Ground Extraction and DTM Generation

If you have ever looked at a terrain model and found a building in it, you have already met the subject of this chapter. Somewhere upstream, a filter kept a roof return as ground. The interpolator used that label, and the roof became a hill. Nothing later in the pipeline could undo it, because the DTM is interpolated from exactly the points this step labels. The contours drawn on that model become the official map of the territory. Of all the decisions in this book, this is the one that reaches a published product most directly.

How often does it happen? More often than you might expect. The ISPRS Commission III filter test (Sithole and Vosselman, 2004) ran eight filters over fifteen reference samples and counted both kinds of mistake. Type I error, ground points removed by mistake, reached 64 per cent on the hardest samples. Type II error, object points kept as ground, reached 19 per cent. Most of the eight filters were parameterised to keep objects out rather than to keep ground in, which is why Type I is the larger rate. The progressive TIN densification of Axelsson (2000) is one of the two exceptions (Hu and Yuan, 2016). Now attach a point density to those percentages. At the two pulses per square metre of the USGS standard quality level they become hundreds of thousands of misclassified points per square kilometre. At the ten pulses per square metre of the French LiDAR HD programme, millions. No method in this chapter handles dense canopy, steep slopes, low scrub and flat roofs at one parameter setting, so production work pairs the algorithm with a manual quality-assurance pass. I treat that pass as part of the method, and so do the national programmes of section 9.10.1, which plan and cost it in exactly that way.

Notice what kind of output this is. The cleaned coordinates of chapter 6 and the normals of chapter 7 showed you where a point lies and how its surroundings are shaped. Neither showed you whether the surface beneath it is the terrain or the top of something resting on the terrain. Ground and non-ground are the first semantic labels in the pipeline, class memberships rather than measurements, and that changes the consequences of an error. A slightly wrong coordinate stays where it is. A roof accepted as ground enters the DTM, survives into every height normalised against it, and carries on into the land-cover classification of chapter 12 and the surface differencing of chapter 14.

So what assumptions about terrain does a ground filter actually use? Every classical filter is model-driven: it is a rule about the shape of terrain, written down before a single point is classified. And its scale of support is the window, the seed tile or the search radius. That scale fixes the largest object the filter is able to remove, which is why it is the first number I check when a building has survived into a terrain model. Let's start with the surfaces themselves, since which of the three you need determines what ground has to mean.