You are on page 1of 5

IEEE TRANSACTIONS ON GEOSCIENCE A N D REMOTE SENSING, VOL. 32, NO. 5 . SEPT.

1994

Ill0

Communication
The Principle of Speckle Filtering in Polarimetric
SAR Imagery
R. Touzi and A. Lopes
Abstract-The principle of speckle reduction in polarimetry is reconsidered. It is shown that polarimetric data can be speckle reduced if
and only if all the elements of the Mueller matrix are filtered, which is
equivalent to filtering the scattering vector covariance matrix. Assuming that speckle is multiplicative and stationary, the algorithms proposed in [l] and [2] are extended to filter the covariance matrix of reciprocal and nonreciprocal targets on one-look and multilook images.
The problem of estimation of the first- and second-order statistics of
the four-channel speckle vector is discussed, and a solution is proposed
for one-look and multilook images.

I. INTRODUCTION
For a better discrimination of scene targets and an easier automatic image segmentation, speckle reduction is required on radar
images. Several filters (Novak and Burl [3], Lee et al. [l]) have
been recently proposed for speckle reduction in polarimetric images. Combining the three' (complex or intensity) measured channel (HH, HV, VV), these filters output one-channel intensity image
with reduced speckle. In addition to these scalar filters, vectorial
filters have been developed for polarimetric data in [ l ] and [2].
Under the multiplicative noise speckle assumption, and using only
the intensity of the three measured channels, the Lee et al. algorithm [ 11 outputs a tridimensional vector whose components are the
filtered intensity channels (HH, HV, VV). Goze and Lopes [2] extended, for one look image, the Lee's study to filter in addition to
the three intensities, the three double products of the complex scattering channels (HH . HH*, HH . HV*, HV . VV*), thus preserving the phase information. However, all the scalar filters and
most of the vectorial filters do not preserve the polarimetric information, and the concepts of speckle filtering in polarimetry have
to be redefined.
In the following, the principle of speckle filtering in polarimetry
is reconsidered. It is shown that all the usual polarimetric entities
(such as the received intensity for any combination of transmitting
and receiving antenna polarizations, the scattered intensity, the degree of polarization) can be speckle reduced if and only if all the
elements of the Mueller matrix are filtered. A minimum mean
square error (MMSE) filter, which is an extension version of the
algorithms developed in [ l ] , [2], is introduced in Section 111 for
one-look and multilook images. The filter which does not assume
target reciprocity can easily be adapted (and thus simplified) to the
reciprocal case. Finally, the problem related to the estimation of
signal and speckle vector local statistics of the first and second
order is discussed, and a solution is proposed for one-look and multilook images.
Manuscript received October 26, 1993; revised March 29, 1994.
R. Touzi is with Colltge Militaire Royal de Saint-Jean, Qutbec, Canada.
A . Lopes is with Centre d'Etudes Spatiales de Rayonnement, CNRS/
UPS, Toulouse, France.
IEEE Log Number 9403640.
'Target are assumed of reciprocal backscattering (HV = VH).

11. PRINCIPLE
OF SPECKLE
FILTERING
I N POLARIMETRIC
SAR

IMAGERY
An imaging radar polarimeter is a system which permits measurement of the full polarization signature of every resolution element. Four complex scalars are measured almost simultaneously at
the linear polarizations (HH, HV, VH, VV). These four elements
of the scattering matrix [SI permit the calculation of the received
power for any possible combination of transmitting and receiving
antenna polarization [4], [5] :

P,,

K (,?rec[S],?tr]2

(1)

where K is fu_nction of+the antenna gains, the wavelength and radar


parameters. E"' and E" are the electric field of the receiving and
transmitting antenna, respectively.
The quadratic form of the P,, expression can be written under
the following linear form [4]:

