SPECKLE FILTERING OF DUAL-POLARIZATION AND POLARIMETRIC SAR DATA
BASED ON IMPROVED SIGMA FILTER
Jong-Sen Lee1,2, Thomas L.Ainsworth1, Kun-Shan Chen2
1
Remote Sensing Division, Naval Research Laboratory, Washington DC, USA
2
Center for Space and Remote Sensing Research, National Central University, Chung-Li, Taiwan
ABSTRACT
The advancement of SAR technology with high resolution
and multiple polarization data demands better and efficient
speckle filtering algorithms. In this paper, we developed an
effective speckle filtering algorithm for dual-pol and fully
polarimetric high resolution data. The proposed algorithm
is effective and computational efficient based on the
improved sigma filter [1]. ALOS/PALSAR and JPL
AIRSAR data were used for demonstration.
Index Terms— Polarimetric SAR, Speckle reduction,
Dual Polarization, Spaceborne SAR
1. INTRODUCTION
The advancement of SAR technology with high resolution
and multiple polarization data demands better and efficient
speckle filtering algorithms. Recently, several PolSAR and
Pol-InSAR speckle filtering algorithm has been developed,
notably by Lee at al [2, 3], Vasil et al. [4] and Lopez-
Martinez et al (IGARSS 2007), for speckle reduction of
polarimetric SAR and Pol-InSAR data. In this paper, we
developed an effective speckle filtering algorithm that is
computational efficient based on the improved sigma filter.
The original Lee sigma filter [5, 6] was developed, more
than 20 years ago, for a single SAR image based on the
simple concept of two sigma probability, and is reasonably
effective in speckle filtering. However, deficiencies were
discovered in producing biased estimation, in blurring and
in depressing strong reflected targets. Recently, the original
sigma filter has been improved by eliminating these
deficiencies. The bias problem was solved by redefining
the sigma range based on the speckle probability density
functions. To mitigate the problems of blurring and
depressing strong reflective scatterers, a target signature
preservation technique was developed. In this paper, we
extended the improved sigma filter to filter the dual-
polarization and polarimetric SAR data.
2. BASIC PRINCIPLE OF SPECKLE FILTERING OF
DUAL-POL AND QUAD-POL SAR DATA
978-1-4244-2808-3/08/$25.00 ©2008 IEEE
Authorized licensed use limited to: UNIVERSITY TECNOLOGI MARA. Downloaded on December 29, 2009 at 15:21 from IEEE Xplore. Restrictions apply.
The principals [2, 3] underlined dual-polarization and fully
polarimetric SAR speckle filtering are
1) To preserve the statistical characteristics, each term
of the covariance or coherency matrix should be
filtered in a manner similar to multi-look
processing by averaging the matrix. In other
words, all terms of the matrix should be filtered
identically.
2) To avoid introducing cross-talk between
polarizations, each term should be filtered
statistically independent of other terms. Filtering
algorithms exploiting the degree of statistical
dependence between polarizations will introduce
cross-talk.
3) To preserve polarimetric properties, only
neighboring pixels possessed similar scattering
mechanism should be included in the filtering.
When filtering polarimetric SAR data, the first priority is
not to destroy the nice statistical structure inherent in SAR
data. If we filter each element of the coherency matrix
separately, for example, the correlation between
polarizations would be affected producing the bad effect
that the correlation coefficients are no longer preserved,
and, in the worst cases, they may be greater than the value
of 1. Also, following these principals, we will ensure the
filtered coherency matrix remains complex Wishart
distributed.
3. THE IMPROVED SIGMA FILTER
In principle, speckle reduction for distributed scatterers
requires averaging pixels within a homogeneous area. The
sigma filter was introduced by Lee [5, 6] in early 1980 as a
way of selecting homogeneous areas. During that time
frame, computers were expensive and the computational
efficiency and the memory requirement were the main
concerns. The Lee sigma filter has been implemented in
several GIS software, because of its simplicity, its
effectiveness in speckle reduction, and its computational
efficiency. However, the following deficiencies have been
discovered: 1) it fails to maintain the mean value, especially,
IV - 21 IGARSS 2008
 if the number of looks of the original SAR data is small
(less than 4 looks), 2) the highly reflective point targets are
blurred and their powers reduced, and 3) dark spotty pixels
are not filtered.
