How Is Building Height Data Extracted for RF Planning: LiDAR vs Photogrammetry vs Satellite?

Building height data for RF planning is extracted using three primary methods: airborne LiDAR, aerial photogrammetry, and satellite-based techniques (stereo photogrammetry or AI extraction from imagery). LiDAR offers the highest raw accuracy (5 to 15 cm) but limited global coverage and high cost. Satellite AI extraction delivers 1 to 2 m accuracy at global scale with on-demand availability, making it the practical choice for most telecom deployments.

Why Does Building Height Accuracy Matter for RF Planning?

Building height directly controls three critical propagation calculations:

1. Rooftop diffraction: the diffraction loss over a building edge is computed from the height difference between the antenna, the building top, and the receiver. A height error of 2 m changes the diffraction angle and the associated path loss, shifting predicted coverage by several dB.
2. Shadow boundaries: the geometric shadow behind a building is determined entirely by building height. Incorrect heights create predicted shadows in the wrong locations, producing phantom coverage holes or unexpected coverage where the model shows obstruction.
3. Fresnel zone clearance: for point-to-point and FWA link analysis, Fresnel zone clearance calculations require accurate heights of all obstructions along the path.

At 5G sub-6 GHz, a 2 m height error is tolerable in many planning contexts. At 26 GHz mmWave, even 1 m matters because the wavelength is 11 mm and the Fresnel zone is correspondingly narrow.

Method 1: Airborne LiDAR

How It Works

An aircraft equipped with a laser scanner emits thousands of pulses per second toward the ground. Each returned pulse carries time-of-flight data that translates to precise 3D coordinates. Modern systems fire multiple pulses simultaneously and record multiple return echoes per pulse, enabling vegetation penetration and separate classification of ground and above-ground features.

Accuracy

• Vertical accuracy: 5 to 15 cm RMSE for modern airborne systems at typical survey altitudes
• Point density: 4 to 20+ points per m² depending on survey specification
• Horizontal accuracy: Sub-meter, dependent on aircraft GPS/IMU quality

Strengths for Building Height Extraction

• Resolves individual parapets, rooftop structures, elevator shafts, and antennas
• Full-waveform systems can distinguish first and last returns, enabling clean DTM extraction under trees
• Highly consistent accuracy across the survey area regardless of surface reflectance

Limitations

• Geographic coverage is limited to areas where government programs or commercial surveys exist
• Cost of new airborne surveys: approximately €500 to €1,500 per km² at standard density
• Rapid obsolescence, as new buildings appear between survey cycles (typically 5 to 10 years for government programs)
• Significant post-processing required before data becomes planning-ready

Method 2: Aerial Photogrammetry

How It Works

Overlapping aerial photographs taken from a fixed-wing aircraft or UAV are processed using stereo photogrammetry algorithms (Structure from Motion, dense matching) to reconstruct a 3D point cloud and surface model. Building heights are then extracted from the difference between the Digital Surface Model (DSM) and Digital Terrain Model (DTM).

Accuracy

• Vertical accuracy: 10 to 30 cm RMSE for high-resolution aerial imagery (typical GSD 5 to 15 cm)
• Building height accuracy: Typically 0.3 to 0.5 m RMSE after DTM subtraction
• Horizontal accuracy: 5 to 20 cm depending on flight height and control point density

Strengths for Building Height Extraction

• Better than LiDAR for detailed facade texture and building geometry reconstruction
• Photogrammetric meshes enable LOD2 building model reconstruction with roof geometry
• Useful for urban digital twin applications requiring visual fidelity in addition to height

Limitations

• Requires cloud-free conditions; aerial campaigns must be weather-dependent and scheduled
• Cannot penetrate vegetation canopy: DSM under tree cover is a canopy surface, not the ground
• Higher mobilization cost than satellite for large areas
• Geographic coverage limited to commissioned surveys; not globally available

Method 3: Satellite-Based Methods

3a: Satellite Stereo Photogrammetry

High-resolution commercial satellites (such as Pléiades, WorldView, or SPOT) can acquire stereo image pairs over the same area from different angles. Stereo matching algorithms reconstruct 3D surface geometry similar to aerial photogrammetry, producing DSM and DTM products.

Accuracy:

• Vertical accuracy: 1 to 3 m RMSE for 50 to 70 cm GSD satellite imagery
• Building height accuracy: 1 to 2 m RMSE after DTM subtraction
• Coverage: Global, on-demand tasking

Practical advantages:

• No mobilization cost, as satellite archives may already contain stereo pairs over the target area
• Consistent methodology globally, enabling comparable data quality in all markets
• Refreshable whenever new imagery is available

3b: AI Building Extraction from Single or Stereo Imagery

Machine learning models (particularly convolutional neural networks) trained on large labeled datasets can infer building footprints and height estimates from single satellite images or from DSMs derived from stereo pairs. This approach works at scale with high automation and consistency.