P,,

= KZEC. [WI

Ztr

(2)

where is given as a function of the electric field components of


the transmitting and the receiving antennas by: g 7 = ( E ; E;,
E ; ) . The 4 x 4 matrix [W] is given by: LW]
ET+E;,+ET E;,
= S . S * T , where S is the vectorial form of the matrix [SI: S *'
= ( S H H ? Sfw, S V H , S v v ) .
Due to the coherent nature of the radar signal, the measured scattered signal suffers from fading caused by the constructive and destructive interference of the signal component reflected by individual scatters within a resolution cell. The effect of this fading is to
cause speckle in the four channel measurements of the radar polarimeter. Generally, the bandwidth of the signal can be considered
small compared to the bandwidth of radar, and speckle behaves like
a multiplicative stationary independent and white zero-mean complex circular Gaussian noise. If [ X I is a 4 x 4 diagonal matrix
whose diagonal elements are the four channel complex speckle realization elements of the veztor < ; = (nHH,nHV,nVH,nvv)', the
measured scattering vector 0 1s the product of the speckle matrix
[ X I by the scattering vector S. To take into account the speckle
effect, the matrix [W] in the expression above (2) should be replaced by the following one:
[w,] =

[XIS. S*T[321*7.

(3)

An ideal polarimetric filter should reduce speckle of the received


power for any polarization combination of transmitting and receiving antennas. If the received power expression is now analyzed,
the following points can be noticed:
The speckle can be filtered if [W,] is filtered and replaced in
(2) by the estimated matrix [I@] of the unspeckled signal matrix [W].
The multiplicative speckle model is suitable for the four
measured linear channel intensities (HH, HV, VH, VV). For
the other possible polarization combinations of transmitting
and receiving antenna polarizations, the received intensity P,,
expression is a linear combination of the [W,] elements, which
cannot be written under the signal-speckle product form.
Hence, the multiplicative speckle assumption can only be used

0196-2892/94$04.00 0 1994 IEEE

IIII

IEEE TRANSACTIONS ON GEOSCIENCE A N D REMOTE SENSING, VOL. 32, NO. 5 . SEPT. 1994

to filter the intensity received at linear horizontal-vertical polarization combinations of transmitting and receiving antenna.
Consequently, with the exception of the intensity images
measured at linear horizontal-vertical polarization combinations, any classical filter based on the speckle multiplicative
noise and applied on the P,,, image, should be less efficient
than an operator which filters [W,], using the multiplicative
speckle model (3), and deduces by (2) the filtered channel
Prec.Besides, it is more convenient to filter [W,] and to obtain
a filtered received power image for any possible combination
of transmitting and receiving antenna polarizations, than to
apply a classical filter to each received image at the desired
pair of polarizations.
When the filtered matrix [k]is estimated, it is trivial to deduce the corresponding filtered Mueller matix. The filtered
Stokes vector of the scattered wave can then be deduced using
[A?], and the filtered received power can also be determined
for any combination of transmitting and receiving antenna polarizations.
The relation (2, 3 ) can be extended to a multilook image by
replacing [W,,] by the average of the single look matrices
( [ W , , ] ) = ( [ 6. 6*]),[W] by ( [ W ] ) =
. ?*I), and
[XI by the multi-look equivalent speckle matrix. In the following, we will use the same notations [W,,], [ W ] , and [XI
for I-look and multilook_ images. The matrices+[W,,] = and
[W] are defined as the 0 covariance, and the S covariance,
respectively.
The speckle filtering can also be performed, on one-look i r r ages, by filtering the scattering complex vector 0 = [ X I S .
However, for a better speckle filtering the channel correlation
should be used [9], and the four channels cannot be filtered
independently. Furthermore, if any local spatial pixel value
averaging is involved in the filter processing, this averaging
should be done in Mueller matrix not in scattering matrix [4]
to preserve the whole polarimetric information.

([S

In summary, to preserve the whole polarimeteric information


while taking advantage of spatial averaging, speckle filtering should
be done imperatively, for one-look and multilook images, by estimating the unspeckled Mueller matrix. This new concept of speckle filtering eliminates many filters currently used for speckle reduction of polarimetric data such as the scalar and vectorial Lee et
al. filter [ l ] , the Novak et al. filter [3], and the span filter [ l ] .

