IEEE Geoscience and Remote Sensing Magazine - June 2021 - 37

stacks of SAR images, i.e., SAR images of the same scene
acquired at different times. The creation of such data sets is
critical. In particular, one of the main difficulties is to retrieve
a sufficient number of multitemporal images. Some
of the data sources presented in Table 1, such as Sentinel-1
and TerraSAR-X, provide acquisitions of the same scenes at
different moments in time and enable the creation of large
multitemporal data sets, which can be effectively used for
training. For example, Sentinel-1 has a revisit time of six
days, which supports fairly dense temporal sampling.
Moreover, several deep learning methods for SAR despeckling
use optical images to train the model in a supervised
fashion, as explained in the previous section. In this case,
general-purpose data sets used for visual problems, such
as BSDS500 [41], can be employed. Aerial images are also
widely used for training SAR despeckling methods, such
as those of the Google Maps [53], University of California,
Merced [54], and Northwestern Polytechnical University
data sets [55]. The images contained in such collections are
more similar to actual SAR images; therefore, including
them in the training set will help in the learning of priors
that can better match the characteristics of SAR data and
result in a reduced domain gap.
Even though many SAR data sets are available and can be
used for training and testing, a standardized procedure for
the experimental evaluation of the despeckling methods is
still lacking: every work considers a different data set, and,
even when the data set is the same, the partition between
training and test sets can change. Moreover, it is often the
case that the data sets used for training and testing are not
clearly described. This severely hinders the reproducibility of
the results and precludes the possibility of a fair comparison
between different methods. These issues represent a serious
problem for the advancement of the research in this field
since the lack of a rigorous performance assessment creates
TABLE 1. SAR DATA AVAILABILITY.
PLATFORM
YEAR
Sentinel-1
TerraSAR-X
Constellation of Small Satellites
for Mediterranean Basin Observation
(COSMO)-SkyMed
Radarsat-2
Airborne SAR
Advanced Land Observing
Satellite-Phased Array
L-Band SAR
European Remote Sensing
Satellite-2
2007-present
confusion and makes it impossible to discern the valid contributions
that provide advances with respect to the state of
the art. For these reasons, there is a strong need for standardized
data sets with predefined testing and training splits that
can lead to a fair comparative evaluation and reproducible
results. This is the only way the research in this field can continue
to make progress, and this aspect is discussed further
in the " Future Directions and Open Problems " section.
SUPERVISED MODELS: ARCHITECTURES
We now review despeckling methods based on supervised
deep learning models, focusing first on architectures (this
section) and then on training (in the " Supervised Models:
Training and Testing " section). The techniques are divided
in two large families, depending on their architecture.
The first includes " direct " methods based on a plain deep
network that accepts as input the noisy image and outputs
a denoised version. In the second, we cluster more elaborate
methods in which deep networks are used in the
context of a larger despeckling procedure, for example,
to estimate parameters of a model-based denoising engine.
For direct methods, special attention is devoted to
the fundamental choice of the noise model, explaining its
tight relationship with the filtering architecture. A summary
of the analyzed methods with their main features
is in Table 2.
DIRECT DEEP LEARNING-BASED DESPECKLING
Most methods proposed in the literature belong to this
family and, even in the context of direct filtering, display a
large variety of architectural solutions. So we further divide
them depending on two fundamental choices (an overview
appears in Figure 5): 1) working in the original or in the
log domain and 2) using a residual or a plain network. Indeed,
these choices are related, more or less explicitly, to
DESCRIPTION AVAILABILITY
LINK
2014-present [46]
[47]
2008-present [48]
2008-2019
1990-2004
2006-2011
[49]
[50]
[51]
Available upon registration
Available upon registration
Limited availability for users in ESA member
states, European Commission member states,
Canada, Africa, and China
Limited availability for users in ESA member
states and European Commission member states
Available upon registration
Available upon registration
https://scihub.copernicus.eu/
https://tpm-ds.eo.esa.int/oads/
access/collection/TerraSAR-X
https://tpm-ds.eo.esa.int/oads/
access/collection/CosmoSkyMed
https://tpm-ds.eo.esa.int/oads/
access/collection/Radarsat-2
https://asf.alaska.edu/data-sets/
sar-data-sets/airsar/
*https://alos-palsar-ds.eo.esa.int/
oads/access/
†https://asf.alaska.edu/data-sets/
sar-data-sets/alos-palsar/
1995-2011
[52]
Available upon registration
*Spatial coverage: Europe, Africa, and the Middle East plus some worldwide products received from the Japan Aerospace Exploration Agency.
†Spatial coverage: the Americas, Antarctica, and selected global sites.
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
37
https://asf.alaska.edu/data-sets/
sar-data-sets/ers-2/

IEEE Geoscience and Remote Sensing Magazine - June 2021

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