Building height from AI extraction is derived either from DSM-minus-DTM (nDSM) calculations from stereo imagery, or from height regression models trained on imagery appearance features correlated with known heights.

Accuracy benchmarks (LuxCarta):

• Building footprint precision: 87.67%, recall: 88.44%
• Overall capture rate: 93%+
• LOD2 building model accuracy from textured meshes: 90%+

Direct Comparison: Which Method Is Best for Telecom Planning?

For 5G macro cell and FWA planning, which constitutes the majority of global 5G investment, satellite AI extraction's 1 to 2 m building height accuracy is entirely sufficient. The propagation models themselves (empirical or semi-deterministic) have calibration uncertainties in the same order of magnitude, meaning that higher-accuracy input data does not produce proportionally higher prediction accuracy.

What Is LOD1 vs LOD2 Building Data, and Does It Matter?

The CityGML Level of Detail (LOD) standard defines building model complexity:

• LOD1: Extruded footprint building represented as a box with uniform height (typically roofline or eave height). Sufficient for most macro cell planning.
• LOD2: Full roof geometry building includes roof slopes, dormers, overhangs, and distinct roof surfaces. Required for accurate diffraction calculations in dense urban areas and for mmWave planning.

LiDAR and aerial photogrammetry can both produce LOD2 models. Satellite AI extraction has historically produced LOD1 (box models), but newer techniques, including LuxCarta's deep learning method for 3D mesh building extraction (published at SPIE 2023), achieve LOD2 reconstruction with 90%+ accuracy from satellite-derived textured meshes, at 4x the productivity of manual digitization.

How LuxCarta Addresses This

LuxCarta's building extraction pipeline is fully satellite-based, using AI/deep learning (U-Net CNN architecture) to process satellite imagery into building footprints, heights, and 3D models at scale. The system achieves 93%+ building capture rate globally, enabling consistent data production for any market, including regions where LiDAR and aerial survey data have never been collected.

For LOD2 applications, LuxCarta's technique for extracting 3D building meshes via semantic segmentation from textured meshes (SPIE 2023 publication) extends the automation frontier, making high-fidelity 3D city models achievable at national scale without the cost of manual modeling or LiDAR acquisition.

The BrightEarth platform delivers these building datasets in formats native to Forsk Atoll, InfoVista Planet, and other major planning tools, including SHP, GeoJSON, GeoPackage, and GeoTIFF, without requiring post-processing by the end user. For engineers who need to bring their own sub-meter licensed imagery, BrightEarth's BYOI (Bring Your Own Imagery) capability allows AI extraction to be run on customer-provided imagery for maximum currency and spatial resolution.

Frequently Asked Questions

Which building height extraction method should I choose for 5G mmWave small cell planning?

For mmWave small cell planning in a limited geographic area (a downtown district or campus), LiDAR or aerial photogrammetry provide the height accuracy to resolve parapets and rooftop structures that affect link budgets at 26 GHz. If LiDAR data is already available for the area, it is the most efficient choice. If it must be newly acquired, compare the survey cost against the scale of the deployment.

Can AI extract accurate building heights in areas with dense vegetation canopy?

Satellite imagery cannot see through tree canopy, so buildings obscured by dense canopy present challenges for AI extraction. Airborne LiDAR with full-waveform capabilities can penetrate canopy and recover both tree height and building height separately. In practice, most urban areas have sufficient open sky view of rooftops for satellite AI methods to function effectively; only heavily tree-covered suburban areas show meaningful extraction gaps.

How is building height defined in the extracted data: is it roof height or eave height?

This depends on the extraction method and the specification. Most telecom-oriented building datasets provide the maximum roof height (the highest point of the structure) because this is the relevant metric for diffraction and shadow calculations. Eave height (the height of the wall before the roof slope begins) is sometimes also provided for LOD2 models. LuxCarta's building data provides building heights calibrated for use in propagation tools.

Can building height data become outdated quickly?

Yes, particularly in fast-growing cities. New high-rise developments fundamentally change the propagation environment in their vicinity. The advantage of satellite-derived extraction is that it can be re-run from new imagery as new construction appears, keeping the planning dataset current without commissioning a new LiDAR survey. Most operators refresh city-wide geodata every 1 to 3 years as part of their network optimization cycle.

What is the difference between nDSM and building height data?

The normalized Digital Surface Model (nDSM) represents the height of above-ground objects above the terrain surface: it is the DSM minus the DTM. Building heights can be read directly from the nDSM at each building footprint location. Individual building height attributes in vector datasets are typically computed from nDSM statistics (mean, maximum, or 90th percentile) within each building footprint polygon.

LuxCarta provides AI-powered 3D geospatial data solutions for telecom, simulation, and smart city applications worldwide. Learn more at luxcarta.com or explore on-demand extraction at BrightEarth.