Recently, Lee et al [1] improved the original Lee sigma
filter by eliminating these deficiencies. The bias problem is
solved by redefining the sigma range based on the speckle
probability density functions. To mitigate the problems of
blurring and depressing strong reflective scatterers, a target
signature preservation technique is developed. In addition,
we incorporate the minimum mean square error estimator
for adaptive speckle reduction.
3.1. Bias compensation and dark pixels removal
For the single look amplitude and intensity SAR data, their
probability distribution is far from symmetrical, because
they have the Rayleigh and the Negative Exponentially
probability density functions (PDF), respectively that
produce biased estimation. Consequently, to remove the
bias, the sigma ranges were recomputed based on the
corresponding distribution functions. The idea is to preserve
the mean value. For the general case of the mean value of
1, the sigma range of the intensity SAR data is computed.
We have found that 90% sigma range works better than the
typical two sigma range of 95.5%. The result is shown in
Table 1.
Table 1 The revised 90% sigma range ( I1 , I 2 ) for 1-look to 4-
look intensity data
I1 I2
# of looks
1 0.084 3.941
2 0.221 2.744
3 0.313 2.320
4 0.378 2.094
We note that, due to the asymmetrical PDFs, the interval
( 1  I1 ), between I1 and the mean value of 1, is always
smaller than the interval ( I 2  1 ). The new sigma interval,
for any a priori mean value x ~ can be easily computed from
~ , I x~ ). The original sigma filter
these two tables as ( I 1 x
2
~ that
used the center pixel in a moving window as x
produces the isolated dark pixel problem. The improved
sigma filter applied the minimum mean square estimation
(MMSE) filter in the 3x3 window as estimate of the a priori
mean x ~ that effectively reduced the isolated dark pixels
and improved overall speckle filtering effect. The minimum
mean square filter has the form,
xˆ (1  b ) z  bz (1)
Authorized licensed use limited to: UNIVERSITY TECNOLOGI MARA. Downloaded on December 29, 2009 at 15:21 from IEEE Xplore. Restrictions apply.
Var( x)
with b , (2)
Var( z)
Var ( z )  z 2K 2
and Var ( x ) v (3)
(1  K ) 2
v
In (1), (2), and (3), the local mean z and the local
variance Var (z ) are the mean and the variance of pixels in
the sigma range, respectively, and Var ( x ) is the variance
of x , which is computed by (3). The parameter Kv is the
standard deviation of multiplicative noise of v , and is a
function of number of looks. For example, for 1-look
intensity, K 1 . For N-look SAR data, Kv is 1 / N .
v
In a moving window of 9x9, 11x11 or larger, pixels within
these new sigma ranges are included in the application of to
retain edges and fine details, and then the MMSE filters are
again applied to the selected pixels. However, when
applying the MMSE filter in the final filtering, we found
that the speckle noise standard deviation K v has to be
adjusted, because the valid range of PDF is limited by the
sigma range. The revised K~ as shown in table 1, has a
v
value smaller than the original K v . The data would be over
smoothed, if the original K v were applied.
~
K
v 3.2 Preservation of point targets
0,8191
0.5699 Preservation of signatures from point targets and manmade
0.4624 structures are desirable for image interpretation and other
0.3991 applications. These high return pixels are generally
produced by the double bounce scattering mechanism, or by
direct specular reflection. Backscattering signatures of
point targets are significantly different from that of
distributed media. We re-emphasize that backscattered
signatures from point targets are dominated by a small
number of strong scatterers, contrasting to the speckle
characteristics of distributed media. In other words, they do
not possess the typical speckle characteristics.
Consequently, they have to be processed differently from
backscattered signatures of distributed media.
The algorithm is based on the detection of strong scatterers
of several pixels in size. The first step of this algorithm is
to compute the 98 percentile of all pixels of the SAR data to
be filtered. We denote the 98 percentile as Z 98 . If the pixel
to be filtered has its value greater than Z 98 , we check the
number of pixels in a 3x3 window greater than Z 98 . If
IV - 22
 the total number greater than a threshold Tk , we consider
all those pixels belong to the point target, and their values
will be retained. In other words, they will not be filtered.
The threshold Tk is typically set between 5 and 7 for
general applications.

