In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
285 
shape result in one segment, whereas large deciduous trees with 
multiple tree tops are represented by multiple segments (one per 
convex tree top). For selected 181 alley trees the average 
number of segments per tree is 2.6, indicating over 
segmentation and trees with multiple tops, respectively. Within 
the park areas with larger deciduous trees the number of 
segments per tree lies clearly higher but could not be assessed 
due to missing reference positions. Over-segmentation can also 
be reduced by prior filtering (e.g. Gauss filter) of the DSM in 
order to reduce canopy roughness and suppress small structures 
(cf. Hirschmugl et al., 2007). 
7 - Non-Vegetation 
Perc. of echoes above min. height (2 m) [%] 
>99.0 
Mean amplitude of first echoes [DN] 
>50 
Mean amplitude of last echoes [DN] 
>100 
Mean echo width of first echoes [ns] 
<4.1 
Std.dev. of echo width of first echoes [ns] 
<0.35 
Std.dev. of heights of first echoes [m] 
> 10 
Std.dev. of heights of first echoes [m] 
<0.2 
ER of segment [%] 
<5 
Compactness && 
> 1.7 
Mean nDSM height [m] && 
>3.0 
Stddev. of height of first echoes && 
> 1.0 
Perc. of boundary covered by neighbors 
< 60% 
6 - Shrubs 
Mean nDSM height [m] && 
<3.0 
Std.dev. of heights of first echoes [m] 
< 1.0 
5 - Stem (reference) 
Distance to reference tree position [m] 
< 1.0 
1 - Detached 
No. of adjacent segments 
<0 
2 - Semi-detached 
Pere, of boundary covered by neighbors 
<20 
4 - Potential stem segment 
Number of echoes in height interval 1.0 - 2.5 m 
> 10 
3 - Crown segment 
All remaining segments 
Table 1. Rules and thresholds on segment attributes for 
detecting vegetation and further characterization 
Although the constraint on ER already excluded the majority of 
non-vegetation objects, the segments still contain non 
vegetation objects such as building walls, roof overhangs, 
transparent roofs and power lines. Thus, full-waveform point 
cloud information derived on segment level is valuable for 
separating vegetation from non-vegetation. Particularly, echo 
width and signal amplitude show clear signatures for vegetation 
(refer to Fig. 4). Vertically extended objects with a multitude of 
small scatterers (e.g. branches) exhibit larger echo widths and 
lower amplitudes due to the relatively small target areas 
contributing to each echo (cf. Wagner et al., 2008). The main 
part is to exclude non-vegetation segments. Through 
exploratory data analysis and visual inspection suitable 
attributes and thresholds could be obtained. Table 1 shows the 
applied rules to the rule base defined in Fig. 5 and Fig. 8 the 
resulting classified segments. For example, the percentage of 
echoes above the min. tree height of 1 m indicates a low ground 
penetration, which mainly occurs at building walls not 
connected with the ground, transparent roofs, antennas on roofs, 
and even vegetation on top of buildings, exhibiting an ER 
above 5%. Building facades can be excluded by high 
compactness, as they are elongated, and high std.dev. of first 
echo heights together with low coverage of adjacent segments 
and relatively high mean nDSM heights. High nDSM heights 
for vertical walls are also due to the generation procedure of the 
DSM, where the maximum height value per cell is taken. 
Shrubs are distinguished by using mean nDSM height and 
std.dev. of first echo heights, which is assumed to be lower than 
for trees. 
Figure 8. Classified vegetation segments further separated into 
six sub-classes. 
This manual set up and selection of attributes and thresholds 
will be replaced by automatic classification procedures in future 
such as statistical classification trees (Rutzinger et al., 2008) or 
Support Vector Machines (Mallet et al., 2008), which have 
already been applied to classify single laser points of large 
point clouds. The segment-based approach may lead to more 
stable features for classification (e.g. mean echo width per 
segment; cf. Rutzinger et al., 2008) but is strongly dependent on 
the quality and delineation accuracy of the segmentation. The 
current state of segmentation and classification provides the 
necessary input for vegetation mask generation and derivation 
of tree segment parameters (e.g. height, diameter, position) for 
visualization purposes by reconstructing artificial tree objects 
(cf. Vosselman, 2003). However, for urban tree inventory single 
tree detection is required and tree positions should be derived 
from stem detection. The class of potential stem segments could 
be a starting point for further point cloud based stem extraction. 
Multi-level/scale segmentation, i.e. further segmentation on the 
derived segments, could solve the problem of over 
segmentation (Blaschke, 2010) and join segments belonging to 
one tree. Promising studies have already shown the potential of 
point cloud based single tree detection for airborne (Reitberger 
et al., 2009) and mobile LiDAR data (Rutzinger et al., 2010). 
No point cloud segmentation is required for the class of 
detached tree segments, representing a single tree object. But 
for trees with multiple tops and no distinct crown shape 3D 
point cloud segmentation shows great potential, providing the 
required information inherent in the vertical sampling of the 
objects by airborne LiDAR. Multi-level LiDAR analysis, such 
as prior image based detection of candidate regions with 
following point cloud based object detection increasing 
delineation and classification accuracy offers the possibility to 
process large areas even with very high point densities in an 
operational manner without major loss in classification 
accuracy, if the pre-selection has high completeness (cf. Hofle 
et al., 2009). 
For evaluation the alley tree inventory (i.e. tree positions) is 
used. Out of 668 alley trees 639 (95.7%) could be successfully 
detected and included in the final vegetation mask. The missing 
trees are mainly young trees with low diameter and crown area,