111. MMSE POLARIMETRIC


SPECKLE
FILTERFOR LOOK
MULTILOOK
IMAGES
A . Principle

AND

Given a speckled image in one-look scattering matrix format or


in multilook Mueller matrix format, the objective is to conceive an
output filter constituted of unspeckled Mueller matrix. Unspeckled
polarimetric parameters can then be deduced by combination of the
Mueller matrix obtained with the desired polarizations of the transmitting and the receiving antennas.
For practical reasons, it is easier to estimate in a first step the
unspeckled covariance matrix [ k ] , to deduce in a second step the
unspeckled Mueller matrix [A]. The covariance matrix can be deduced from scattering matrix data (one-lock image) by calculating
the complex dot product vector [ W ] = S . ?*. For multilook
images, data are in Mueller format, and the covariance matrix can
be easily deduced. Fig. 1 shows a simplified block diagram of the
polarimetric filter proposed.

unspeckied

unspeckiee

Fig. 1. Block diagram of a polarimetric filter.

B. Calculation of the MMSE Unspeckled Covariance


Lee et al. developed in [ 11 a MMSE vector filter for polarimetric
images. Assuming that speckle is multiplicative, the filter calculates (under target reciprocity assumption) the first- and secondorder statistics of the measured tridimensional intensity vector
((HHI2, IHV(*, IVV() to estimate the unspeckled tridimensional
vector (ISHH/, lSHVJ2,JSvvJ2)which minimizes the mean square
error. The three components of this vector are the diagonal components of the 3 x 3 unspeckled covariance matrix [k]under target reciprocity assumption.
Goze and Lopes [2] extended, for one-look image, the Lees
study to filter in addition to the three intensities, the three double
products of the complex scattering channels (HH . HH*, HH *
HV*, HV . VV*). The six filtered components permit the ctlculation of the remaining components of the hermitian matrix [W].
For one-look images, with only targets of reciprocal backscattering, the algorithm proposed in [2] can be applied to estimate the
unspeckled covariance matrix for each resolution cell. The unspeckled Mueller matrix should then be deduced for each pixel (cf. Fig.
1) and the method proposed in [6] can be used to obtain a compressed unspeckled Mueller matrix file.
In the following, the Goze and Lopes algorithm [2] is extended
to filter all the elements of the covariance matrix [W,,],deduced
from one-look scattering matrix or multilook Mueller matrix, without assuming target reciprocity. The unspeckled covariance matrix
will be determined by calculating the linear estimation which optimizes the minimum mean square error (MMSE) criteria. The algorithm proposed can easily be adapted to the reciprocal case if the
user wishes to filter a one-look or multilook scene with only reciprocal targets.

C. Vectorial Multiplicative Speckle Model


The following notations will be used:
+

Yc: vectorial form of the speckled covariance matrix [W,,].


Since [W,,] is an hermitian matrix, only ten elemenjs are
needed to determine all the matrix elements. Hence, Y,. is a
ten dimensional complex vector:
=

[(OHHO:H)r

(0HVO:V)r

(0VHO:H)r

I I12

where 0,, is the speckled target scattering matrix elements,


(0,O:) is the scalar product 0,O: for one-look images, or
the 0,O: average over the different one-look measurements
for a multil2ok image. If target reciprocity is assumed, YHv
= YvH and Y,' is a six-dimensional vector.
X,: veztorial form of the unspeckled target covariance matrix
[ W ] .X , is also a ten-dimensional complex vector defined similarly as a function of ( S , s z ) , where S,, is the real target
scattering matrix elements, and ( p , q , r, s) are the H or V
p_olarizations.
V,: Ten-dimensional vectorial form of the speckle covariance
matrix: [ N ] = <ii,$'>.
The three vectors Y,, X,, and pc will be called the complex
observed signal vector, the complex signal vector, and the
complex speckle vector, respectively.

As speckle is assumed to be multiplicative, each measured element 0,, of the scattering matrix is related to real Spqvalue by: 0,,
= n,,S,, where npqis the stationary white zero-mean complex circular gaussian noise in the channel pq. Since+speckle is also assumed to be @dependent of the signal, the Y, elements ( y , ) are
related to the X , elements (x,) by (i = 1, IO):
Y , = v,x,

(4)

where y z = (0, O:'), and v, = ( n,n:').


