IEEE Geoscience and Remote Sensing Magazine - September 2013 - 31

(a)

(b)

(c)

(d)

FIgURE 16. (a) Original one-look COSMO-SkyMed image ( t - 0.29 ); (b) MAP-GG-S [83] of original (ENL = 27.90, n r = 0.940, v 2r = 0.702 );

(c) whitened [140] ( t - 0.05 ); (d) MAP-GG-S of whitened (ENL = 142.29, n r = 0.997, v 2r = 0.936 ). t is the CC of speckle measured on the
complex image [29]; ENL, n r and v 2r are calculated on a homogeneous patch after despeckling.
of one scale in a unique plane; thus the number of coefficients to be despeckled, and hence computing times, would
be three times lower. However, the adaptivity with orientation featured by UDWT would be lost with ATWT.
The assessment of the performances of despeckling filters on real SAR data is often problematic due to the lack
of with-reference indexes. In order to overcome such problems, a possible idea is to use electromagnetic SAR image
generators [23]. Such simulators are based on more physical-oriented models, which consider the propagation of the
electromagnetic wave and its interaction with targets and
surfaces, and usually require a more detailed parametric
description of the target scene with respect to the models
used in signal processing applications. In [22], the authors
use an electromagnetic SAR image generator to simulate
several independent acquisitions of the same scene. If the
number of acquisitions is sufficiently high, their average
can be considered as a good approximation of the noise-
free reflectivity and can be used to compute with-reference
indexes. The advantage of this technique is that the simulated images do not necessarily obey the fully developed
speckle model and provide insights on the behavior of the
filter on point targets and highly heterogeneous areas. On
the other hand, the underlying reflectivity follows a synthetically generated pattern, which may not be fully representative of the reflectivity usually encountered in real SAR
images, especially in complex scenes, due to the ideal models of objects fed to the simulator.
Another viable approach to devise a fully automatic
method for quality assessment of despeckled SAR images
was recently proposed by the authors [144]. The rationale of
the new approach is that any structural perturbation introduced by despeckling, e.g., a local bias of mean or the blur
of a sharp edge or the suppression of a point target, may
be regarded either as the introduction of a new structure
or as the suppression of an existing one. Conversely, plain
removal of random noise does not change structures in the
image. Structures are identified as clusters in the normalized
scatterplot of original to filtered image. Ideal filtering should
produce clusters all aligned along the main diagonal. In pracSeptember 2013

ieee Geoscience and remote sensing magazine

tice clusters are moved far from the diagonal. Cluster centers
are detected through the mean shift algorithm. A structural
change feature is defined at each pixel from the position and
population of off-diagonal clusters [145]. Such a feature may
be regarded as a spatial map of filtering inaccuracies. A preliminary validation has been carried out on simulated SAR
images, with a good correlation between feature and objective filtering error. In experiments on COSMO-SkyMed
images, the automatic ranking of filters matches the subjective trials of experts. The proposed feature detects filtering
impairments but is unable to measure the overall effectiveness of filtering. Therefore, its use must be coupled with
another index measuring the effectiveness of cleaning, e.g.,
ENL, regardless of its accuracy.
IX. CONCLUSIONS AND PERSPECTIVES
This tutorial has demonstrated that despeckling of SAR
images takes into account several issues related to signal and
noise modeling, signal representation, estimation theory and
quality assessment. Concerning Bayesian estimation, starting from Lee filter, local-window adaptive filtering has been
progressively enhanced, up to a saturation of performances,
due to the trade off of using windows small enough to retain
edges textures and fine details and large enough to allow a
consistent and confident statistical estimate to be achieved.
In the last two decades, the introduction of multiresolution analysis has been found to boost despeckling
algorithms performances. Key points of wavelet-based
despeckling is the modeling of the reflectivity and of the signal-dependent noise in the wavelet domain and the choice
of the estimation criterion to achieve the noise-free wavelet
coefficients. While several authors have chosen overfitting
models sacrificing space adaptivity, others have tried to
keep the advantages of an adaptivity in both scale and space
by using pdfs with few parameters to be estimated locally on
subbands/frames. A preprocessing step of point targets that
must retain their radiometry after despeckling and a segmented approach, in which sample statistics are calculated
on homogeneous segments, complete Bayesian despeckling
in wavelet domain.
31

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2013