International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
borders of buildings and for smaller segments. Tests using
bilinear or cubic interpolation were performed, but show
only minor changes of the results. The main problem is the
separability of the classes gravel, stone, and slate, which
also becomes obvious checking the stability of the classifi-
cation results (cf. Fig. 14). Most of the roof segments with
unstable result — i.e. second best classification result has
only small difference in its membership value compared to
the best — belong to the above mentioned classes. These
segments are shown in red. In case the fuzzy or(max) is
used, already one feature with high membership value is
sufficient for classification. If we use the fuzzy and(min),
all feature membership values have to be high for a class to
be selected. Fig. 15 shows the results for fuzzy and(min).
A number of segments are only classified as stone like with
higher classification stability.
Fig. 16 shows the result of classification based on our seg-
mentation using last pulse laser scanning data as input. A
visual comparision with the result in Fig. 13 — both based
on fuzzy or(max) — does not show large differences in clas-
sification. Differences occur in case the material in a pla-
nar patch changes or two roof surfaces are segmented as
one segment, because the change in geometry is only small
(only small height differences, smooth transition from one
roof plane to another), thus indicating that a refinement by
using the spectral information as described in Section 4.1
is mandatory.
Up to now, no geometric information has been used for
the classification of the roof surface materials. We ex-
pect that introduction of gradient information as additional
clue may help to discern at least slate from gravel and
stone. First tests based on gradients directly derived from
the laser scanning data indicate that gradient information
should not be extracted directly from the laser scanning
data, but from roof planes or segments to give reasonable
results.
6 CONCLUSIONS
In this contribution we presented our approach for the char-
acterization of urban surfaces, focussing in a first step on
roof surfaces. Input data are laser scanning and hyperspec-
tral data, which are analysed using the software package
eCognition and our own software for the segmentation of
laser scanning data. First results are presented, which show
in principle the feasibility of our approach. The main prob-
lems with respect to classification of surface materials are
the variability of the materials on one hand and the simi-
larity of some materials’ spectra on the other hand. A clas-
sification based only on the hyperspectral data is difficult,
although the data provides high spectral resolution. We
therefore intend to include geometric properties, namely
the slope of roofs, into our approach. Furthermore, a quan-
titative evaluation of the results is necessary. Up to now
our reference data is only coarsely related to the buildings
and has to be improved to serve as reference data for sin-
gle roof segments. The ongoing research by the Engler-
Bunte-Institute, Chair of water chemistry, on the chemical
processes on roof surfaces, will also influence our work,
because this research will indicate, which surface materi-
als have to be discerned and which may be grouped with
respect to the resultant pollution, thus the requirements on
the classification may still change.
ACKNOWLEDGEMENTS
The project is funded by Ministerium fiir Wissenschaft,
Forschung und Kunst Baden-Württemberg — Forschungs-
schwerpunktprogramm Kapitel 1423 Titelgruppe 74,
Quantitative Assessment of Pollutants on Urban Surfaces
by Chemical Analysis and Image Processing Methods. The
authors also would like to thank FGAN-FOM, Ettlingen,
for their co-financing of the hyperspectral data aquisition
and Eberhard Steinle, who provided the software for the
segmentation of roof segments.
REFERENCES
Bochow, M., Greiwe, A. and Ehlers, M., 2003. Ein
Prozessmodell zur Analyse von Hyperspektraldaten in
urbanen Gebieten. In: Vorträge 23. Wissenschaftlich-
Technische Jahrestagung der DGPF, Bochum, pp. 255 —
264. ‚
Butz, J. and Fuchs, S., 2003. Estimation of sealed surfaces
in urban areas and the impact on calculated annual pollu-
tion loads due to combined sewer overflows. In: C. Jürgens
(ed.), Remote Sensing of Urban Areas, IAPRS, XXXiV-
7/W9, pp. 35 — 40.
Dell’Aqua, F. and Gamba, P., 2003. Using image magni-
fication techniques to improve classification of hyperspec-
tral data. In: IGARSS 2003, Toulouse. CD-ROM.
Gamba, P. and Houshmand, B., 2000. Hyperspectral and
[FSAR data for 3d urban characterization. In: IGARSS
2000, Honolulu. CD-ROM.
Heiden, U., Roessner, S. and Segl, K., 2001. Potential of
hyperspectral hymap data for material oriented identifica-
tion of urban surfaces. In: Remote Sensing of Urban Areas,
Regensburger Geographische Schriften, Vol. 35, pp. 69 -
77.
Madhok, V. and Landgrebe, D., 1999. Supplementing hy-
perspectral data with digital elevation. In: IGARSS 1999,
Hamburg. CD-ROM.
Simental, E., Ragsdale, D., Bosch, E., Dodge, R. and
Pazak, R., 2003. Hyperspectral dimension reduction and
elevation data for supervised image classification. In: 14th
ASPRS Conference, Anchorage, AK.
Vógtle, T. and Steinle, E., 2000. 3d modelling of buildings
using laser scanning and spectral information. In: IAPRS,
Vol. 33, Part B3, pp. 927 934.
Vógtle, T. and Steinle, E., 2003. On the quality of ob-
ject classification and automated building modelling based
on laserscanning data. In: ISPRS WG III/3 Workshop
- 3D Reconstruction from Airborne Laserscanner and In-
SAR Data. CD-ROM.
1016
fs
KE
AB!
The
[KC
curr
with
asse
Qui
appl
reso
phot
has
asse:
hanc
outli
Mori
live
fourt
incre
are tl
by ai
decis
wide
most
Trad
using
order
areas
succe
1999
imag
and I
satell
imag:
photo
cartos
resol
Quick
have
resolt
image
inforr
metre
metre
