IEEE Geoscience and Remote Sensing Magazine - June 2021 - 42

Unfortunately, no such thing as a clean SAR image exists,
as discussed in the " Data Availability " section. Therefore,
approximate solutions are necessary to provide the network
with an adequate number of reference images.
TRAINING PROCEDURES
In principle, one could remove speckle, and hence obtain
fully meaningful references, by means of temporal multilooking,
that is, by averaging a large number of coregistered
images of the same scene that have the very same signal
component but independent realizations of noise. Indeed,
this approach is followed in [56], [79], and [80], where a
stack of 26 single-look Constellation of Small Satellites for
Mediterranean Basin Observation (COSMO)-SkyMed images
is used for training with a leave-one-out strategy: 25 images
are multilooked to provide the desired reference for the
remaining one. A similar procedure is used in [76] with a
stack of 52 TerraSAR-X images. This approach, though appealing,
has two obvious limits. First, 25 or even 50 images
are not enough to adequately approximate a clean infinitelook
reference. A good despeckling filter can generate images
that are very smooth in homogeneous areas, with an
ENL easily exceeding 100. Therefore, for these areas, a 25look
reference represents quite a poor example. Not surprisingly,
under visual inspection, the images output by CNNs
trained on multilook references are more effective at preserving
high-frequency details than at suppressing speckle.
A second problem is that the signal must not change across
the multitemporal stack. This condition can be checked in
advance, keeping for training only those regions that pass
a suitable test. Of course, as more images are used, the
spanned temporal arc becomes longer, and the less likely it
is to find unchanged regions, so the two requirements are
somewhat at odds with each another.
Another alternative is to use images despeckled by other
methods as clean references. This approach does not really
make sense if applied to individual images. The trained network
could, at best, mimic the original algorithm. Moreover,
state-of-art methods cannot really provide filtered images
of adequate quality for training since speckle rejection
is often obtained at the cost of resolution loss and filtering
artifacts. Instead, this procedure can be effective when used
with data that have already been filtered in the temporal
domain. Given a carefully multilooked image, one can apply
a conservative despeckling filter to improve speckle rejection
in homogeneous regions, without significant side
effects. This approach is followed in [57] and [58], where
a relatively large stack of Sentinel-1 images is filtered in the
temporal dimension, and the result is then filtered in space
by applying the MuLoG and BM3D. A good reference is
eventually obtained. Contrary to previous methods, however,
this clean reference is not paired with a noisy image
of the stack for training but with a simulated noisy image
obtained by injecting speckle on the reference itself. While
this procedure provides many degrees of freedom for the experimental
phase, it creates poor, noisy images with speckle
42
that is fully uncorrelated, and it affects regions where the
fully developed model does not hold. Something similar is
done in [66], where simulated speckle is injected into multilooked
SAR data, without spatial filtering, to obtain noisy
data for controlled experiments.
Apart from the preceding exceptions, the vast majority of
the methods proposed in the literature adopt a fully simulated
training procedure. Noiseless optical images are used
as clean references paired with simulated noisy images obtained
through the injection of white speckle. The underlying
assumption is that the trained models will eventually
transfer well to the target domain. However, the simulation
procedure is flawed by several sources of inaccuracy, as discussed
in the " Data Availability " section, and, in fact, experimental
results provide only partial support for the use
of fully simulated training. In particular, the numerical
performance observed on simulated test data does not seem
to be a good predictor of despeckling quality in real SAR
images. Often, methods trained on synthetic data provide
large improvements over the state of the art on aligned data
but no apparent improvement on real-world SAR images.
Moreover, experiments carried out in [58] confirm that, despite
all the difficulties described previously, models trained
on real SAR data guarantee better despeckling quality than
comparable ones trained only on simulated data.
LOSS FUNCTIONS
Let us now briefly analyze the loss functions proposed in
the literature for the purpose of image despeckling. Basically,
all methods include a standard data fidelity term, in
most cases the 2
, distance [the Euclidean distance, mean
,
, and 2
,
square error (MSE), and despeckling gain] as well as the 1
and smoothed 1, distance. The relative merits of 1
losses are well known, with the former penalizing small
errors more harshly and the latter focusing more on large
IEEE Geoscience and Remote Sensing Magazine - June 2021

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