Let's d5fine a IO X IO diagonal matrix [ V ' ] , whose diagonal is
equal to V,. The speckled and the unspeckled covariance vectors
are then related by

An equivalent equation was first derived by Lee et al. [ I ] for tridimensional vector, then extended to six-dimensional vector in [ 2 ] .

[ V I : the 16 x 16 real matrix equivalent to the I O x IO complex matrix [V'].

- To calculate the [VI elements, the real and imaginary part of


Y, complex elements can be expressed, using the statktical independence of signal and speckle, as functions of the V, elements
(Y,), and the X , elements ( x , ) . Each diagonal complex element v,,
of [ V ' ] can then be replaced in [VI by the following 2 X 2 submatrix:

Since all the entities are now real, the equation established in
[ I ] , [ 2 ] can be used to derive the linear MMSE covariance vector
expression:
=

E(i)

+ ([Cov (?)]E([V]')[Cov ( ? ) I - ' ( ?

- E([V])E(k))

(7)
where Cov denotes the covariance matrix. In the following, all the
first- and second-order statistic parameters involved in (8) will be
expressed as a function of local statistics which can be calculated
adaptively over the moving processing window.

IV. LOCALSTATISTIC
COMPUTATION
A . First-Order Local Statistics
1) Mean of the Signal Vector: Th+e MMSE signal vector estimation needs the calculation of E ( X ) . Using the multiplicative
speckle model, the+complex elements of X , can be expressed as a
function of the E(Y) elements as follows:

(8)

E(x,) = ( y , ) / E ( u , )

D. MMSE Covariance Estimate


The linear MMSE filter for multiplicative noise is first proposed
in scalar form in [ 7 ] , [ 8 ] , and generalized later in vector form by
Lin and Allebach [ 9 ] .If we process similarly, the unspeckled signal-vector estimate Xc is a linear combiyation of the mean vector
E ( X , ) and the observed speckled vector Y,:
Xc

[ A ] E(ic)[ B ]?<

where [ A ] and [ B ] are-I0 X-IO complex matrices which are to be


determined so as E[llX, - X,.ll*] is minimum. This entity is minimum if its partial derivative on [ A ] and on [ B ] are equal to zero
[ I ] . This supposes that the two complex matrices are holomorphic
which is a restrictive condition hard to be satisfied. The calculation
is easier if all the complex entities are transformed (as in [ 2 ] ) as
real ones in terms of real and imaginary parts. The following real
entities are then defined:

-Y: the 16-dimensional real vector equivalent to the IO-dimen-

sional complex Y, (four r5aI elements + six complex elemen_ts).The Y elements of Y are deduced from the y elements
of Y, as follows:
1) For 1 5 i 5 4, Y, = y ,
2 ) F o r i = 5 , 7, 9 . . . , 15, a n d j = 5 . . . IO, Y, = Re
(Y,)and Y , + I = Im (Y,)
2: the 16-dimension_al real vector (equivalent to the 10-dimensional complex X ( ) .
The unknown [ A ] and [ B ] matrices are now 16 X 16 real matrices.

where ( y, ) is the y , average over the processing window and E ( v , )


is an element of t)e mean of the speckle covariance vector E ( V , ) .
The real vector X can then be deduced from the complex vector
X,. As it can be noticed, it is more simple to do the calculations in
the complex field and f,o deduce the real expressions, than to express the element of E ( X ) (real and iTaginary part of X,) as a function of the elements yf E(Y)_and E ( V ) (real and imaginary part of
the complex vectors Y, and V J , as done in [2].
2) Mean of the Speckle Vector: In order to preserve target radiometric information (mean of backscattering coefficient), the four
channel speckle intensity E(n,,,n,*,) can be assumed to be equal to
I(E(v,) = 1 f o r i = 1, 4).+
To estimate the order V, mean terms (E(n,,n:)) which correspond to channel speckle correlations, the following relationship
between the observed and the real channel correlation can be analyzed in terms of degree of coherence:
YO =

Yn

'

YS

where
with z
yZ = E(z,z:)/.JE(Iz,~~)E(~~,~~),

0, n , or S.

