IEEE Geoscience and Remote Sensing Magazine - June 2021 - 34

paramount importance is to understand what data have
been used in training. If significant discrepancies exist between
training and testing data, deep models can perform
poorly and provide unreliable results. In this article, we
point out that this is actually the case for deep despeckling
methods that are trained on synthetically speckled optical
images and then tested on real SAR data, but the same could
be argued for other tasks where images are collected from
different sensors, acquisition geometries, and so forth. This
gap between the training and testing data distributions can
be fixed by fine-tuning the pretrained model on more representative
data and by means of techniques such as domain
adaptation [39]. Finally, we also want to suggest that the neural
network design may include prior knowledge of the task
that is able to regularize the model and extract better performance.
This can be done at several levels, and examples of
the despeckling problems that are discussed in the following
sections include network architectures matching noise modeling;
loss functions that specifically enforce known statistical
properties of data; layers implementing operations that
exploit priors, such as nonlocal self-similarity; and explicit
prior models that reduce the need for annotated data.
SAR DATA
This section discusses the key component of any data-driven
algorithm, i.e., the data. First, we delve into the challenges
that data-driven despeckling techniques face, discussing
the various data exploitation options. Then, we analyze
which data sets are available to tackle the SAR despeckling
problem, emphasizing whether they can be retrieved from
the public domain or are difficult to access. In this treatment,
we also discuss how the available data have been
used by the remote sensing community so far, highlighting
a lack of reproducibility and standardized testing procedures
for fair evaluation.
DATA USE
The success of supervised learning with problems such as
image classification, segmentation, and so on has quickly
led to the development of supervised denoisers [24]-[27],
[42], [43], i.e., CNNs that learn a mapping from noisy-todenoised
images by using clean images of the same class
as ground truth in the training process. This setup is appealing
in the classical setting of additive Gaussian noise
removal because it is not difficult to retrieve a large quantity
of virtually noiseless images (e.g., high-quality photos
with very low levels of camera noise) and create an arbitrary
number of noisy images by sampling realizations of
Gaussian RVs. It is only natural that several works in the
SAR despeckling literature tried to replicate this setting.
However, any SAR despeckling technique must face the
fact that no " clean SAR images " exist.
As discussed in the " Primer on SAR Speckle " section, it
is not possible to generate clean references by multilooking,
as this process degrades the spatial resolution. Therefore,
the literature addressing the despeckling problem through
34
supervised deep learning (see the " Supervised Models: Architectures "
and " Supervised Models: Training and Testing "
sections) has essentially presented two methods to create
reference images to be used as ground truth: synthetic
speckle generation and multitemporal fusion. Each of these
has its own advantages and disadvantages, as discussed in
the following.
The synthetic speckle generation approach starts from optical
images, either satellite data or plain photographs [e.g.,
the Berkeley Segmentation Dataset 500 (BSDS500) [44] is
commonly used], where the amount of noise can be considered
negligible. It then uses a model of speckle, such as the
one in (1), and a statistical characterization of the speckle,
e.g., as in (4), to sample realizations of the speckle process
and superimpose these on the clean optical data. This procedure
is a simple means to generate data that, at first glance,
resemble speckled SAR images and can be used to train denoisers
for multiplicative, instead of additive, noise. However,
this simplicity comes at the cost of several disadvantages.
First and foremost, while synthetically speckled optical images
may vaguely resemble actual SAR images, due the different
properties of materials, they match neither their spatial
structure nor their radiometric statistics, which is not surprising
given the completely different nature of the two imaging
mechanisms. In particular, the existence of strong reflectors
is not modeled by using optical images; moreover, priors on
texture patterns and edges learned from optical data may not
match the characteristics of actual SAR data.
Figure 3 illustrates the differences between a real
SAR image and one with synthetically generated speckle.
This mismatch is the source of a problem often referred to
as the domain gap [45], whereby the features of the test set
(real SAR images) may differ from those of the training set
and induce anomalous inference behavior. This can appear
as artifacts, hallucinations of patterns that were prevalent
in the training set, and oversmoothing. Figure 4 demonstrates
how a CNN trained on synthetically speckled optical
images tends to produce a cartoon-like despeckled image,
either oversmoothing regions or hallucinating patterns in
supposedly flat regions. The domain gap is a serious problem
that should not be disregarded, as it reduces the final
users' confidence in the generated products. In addition,
the denoiser will be only as good as the model it relies on;
there are several details that are overlooked by simple models,
such as those presented in the " Primer on SAR Speckle "
section. An important example is the assumption that the
speckle process is spatially uncorrelated. This is an almost
universally used assumption, but it is hardly satisfied in
practice, as it has been observed that the point spread function
of the SAR focusing algorithms generates spatially correlated
speckle [31].
The multitemporal fusion approach uses a stack of actual
SAR images acquired at different temporal instants and exploits
the temporal incoherence of the speckle to effectively
suppress it and generate a clean ground truth. The clear advantage
is that a denoiser trained with this approach would
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2021
IEEE Geoscience and Remote Sensing Magazine - June 2021

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