Building-height extraction for millimetre-wave network planning
Very-high-resolution stereo imagery and airborne lidar can extract per-building heights to roughly 1.5 m RMSE, precise enough to model first-Fresnel-zone clearance for 26/28 GHz 5G small cells. This page explains the photogrammetric pipeline, what it delivers, and where it fails.
Sensors
- WorldView-3 (Maxar): 0.31 m panchromatic, 1.24 m multispectral. In-track stereo pairs collected on a single pass; base-to-height ratios of 0.4–0.6 are typical. The primary commercial workhorse for dense-urban normalised DSM generation.
- WorldView Legion (Maxar): Six-satellite constellation targeting 0.30 m resolution with revisit of 15 minutes or better over priority cities. Useful for change-detection updates to an existing height model when new construction alters the roofscape.
- SPOT 7 (Airbus): 1.5 m panchromatic stereo. Covers larger footprints per tasking than WorldView at lower cost, but the coarser resolution makes sub-2 m height accuracy difficult in dense blocks with narrow street canyons.
- Airborne Leica ALS lidar: Point densities of 4–20 points per square metre at typical survey altitudes. Delivers ground truth to 0.1–0.15 m vertical RMSE and resolves facade overhangs that photogrammetry cannot see. The reference standard when regulatory compliance demands sub-metre accuracy.
Why millimetre-wave propagation punishes vague height data
At 28 GHz the first Fresnel zone radius between a rooftop small cell and a street-level user 200 metres away is roughly 1.7 metres. Any obstruction penetrating that zone introduces diffraction loss that climbs steeply: a parapet or plant-room enclosure that a 5 m resolution clutter model rounds away can add 10–20 dB of additional path loss in practice. That is the difference between a cell that serves the pavement and one that serves only the kerb.
Legacy clutter models built from 10 m or even 5 m resolution data were adequate for 3G and 4G macro planning, where link budgets had tens of dB of margin and diffraction geometry was forgiving. Small-cell 5G at millimetre-wave has neither luxury. The planning tool needs per-building eave heights, parapet heights, and rooftop obstacle positions to within about 1.5 m if the propagation model is to be trusted.
From stereo imagery to normalised DSM: what the pipeline actually does
A WorldView-3 in-track stereo acquisition produces two images of the same scene from slightly different angles, typically within seconds of each other to avoid cloud movement artefacts. Semi-global matching or similar dense image-matching algorithms compute disparity between the two images pixel by pixel, then convert disparity to elevation using the rational polynomial camera models supplied with the imagery. The result is a digital surface model (DSM) that includes building tops, tree canopy, and ground.
Subtracting a bare-earth digital terrain model (DTM), derived either from ground-filtered lidar or from an open source such as the Copernicus GLO-30 DEM where terrain is relatively flat, yields the normalised DSM (nDSM). Building objects are then segmented from the nDSM using height thresholds and planimetric footprint layers. Each building polygon receives a statistical summary: median eave height, maximum height (which catches lift shafts and plant rooms), and height standard deviation across the footprint.
Photogrammetric accuracy in dense urban blocks is typically 1.0–1.5 m RMSE at eave level when checked against independent lidar surveys. Errors are not random. They concentrate at building edges where the matching algorithm sees occlusion, at glazed facades that confuse texture-based matching, and wherever one building casts a hard shadow across an adjacent rooftop. These systematic biases matter more than the headline RMSE figure.
What a floating roof gives away, and what it hides
Rooftop morphology carries information beyond height. A flat roof with a 1.2 m parapet and a 3 m lift-shaft enclosure is a very different propagation obstacle from a pitched roof of the same ridge height. The nDSM captures both features if the imagery resolution is sufficient: WorldView-3 at 0.31 m resolves individual air-handling units and satellite dishes. This matters for small-cell siting because a planner can identify candidate mounting positions where the parapet itself provides interference shielding toward an adjacent cell.
The hard limit is tree canopy. Dense deciduous canopy in leaf absorbs matching texture and returns a smooth surface that the algorithm interprets as a solid obstacle at canopy height. Facades beneath overhanging trees are invisible to stereo photogrammetry. In residential streets with mature street trees this can affect 20–40% of building frontage. Airborne lidar with its discrete-return or full-waveform capability can partially penetrate canopy gaps, but even lidar at 10 points per square metre leaves significant facade occlusion beneath dense cover. There is no satellite-based solution that fully resolves this; the honest answer is to flag canopy-occluded building segments and treat their heights as uncertain.
Airborne lidar as the accuracy ceiling
Where a mobile network operator needs heights accurate to 0.5 m or better, perhaps to satisfy a national frequency regulator's interference coordination process, stereo photogrammetry is not the right primary source. Airborne lidar flown at 1000–1500 m altitude with a Leica ALS80 or equivalent system delivers point densities of 8–20 points per square metre and vertical accuracy of 0.10–0.15 m RMSE against ground control. That is an order of magnitude better than stereo imagery.