Speckle is assumed in [l] to be nonstationary and its statistics


are adaptively estimated from the observed channel correlation over
the homogeneous area to be filtered. This supposes that the area is
perfectly homogeneous: SI, = constant = E ( $ , ) for i = H or V ,
and consequently (ysl = 1. Hence, the measured degree of coherence yo and the speckle degree of coherence yn have the same amplitude, and the filtered area will be completely polarized. This
method fails in practice for the two following reasons:

IEEE TRANSACTlONS ON GEOSCIENCE A N D REMOTE SENSING, VOL. 32


0

The degree of polarization p of an homogeneous area varies


greatly from an area of surface scattering ( p of high value) to
an area of volume scattering ( p of low value). If this variation
is assigned to speckle (i.e., y,! is not assumed to be a constant), this important polarimetric information will be lost after
speckle filtering y j = 1.
For a textured area y j # 1, and the problem of separation of
speckle-channel correlation from the target signal correlation
is no longer obvious.

For these reasons we will assume, for one-look as well for multilook images, that speckle is
1) stationary
2) completely polarized: Iy,,\ = 1. The mean speckle phase difference of two channel is also assumed to be equal to zero:
(arg ( y n ) = 0 ) .
Co"equently, all the elements of the speckle vector's mean
E(VJ are taken equal to 1 and the mean of the 16 X 16 [V] matrix
appearing in equation (8) is the identity matrix.

B. Second-Order Local Statistics


1 ) Calculation of the Covariance Matrix of Real Vector Using
the Covariance Matrix of the Complex Vector: The filter equation
(8) involves :he calculatio_n of the two real vector covariance ma$cy: C t v ( X ) y d Cov ( Y ) . The covari_ance matri_x of a ler:
ve_ctor
Z (Z
, = X , or Y ) is given by: Cov ( Z ) = E [ ( Z - E ( Z ) ) ( Z E ( Z ) ) ' ] . All the elements of this matrix can be-deduced f r o p the
co?plex_ vector cyariance matrix: Cov (ZJ = E[(Z,. E(Z,.))( Z c - E(Z,.))*T] and the pseudocovariance matrix
covp(Z:) = E[(?, - E(?,.))*(Z,. - E(.?,.))*'].
On the contrary to the real vector covariance matrix, the two
complex vector covariance matrices can be easily written under an
analytical form. The following relations can be used to deduce from
the covariance of two complex scalars cov ( x , y ) (where cov denotes scalar covariance), the different covariance involving the real
( x r , y , ) and imaginary parts (x,, y , ) of the two scalars:

NO. 5.

SEPT.

1994

1113

The observed covariance matrix elements (P,,,


P,',)can be measured adaptively over the moving processing window. The elements of speckle covariance matrices ( R , J ,R;J)also have to be determined. In the following, two methods are proposed for estimation of speckle covariance matrices for one-look or a multilook
image.
3 ) Speckle Covariance Matrix: The elements of speckle covariance matrices R,, are given by the following expression:

where ( npqnr\) = npqnr,for a one-look image, and ( npqnr,)is the


average of the one-look realizations npqnr\for a multilook image.
Goze and Lopes [2] proposed a method, under the reciprocal
assumption, for the calculation of the R , , terms in a one-look image. However, this method which can easily be extended to the
nonreciprocal case, is not suitable to multilook data. In the following, a new method is proposed for the estimation of the speckle
vector covariance matrices for one-look or multilook images. This
method is an extension of the algorithm proposed for a tridimensional real vector in [ 11. In contrast to [ 11, speckle is assumed here
to be stationary (cf. Section IV-A.2), and one area (the most homogeneous area) should be selected in the_scene and use$ to estimate the speckle covariance matrices Cov ( V , ) and Cov p (V:). This
area should have the following:
1) The highest degree of polarization for the different channel
combination.
2) Coefficient of variation value within the class of homogeneous area [4], for each channel.

For this kind of area, the real scattering matrix element does not
vary very much (Spq= E(S,,) f o r p , q = H or V ) , and (E(S,,S,,)
= E(S,,)E(S,,).
+
+
The covariance matrices Cov ( Y , ) = (P,,),
and Cov p ( Y : ) =
(P;,)measured over such an area allow the calculation of the speckle covariance matrix elements using one of the two following
equations:

+ cov (x*, y ) ]
cov (x,, y I ) = i Re [cov (x, y ) + cov (x*, y)]

cov (xr, y,) =

COY

(x,, y I ) =

cov (xI,y,) =

Re [cov (x, y )

t Im [cov ( x , y) + cov (x*, y ) ]

i Im [cov (x, y ) - cov (x*, y)].

