IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 20

A Survey on the Applications of Convolutional Neural Networks for Synthetic Aperture Radar: Recent Advances
Figure 1.
Overview ofthe survey paper. CNN applications to SAR (left side) and related topics to CNN and SAR (right side) together with corresponding
keywords to each item.
vision field. The main advantages of CNNs in comparison
with their predecessors, i.e., classical regression and classification
methods, are automatic feature extraction using
hierarchical representations, weight sharing using convolution
layers, and spatial invariance using pooling operations.
CNNs have been proposed in the 1990s [3],
however, due to limited computational and processing
capacity and also the scarcity of annotated training data,
they fell out of fashion. Thanks to the availability of
large-scale dataset of optical images, CNNs have rapidly
developed in recent years. In 2012, Krizhevsky et al. [28]
succeeded to attract the attention to CNNs again by proposing
AlexNet, as it won the ImageNet Large-Scale
Visual Recognition Challenge (ILSVRC-2012). AlexNet
has five convolutional layers and three fully connected
layers. Motivated by this, many other powerful CNN
architectures, such as VGG-16 [29], GoogLeNet [30], and
Res-Net [31] have been proposed to gradually decrease
the top five error rates in the ImageNet challenge from
7.3%, 6.67% to 3.6%, respectively.
CNN MEETS SAR
CNNs have been extensively used to solve many SAR
problems such as ATR, scene classification, noise
removal, image formation, change and anomaly detection,
land use and land cover (LULC) classification, and so
forth. Moreover, CNNs showed their applications to
20
different SAR techniques such as inverse SAR (ISAR),
polarimetric SAR (PolSAR), and interferometric SAR
(InSAR). Apart from the details of CNN architectures,
such as the number of hidden layers and the kernel size of
convolution layers, there are some noticeable differences
between different state-of-the-art methods in SAR
domain. As a prominent example, some studies exploited
complex-valued architectures. Generally, both the input
data (including training, validation, and test sets) and the
network coefficients can be real or complex values. Hereafter,
we use the terms RV and CV for real-valued and
complex-valued CNN architectures. Complex input data
can be considered as two independent channels of real and
imaginary parts or magnitude and phase. Since usual computer
vision applications seek for classification of real
world images, like the well-known cats and dogs classification
example, CV-CNNs are uncommon. However, they
were introduced recently to exploit the phase information
embedded for example in SAR images. In radar imaging
scenarios, typically, only the amplitude information is
used. However, phase information can be employed for a
number of applications. We will thoroughly analyze the
applications of complex-values CNNs to SAR systems in
the section " Related Topics to CNN and SAR. "
CNNs can be used for both regression and classification
tasks. These two tasks are subsets of supervised ML.
The main goal of a regression task is to predict continuous
(numerical) variables, whereas classification is used for
the prediction of discrete (categorical) variables. The
IEEE A&E SYSTEMS MAGAZINE
MAY 2022

IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV