International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
face Model (nDSM) - the difference between a DSM and
a DTM approximating the ground surface — with a plani-
metric resolution of 0.5 m and hyperspectral data taken by
the airborne DAIS 7915 sensor and interpolated to 0.5 m.
The surface model was derived from HRSC-A data and the
non-building areas were masked by building outlines from
digital cadastral data. They investigated two approaches
for the fusion of the data - first, a fusion on signal level and
applying Spectral Angle Mapper (SAM) for classification
based on 16 channels of a minimum-noise-transformed data
set, and second, on a decision level using a binary decision
tree.
Our approach differs from the above with respect to the
input data, in particular the laser scanning data. We use
eCognition, which allows a hierchical classification and in-
troduction of knowledge by using the different information
sources for different decisions within a fuzzy classification
scheme. Details are given in Section 4.
3 DATA
For the characterization of urban surfaces with respect to
their geometry and their materials, two different data sets
are combined: a DSM and hyperspectral data.
The DSM was acquired in March, 2002, with the TopoSys
system using the first (cf. Fig. | and 3) and the last pulse
modes. For ease of use within different software pack-
ages, | m x 1 m raster data sets were generated. These
data sets differ not only concerning the objects included,
but also showing systematic effects: surface patches ap-
pear smoother and building footprints are systematically
“smaller in the last pulse data. The impact of these differ-
ences on the analysis will be discussed in Section S.
The hyperspectral data was acquired in July, 2003, with the
HyMap sensor during the HyEurope campaign organized
by the DLR (German Aerospace Center). Figure 2 dis-
plays a band combination ranging from the visible to near
infrared spectrum (cf. Fig. 3). The white line indicates
the central campus area. The data was preprocessed (at-
mospheric corrections, geocoding) by the DLR, Oberpfaf-
fenhofen, using the DSM. The original data has a ground
resolution of 4 m x 4 m. In order to use the data in combi-
nation with the DSM, the data was resampled to a resolu-
tion of 1 mx 1m. We applied different standard techniques
like (Dell' Aqua and Gamba, 2003) and their impact on the
results of our approach will also be discussed in Section 5.
Dimensionality of hyperspectral data is always of interest.
In order to get a first insight, we tried different techniques
for band reduction. We applied standard principle com-
ponent analysis (PCA), minimum-noise-fraction transfor-
mation (MNF), and manual selection of bands based on
the spectra of surface materials (Fig. 4). The same train-
ing sites were used to analyse the class separability using
the Batthacharyya distance. For the PCA and MNF data
one band after the other were included. Already 12 MNF-
bands and 15 PCA-bands lead to a high separabilty based
on this distance measure.
Figure 3: nEnlargement of subset: nDSM from first
pulse laser scanning data (left), HyMap data RGB-25/15/5
(right)
——- copper -— -brick2 —— brick1 —— slate1
—-— slate2 zinc —-—- stone plates —- roofing felt
vegetation —— asphait —— gravel
7000
6000
5000 --
4000 +
reflection
3000 -————74
2000 -
1000 KA
106 3
113 3
120 3
T uu
+ e w oN o «e e eo ~~ - co U oN o
- Le Q e "€ 0 Lo] ~ ~ eo e
channel number
Figure 4: Spectra of selected surface materials
1012
Intei
Oui
bast
dat:
is d
surf
scal
spe
clas
the
ofn
chai
also
fore
the
fere
The
war
data
of tl
cruc
We 3
data
Ma
E
Sla
Sto
Gr:
| Ro
Zin