(9)

2) Signal Covariance Matrix: The following notations will be


used:
Cov ( X , ) = ( M , J ) with
,
i , j : 0 . . . 16,
Cov p$!)
= (M:J).
Cov ($) = ( P , J ) ,and C o v p ( $ = (P:,).
Cov (V,) = ( R , J ) ,and Cov p ( V T ) = (R:,).
+

To calculate the signal covariance mat$x Cov ( X ) , the-two complex vector covariance matrices Cov (&), and Cov p(X,*) should
be determined. The elements of these matrices can be expressed as
a function of measurable entities: the elements of the observed co, p,!,), and the elements of speckle covariance
variance matrices (P,,
matrices ( R , J ,R:,):

The method proposed here can be applied to one-look images as


well as to multilook (correlated or uncorrelated looks) images.

V . CONCLUSION
Speckle reduction of polarimetric images requires speckle filtering of the Mueller matrix of each resolution cell. The multiplicative speckle model valid only for the four measured linear channels
can be applied to all the Mueller matrix elements (each element is
a combination of the four channel measurement correlations) to
determine the MMSE linear estimate of the Mueller matrix. The
filtered received intensity, and the filtered scattered wave's Stokes
vector can then be deduced, from the filtered Mueller matrix, for
any combination of transmitting and receiving antenna polarizations. The linear relationship between the Mueller matrix and the
covariance matrix allows one to perform the filter processing on
the covariance matrix elements. The filter proposed here for onelook and multilook images does not assume target reciprocity. It
can easily be adapted to the reciprocal case for a benefit in computing time.

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING. VOL. 32, NO. 5 , SEPT. 1994

1114

The polarimetric filter as well as the methods proposed for the


estimation of the speckle vector covariance, will be validated on
one-look and multilook polarimetric data including nonreciprocal
data.

VI. NOTATION
A . Real Scattering Signal
[SI: scattering matrix.
Four-dimensional vector form of [SI.
[MI:Mueller matrix.
Covariance matrix of
[ W I=
. ;*I>.
[ g :Estimation of real covariance matrix [W].
[_MI: Estimation of real Mueller matrix [ M I .
X,: 10-dimensional vector which ch2racterizes completely the
Hermitian covariance matrix [W]: X , = (xi i,= 1, 10).
X. 16-dimensional real vector equivalent to X,.
(M,,):
Element of X , covariance matrix.

S:

[y]:

(M;,): Elements of

S,

2:

([S

[5] M. Born and E. Wolf, Principles of Optics. Oxford: Pergamon,


1989.
[6] P. C. Dubois and L. Norikane, Data volume reduction for imaging
radar polarimetry, in Proc. IGARSS87Symp., Ann Arbor, MI, May
1987, pp. 691-694.
[7] J . S. Lee, Digital image enhancement and noise filtering by use of
local statistics, IEEE Trans. Pattern Anal. Machine Intell., vol.
PAMI-2, Mar. 1980.
[8] D. T. Kuan, A. A . Sawchuk, T. C. Strand, and P. Chavel, Adaptive
noise smoothing filter for images with signal-dependent noise, IEEE
Trans. Patfern Anal. Machine Intell., vol. PAMI-7, pp. 165-177,
Mar. 1985.
191 Q. Lin and J. P. Allebach, Combatting speckle in SAR images:
Vector filtering and sequential classification based on a multiplicative
noise model, IEEE Trans. Geosci. Remote Sens., vol. 28, pp. 647653, July 1990.
[lo] F. T . Ulaby and C. Elachi, Radar Polarimetry f o r Geoscience Applications. Norwood, MA: Artech House, 1990.

Image Processing of Synthetic Aperture Radar Range


Ambiguous Signals
pseudo-covariance matrix defined by:

K. Tomiyasu

B. Observed (Measured) Scattering Signal


[O]: Measured scattering matrix.
[M,]: The measured Mueller matrix.
+

0: Four-dimensional vector form of [ O ] .


