'gions on
)rocesses
ather and
e of high
ialysis of
nd quad-
erospace
“built-up
cedure in
heme for
y results.
ly of the
Ig parks,
Ryherd,
based on
es have
De Kok,
m, 2002;
; provide
tural and
cts 1n the
te (Roth,
t for the
solution,
Is placed
h respect
and the
1 mostly
first part
ncept is
oach are
en.
o detect
irised X-
International Archives of the Photogrammetry, Remote Sensing and Spau! Information Sciences, Vol XXXV, Part B7. Istanbul 2004
2.1 Basic Concept
The methodology consists of two main steps:
e the preparation of the initial data including speckle
suppression and the generation of an additional texture
layer
e the image segmentation and classification.
Subsequent to an initial speckle suppression by means of an
optimised filtering technique a texture layer is generated.
Along with the despeckled intensity image this texture layer is
used for the segmentation of the data set. The classification
itself is based on the despeckled intensity image and the
original data. Both, the image segmentation and the
classification are performed on the basis of the object-oriented
image analysis software eCognition (Baatz & Schäpe, 1999).
This technique partitions the complete scene into groups of
spectrally similar pixels by means of an image segmentation to
form image objects on an arbitrary number of scale levels. The
analysis of these objects instead of single pixels is particularly
applicable for the classification of very high resolution imagery.
It facilitates the distinction between different structures and
their function by considering spectral, geometric and textural
characteristics along with information about properties of the
area surrounding the image objects. Moreover the object-
oriented approach is capable of displaying different objects in a
single image at various scales. With respect to radar imagery
the initial segmentation also helps to reduce the effects of radar
speckle because the intensity of the generated segments results
from an averaging of the underlying pixel values.
Subsequent to image segmentation a "knowledge base" is
created. This knowledge base defines both, the classes to be
identified and the according features for their description and
classification. The rules for the class description might include
spectral and textural characteristics of the objects, the
hierarchical context of the segments or the relationship between
neighbouring objects. According to these predefined rules the
image is analysed and classified by means of fuzzy logic or a
nearest neighbour algorithm (Baatz & Schápe, 1999). Each step
performed during the image analysis can be stored in a separate
protocol which might then be applied to other data
automatically.
2.2 Test Site and Data Set
The X-band SAR imagery was acquired over the German cities
of Ludwigshafen and Mannheim in May 2003 by the airborne
Experimental Synthetic Aperture Radar (E-SAR) system of the
DLR (Moreira, Spielbauer, Pótzsch, 1994). Both cities are in
the Rhine-Neckar region, which represents Germany's 7"
largest conurbation.
During the flight campaign single polarised X-band, dual-
polarised C-band and fully polarimetric L-band data was
recorded along three flight tracks featuring a depression angle
of 20° - 60°. Each track covers an area of 3x10km in a spatial
resolution of 2-3m. The recorded images feature a large variety
of urban, suburban and agricultural structures.
In order to validate the image classification some classes of a
biodiversity GIS vector layer updated in 2000 were merged into
a data set showing the built-up areas within the specified
region. This data base was complemented by additional aerial
photographs recorded during the radar flight campaign.
479
3. SETTLEMENT DETECTION
3.1 Data preparation
In initial experiments it became apparent that a segmentation
based solely on the initial intensity image is difficult. First,
significantly textured medium- or small-scale structures are
often not recognised or reconstructed by the segmentation. At a
low segmentation level consisting of small image objects these
structures are often split up into several individual segments
with distinctively differing backscatter values. Increasing the
size of the objects frequently results in a fusion of the textured
segments with those adjacent objects, whose backscatter is quite
similar. Sometimes the structure isn't even noticed at all.
Moreover the shape of the resulting image objects is frayed and
does not follow the actual boundaries satisfactory.
As a proper image segmentation is crucial for the subsequent
classification we improved this work step by supplementing the
intensity information with an additional texture layer (Mean
Euclidian Distance) calculated for the intensity image. Since
the speckle effect clearly constricts the computation of a
meaningful texture the speckle of the initial intensity image has
to be minimised first. Strong speckle suppression with
conventional speckle filters smoothes edges to a certain degree.
Thus, we developed an optimised filtering technique to preserve
major edges while significantly smoothing homogeneous areas.
This moving window filter represents a combination of a
"selective" mean filter and a conventional Lee Sigma approach.
The effect of this filter is shown in Figure 1.
Figure 1. Speckle Suppression (A: original ; B: filtered)
