| =
LT 0 A need RAA emo
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
4 APPROACH FOR DATA ANALYSIS
Our approach for the characterization of urban surfaces is
based on the analysis of laser scanning and hyperspectral
data as depicted in Fig. 5. The geometry of surface patches
is derived using a DSM from laser scanning, whereas the
surface material information is obtained from both, laser
scanning and hyperspectral data. Of course, the hyper-
spectral data is the main source for the surface material
classification, but the used surface material also restricts
the geometry or vice versa, the geometry restricts the use
of materials. Table 1 shows some examples of roof surface
characteristics, grouped with respect to similar spectra, and
also indicating qualitatively the surface geometry. There-
fore, this information can be used as additional clue within
the classification in case the spectral characteristics of dif-
ferent surface materials are almost similar (see Fig. 4).
The main part of our analysis is performed using the soft-
ware package eCognition. In this software the first step of
data analysis is a segmentation, followed by classification
of the segments. Therefore, the quality of segmentation is
crucial for the quality of classification. In the following,
we will describe both steps in detail using a subset of the
data as example (white dashed line in Fig. 3).
Laser data Hyperspectral data
| 0.
Surface geometry 4 Surface material
slope ? sealed surface ?
orientation ? building surface ?
size ? material ?
Figure 5: Flow chart of approach
Material Geometry Remarks
flat | sloped
| Brick | dio dba |
Slate - + spectrum similar to stone
plates, gravel, roofing felt
Stone plates | + - spectrum similar to slate,
gravel, roofing felt
Gravel + - spectrum similar to stone
plates, slate, roofing felt
Roofing felt | + + spectrum similar to stone
plates, slate, gravel
Copper + + both possible; sometimes
just facing at roofs’
outlines with other
material like gravel
for the main part;
see remarks for copper
bs jl
| Gras Leu
Table 1: Examples of roof surface characteristics
limited slope
1013
J
4.1 Segmentation
The segmentation procedure within the eCognition soft-
ware is based on a region growing algorithm. The crite-
rion for the growing combines three different quantities:
the homogeneity of the segment, the shape of the segment
measured by its compactness, and the smoothness of its
boundary. The homogeneity of the segment takes the de-
viations from the mean of each channel used for segmen-
tation into account. Thus, the underlying model assumes
constant values for each segment’s channel, which is only
adequate when dealing with flat roofs, but not when deal-
ing with roofs consisting of planar faces, which is our as-
sumed model, and using the laser scanning data as main
information for the segmentation. Aware of this problem,
we nevertheless tried the segmentation procedure of eCog-
nition. Examples of these segmentations are given in Fig.
6 and 7. For these segmentations first and last pulse data
and a NDVI (channels 25 and 15 of the HyMap data) are
used. Emphasis was on the geometry data (each channel
with weight 4), and less on the NDVI data (weight 1). The
segmentations are based on two different scale parameters.
A visual inspection of the results indicates what was al-
ready expected: The gable roof of a building in the lower
left corner (cf. Fig. 8) was segmented into several slight
elongated segments in the main roof directions, just ap-
proximating the sloped surface by segments with constant
heights - independent from the choice of scale parameter.
In case of flat roofs, e.g. building the upper middle, the
segmentation resulted in reasonable segments, when con-
sidering, that there are smaller extensions on this roof (cf.
Fig. 8).
Instead of the segmentation by eCognition, our segmenta-
tion procedure for laser scanning data searches for planar
faces. It follows the region growing principle taking the
deviation from a plane in 3D into account. Details of the
algorithm are given in (Vógtle and Steinle, 2000). Fig. 9
shows the result of the algorithm for the subset based on
the last pulse laser scanning data, thus only the geome-
try is taken into account during segmentation. Parameters
were set to include smaller roof extension in the surround-
ing larger surface patch. The use of geometric data only
may lead to problems, when one planar roof surface patch
consists of areas with different surface materials. In order
to overcome this drawback, the segmentation may be intro-
duced into eCognition and a second step of segmentation
using the spectral data to split up the initial segments may
be performed if needed. In this case, segmentation and
classification are closely related, because those channels
carrying the information for classification should also be
used for the segmentation. For the classification described
in the next section, we used the results of the eCognition
segmentation with scale parameter 50 shown in Fig. 6 and
the initial segments without refinement of our segmenta-
tion (Fig. 9).
4.0 Classification
Fig. 4 displays example spectra of materials to be classi-
fied. A closer look reveals the following:
REASONS DIE
