Denoising Techniques For Synthetic Aperture Radar Data - A Review
Denoising Techniques For Synthetic Aperture Radar Data - A Review
Denoising Techniques For Synthetic Aperture Radar Data - A Review
1. INTRODUCTION
SAR and its active mode of signal acquisition entails a very complicated sensor
design, signal processing, image processing and its interpretation [1-3]. The
fundamental theory behind SAR needs transmission and reception of linearly
frequency modulated chirp signals, and Doppler processing of the encoded returned
echo [1],[4]. Coherent signal processing is done to attain high spatial resolution.
However due to this coherent nature of signal, a grainy type of noise is inherent in the
http://www.iaeme.com/IJCIET/index.asp
image, which degrades the radiometric quality of the data. This is known as the
speckle noise whose characteristics are described in detail in [1] [2] [4] [5]. A variety
of application in the areas of agriculture, land use type discrimination, forestry,
biomass estimation, geology, flood mapping, disaster zone mapping, oil spill
detection etc. is possible with SAR data. Due to speckle noise, the accuracies of
image analysis involving classification, segmentation, texture analysis, target
detection etc however are severely affected. In the field of image processing,
denoising is still a challenging task to the researchers. Reducing noise always comes
at a cost of some other aspect of image degradation such as blurring, spatial resolution
degradation and edge smearing. Apart from that, the type of noise present in the
image is a critical factor in formulating the denoising scheme. Most of the images in
the optical domain are afflicted with additive white Gaussian noise. Images produced
by coherent processing such as those of SAR, USG etc are however, tarnished with
speckle noise, which is more difficult to remove. Here, we discuss briefly about the
characteristics of speckle noise followed by some of the well known speckle filters,
which are being used by SAR users. Then we give a review of some of the wavelet
based image denoising techniques, with special emphasis on SAR images.
2. SPECKLE NOISE
Speckle noise is multiplicative in nature. This type of noise is formed as a result of the
random interference between the coherent returns from active imaging sensors such as
LASER (Light Amplification by Stimulated Emission of Radiation), USG (Ultra
SonoGram) and SAR. Its response is represented as:
S=x*
(1)
where S is the output scattering coefficient of a target, x is its true value, and is
the speckle noise affecting the input signal. Fully developed speckle noise has the
characteristic of multiplicative noise as shown by G April et al [4].
Speckle model is explained by using the signal geometry for a SAR sample in
Fig:1.
The basic assumption is that, one resolution cell contains the scattering from a
large number of scatterers with a wavelength which is comparable to the roughness of
the terrain or object being imaged. Hence the response from one cell is a coherently
summed signal from many scatterers.
http://www.iaeme.com/IJCIET/index.asp
another adaptive filter, replaces the pixel with a weighted sum of the values within the
window. The weighting factor decreases with distance from the central pixel. Leesigma filter had some shortcomings, such as blurring of targets, and depressing strong
targets. This was removed in the extended Lee filter [12], whereby the sigma range
was estimated using the probability density functions, and by incorporating minimum
mean square error (MMSE) estimates for adaptive speckle reduction. Extended Lee
and Frost filters were introduced [13], to alter the performance locally according to
three cases. Pure averaging is done when the local coefficient of variation(LCV) is
below a low threshold, selected. Above a higher threshold the filter performs as an all
pass filter, while for other cases lying in between, a balance between these two
filtering is done as is done in standard Lee and Frost techniques. However these also
have limitations of window size, and directional sensitivity and do not enhance the
edges. The hard thresholds many a times lead to artifacts, and noisy boundaries by
leaving the edges unfiltered. Gamma-MAP filter also gives good denoising
performance with preservation of edges. This assumes that natural targets have
Gamma distributed cross section, rather than Gaussian distribution, which is a better
approximation for most SAR images. It also assumes non-stationary mean and
variance parameters. Though a good filter, it produces artifacts in many of the cases,
for different window sizes, and has to be used with caution, and many a times several
iterations are needed to get good results, thereby increasing the computation time.
Another method to produce sharp edges, is the simulated annealing technique [14].
