JRBAN
‚SER
of urban
ed as an
rmation.
res (e.g.,
ction. À
cenes is
matically
pression
sion rate
Ived in a
y sensed
in urban
research
eral, and
“teristics,
'eometric
ed higher
irks etc.).
of limited
itial, it is
lata from
nation of
esents an
'rves that
nsors.
ast to the
, wavelet
d that by
ains etc.),
wavelet-
John Bosco Kyalo Kiema
Against the above background, the main objectives of the study presented here are twofold. Firstly, to consider within
the perspective of data fusion the automatic classification of urban environments from high-resolution remotely sensed
imagery and ALS data. And secondly, to investigate the influence of wavelet compression on the classification accuracy
of urban objects.
2 CLASSIFICATION OF URBAN ENVIRONMENTS
2. Study area and data used
The selected test area for this study covers an area of about 400m x 400m in the city of Karlsruhe which lies in the
south-western part of Germany. This includes the famous Karlsruhe castle and part of the main campus of the
University of Karlsruhe. The principal data used is airborne Daedalus scanner imagery. Marketed under the acronym
.Daedalus ATM (Airborne Thematic Mapper), this is basically an opto-mechanical line scanner with 11 different
spectral channels and 8 bit data quantisation ideally developed to match the successful Landsat TM. The system
specifications for this sensor are given in (Kramer, 1996). In addition, ALS data is also employed. In general, two ALS
sensor types can be distinguished: pulse and continuous wave laser scanning systems. The operating principle and basic
system overview for ALS is discussed in (Wehr and Lohr, 1999).
2.2 Fusing the different datasets
Data fusion is defined as a formal framework in which are expressed means and tools for the alliance of data originating
from different sources (Wald, 1999). This aims at obtaining information of greater quality. Hence, data fusion ought not
to be viewed as a mere collection of tools and means but rather, a comprehensive framework through which the
integration of spatial data is realised. Although the concept of data fusion is easy to understand, its exact meaning and
use often varies from one scientist to another. Different approaches to this may theoretically be employed (e.g., RGB
colour composites, IHS transformation, Principal Component Analysis, wavelets etc.) (Pohl, 1999). The particular
method adopted depends on several factors including: the structure of the data to be fused; the specific image
characteristics that need to be enhanced or preserved etc.
Two different approaches to the fusion of multi-sensor data may be adopted for applications like the one described in
this study. It is possible to use either the hierarchical classification approach or the additional channel concept. The
hierarchical or layered classification approach is basically a structured technique through which the different datasets to
be fused are applied in such a way as to successively divide the working area into more detailed object classes (Savian
and Landgrebe, 1991). In principle, this begins with basic object classes before progressively zeroing in on more
detailed ones. On the other hand, in the additional channel method the different datasets to be fused are introduced as
separate channels in an integrated fashion within an expanded dataset. This of course necessitates the co-registration of
the different datasets to a uniform georeference system.
Comparing the above two methods, the main disadvantage of the hierarchical classification approach is the propagation
of the classification errors in the subsequent classification steps. On the other hand, the main advantage of the additional
channel procedure is the enhanced flexibility in the data processing. For this reason, the additional channel concept is
adopted for the data fusion in this study. Hence, the normalised ALS data is introduced as an additional channel
alongside the multispectral channels of the Daedalus ATM imagery.
2.3 Classification Approach
Spectral information is conventionally employed in the
classification of multispectral imagery. The inadequacy
spectral features spectral signature of this in the extraction of urban objects has been
texture acknowledged in many studies ((Gong and Howarth,
spatial features structure 1990); (Fung and Chan, 1994); (Haala and Brenner,
size 1999)) etc. In applications where high accuracy is
shape/contour required, it is imperative to enhance the object feature
topology base. This needs to be expanded to include both spectral
and spatial feature characteristics ((Schilling and Vogtle,
1996); (Bähr and Vôgtle, 1998)). Table 2 shows some of
the variables that need to be considered. These may be
included either, explicitly through the integration of
different multi-sensor data source(s) or, implicitly through the use of appropriate segmentation methods. For example, it
Table 1: Variables for integrated image segmentation
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 489
