In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
283 
The edge detector is based on applying a threshold to the 
minimum curvature in direction perpendicular to the direction 
of maximum curvature in a certain window (e.g. 7x7 cells), i.e. 
curvature<0.0 for concave areas (cf. Fig. 3c), and skeletonizing 
the potential edge areas to reach a final edge map. These edges 
correspond to the segment boundaries between adjacent objects. 
By combining the edge map with areas fulfilling height (e.g. 
>1.0 m; Fig. 3a) and ER threshold (e.g. >5%; Fig. 3b), the final 
segment raster is derived. In order to derive the segment 
polygons, a connected component labeling and vectorization of 
the connected region boundaries are applied (Fig. 3d). 
segment features could be obtained for each segment with 53 
point cloud based values. 
Echo width [ns]: Mean of first echoes 
transp. roofs 
building edges 
temporary booths 
Amplitude [DN]: Mean of first echoes 
m 0 000000- 17 746000 
« 17.746000 - 22.915000 
Id 22 916000 - 31.154000 
« 31.154000 -41 359000 
I« 41 359000 - 9697 000000 
Figure 3. Input layers for segmentation of convex regions in 
the nDSM having high transparency (i.e. echo ratio) 
3.3 Segment feature calculation 
For classification of the derived segments the available segment 
features (i.e. attributes attached to the segment polygons) are 
essential. In this step an extensive segment feature database is 
generated, considering segment features based on the point 
cloud, segment geometry and topology (Fig. 4). The derived 
features are attached to the attribute table of the GIS polygon 
layer. By Point-in-Polygon-Test the point attributes (normalized 
height, amplitude, echo width) stratified by laser echo classes 
(all, first, multi echoes = first and intermediate, last and single) 
are aggregated (number of echoes, min., mean, max., standard 
deviation) per point attribute and segment. In this aggregation 
step the laser echoes are filtered by the minimum vegetation 
height value of 1 m above ground, in order to exclude the 
terrain signature from the segment statistics, except the 
descriptive statistics for the "all" echo class. Additionally, the 
number of points falling within a potential height interval for 
tree stems (i.e. between 1.0 m and 2.5 m) are counted per 
segment. Furthermore, the ER on segment basis is derived 
(Eq. 1), and the percentage of points below the minimum 
vegetation height (i.e. 1.0 m) and above are attached. To 
include surface information, the statistics of the nDSM cell 
values are also calculated per segment (e.g. mean nDSM 
height). Based on segment polygon geometry i) area, ii) 
perimeter, iii) compactness (perimeter / (2 * sqrt(7t * area)) are 
derived. Due to the topological vector data model in GRASS 
GIS, topological information can easily be assessed and 
attached to the segments: i) number of adjacent polygons and ii) 
percentage of boundary shared with neighbors. All together 66 
Figure 4. Segments colored by mean echo width (top) and 
signal amplitude (bottom) of all first echoes within a segment. 
Non-vegetation segments can clearly be identified by low echo 
widths and higher amplitudes 
3.4 Classification 
Exploratory data analysis was performed, in order to set-up a 
rule-base for supervised classification. A logical rule base for 
classification was developed, which in a first step aims at 
separating vegetation from non-vegetation segments. Tree 
positions from the reference map are included in the 
classification process. Final classes and classification hierarchy 
are shown in Fig. 5. The GIS environment easily allows to 
perform the final classification using SQL statements in the 
attribute database. 
Figure 5. Classification scheme including reference tree 
positions. Target classes are numbered from 1 to 7.