IEEE Geoscience and Remote Sensing Magazine - December 2021 - 153

were produced by ice experts via visual interpretation. The
algorithm was tested on dual-polarization RadarSat-2 data.
Since the problem under consideration concerns the regression
of a continuous value, the mean square error is selected
as the loss function. Experimental results demonstrate
that CNNs can offer a more accurate result than comparative
operational products.
In a different application, Song et al. [130] used a deep
CNN, including five pairs of convolutional and maximum
pooling layers followed by two fully connected layers, for inverting
rough surface parameters from SAR images. The network
training was based on simulated data, due solely to the
scarcity of real training material. The method was able to invert
the desired parameters with a reasonable accuracy, and
the authors showed that training a CNN for parameter inversion
purposes could be done quite efficiently. Furthermore,
Zhao et al. [131] designed a CV-CNN to directly learn physical
scattering signatures from PolSAR images. The authors
notably proposed a framework to automatically generate
labeled data, which led to a supervised learning algorithm
for the aforementioned parameter inversion. The approach
is similar to the study presented in [132], where the authors
used deep learning for SAR image colorization and for learning
a full PolSAR image from single-polarization data. Another
interesting application of deep learning in parameter
inversion was recently published in [133]. The authors propose
a deep neural network architecture containing a CNN
and a GAN to automatically learn SAR image simulation
parameters from a small number of real SAR images. They
later feed the learned parameters to a SAR simulator, such as
RaySAR [134], to generate a wide variety of simulated SAR
images, which can increase training data production and
improve the interpretation of SAR images that have complex
backscattering scenarios.
On the whole, deep learning-based parameter estimation
for SAR applications has not yet been fully exploited.
Unfortunately, most of the remote sensing community's
focus has been devoted to classical problems, which overlap
with computer vision tasks, such as classification, object
detection, segmentation, and denoising. One reason
for this might be that, since parameter estimation usually
requires the incorporation of appropriate physical models
and tackles the problem at hand as regression rather than
classification, domain knowledge is quite essential for
applying deep learning for such tasks, especially for SAR
images, with their peculiar physical characteristics. One
interesting study [87], described in detail in the " Terrain
Surface Classification " section, designs discriminative features
through the spectral analysis of complex-valued
SAR data and is an important work toward including deep
learning in parameter inversion studies using SAR data. We
hope that, in the future, more studies will be carried out in
this direction.
DESPECKLING
Speckle, which is caused by the coherent interaction among
scattered signals from subresolution objects, often makes
processing and interpreting SAR images difficult. Therefore,
despeckling is a crucial procedure before applying SAR
images to various tasks. Conventional methods aim at removing
speckle either spatially, where local spatial filters,
such as Lee [135], Kuan [136], and Frost filters [137], are
employed, or by using wavelet-based methods [138]-[140].
For a full overview of these techniques, the reader is referred
to [141]. During the past decade, patch-based methods for
speckle reduction have gained popularity due to their ability
to preserve spatial features while not sacrificing image resolution
[142]. Deledalle et al. [143] proposed one of the first
nonlocal patch-based methods applied to speckle reduction
by taking into account the statistical properties of speckle,
combined with the original nonlocal image denoising algorithm
introduced in [144]. A vast number of variations of the
nonlocal method for SAR despeckling have been proposed,
with the most notable ones included in [145] and [146].
(a)
(b)
FIGURE 5. (a) A VHR TerraSAR-X image of Berlin and (b) the predicted building mask [120] (right).
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
153

IEEE Geoscience and Remote Sensing Magazine - December 2021

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