The trade-off is cost and coverage. A lidar survey of a single city costs significantly more per square kilometre than a WorldView-3 tasking, and national coverage is rarely available at useful point density. The practical workflow for most operators is to use stereo-derived nDSMs for city-wide initial planning, then commission targeted lidar surveys over the densest urban cores where small-cell economics justify the investment and where propagation modelling is most sensitive to height error.
Delivering a height model the radio planning tool can actually ingest
A normalised DSM as a raw GeoTIFF is not a radio planning input. The propagation tools used by most operators, whether Atoll, Planet, or similar, expect vector building layers with height attributes, or rasterised clutter files at specific resolutions such as 1 m or 2 m per cell. The photogrammetric output therefore requires building segmentation, attribute assignment, and format conversion before it is useful.
Segmentation quality degrades in terraced rows where buildings share party walls and the nDSM shows no gap between adjacent rooftops. Automated segmentation typically under-segments such rows, producing single polygons with height variance that masks individual building differences. Manual editing or cadastral footprint overlay is needed to split them correctly. This is a labour cost that project scoping must account for honestly.
Satellize structures its height-extraction analytics to deliver both a 1 m resolution raster nDSM and a vector building layer with eave height, maximum height, and a canopy-occlusion flag per polygon, formatted for direct import into standard radio planning environments. The workflow draws on the same photogrammetric pipeline used in the Tonga crop-estimation programme, adapted from agricultural canopy height to urban building height.
Keeping the model current: when to re-survey
A building-height model degrades as the city changes. A new residential tower or a rooftop plant-room addition can invalidate small-cell coverage predictions for surrounding sites within months of construction completion. WorldView Legion's high revisit rate makes change detection practical: a monthly or quarterly comparison of panchromatic imagery against the baseline nDSM can flag construction activity that warrants a local height update.
The trigger threshold matters. A 2 m height change over a 10 m by 10 m footprint is detectable in a well-registered stereo pair; a single-storey extension to a terraced house is at the edge of detectability and may require a field check. Operators should define a minimum change threshold tied to their propagation model's sensitivity before commissioning a monitoring programme, otherwise the alert rate will overwhelm the planning team.
Typical figures
| Primary imagery resolution | 0.31 m pan (WorldView-3); 1.5 m pan (SPOT 7) |
| Stereo nDSM output resolution | 0.5–2 m per pixel, depending on source imagery and matching parameters |
| Vertical accuracy (stereo photogrammetry) | 1.0–1.5 m RMSE at eave level in dense urban blocks; worse at building edges and beneath canopy |
| Vertical accuracy (airborne lidar, 8–20 pts/m²) | 0.10–0.15 m RMSE against ground control |
| Tasking revisit (WorldView Legion) | 15 minutes or better over priority cities |
| Typical processing latency (stereo to nDSM) | 3–7 days from imagery acquisition to validated height layer |
| Minimum detectable rooftop feature | Approximately 1 m height, 2 m footprint with WorldView-3 stereo; smaller features require lidar |
| Archive depth (WorldView constellation) | WorldView-1 archive from 2007; stereo pairs available over most major cities from approximately 2010 |
| Delivery formats | GeoTIFF nDSM, GeoPackage or Shapefile vector building layer, CityGML LoD1/LoD2 on request |
| Canopy occlusion flag | Per-building polygon attribute; derived from NDVI thresholding of multispectral bands |
Analytics Satellize can run
| Normalised DSM (nDSM) | Semi-global matching on WorldView-3 stereo pair; bare-earth subtraction using Copernicus GLO-30 or client-supplied DTM | 1 m resolution GeoTIFF covering the tasked area, with per-pixel confidence layer |
| Vector building height layer | Watershed or graph-cut segmentation of nDSM; attribute join with cadastral footprints where available | GeoPackage with eave height, maximum height, height standard deviation, and canopy-occlusion flag per building polygon |
| Rooftop morphology classification | Height gradient and planimetric shape analysis to distinguish flat, pitched, and complex rooftops; plant-room and parapet detection by local maxima | Attributed vector layer with roof-type classification and obstacle position points, formatted for Atoll or Planet import |
| First-Fresnel-zone clearance map | Link-geometry calculation between candidate small-cell positions and street-level receiver grid, using nDSM obstruction heights and ITU-R P.526 diffraction loss curves | Raster clearance map and per-link clearance report for candidate site list |
| Construction-change alert | Monthly panchromatic change detection against baseline nDSM; height-change threshold filtering to suppress noise | Quarterly GIS alert layer flagging polygons with height change exceeding client-defined threshold, with before/after image chips |
| Lidar-stereo accuracy reconciliation report | Point-to-surface comparison between client-supplied lidar ground truth and stereo nDSM; bias and RMSE by building type and street-tree density class | PDF accuracy report with spatial error maps; used to define zones requiring lidar re-survey before regulatory submission |
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.