Ground Extraction

Of all the steps in a point cloud pipeline, ground filtering is the one whose mistakes end up on public record. When a national mapping agency gets it wrong on a single tile, the Digital Terrain Model it publishes carries phantom buildings, or mountains that are really tree clumps, and the contour lines drawn on top of that DTM become the official map of the territory for anyone who consults it. The error rates are not negligible. On the ISPRS Commission III filter test (Sithole and Vosselman, 2004), the best classical methods, the progressive TIN densification of Axelsson (2000) among them, reach Type I error rates of roughly 5-15% and Type II error rates of 1-6% on the harder reference tiles, which amounts to tens of thousands of misclassified points per square kilometre at typical ALS densities. No current method handles dense canopy, steep slopes, low scrub, and flat roofs well at one parameter setting, which is why the production world still pairs the algorithm with a careful manual quality-assurance pass.

Ground filtering, also called ground extraction, is the binary classification of every point in a cloud acquired over terrain as either ground, a return from the bare-earth surface, or non-ground, a return from an object above it such as a building, vegetation, a vehicle, or a power line. The Digital Terrain Model is then interpolated from the points labelled ground. This is the pipeline's first semantic decision, the moment a point stops being a bare coordinate and becomes either terrain or object. It consumes the cleaned clouds of Chapter 6, so the points reaching the filter are plausible returns rather than sensor artefacts, and the slope criteria at the heart of several filters are close relatives of the surface normals of Chapter 7. Downstream, the stakes are wide: the normalised heights that feed land-cover classification (Chapter 12) are measured against the DTM, and change detection between epochs (Chapter 14) differences surfaces built from it. The chapter separates the three elevation surfaces (DTM, DSM, nDSM), states the filtering problem and its failure modes, works through the four classical families (morphological filters, progressive TIN densification, cloth simulation, slope-based methods), and closes with evaluation on the ISPRS benchmark, learned filters, and the interpolation that turns ground points into a DTM.

Before any single method, it helps to see the whole field at once. The figure below previews the four canonical ground filter families covered in this chapter and the data conditions under which each is preferred.

DTM, DSM, and nDSM

Three elevation surfaces arise from LiDAR data:

• The Digital Surface Model (DSM) represents the highest visible surface: rooftops, tree canopy tops, bridge decks, and bare ground where no objects exist. It is computed by selecting the maximum elevation in each grid cell.

• The Digital Terrain Model (DTM) represents the bare-earth surface, with all above-ground objects removed. Constructing it requires ground filtering, the subject of this chapter.

• The normalised Digital Surface Model (nDSM), also called the canopy height model in forestry, is the difference:

The nDSM gives the height of objects above ground. It is useful for estimating building heights, tree canopy height, and detecting changes between epochs (Chapter 14). The figure below shows how the three surfaces relate on a single terrain profile.