When this algorithm is applied to despeckling SAR data, the RCS (Radar Cross
Section) is modeled as a Markov random field that is governed by the Gibbs
distribution of the conditional probability of the RCS. Simulated annealing filters
produce relatively good results if parameters are selected appropriately. However,
processing times are very long when compared to other filters, and the results are
dependent on the judicious choice of iterations and the filtering parameters.
Another breakthrough in the domain of data denoising came with the introduction
of speckle reducing anisotropic diffusion (SRAD) technique by Y Yu et al [15]. This
method introduced a diffusion method which is specifically tuned to denoise USG and
radar images which are afflicted with speckle noise. This paper deals with the edge
sensitive diffusion for speckled images, using a method which is similar to the
conventional anisotropic edge sensitive diffusion for images which are corrupted by
additive noise, such as in optical images. Here it is shown that, in the case of Lee and
Frost filter, adaptive filtering is done using the coefficient of variation, whereas the
SRAD algorithm exploits the instantaneous coefficient of variation, which is shown to
be function of the local gradient magnitude and Laplacian operators. The technique is
based on the minimum mean square error (MMSE) approach for filtering, as in Lee,
and Frost, but is the edge sensitive extension of the conventional adaptive speckle
filters. Here it is mentioned that the existing diffusion techniques use a homomorphic
approach [16], while this work deals with linear domain SAR data thereby useful
information with regards to SAR data is preserved.
A denoising technique which reports a concept of block-matching and 3D
algorithm (BM3D) which is modified to suit SAR image noise, is reported in [17].
Research work in the BM3D was initially reported by Dabov et al[18]. This technique
exploits the specific non-local image modeling through a process called, grouping and
collaborative filtering. The grouping process finds mutually similar 2-D image blocks
and stacks these together in 3D arrays, while collaborative filtering produces
individual estimates of all the blocks, by filtering them jointly, by applying transform
http://www.iaeme.com/IJCIET/index.asp
domain shrinkage of the groups. Hence both local and non local characteristics of
natural images are taken into consideration. Finally the effectiveness of the
subsequent shrinkage is related to the sparsity of the real signal [19]. This method is
supposed to be a state of the art denoising for images with additive noise. However it
is observed by Hasan et al [20] that for images rich in texture or those corrupted with
higher level noise, this technique does not give the best results. So scope exists to
make more improvement. It is to be noted that most of these techniques have been
tested on images corrupted by additive white Gaussian noise (AWGN). For speckle
noise removal, this technique is not so widely applicable, and Parilli et al have used
this concept with the wavelet domain shrinkage to denoise SAR data in [17]. A recent
work for SAR image denoising is reported by J Liu et al [21]. Here a method called
total variation (TV) regularization has been resorted to, in order to filter multiplicative
noise. It is reported here that though TV means is highly researched area, this can
sometimes give rise to undesirable staircase artifact. To circumvent this problem, high
order TV norms have been judiciously used to balance the edge and smoothness
regions. This method has been shown to give better efficiency in terms of SNR and
structure similarity index. In a recent work by Zhang et al [22], it is pointed out that
with self adaptive sliding window based methods, the local stationary assumption of
classical speckle filters can be satisfied better than those with fixed windows. But
region variance is always affected by speckle and is not accurate enough specially for
textures, edges and point type targets. Hence these are only effective for homogenous
regions. In this paper, trade-off between homogeneous areas and point target
preservation is resorted to, by an adaptive threshold technique which is treated as a
linear function of saliency map, and the threshold range is estimated by Monte Carlo
algorithm based on a given significance level and the number of looks in the image.
This technique claims to give good noise removal without sacrificing the sharp edges.
