IEEE Geoscience and Remote Sensing Magazine - June 2021 - 47

as we highlighted in this article, self-supervised methods
are promising for the exploitation of the large collections
of real SAR images that are available and that remove the
need for a clean ground truth. This is an inescapable problem
that characterizes SAR despeckling, and the ultimate
goal would be to fully learn the specific characteristics of
the noise and the images from noisy data alone. The early
techniques we presented must be refined in terms of how
they handle temporal variability and the assumptions they
make about noise statistics, such as spatial independence.
More sophisticated priors could be integrated to increase the
representation capabilities of such models. Nonlocal selfsimilarity
is a prior that has been proved time and again to
be well suited for remote sensing imagery, and there is currently
great interest in neural network architectures that can
fully exploit it [27], [117]. Moreover, guiding the denoising
task with side information is useful to combat the difficulty
of extracting good features from images that are severely
affected by speckle, such as single-look ones. In this sense,
a few works [118], [90] explored using a coregistered optical
image, where salient features of the scene are more easily
identified, to guide the despeckling filter. Bringing this
technique to the deep learning world opens a number of
interesting questions, such as how to design an architecture
that effectively fuses the information.
COMMON DATA SETS
Even when deep models from existing papers are available,
a solid assessment of their performance should be carried
out on images that are publicly available to all researchers.
As for training, using private data makes full sense,
as the information can be considered part of the authors'
scientific choices. However, if one is interested in comparing
architectural solutions, without biases from different
training strategies, again, the very same training data and
procedures should be used. In general, we find it advisable
to employ public data sets and to clearly specify how
they have been used (e.g., training/validation/test splits).
This also applies to the fully simulated approach, where it
would be invaluable for the community to adopt common
data and procedures for model training and testing.
Despite the inherent problems and results that are still
inconclusive, we believe that using real SAR data is the
main avenue to eventually generating good reference data.
Toward this end, one should exploit all available resources.
Multitemporal data are now freely available for some sensors,
e.g., that of Sentinel-1, and collecting a large number of
coregistered images should be relatively easy. Moreover, optical
remote sensing images coregistered with SAR images
are also available, e.g., Sentinel-2 [97], and can be factored
in through suitable fusion. A careful combination of temporal
multilooking, spatial despeckling, and optical SAR
information fusion can certainly lead to high-quality data
sets for the reliable training of deep learning despeckling
methods. We encourage owners to consider making such
derivative data sets available to the community.
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
COMMON EVALUATION FRAMEWORKS
We acknowledge that, in some cases, making software
available may not be feasible (e.g., due to company policy).
Still, if researchers adopt common evaluation frameworks,
it is possible to compare the results obtained by their
methods with published outcomes of other approaches, as
long as these results have been worked out in a standardized
way on the same data set and with the same quality
metrics. An example is provided in [102]; it includes a set
of canonical scenes and corresponding simulated SAR images
along with corresponding objective measures on the
SAR images that account for speckle suppression and feature
preservation. This is an initial step toward standardizing
performance evaluation, and, while limited to a few
specific cases, it enables fair comparisons among different
methods. More generally, a first step toward building such a
common framework lies in the definition of a set of suitable
quality metrics to be adopted, as the works in the available
literature often employ different criteria.
SOFTWARE IMPLEMENTATION
Many papers claiming significant performance improvements
over the state of the art do not include a freely available
software implementation. Given the complexity of the
most recent methods as well as the effort to train a deep
neural network in terms of time, computational resources,
and data, it is difficult for researchers to implement, from
scratch, several novel methods to properly assess their
own technique. Authors should strive to make at least
the trained models for their approaches available so that
comparisons are possible. We note that, while this requires
more effort, it typically leads to a greater impact. Early dissemination,
e.g., in the form of a preprint, also helps to
broaden the impact of a work and to speed up the research
pace, and, in most venues, it does not prevent conference
and journal publication.
AUTHOR INFORMATION
Giulia Fracastoro (giulia.fracastoro@polito.it) is an
assistant professor in the Department of Electronics and
Telecommunications, Politecnico di Torino, Turin, 10129,
Italy. Her research interests include graph signal processing,
image processing, and deep learning. She is a Member
of IEEE.
Enrico Magli (enrico.magli@polito.it) is a full professor
at Politecnico di Torino, Turin, 10129, Italy. His research interests
include compressive sensing, image and video coding,
and vision. He is an associate editor of IEEE Transactions
on Circuits and Systems for Video Technology and EURASIP Journal
on Image and Video Processing and a former associate editor
of IEEE Transactions on Multimedia. He was an IEEE Distinguished
Lecturer from 2015 to 2016. He was the recipient
of the IEEE Geoscience and Remote Sensing Society 2011
Transactions Prize Paper Award and the 2010 and 2014 IEEE
Transactions on Circuits and Systems for Video Technology
Best Associate Editor Award. He is a Fellow of IEEE.
47
IEEE Geoscience and Remote Sensing Magazine - June 2021

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