysts and planners typically want derived metrics: floor area ratio (FAR, the ratio of total floor area to plot area), plot coverage, gross above-ground volume, or storey count. Each requires a building footprint layer to accompany the height data. Footprints can come from OpenStreetMap, national cadastre, or automated segmentation of the same VHR optical imagery used for stereo. Automated segmentation at 30 cm resolution achieves intersection-over-union scores of around 0.7 to 0.85 in well-contrasted suburban areas, lower in dense historic cores where rooflines are irregular.
Once height and footprint are combined, FAR and volume can be computed at parcel or block level and compared across a city or against planning consent limits. A block where measured volume exceeds the permitted envelope by more than the height uncertainty budget is a candidate for compliance investigation. A district where average FAR has risen measurably between two acquisition dates signals densification that may not yet appear in official statistics.
Satellize runs this kind of nDSM differencing and density derivation as part of its analytics stack. The methodology is the same published photogrammetric pipeline used across the remote-sensing community; the value is in integrating tasking, processing and delivery into a single workflow for clients who lack the in-house capacity to manage commercial satellite archives.
What to ask before commissioning a height survey
Four questions determine whether satellite-derived height data will be fit for purpose. First, what is the minimum height difference that matters for the decision? If the answer is less than one metre, stereo photogrammetry over dense urban areas will not reliably resolve it. Second, how much of the area of interest is in deep urban canyon, where occlusion will degrade accuracy? Third, is the climate conducive to obtaining a cloud-free tri-stereo collect within the project timeline? Fourth, does a usable bare-earth DTM already exist for the area, or does one need to be acquired or approximated?
If the answers suggest that optical stereo will struggle, the honest recommendation is to combine it with a coarser SAR-derived baseline and to quantify the uncertainty explicitly in any downstream analysis. Height data presented without an accompanying accuracy assessment is not analysis; it is decoration. Ask for the RMSE, the occlusion mask, and the co-registration report before signing off on any deliverable.
Typical figures
| DSM spatial posting (VHR stereo) | 50 cm (Pléiades Neo, WorldView-3 tri-stereo) |
| Height RMSE, open rooftops | 0.5 to 1.5 m (VHR stereo, published validation range) |
| Height RMSE, dense urban canyon | 3 to 5 m or worse due to stereo occlusion |
| SAR DEM posting (TanDEM-X / Copernicus GLO-30) | 12 m / 30 m; vertical LE90 better than 10 m absolute |
| Revisit for tasking (Pléiades Neo constellation) | Theoretically daily at mid-latitudes; cloud and scheduling reduce effective revisit |
| Shadow-height accuracy (isolated buildings) | 1 to 3 m RMSE; degrades with overlapping shadows or sloped ground |
| Archive depth (commercial VHR) | Pléiades 1A/1B from 2012; WorldView series from 2007 onwards |
| Minimum resolvable building footprint | Approximately 2 m x 2 m at 50 cm posting, subject to contrast |
| Typical delivery formats | GeoTIFF nDSM, vector building footprint (GeoPackage / Shapefile), parcel-level attribute table (CSV / GeoJSON) |
Analytics Satellize can run
| Normalised digital surface model (nDSM) | Tri-stereo photogrammetric DSM differenced against Copernicus DEM GLO-30 or client-supplied LiDAR DTM, with ground-control co-registration | GeoTIFF raster at 50 cm posting with per-pixel uncertainty estimate |
| Building height layer | nDSM segmented by building footprint polygons; median and 90th-percentile height extracted per footprint | Vector layer (GeoPackage) with height attributes and occlusion-flag field |
| Floor area ratio (FAR) by parcel | Estimated storey count (height divided by assumed floor-to-floor height, typically 3 m) multiplied by footprint area, divided by plot area from cadastre or OSM | Parcel-level attribute table joinable to client GIS or planning database |
| Temporal density change report | Differencing of nDSMs from two acquisition dates; change pixels filtered by height uncertainty threshold to suppress noise | Change raster and summary statistics report (PDF + CSV) per defined zone |
| Shadow-based height cross-check | Automated shadow-length extraction from single-date VHR imagery using solar elevation metadata; height inversion per building | Tabular comparison of stereo-derived and shadow-derived heights with discrepancy flags |
| Planning compliance screening | Measured building volume compared against permitted envelope from planning consent records; exceedances ranked by magnitude relative to height RMSE | Ranked alert list of candidate non-compliant parcels for field verification |
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.