In one of the most recent papers on SAR denoising, it is reported that nonlocal
grouping and transformed domain filtering has led to the state of the art denoising
techniques, and that this has been extended to SAR images based on disjoint local
regions with similar spatial structure[23]. Each such region is then denoised by linear
minimum least square error (LMMSE) filtering in principal component analysis
(PCA) domain. Both clustering and denoising are done on image patches. This
method reportedly gives good performance in terms of noise reduction and detail
preservation. Here the technique has been done based on additive signal dependant
noise (ASDN) model.
http://www.iaeme.com/IJCIET/index.asp
extremely suitable for filtering noise. Wavelet based denoising schemes involve non
linear thresholding of wavelet coefficients in time-scale transform domain, whereby
the noise frequencies are suppressed. One method for finding the mother wavelet and
scaling coefficients was proposed by I Daubechies[25], which ushered in a whole
domain of mother wavelet generation for different research purposes. In the domain
of optical images, wavelet and its derivatives have been in use since the inception of
wavelet based techniques [26]. Donoho in 1995[27] had proposed thresholding
techniques for denoising, which had paved the way for research on various means of
thresholding for filtering out different kinds of noise in images. Research in the field
of speckle noise removal using wavelet based technique gained momentum, very
soon, as is discussed by Gagnon et al[28], Argenti et al[29]. Many types of wavelet
functions have been in use, out of which the discrete wavelet transform (DWT) which
transforms digital signals to discrete coefficients in the wavelet domain, has good
capability for both signal as well as image processing applications. Hou et al[30]
reported speckle reduction on SAR images based on Bayesian MAP estimation in the
wavelet domain.
Many papers during this period reported results of wavelet based filtering on SAR
as well as medical data sets using different techniques developed, and by extending
the original ideas. Some of the results have been reported by P.U Fangling et a l[31],
Solbo et al[32], Gleich et al[33]. S Parrilli et al in 2012 proposed a nonlocal SAR
image denoising algorithm based on LMMSE wavelet shrinkage [17]. A novel
despeckling algorithm for SAR images was proposed here, based on the concepts of
nonlocal filtering and wavelet-domain shrinkage. It follows the structure of the blockmatching 3D (BM3D) algorithm, which had been proposed, for additive white
Gaussian noise denoising[18]. At present, BM3D is considered to be the state of the
art for AWGN denoising, as has been mentioned in the previous section.
Most of the filtering techniques in the wavelet domain operate in a homomorphic
way for the SAR images [32]. Log transform for SAR data, however changes the
basic characteristics of the SAR image, and after speckle removal and inverse
logarithm, cannot be retrieved back, as has been reported by Ulaby et al [34]. Hence
wavelet based SAR denoising is still an open challenge in the power domain of data.
Alparone et al[35] in a recent work have reported the results of evaluation of Bayesian
estimators and PDF(Probability Density Function) models for despeckling in the
undecimated wavelet domain. According to their study the best suited method comes
from maximum-a-posteriori (MAP) estimation of the wavelet coefficients.
Since SAR is corrupted by multiplicative noise, the methods adopted in all the
wavelet based work is that of homomorphic approach whereby the logarithm of the
data is taken, to force the data into the additive domain. Subsequently the generic
denoising filters are applied on those transformed data.
Denoising techniques developed, or the thresholding methods developed are of
interest to all researchers working in the image processing domain. Different mother
wavelets have evolved, which are used to convert the images from the linear domain
to the wavelet domain. One such technique was proposed by Daubechies [25]. DWT
is used to transform the digital signals in the image to discrete coefficients in the
wavelet domain. The output of the low pass filter are known as the approximation
coefficients, while those from the high pass filter are called the detailed coefficients.
Wavelet transform is used to decompose the image into sub-bands, and this can be
repeated up to multiple levels, depending on the dimension of the image[24]. At each
level of decomposition, the signal is decomposed into four sub-bands, consisting of
http://www.iaeme.com/IJCIET/index.asp
Discrete wavelet transform is applied on the data sets to get the sub-bands. The
detailed wavelet coefficients are taken for denoising at any decomposition level.
Thresholding criteria is used to suppress the noise. There are two methods of
thresholding such as the hard thresholding given in equation 2 or soft thresholding
given in equation 3[27].
After thresholding on one level, further levels of decomposition may be attempted
based on the denoising desired, and the above procedure is repeated on the
approximate coefficients.
After going through the desired number of decompositions, the thresholded values are
taken through inverse wavelet transform, through all the levels. Final reconstructed
image gives the noise free image.
