IEEE Geoscience and Remote Sensing Magazine - June 2021 - 43

divergence-between the empirical and theoretical speckle
pdfs. Ablation studies show some improvements, especially
for real SAR images. This approach is taken to the extreme
in [60] where, as mentioned in the " Supervised Models:
Architectures " section, the SAR signal is also modeled as
an RV. Accordingly, the loss function measures the Kullback-Leibler
divergence between the pdfs of the observed
and estimated SAR images (all in the log domain), the latter
computed as the convolution between the pdfs of the clean
SAR image and speckle.
TESTING PROCEDURES
We conclude this section with a few words about testing
procedures. Almost always, the proposed methods are tested
in two steps, first on simulated images and then on real
SAR images. Considering that SAR image despeckling is the
actual goal, the first step has limited significance due to the
poor transferability of models to the real target domain. On
the positive side, however, the presence of clean references
enables a sound performance assessment based on widespread
full-reference measures, such as the MSE, the PSNR,
and structural similarity. Unfortunately, experiments are
carried out on a wide variety of different test images, preventing
a direct comparison of numbers, which is why we
do not report any numerical results in this work. An indirect
(but somewhat shaky) comparison among methods could
be established based on results on common baselines with
available code, such as PPB and SAR-BM3D. With a few notable
exceptions, few or no ablation studies are carried out,
the experiments cover a limited range of cases (sometimes
not even the single-look case), and the software is rarely
published online. In general, there seems to be little attention
to the reproducibility of the research, which guarantees
a negative impact on the speed of progress in this field.
All these problems only worsen when considering real SAR
data, given the absence of widespread data sets. Nonetheless,
they become irrelevant with respect to the lack of clean
images, which prevents the use of full-reference measures.
Therefore, assessing despeckling performance becomes a
problem of its own, with many different solutions adopted
in the literature.
The visual inspection of despeckled images is unanimously
recognized as the foremost way of assessing image
quality. In fact, a despeckling algorithm should not only
remove speckle but also preserve image features (texture,
edges, point targets, and urban areas) and avoid introducing
artifacts. All these aspects are important, and not even
a full-reference measure could entirely capture them. Notably,
visual inspection is used in all the papers reviewed
here. Unfortunately, with applications to different images
that have various natures and may be cherry-picked to
underline some phenomena of interest, visual inspection
becomes subjective and useless for comparison purposes.
Also very popular and useful is the ENL, computed as the
local squared mean-to-variance ratio. Although it measures
only speckle suppression, it does so in a very stable
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
and reliable way, provided that a large homogeneous region
exists and is selected for measurements.
The visual inspection of ratio images, obtained as the
pointwise ratio between original and despeckled images, is
also fundamental. Ideally, in homogeneous regions, the ratio
image should contain only gamma-distributed, weakly
correlated speckle with no traces of the original image. Under
visual inspection, signal leakage and noise correlation
are easily detected in the ratio image, providing a subjective
but reliable quality assessment tool. Starting from the ratio
image, one can also compute simple numerical indexes, such
as the Kullback-Leibler divergence between the empirical
and theoretical distributions of speckle and the homogeneity
index proposed in [101] to measure spatial correlation.
Many other ways to measure despeckling performance
have been proposed in the literature, and their thorough
analysis extends beyond the scope of this article. We mention
only briefly the benchmarking framework proposed
in [102] and available online. Although relying on (physics-based)
image simulation, it tries to explicitly address
various aspects of despeckled image quality, from speckle
rejection to edge and texture preservation to radiometric
fidelity. Toward this end, a set of canonical scenes is
provided, both clean and noisy versions, together with a
number of numerical measurement tools. The weak point
is that the selected scenes cannot account for the complexity
of real SAR images. However, the availability of online
images and tools for objective assessment supports meaningful
comparisons.
We conclude this section by presenting the results of an
experiment aimed at providing some insight into the potential
of despeckling methods based on supervised deep
learning. Figure 6 presents a thin strip of a single-look
COSMO-SkyMed image acquired over Caserta, Italy, together
with a reference obtained by temporal-multilooking
25 coregistered images. Then, we show the output of three
model-based baselines-PPB, SAR-BM3D, and NL-SAR-
and two deep learning-based methods: the SAR-CNN and
CNN-NLM. It is clear that the latter provide a significant
quality improvement over model-based methods. PPB
and NL-SAR ensure good speckle rejection but introduce
patterned artifacts and tend to oversmooth the image,
sometimes losing structural details. On the contrary, SAR-
BM3D preserves all details well but does not reduce speckle
much. Deep learning-based methods seem to provide a
better compromise, with the accurate reproduction of informative
details and significant suppression of speckle in
homogeneous areas. However, a comparison with the 25look
reference shows that many details have been lost in the
process, especially in textured areas. This cannot be surprising,
considering the extremely noisy input, where the weakest
signal components are overwhelmed by speckle, which
makes their recovery nearly impossible. It is also worth
underlining that the two deep learning-based methods,
trained on the same data but having very different architectures,
provide similar results. This fact suggests that we may
43
IEEE Geoscience and Remote Sensing Magazine - June 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2021
