Schardt, Mathias
800
E
= a
2
=
3 600 — .
3 5
=) 400 — d pi a
> n
8 20
200 — pe? s 9s
9 i | I
0 200 400 600 800
predicted timber volume
Figure 6: Timber volume from forest inventory vs. predicted timber volume - hilly test site
Figure 7: Tree height image (left) and segmentation result (right)
5 SINGLE-TREE SEGMENTATION BASED ON LASER SCANNER DATA
In this section a method for delineating single trees in laser scanner data is presented. The key part of this method is
a watershed segmentation algorithm. In order to get meaningful segmentation results, the processed tree height map
demands some further preprocessing before being applied to the watershed algorithm. The whole processing chain and
the constraints implied by the used algorithms is explained in (Ziegler et al., 2000).
Results
Figures 7 and 8 represent segmentation results from the Alpine test site. Figure 7 shows a 40 x 40 meter section of the
tree-heights image generated from the laser scanner data and the result of the single tree segmentation method presented
in (Ziegler et al., 2000). For validation purposes the same plot was surveyed by ground measurements. The exact position
of the single trees was determined by the use of a differential GPS system and accurate terrestrial measurements (within
15 cm accuracy in x, y). For every tree in the plot a map of the tree crowns was created (reference tree crowns). In
figure 8 these crown maps were overlaid to the original tree height image. As it can be seen, the maps do not exactly
match with the tree height image. The determination of the size and shape of a crown map is based on terrestrial distance
measurements but also on visual interpretations. Therefore a typical error for terrestrial measurements of about 2096 has
to be assumed. Although this is not very accurate the maps provide a good estimation, which is useful for verifying the
segmentation results. The second image of figure 8 compares the outlines of the segments calculated by the watershed
approach (light lines) with the crown maps (dark lines).
The segmentation method found 17 out of 19 trees in this plot. One very small tree in the upper left quarter of the image
was merged with the neighboring trees. Examining the same area in the tree heights image it is also impossible to find
this tree by visual interpretation. In this case the resolution of the laser scanner data was to low or the one meter raster
grid to coarse to generate a image feature for this tree. The top left quarter also shows two trees merged. In the tree height
image these two trees are separated only by a hardly identifiable valley. This valley gets blurred in the smoothing step and
the regions representing the two trees merge. Choosing the best smoothing scale is therefore the most sensible part of the
method. A high smoothing scale will end up in a merging of trees standing dense but too low smoothing scales will split
up single trees into several segments.
1322 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.
Fur
seg
par
exa
futi
Ad
It is
disc
mos
true
dat:
Ins
tabl
pos:
exis
Ad
Fore
clas
of s:
rem
of re
has .
REI
Hos.
würt