In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
3.5 Tree segment parameters 
4. RESULTS AND DISCUSSION 
The city administration is mainly interested in updating tree 
parameters for their available tree positions. These parameters 
serve for the city tree inventory but also for visualization 
purposes (i.e. virtual reality/city). Tree position, tree height and 
crown diameter are derived for the classified stem and potential 
stem segments. Due to the occurrence of power lines, adjacent 
buildings/roofs a robust calculation of tree height is introduced. 
The robust tree height is defined as the highest laser point 
fulfilling the criterion of being within a maximum height above 
the mean height of the k highest points of a segment (Fig. 6a). 
Tree crown diameter is approximated by the smallest enclosing 
circle (SEC) to the segment boundary including calculation of 
the mean deviation to the maximum diameter, in order to 
provide a quality measure of the derived diameter value 
(Fig. 6b). Different methods for tree position estimation are 
implemented: i) position of robust highest point, ii) circle center 
of SEC, iii) center of gravity of segment boundary vertices 
(Fig. 6c). The tree position estimation results in a new point 
GIS vector layer holding all segment attributes as well as crown 
diameter and tree height attributes. 
Figure 6. (a) Robust highest point/segment, (b) crown diameter 
by smallest enclosing circle and (c) different methods of tree 
position estimation compared with reference position (left) and 
point cloud cross-section indicating highest point (right) 
3.6 Derivation of vegetation mask 
Vegetation mask derivation is straightforward, due to the GIS 
vector topology. Common boundaries between vegetation 
segments are dissolved, i.e. adjacent polygons are merged. This 
procedure is followed by a generalization step, where small 
isolated polygons are removed (e.g. 20 m 2 ) and small holes (i.e. 
islands; e.g. 20 m 2 ) are closed. Furthermore, the vegetation 
mask boundary can be generalized by line simplification (e.g. 
Douglas-Peucker) or smoothing (e.g. Snakes) readily available 
in the GRASS GIS framework. 
The defined echo ratio (ER) raster layer based on echo classes 
(e.g. first, last echo) clearly shows a very good agreement with 
potential urban vegetation areas (Fig. 3b). Compared to other 
definitions of echo ratios, such as a purely geometric 
computation (e.g. Rutzinger et al., 2007; Höfle et al., 2009) 
where number of neighbors are counted in a defined 2D and 3D 
neighborhood, the high point density and the increased echo 
detection sensitivity and echo labeling provided by full, 
waveform LiDAR allows this faster ER derivation without the 
need for computationally intensive 3D point cloud 
neighborhood analysis. Furthermore, the full-waveform based 
ER is more robust against multi-temporal effects caused by 
temporary objects (Fig. 7). For example, a truck scanned in one 
flight strip but not present in a second strip covering the same 
area will cause an artificial object with a certain height above 
terrain and a very high transparency but will still have low echo 
widths and a low number of multiple echoes per shot. 
(a) Full-waveform based ER (b) Geometry-based slope-adaptive ER 
temp, booths cars on street 
Figure 7. (a) Full-waveform based echo ratio (ER), (b) 
geometry-based echo ratio (Hofle et al., 2009). Temporary 
objects can clearly be identified in the geometry-based ER 
In order to evaluate the separation capability of vegetation from 
non-vegetation, the optional building mask was not included for 
segmentation. Our experiments have shown that optional 
including of a building mask achieves best results when used as 
"soft mask", due to the occurrence of buildings below 
vegetation such as smaller buildings in parks or subway stations 
below alley trees. A "hard mask" would exclude building areas 
found in the cadastre from further processing and thus finally 
classify these areas as non-vegetation. A soft mask means a 
stricter threshold on ER than for the non-building areas (e.g. 
ER>50%), which should exclude most of structures related to 
free-standing buildings (e.g. walls and antennas) but still 
enables the detection of vegetation overtopping low buildings. 
These trees are mainly deciduous species and hence transparent 
under leaf-off conditions with a high number of laser shots with 
multiple echoes (i.e. high ER). The choice of window size for 
curvature calculation controls the degree of detail reflected in 
over- or under-segmentation (cf. Hofle et al., 2008). The larger 
the window size, the less segments are found. The threshold on 
curvature additionally controls the sensitivity for edge 
detection. The more moderate the threshold is chosen, the more 
edges can be detected, but also the potential edge zones become 
wider (Fig. 3c) and thus the location accuracy of the "thinned" 
edge line gets more uncertain. The location accuracy of the 
depression line between two objects could be improved by 
weighted skeletonization (by curvature and height) or drawing 
profiles for detecting the lowest point or the point with highest 
concavity in the boundary zone of two segments. 
The segmentation step results in 23788 segments (compared to 
67 Mio. laser echoes within the study site) with an average size 
of 10.6 m 2 . A window size of 7x7 pixel (i.e. 0.5 m resolution) 
for curvature calculation with curvature constraint of lower than 
-0.2 were chosen. Trees with a compact and convex crown