You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Fusion of sensor data, knowledge sources and algorithms for extraction and classification of topographic objects
Baltsavias, Emmanuel P.

International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
original CIR imagery based on the available DSM. Figure 5
shows the resulting ortho-image of the small test area.
3.2. Combination of Multisource Data
A number of techniques, aiming at the combination of
multisource data for scene labelling, are available. A review of
concepts and ideas for utilization of additional datasets in
multispectral classification procedures is for example given in
(Hahn and Stätter (1998)). Generally, for our application two
approaches are feasible. The available multispectral and
geometric information can be combined applying the additional
channel concept or the hierarchical classification approach.
In hierarchical classification, the different types of data are
applied in order to successively divide the working area into
more detailed object classes (Savian and Landgrebe (1991)).
Figure 6 shows the subdivision of the scene into vegetation
(black) and non-vegetation (white) areas. This step was
performed based on the analysis of the CIR ortho-image.
Fig. 6. Vegetation and non-vegetation regions extracted
from CIR aerial image.
In the next step of the hierarchical classification, these areas can
be further subdivided based on the laser scanning data using the
information on the local height above ground, which is
provided by the normalized DSM. The vegetation regions can
be separated into tree regions (high values of the normalized
DSM) and other vegetation like grass-covered areas (low values
of the normalized DSM). Accordingly, the non-vegetation areas
can be differentiated into buildings (high areas) and non
building regions like streets (low areas).
The main problem of this hierarchical or layered classification is
that classification errors of the first step are propagated to the
subsequent steps. Furthermore, the additional channel concept
enables a more flexible processing of the data. This is the
reason, why we prefer the use of this concept in our approach.
3.3. Additional Channel Concept
The main objects we are interested in are buildings, streets, trees
and grass-covered areas. In order to demonstrate the
insufficiencies of a standard classification, which is restricted to
the analysis of multispectral information, a maximum likelihood
classification was applied to the CIR ortho-image. The result of
the separation into the required object classes is depicted in
Figure 7. For this and all other classification examples, the
training areas required in order to obtain the spectral
characteristics of the different landuse classes were digitised
manually. Based on this information, the pixels are assigned to
one of the predefined classes in the classification stage.
Alternatively, the training can be derived automatically from
already existing GIS databases (Walter (1998)).
Buildings HH Trees
Roads 1 1 Grass-covered
Fig. 7. Maximum likelihood classification based on CIR
Figure 8 depicts the classification result again by a standard
maximum-likelihood classification. However, in this case, the
normalized DSM was introduced as an additional channel in the
classification and thereby combined with the multispectral
channels. Figures 7 and 8 demonstrate very well that in an