ST = 0
Cc
Yt
Xt
Iv
1
an
George Vosselman
4 EXPERIMENTS
Several experiments have been performed to obtain insight into the behaviour of the developed filter method. The next
paragraphs discuss the used dataset, the error measures, and the derived filter functions.
4.1 Lisselstreek data set
The data used in this test has been recorded by the helicopter based FLI-MAP system [Huising and Gomes Peirera,
1998, Baltsavias, 1999]. The point density varies between 5 and 7 points per m'. The FLI-MAP system is primarily
designed for monitoring roads, rail-roads and power-lines and, thercfore, only records the first returning lascr pulsc.
Consequently, the penetration rate in arcas with vegetation is relatively low. A small part with high vegetation, a
meadow, and a dike with about 1.3 million points was selected (figure 3, upper left image). The average point density in
this arca was 5.6 points per m”. Roughly half the points arc ground points.
4.2 Ground "truth"
Unfortunately, it is. virtually impossible to establish accurate ground truth for laser altimetry data. This holds in
particular for areas with dense vegetation. Ground truth for the test area was not available. For the analysis of the filter
results, we made use of the filter property that the classification results improve with the point density. E.g. if there is a
height difference of 1 m between two points at a distance of 0.5 m, the higher point is most likely not a ground point.
If, however, the point distance 1s 4 m, a height difference of 1 m could as well be explained by variations in the terrain
height. The filter results obtained with the original dataset were therefore used as a reference for the filter results of
datasets with a reduced point density. From the original dataset, datasets with average point densities of one point per |,
4, and 16 m! were derived.
4.3 Derived filter functions
A training set with a part of a dike and some vegetation was sclected for the derivation of the filter functions described
in section 3.2 and 3.3. These filter functions will be called the maximum filter and the probabilistic filter. The training
sct contained 43000 points from which about 69 million height differences between points within a distance of 10 m
could be computed. The resulting functions are shown in figure 2. It was found that the standard deviations of the
maximum height differences were quite small, usually below I cm. Thus only a small confidence interval was added to
the maxima found in the training set. Figure 2 clearly shows that the maximum filter allows larger height differences
than the probabilistic filter.
5m 4-
x — Maximum
o
Ah E — Probabilistic
L3
el
©
= # Feud
0m E n
10m = Fr Le
= je
0 d
0 2 4 6 8 HU
Sm Distance between points (11
Figure 2. Filter functions derived from training data. Left: Probabilities P(p, € DEM | Ah, d, p; € DEM). Black is 0.0,
white is 1.0. Right: The upper curve shows the maximum height differences between ground points. The lower curve
shows the height differences for which P(p; e DEM | Ah, d, p, € DEM) = 0.5.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 939
