——— zum
sumption of a homogeneous area is given, which contains just
variance caused by speckle. From that area we simply choose
the mean as a new value for the filtered image. In order to de-
termine homogeneity, the a priori knowledge of the coefficient
of variation — that is the standard deviation related to the
local mean — is necessary. This coefficient remains constant
in homogeneous regions, where it is fully determined by the
amount of speckle within the image. To find homogeneity
even in heterogeneous regions and near edges, the averag-
ing area is constructed from eight non-overlapping triangles
around each pixel. This is done by successive elimination of
triangles with the highest coefficient of variation, i.e. those
containing an edge. If no homogeneous triangle was found,
the observation window is reduced in its size and the pro-
cedure starts again. For single scattering peaks the area is
reduced to one pixel and therefore no filtering is applied to
those pixels. This enables a strong reduction of variation even
in the neighborhood of edges and the preservation of edges
as well as single scattering targets. The initial size of the
observation window may be large, because it is reduced by
the algorithm if necessary. Therefore the initial size of the
window has less influence on the filtering result, if it is just
large enough to enable an efficient reduction of the variation.
This filtering algorithm will be compared with other filters
in order to demonstrate the efficiency of different filters for
individual tasks.
Most papers dealing with comparison of speckle filters use
subjective criteria in order to compare the algorithms. Ob-
jective criteria are very hard to find, since the filters are adap-
tive to the signal and therefore measurements on a standard
signals, as the impulse response of the filter, are not char-
acteristic for the performance of the algorithm. In order to
approximate an objective performance criterion we first an-
alyze the requirements to speckle reduction. Since the dis-
tortions of edges and points within a filtered image increases
with the decrease of noise, the amount of signal variation
found in the filtered image is adjusted to a similar value for
all algorithms. Most of the papers dealing with comparison
of filters show the decrease of noise and the preservation of
the image contents separately, thus an objective comparison
is not provided. 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 edges, lines and points are used to calculate the RMS. In
addition different signal to noise ratios are used for the com-
parison to achieve an exhausting overview of the performance
of the algorithms.
2 REQUIREMENTS OF SPECKLE FILTERING
As mentioned in the introduction, the rating of speckle filter
performance using objective criteria is quite difficult, since
the behavior of the adaptive filters used is extremely sensitive
to the image contents. This results in a wide field of possible
measurements which may be used as comparison criterion.
Thus we first have to analyze the requirements to filter algo-
rithms and derive comparison rules, in order to create rating
criteria useful for practical applications.
In this paper we deal with landuse mapping as a frequently
occuring remote sensing application. The main problem of
the classification using SAR images is the spectral similarity
of several classes. Meadows and water have similar signa-
tures if the water surface is rough. Also the signatures of
some loosely populated areas are similar to forest signatures.
Conditioned by the high amount of speckle noise in SAR im-
ages, those classes may not be separated in the feature space
which leads to an unacceptably high degree of misclassifi-
cation. For this reason we found the main criterion for a
speckle filter is to reduce the amount of speckle variance
drastically. Using a mean filter with a 10 by 10 pixel averag-
ing region will produce a higher accuracy as the classification
of the original, speckled data. The loss of some geometric
details is compensated by a much better distinction of class
signatures, they appear more compact in the feature space.
Therefore the radiometric image quality proofs to be the most
important criterion for landuse mapping applications. On the
other hand geometric distortions decrease the classification
accuracy especially in heterogeneous regions containing rela-
tively small fields of common semantic on the earth surface.
Geometric objects may be grouped in areas, lines and points.
Areas are typically build by classes as forest, water, meadow
and agriculture, lines results from roads, railways and rivers.
Points appear in urban areas as a result of double-bounds
reflections and—depending on the resolution of the SAR—in
other textured regions. Thus it is obvious that areas cover
most of a SAR scene, followed by points and lines. Accord-
ing to the above geometric primitives, geometric distortions
appear at edges between areas, lines and points within the
image. Regarding just the edges of areas instead of the ar-
eas itself will also result in the fact, that edges cover much
more of the image than lines and points. Of course this fact
depends strongly on the area mapped, but it holds in large
regions not just containing a local phenomenon like an urban
area. Furthermore the distortions located at points may be
compensated by the use of texture features in the classifi-
cation process. Texture features are described in (Haralick,
1978), (Hagg et al., 1995) and many other publications.
Recapitulating this section, the radiometric enhancement of
SAR images by reducing the speckle variance is the primary
task to solve by speckle filter algorithms, while the problem
of geometric distortion proofs to be secondary. The focus of
interest regarding the geometric primitives is the distortion at
edges between areas. Line and point features may decrease
the classification accuracy less and therefore they may be
rated more laxly in a comparison criterion. Depending on
the application, other criteria may be suggestive, as it is for
the extraction of linear features as roads from an image. For
applications dealing with areas, the above mentioned criteria
may hold generally.
3 FILTERING ALGORITHMS
Adaptive speckle filtering algorithms may be separated in two
categories; (1) statistical algorithms, using the local statistic
within the moving window to adapt the filter to the image
contents and (2) geometric algorithms, which take into ac-
count the signal at different angular directions around each
pixel. In opposition to one dimensional signals, where each
sample value has just two neighbors, a pixel in an image is
strong related to its environment. In addition to the distance
relation of a one dimensional signal, the angle is a second
relation for images. At a distance of one pixel, 8 different an-
gels are distinguishable; in general for a n pixel distance 8n
different pixels are related to the central pixel. This strong
embedding of each pixel enables the extraction of information
from different angles in order to optimize the filter adaption
136
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
to 1
In Si
199
pap
bre
algc
EPI
GE
R-L
and
FRI
E-F
LEE
KU
G-N
This
obje
crite
subj
com
sign
on ¢
perf
ima
may
the
tot
obse
tern
geo!
geor
diffe
man
of e
the
4.1
The
smo
ject:
filte
mat
On
is e
the
dist
the
with
try 1
but
orde
one
one
