2004
del is
e. the
owest
-high,
ective
anned
echo-
TOW,
while
[S are
owest
ently,
| the
dges.
c
rough
upper
upper
left,
minus
ensed
"nitive
clude
by a
tation.
ie. ît
model
time.
t also
tation
icular,
on the
ics as
y Thé
(like
s) has
nsible
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
With the advent of multi-sensor systems as presented in the
previous section, the number of extracted scene features
increases. To deal with this large amount of information and
eventually inherent contradictions, a fusion process on various
levels becomes necessary. While most efforts in the past have
been laid upon fusion on signal level, in our context a fusion
on feature level is of great importance in order to integrate the
several information in the sense of perceptual organisation as
described above. Hence, in the following section we want to
demonstrate a concept and an implementation of such a feature
level fusion process for multi-sensor data for the particular
application of extracting and classifying topographic surface
edges.
Scene interpretation ^ Cognitive perception
| knowledge base | memory
| knowl. representation | mem. representation
object recognition (classification)
A A
perceptual organisation
A X
feature extraction
1 t
remote sensing stimuli
suistueqoour xoeqpoa]
scene representation
| signal | [^ stimuli |
Figure 3. Analogy between the processes of remote sensing
scene interpretation and cognitive perception
4. SURFACE EDGE EXTRACTION
4.1 Definitions
For the description and interpretation of topographical surfaces
those curves are of major interest that represent either local
maxima (ridge lines, watersheds), local minima (valley lines)
or the border between surfaces with significantly different
gradients. In the following we will concentrate on the latter
type, which will be termed here surface edges and can be seen
as a subset of the Digital Surface Model (DSM). Hence,
surface edges combine the ^hard" edges of topographical
objects (object edges, like those of buildings or vegetation) and
of terrain edges (as a subset of the Digital Terrain Model,
DTM, like embankments, ditches, etc.). It shall be noted that
the commonly used term of breaklines is strictly not correct
because those represent only particular edges which had been
generated through geomorphologic processes (Brunner, 1985).
Surface edges represent either an abrupt gradient change only,
or they build complex objects which might consist of a lower
and upper edge as well as a surface in between (like with an
embankment). In principle surface edges are modelled by 3D-
vectors, whereas for a couple of applications (like
topographical maps) a 2D-ground plan representation is
sufficient. Unfortunately, from a modelling point of view,
generally accepted quantitative criterions for surface edges (in
particular thresholds for surface gradients) do not exist. As one
example, the German ATKIS system specifies only the height
and length of the object but no gradient value for capturing
embankments.
4.2 Application potential
Surface edges can be seen as value added information and an
improvement to any given Digital Elevation Model (DEM).
Typical applications using edge information are for example
flood prevention and river and drainage management, where
characteristic lines for hydrological/hydraulic models are
needed, the inspection survey of power lines, or the generation
of 3D city models. Furthermore surface edges define the
outline of so-called reduction surfaces which have to be
masked out from DSMs in order to derive DTMs in the process
of a DSM normalisation (Schiewe, 2003). Finally, they can
significantly contribute to a reduction of data amount of very
densely measured or interpolated DEMs.
4.3 Previous work
As regularly mentioned in the literature, the (semi-)automatic
derivation of surface edges from irregularly or regularly spaced
elevation points has led to unsatisfying results so far (e.g.
Petzold et al., 1999; Pfeifer & Stadler, 2001). Obviously this is
mainly due to the still limited quality of the input data in terms
of the spatial resolution or point density as well as the
geometrical accuracy in the vertical and horizontal components.
On the other hand, these limitations will not be valid anymore
in the near future with a certain probability, or can even be by-
passed with some technical efforts nowadays. Hence,
advancements with suitable algorithms for the extraction of
surface edges are of great importance.
One of the major contributions for automatic surface edge
detection in the past came from Wild et al. (1996) who applied
an adaptive edge preserving filter in the process of the DEM
generation before extracting edge pixels through gradient
filters (e.g. a Sobel filter). Brügelmann (2000) used the second
derivatives and hypothesis testing to derive regions of break
points which then had to be further processed. Kraus & Pfeifer
(2001) describe the derivation of 3D structure lines which uses
the pre-knowledge of an approximate ground plan of the edges.
However, it has be stated that these and other algorithms (like
the ones of Chakreyavanich, 1991, or Gaisky, 2000), which are
based upon geometrical information only, cannot compete with
the manual, photogrammetrical measurement of surface edges
in terms of completeness and accuracy.
4.4 Methodology
4.4.1 Core idea and outline: In contrast to other algorithms
our proposed multi-sensor data fusion approach for the
extraction of surface edges differs with respect to the following
aspects:
eo We will not only use one single elevation data set but also
the various multiple reflections from a laser scanning
systems (as presented in section 2.2) in order to increase