(2)
(3)
The coefficients of the wavelet transform are usually sparse. That is, coefficients
with small magnitude can be considered as pure noise and may be set to zero in order
to denoise the data. The detailed wavelet coefficients are compared with a threshold
value in order to decide whether it constitutes a signal or a noise, and is known as
wavelet thresholding. Donoho in 1995 [27], proposed a method to reconstruct a
function f from a noisy data set d. This was done by translating the empirical wavelet
coefficients of d towards 0 by an amount given by:
X = (2log(n)/n)
(4)
where n is the number of data points in the signal, and is the noise or standard
deviation in the data which is given by an estimate in the wavelet domain as :
2 = [(median |Yij |) / 0.6745]2
(5)
where Yij denotes the coefficients in the HH subband.This was found to be quite
effective for Additive White Gaussian noise(AWGN) which is found in optical data.
Subsequently several shrinkage methods for soft thresholding have been proposed by
researchers. SURE(Steins Unbiased Risk Estimate) technique whereby the universal
threshold is replaced by an adaptive SURE based thresholding which is reported to
give better denoising as was reported by Zhang et al[36], and Thierry et al[37]. This
also was found out to be highly efficient for optical images. Chang et al, in 2000
proposed a novel concept of denoising and compression of optical images by an
http://www.iaeme.com/IJCIET/index.asp
adaptive, data-driven threshold for image denoising via wavelets based on softthresholding[38]. The threshold was derived in a Bayesian framework, and the a priori
model used on the wavelet coefficients was the generalized Gaussian distribution
(GGD) which is widely used in image processing applications. The proposed
thresholding is adaptive to each sub-band since it depends on estimates of the
parameters from the data. Experimental results show that the proposed method, called
Bayes Shrink, is typically within 5% of the Mean Square Error (MSE) of the best softthresholding benchmark with the image assumed known. It is reported that it
outperforms Donohos[27] SureShrink method most of the time. Xie et al in 2002
reported speckle reduction for SAR using wavelet denoising and Markov Random
Field (MRF) modelling[39]. Using MRF has been popular in the field of image
processing and this was extended to SAR denoising. In this paper, a technique is
developed for speckle noise reduction by fusing the wavelet Bayesian denoising
technique with Markov random-field (MRF)-based image regularization.
Experimental results showed that the proposed method outperformed standard
wavelet denoising techniques in terms of the signal-to-noise ratio and the equivalentnumber-of-looks (ENL) measures in most cases. It also achieves better performance
than the refined Lee filter. It has been observed that model based despeckling mainly
depends on the chosen models. Bayesian methods have been commonly used in
denoising, whereby the prior, posterior and evidence probability density functions are
modeled as illustrated by Chang et al[38]. A novel approach of using SURE based
shrinkage to optical image denoising was proposed by Thierry et al [37]. Using the
SURE and LET (linear expansion of thresholds) principles, it was shown that a
denoising algorithm merely amounts to solving a linear system of equations which is
fast and efficient. The very competitive results obtained by performing a simple
threshold (image-domain SURE optimized) on the un-decimated Haar wavelet
coefficients showed that the SURE-LET principle has a huge potential. Gagnon has
reported some numerical results based on wavelet filtering of speckle noise[40]. One
multi-scale speckle reduction method based on the extraction of wavelet inter scale
dependencies to enhance medical ultrasound images with multiplicative noise was
done using dual tree complex wavelet transform [41]. This method is reported to give
good performances compared to standard spatial despeckling filters.
7. CONCLUSION
Synthetic Aperture Radar data is increasingly being used in the field of remote
sensing applications due to its all weather, and day and night imaging capabilities.
