edge
hing
ision
men-
ional
sing,
stem.
ence
ata",
ligital
netry
übler,
roject
Work.
Jress,
n Still
sung,
point
t No.
ETH
semi-
metric
ce on
5, 9t
o 3D-
rift für
, 193-
(try for
jects.
emote
D:
on by
ited to
|, 9-19
ctronic
21(4),
anding
Digital
eipke),
ectural
action.
onsing,
ophoto
on the
tations,
THE EPOS SPECKLE FILTER: A COMPARISON WITH SOME WELL-KNOWN SPECKLE REDUCTION
TECHNIQUES
Wilhelm Hagg, Manfred Sties
Institute of Photogrammetry and Remote Sensing
Karlsruhe University
Germany
Email: {hagg,sties}@ipf.bau-verm.uni-karlsruhe.de
Commission Il, Working Group 4
KEY WORDS: Remote Sensing, Radar, SAR, Speckle, Multiplicative Noise, Edge Preserving Filters, Comparison
ABSTRACT
Any classification process using SAR images presupposes the reduction of multiplicative speckle noise, since the variations
caused by speckle does not allow a distinction between neighboring classes within the feature space. This may be done by
smoothing the images with digital filter algorithms, which removes the high frequent noise but also causes distortions at the
high frequent image contents, i.e. sharp edges. Several adaptive filter algorithms have been developed, which aim at the
preservation of edges and single scattering peaks, while homogeneous areas are smoothed as much as possible. This task is
rendered more difficult by the multiplicative nature of the speckle noise: the signal variation depends on the signal itself. The
recently developed EPOS speckle filter is compared with other well-known algorithms in this paper. In order to enable an
objective comparison, the smoothing capability of all filters is adjusted to a similar value. To achieve a measurement for the
quality, speckle is added synthetically to an image, so it is possible to calculate the RMS-error for each filtering method. Since
the RMS differ according to the image contents, typical areas for several geometric objects are used to calculate the RMS.
Also different signal to noise ratios are take into account in the comparison procedure to achieve an exhausting overview of
the algorithm performance.
1 INTRODUCTION solve for an operational SAR image interpretation.
The availability of optical remote sensing data for landuse = Two categories of speckle reduction techniques are distin-
applications is limited by the local weather conditions, espe- ^ guishable; (1) the averaging of several looks of the same
cially by clouds. In addition illumination effects due to differ- ^ scene (multi-look processing), and (2) the smoothing of the
ent sunsets aggravate the interpretation of optical datasets. ~~ image using digital image processing techniques. The multi-
Therefore “Synthetic Aperture Radar” (SAR) systems were look processing is limited by the geometric resolution of the
developed which use their own, well defined, microwave il- SAR instrument and the required resolution of the final im-
lumination to penetrate clouds in the atmosphere. Another age. For the geocoded standard ERS-1 products three inde-
reason to develop SAR systems is the total different backscat- ^ pendent looks are averaged to achieve a final resolution of
ter behavior of microwaves due to the long wavelength of ^ about 25 meters. Both techniques use the fact, that averag-
microwaves relative to optical systems. There is a lot of ing of several independent samples of a measurement reduces
operational SAR systems on satellite platforms available as the variance of a signal. In the first case, several images are
the European ERS-1 and ERS-2, the Japanese JERS-1 and averaged, while in the second case values of neighboring pix-
the Canadian RADARSAT. In addition several experimental els are averaged. The most simple way of smoothing images
systems were developed for a flight on board of a space shut- ^ with digital image processing techniques is to apply a mean
tle. Recently the X-SAR/SIR-C mission resulted in the first filter within a moving window around each image pixel. The
spaceborn multi-frequency images of the earth surface; the result will be a smoothed image with reduced speckle vari-
instruments operate in the L- C- and X-Band. Also there ex- ation, but edges and single point scatterers, as they appear
ists a lot of airborne SAR-systems for experimental purposes. in urban areas, are smoothed as well. Adaptive filter algo-
rithms have been developed which aim at the preservation of
edges and single scattering peaks, while homogeneous areas
are smoothed as much as possible. For that purpose the fil-
tering function has to be adapted to the local image contents
to reduce the geometric distortions. It is obvious that the
correction of a grey value, using the pixel values in the neigh-
borhood, is efficient only in those cases where the neighboring
pixels represents the same object on the ground.
While the SAR-systems and SAR-processors are in an oper-
ational state, the interpretation of SAR images is still under
development due to several problems such as speckle noise
and illumination effects in undulated terrain. SAR-processors
are necessary to reconstruct the image from the synthetic
aperture by coherently processing the returns from successive
radar pulses along the flight path, and therefore they may
be interpreted as the software component of the SAR instru-
ment. Speckle noise is a system made consequence of the Most filter algorithms for speckle filtering are not fully sat-
coherent radar illumination and appears as a granular pattern isfactory for the purpose of SAR image classification, since
in the image. One of the most important feature of speckle the decrease of speckle caused variance is not sufficient for
noise is its multiplicative character which links the amount of a distinction of neighboring classes within the feature space.
noise to the signal intensity. The strong distortions of SAR Thus we developed a filter called EPOS (Edge Preserving Op-
images by speckle noise often prevent an useful application of timized Speckle-filter) which is published in (Hagg and Sties,
SAR images in remote sensing. Thus, an effective reduction 1994). The idea for the new algorithm was to adapt an area
of speckle noise is one of the most important problems to geometry to the image contents in such a kind, that the as-
135
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