LWJ:Covariance matrix of 0, [w,,] = (10 . 6*1>.
Y,: Ten-dimensional vector which characterizes completely
t$e Hermitian covariance matrix [W,,]: Y, =,cy, i = 1 , 10).
Y: 16-dimensional_real vector equivalent to Y,.
(P,,):
Element of ,k covariance matrix.
(P;,):
Elements of Y,* pseudo-covariance matrix.

C. Speckle Parameters

iic:Four-channel complex speckle vector.


[X: 4 X 4 diagonal matrix whose diagonal elements are the
components of ii,.
Covariance matrix of iir, [N] = [ii, . Z r T ] > .
V,: Ten-dimensional vector which chyacterizes completely
the Hermitian covariance matrix [N]: V, = (v,i = 1, 10).
[V,]: 10 x 10 diag_onal matrix whose diagonal elements are
the components of V,.
[VI: The 16 X 16 real matrix equivalent to the complex matrix [ V,].
,
( R ! , ) :Element of
covariance matrix.
(R,!,):Elements of
pseudo-covariance matrix.

PI:

c,

REFERENCES
J . S. Lee, M. R. Grunes, and S. A. Mango, Speckle reduction in
multipolarization, multifrequency SAR imagery, IEEE Trans.
Geosci. Remote Sens., vol. 29, no. 4, pp. 535-544, July 1991.
S. Goze and A. Lopes, A MMSE speckle speckle filter for full resolution SAR polarimetric data, J. Elect. Waves and Applic., vol. 7,

no. 5, pp. 717-737, May 1993.


L. M. Novak and M. C. Burl, Optimal speckle reduction in polarimetric SAR imagery, IEEE Trans. Aerosp. Elect. Syst.. vol. 26,
pp. 293-305, Mar. 1990.
I . J. Van Zyl, C. H. Papas, and C. Elachi. On the optimum polarizations of inwherentlv reflected waves. IEEE Trans. Anfenn. Prop u g . , vol. AP-35, pp.-818-824, July 1987.

Abstract-If wide-range swath scenes are illuminated by a synthetic


aperture radar with high-pulse repetition frequencies, ambiguous scattered signals from an undesired range may appear in its receiver at the
same time from consecutive transmitted pulses. This simplified analysis shows that range ambiguous signal returns result in a defocused
image because they are convolved with the phase history of the desired
range signal. A satellite borne SAR example is presented.

INTRODUCTION
Synthetic aperture radars (SAR) have been found useful for remotely sensing geophysical features from both aircraft and spaceborne platforms [ l ] , [2]. For some applications, very high-area
mapping rates are desired, and these rates can be achieved by using
wide-elevation beamwidth side-looking antenna illuminating wide
ground swaths [3].
The generation of an image requires the collection of numerous
pulses, storage of these time-tagged complex signals, and convolution of these signals with a deterministic phase history [l], [4],
[ 5 ] . This process entails sampling, and the radar parameters are
selected to avoid otherwise inherent ambiguities in the azimuth and
range directions [6], [7].
In order to prevent aliasing in the azimuth direction, the pulse
repetition frequency (PRF) must be higher than the Nyquist sampling rate [ 5 ] , [ 6 ] . If this PRF condition is not met, ambiguous
images displaced in azimuth can occur [ 5 ] . In hardware terms, the
pulses have to be sufficiently frequent that a second pulse after the
first is emitted before the SAR antenna and hence its platform has
moved by one-half the antenna length in the velocity direction [4].
The range swath width is limited by the elevation beam width to
avoid two consecutive pulses in the antenna beam footprint which
are received at the same time [ 2 ] , [4]. Potentially, this may result
in data corruption and range ambiguous images. A commonly stated
and generally accepted design procedure to avoid both azimuth and
Manuscript received October 13, 1993; revised March 23, 1994.
The author is with Martin Marietta Corporation. Philadelphia. PA 19101.
IEEE Log Number 9403638.

0196-2892/94$04.00 0 1994 IEEE

You might also like