However due to the inherent coherent processing, SAR images are corrupted by
speckle noise which needs different ways of filtering. Speckle being signal dependent
multiplicative type of noise, is difficult to remove using techniques which are
generally suitable for additive white Gaussian noise. Over the last decade several
techniques have been developed to suppress speckle noise in SAR as well as medical
data. In this paper we have brought out a comprehensive list of different kinds of
speckle filtering algorithms, which have evolved during the past as well as the recent
times. Initially the multilook techniques were used for speckle reduction, but this
came at a cost of degrading the spatial resolution. The algorithms which evolved
subsequently were based on spatial filtering techniques. Compared to non adaptive
techniques, the adaptive means have shown promising results. A good adaptive
speckle filter should reduce the noise in statistically homogeneous areas, but preserve
the features such as edges and point targets, and also improve the radiometry. Of late
http://www.iaeme.com/IJCIET/index.asp
the wavelet based multi-resolution analysis has shown great promise for image feature
detection at different scales, and using this property several denoising algorithms have
been proposed. Factors influencing denoising of images are the choice of the wavelet
basis, order of the mother Wavelet, decomposition level, thresholding and shrinkage
criteria chosen. Apart from this, the speckle model and statistics play a key role in the
filter performance. There is no best solution to speckle reduction, as each method has
its pros and cons, as has been discussed in the papers reviewed here. Similarly each
technique assumes certain noise model and region based statistics. Thus, denoising of
SAR images offers great scope in the field of image processing research.
8. ACKNOWLEDGEMENTS
The authors would like to thank Dr S Garg, HOD, CSE Department, Nirma
University, Ahmedabad for his valuable support in carrying out this work. They
would like to thank Dr Raj Kumar, GD, GSAG, Dr P K Pal, DD, EPSA and Sri Tapan
Misra, Director, SAC/ISRO for supporting them in carrying out this research activity.
We also thank the reviewers for their valuable suggestions.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
F. T. Ulaby, R.K. Moore, A.K. Fung, Microwave Remote Sensing Active and
Passive, Vol-2, Artech House, 1986.
J. Reeves, Manual of Remote Sensing, American Society of Photogrammetry,
First Edition, 1975, Vol-1.
R. K. Raney, Radar Fundamentals: Technical Perspective, Principles and
Applications of Imaging Radar: Manual of Remote Sensing, 3 Edition, New
York: Wiley Interscience, 1998, Vol- 2, 1998, pp.9-130.
G. V. April, E. R. Harvy, Speckle Statistics in Four-Look Synthetic Aperture
Radar Imagery, Optical Engineering, 30(4), 1991, pp.375-381.
J. Bruniquel, A. Lopes, Multivariate Optimal Speckle Reduction in SAR
Imagery, International Journal of Remote Sensing, 18(3), 1997, pp.603-627.
Buades, B. Coll, J.M. Morel, A review of denosing algorithms, with a new one, A
SIAM Interdisciplinary Journal, 2005, 4(2), pp. 490-530.<hal-00271141>.
J.S. Lee, Speckle Suppression and Analysis for Synthetic Aperture Radar Images,
Optical Engineering, 25(5), 1986, pp. 636-643.
J.S. Lee, Digital image enhancement and noise filtering by use of local statistics,
IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol-2, 1980,
pp.165-168.
V.S. Frost, J.A. Stiles, K.S. Shanmugan , J.C. Holtzman, A model for radar
images and its application to adaptive digital filtering of multiplicative noise,
IEEE Transaction on Pattern Analysis and Machine Intelligence, Vol- 4, 1982,
pp.157-166.
D.T Kuan, A.A Sawchuk, T.C Strand, P Chavel, Adaptive noise smoothing filter
for images with signal dependant noise, IEEE Transaction on PAMI-7, No-2,
1985, pp.165-177.
Baraldi, F. Pannigianni, A refined Gamma MAP SAR speckle filter with
improved geometrical adaptivity, IEEE Transactions on Geoscience and Remote
Sensing, Vol-33, 1995, pp.1245-1257.
J.S. Lee,J.H.Wen,T.L. Ainsworth, K.S. Chen, A.J. Chen, Improved sigma filter
for speckle filtering of SAR imagery, IEEE Transactions on Geoscience and
Remote Sensing, Vol-47, 2009, pp.202- 213.
http://www.iaeme.com/IJCIET/index.asp
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
http://www.iaeme.com/IJCIET/index.asp
10
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
http://www.iaeme.com/IJCIET/index.asp
11