4. FILTERING THE COHERENCY MATRIX

For fully polarimetric SAR data, the filtering is based on the

coherency matrix T . The basic steps of the proposed
filtering algorithm are given as follows:

Step 1: Bright point target preservation

Strong point targets are detected by |HH-VV|2 for double
bounce and |HH+VV|2 for specular returns. The cross-pol
term is not used for point target detection, because point (A) Original (4-look, 256x256)
targets have low returns in HV. In a moving window (9x9,
11x11 or larger), the center 3x3 window is used for point
target detection. The 98 percentile and the threshold
Tk 5 are used for the detection from T11 and T22 images.
Then the detected point targets are combined. These pixels
will not filtered, or

T T (4)
Step 2: Estimate the a priori mean
For pixels from distributed media, the a priori means of
~ ~ ~
T11 , T22 and T33 are estimated by the MMSE filter in 3x3
windows with the regular standard deviation to mean ratio
(for example, Kv = 1 for single look).

Step 3: Select pixels in sigma range

T11 , T22 and T33 are applied to select homogeneous pixels (B) 5x5 Boxcar
within their established sigma boundaries as shown in Table
1. Only pixels selected by all three polarizations are
included in the filtering. Other pixels are considered as
outliers, and will be ignored. All elements of the coherency
matrix are filtered equally and separately by

T (1  b)T  bT (5)
where the weight b is a scalar computed based on the span
of the selected pixels using (2) and (3). The revised K~ in
v
Table 1 is applied here.

5. EXPERIMENT RESULTS

To illustrate the effectiveness of the proposed PolSAR

filtering algorithm, JPL AIRSAR data of San Francisco was
applied. The data were four look processed. For clarity,
only a small area of 256x256 pixels were extracted for (C) The proposed method
illustration. Fig. 1 Comparison of filtering results. The Pauli vector color
display is applied here: for |HH-VV|, green for |HV|, blue for
|HH+VV|.

IV - 23

Authorized licensed use limited to: UNIVERSITY TECNOLOGI MARA. Downloaded on December 29, 2009 at 15:21 from IEEE Xplore. Restrictions apply.
 The color composition using Pauli vector is applied here
with red for |HH-VV|, green for |HV|, blue for |HH+VV|.
The original is shown in Fig.1A revealing various scattering
mechanisms: city blocks shows high returns from double
bounce and specular scatterings, ocean areas from surface
scattering, and park areas from volume scattering. Figure
1B shows the results of a 5x5 boxcar filter revealing
considerable blurring of the PolSAR image. The result
from the proposed method is shown in Fig. 1C. Compared
with the original and the 5x5 boxcar result, the proposed
method clearly shows effective speckle reduction while
retaining distinguishable details and strong point targets.
(A) Original ALOS/PALSAR image (512x512)
(B) Improved sigma filter (9x9 window)
Fig. 2 Speckle filtering of ALOS/PALSAR of Tomakomai, Japan.
Color composition is based on Pauli vector.
We also applied the improved sigma filter to the
ALOS/PALSAR data of Tomakomai, Japan. As shown in
IV - 24
Authorized licensed use limited to: UNIVERSITY TECNOLOGI MARA. Downloaded on December 29, 2009 at 15:21 from IEEE Xplore. Restrictions apply.
Fig. 2, it clearly demonstrated the effectiveness of this
filtering algorithm.
6. SPECKLE FILTERING OF DRAL-
POLARIZATION SAR DATA
The recent successfully launches of ALOS, RADARSAT-2
and TerraSAR-X produce large volumes of high resolution
dual-polarization SAR data. For coherently processed dual-
pol data, such as (HH and HV), (VV and VH) and (HH and
VV), it is important to preserve the correlation between
polarizations just as for fully polarimetric SAR data. The
proposed algorithm can easily adopted for dual-pol case
based on the 2x2 covariance matrix. The point target
detection will be based on co-pol intensity data, and both
polarization data are to be used for filtering pixels from
distributed media.
7. CONCLUSION
We have extended the improved sigma filter for speckle
filtering fully polarimetric and dual-polarization SAR data.
The proposed algorithm will preserve polarimetric property
and retaining strong target signatures. This algorithm is
simple in concept and computationally efficient. It is a
good alternative to the Refined Lee PolSAR filter.
REFERENCES
[1] J.S. Lee et al,“Improved Sigma Filter for Speckle Filtering of
SAR Imagery”, IEEE Transactions on Geoscience and Remote
Sensing (in press)
[2] J.S. Lee, M.R. Grunes and G. De Grandi, "Polarimetric SAR
Speckle Filtering and Its Implication for Terrain Classification" IEEE
Transactions on Geoscience and Remote Sensing, vol. 37, no. 5,
2363-2373, September 1999.
[3] J.S. Lee, D.L. Schuler, M.R. Grunes, E. Pottier, and L. Ferro-
Famil, “Scattering Model Based Speckle Filtering of Polarimetric
SAR Data,” IEEE Transactions on Geoscience and Remote Sensing,
vol.44, No. 1, 176-187, January 2006.
[4] G. Vasile, E. Trouve, J.S. Lee and V. Buzuloiu, “Intensity-
Driven-adaptive-neighborhood Technique for Polarimetric and
Interferometric Parameter Estimation,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 44, no. 6, 1609-1621, June
2006.
[5] J.S. Lee, "A Simple Speckle Smoothing Algorithm for Synthetic
Aperture Radar Images," IEEE Trans. on Systems, Man and
Cybernetics, Vol. SMC-13, No. 1, pp.85-89, 1983.
[6] J.S. Lee, "Digital Image Smoothing and the Sigma Filter,"
Computer Vision, Graphics and Image Processing, Vol. 24, pp. 255-
269, 1983.