The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B4. Beijing 2008
279
levels of abstraction of knowledge (Mayer & Steger, 1998). The
extraction benefits from using different scales, since initial
hypotheses can be efficiently formed in coarse scale, which are
validated and tested in fine scale. Vice versa, intermediate
results obtained in fine scale can be evaluated with respect to
their importance for the global functional properties of the
objects described in coarse scale.
Recently, (Drauschke et al. 2006) analyzed the scale-space
behaviour of building features. They show that there is no
particular optimum scale for building reconstruction, yet they
observe a stability of the features over certain intervals in scale-
space. Thus, a low number of selected scales seems to be
sufficient for building extraction. In the same spirit (Forberg,
2004; Forberg & Mayer, 2006) show how the behaviour of
buildings in different scale-spaces can be used for their
generalization in 3D. They use so-called scale-space events in
the morphological and curvature space to reduce the level of
detail of buildings for cartographic generalization. For road
extraction, usually two (Baumgartner et al. 1999, Hinz &
Baumgartner 2003) but in some special cases also three (Mayer
1998, Hinz 2004a) different scales have been used. While the
coarse scale, i.e., 2m to 4m, allows to extracting roads based on
their fundamental characteristics and functional properties -
such as curvilinear shape, network characteristics, and optimum
routes between certain places - the fine scale, i.e., less than 50
cm, enables detailed analysis and precise geometric delineation
of the road layout. (Heuwold 2006, Heuwold & Pakzad 2006,
Heuwold et al. 2007) aim at generalizing this issue for a quasi-
continuous scale space in 2D. The results shown in (Heuwold
2006) are promising for the case of parallel linear structures as
they often appear in the context of road extraction, but they also
show that this kind of fundamental research is far from being
solved in general.
Remote sensing can contribute significantly to all these
components in different ways, especially because of the large
coverage of remotely sensed imagery and its global availability.
Time, however, is the overall dominating factor once a disaster
has hit a particular region. This becomes manifest in several
aspects: firstly, available satellites have to be selected and
commanded immediately. Secondly, the acquired raw data has
to be processed with specific signal processing algorithms to
generate images suitable for interpretation, particularly for
Synthetic Aperture Radar (SAR) images. Thirdly, the
interpretation of multi-sensorial images, extraction of
geometrically precise and semantically correct information as
well as the production of (digital) maps need to be conducted in
shortest time-frames to support crises management groups.
While the first two aspects are strongly related to the
optimization of communication processes and hardware
capabilities (at least to a large extend), the main methodological
bottleneck is posed by the third aspect: the fast, integrated, and
geometrically and semantically correct interpretation of multi-
sensorial images.
Concerning the automated extraction of objects, special focus is
on the extraction, analysis and characterization of infrastructural
objects like roads and buildings due to their importance for the
immediate planning and implementation of emergency actions.
In the following, we concentrate on the detection of intact road
networks under the constraints of an effective disaster
management. This means, that those parts of the road network,
which are not detected by the system, should be regarded as
destroyed and not accessible for rescue and evacuation teams
anymore. A similar approach for buildings can be found in
(Rehor & Bahr, 2006), where the goal is to decide if buildings
in laser-scanner data are damaged, e.g., by an earth quake, and
what kind of damage has occurred.
Many of the discussed aspects are highly relevant for the current
activities directed towards the design and implementation of an
image understanding system for disaster management at the
German Aerospace Center (DLR). The scientific challenges are
moreover accompaigned with requirements of an operational
environment, reasonable processing times and efficient
workflows. The next section will outline this in more detail.
3. OBJECT EXTRACTION FOR DISASTER
MANAGEMENT
There is no doubt, that remote sensing methods can provide
valuable support for activities directed towards the prevention
of natural hazards and managing the consequences of natural
disasters. Techniques for automatic object extraction and
change detection consequently play a major role in this context.
However, in addition to the scientific and methodological
challenges, the approaches to object extraction must be integral
part of a comprehensive system of actions, documentations and
planning activities during, after and also before a (potential)
disaster. This is only possible if the image understanding
methods are embedded into well-defined strategies for
supporting the preparedness and prevention of hazards, for fast
reaction in case of disasters, as well as for damage
documentation, planning and rebuilding of infrastructure after
disasters.
In the following, we link the description of the system’s
underlying concept and methodologies with the discussion of
the issues of object extraction posed at the beginning: (1)
system design and different levels of automation; (2) automatic
extraction under the light of available sensors and data
characteristics; (3) effectiveness/processing time and the need
for internal evaluation; and (4) quality of the final results and
verification/post-editing.
3.1 System design: Different levels of automation
At the moment, fully automatic approaches for object extraction
must stiil be regarded as a subject of fundamental research, and
they seem not to be able to find their way into operational work
flows in the very near future. On the contrary, semi-automatic
approaches seem more likely to be useful in operational
applications. Automatically achieved results nonetheless may
provide a basis for efficient checking, editing and improving.
Hence, the framework for the automated detection of intact
roads from multi-sensorial imagery is conceptually divided into
three main parts. The first part comprises the (fully-) automatic
extraction of roads. The results will be of course not 100%
complete and correct. It follows therefore an internal evaluation
of the automatically achieved results as second module. This
provides measures about the reliability of road parts, which
should guide an operator during editing the results. This editing
phase - the third part - involves user-assisted tools to support
effective checking and completing the results.