IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 22

A Survey on the Applications of Convolutional Neural Networks for Synthetic Aperture Radar: Recent Advances
convergence and to avoid the vanishing gradient problem.
However, it also has some shortcomings. Since the derivative
of the ReLU function is always zero when the input
value is negative, it is likely that the phenomenon of neuronal
necrosis happens [40]. Neuronal necrosis or dying
ReLU problem refers to the condition in which ReLU
gives zero for any input and neurons become dead or inactive.
In order to solve this problem, Leaky ReLU was first
introduced. It multiplies the negative input by a small
value. Afterward, the Parametric ReLU (PReLU), was
introduced where the slope in the negative region is learnt
during the training phase.
POOLINGLAYER
A pooling layer, typically used after convolutional and
nonlinear layers, decreases the dimension of each feature
map. Pooling is performed by sliding a window (typically
square window) over each feature map, taking only one
value out of each window. Note that the down sampling is
only applied on the width and the height dimensions,
therefore, the number of channels will not be changed
after pooling. Max-pooling and average-pooling are two
most common pooling mechanisms used in CNNs. Moreover,
pooling helps to make the representation become
approximately invariant to small shifts in the input [41].
FULLYCONNECTEDLAYER
The output of the final pooling/convolutional layer is flattened
to a vector and fed to a fully connected layer. A fully
connected layer is to learn nonlinear combinations of previous
high-level features. All the input elements of this
layer are connected to its output elements similarly to a
traditional MLP neural network. One or more fully connected
layers can be used successively.
CLASSIFICATIONMODULE
The classification layer, which is the last module in CNN,
is used to map the output of the last fully connected layer
to the probability domain. Softmax classifier with crossentropy
loss is widely used for multiclass classification.
For the regression-type CNN applications, the classification
module is removed and the output of the last fully
connected layer is compared with the ground truth using
the MSE loss function or other loss functions that are suitable
for regression purposes.
APPLICATIONS OF CNN IN SAR DATA ANALYSIS
In this section, we categorize almost all the SAR tasks that
were tackled by CNNs in recent studies. A few other DL
techniques that incorporate convolution operation are also
mentioned. Note that the different parts of this section are
not semantically independent. For instance, we formed a
22
separate category for image denoising to discuss it in
detail. The emphasis of some papers is denoising, however,
some other papers address the denoising as a preprocessing
step for target recognition and other purposes.
Similarly, there are some parts such as data augmentation
and transfer learning that are related to CNNs and will be
addressed in " Related Topics to CNN and SAR. "
AUTOMATIC TARGET RECOGNITION
In the past few years, many SAR-ATR methods have been
proposed that made use of CNNs. ATR based on SAR
images plays a major role particularly in military applications
such as friend and foe identification, battlefield surveillance,
and disaster relief [42]. Ding et al. [42] utilized a
simple CNN architecture based on three convolutional
layers to implement a SAR-based ATR system. Since their
emphasis is on the data augmentation, we will cover it in
" Related Topics to CNN and SAR. " Chen et al. [43] proposed
a CNN architecture for ATR based on SAR images.
Since the number of training SAR images for each target is
limited, CNNs will suffer from severe overfitting. To
address this problem, the authors of [43] proposed an allconvolutional
network (A-ConvNet) with fewer degrees of
freedom by removing the fully connected layers. These two
papers, i.e., [42] and [43], are two of the most cited ones in
the SAR-ATR context.
Some studies have concentrated on the classifier of
their proposed CNNs for SAR-ATR. Wagner [44] proposed
a combination of CNN and support vector machine
(SVM), instead of a conventional softmax classifier, for
the task of SAR-ATR. Since SVM can discriminate
between two classes, they had to design ten SVMs for
their problem. Inspired by [44], Gao et al. [45] have also
addressed the combination of CNN and SVM for the
SAR-ATR task with a different loss function. Zhang et al.
[46] used an ensemble learning based classifier (linear
combination of many weak classifiers) to replace the softmax
layer in their CNN. Zhou et al. [47] utilized the
large-margin (LM) softmax classifier [48] in the last layer
to increase the divisibility of different target types for the
SAR-ATR task. The LM-softmax classifier was introduced
to increase the restriction of the traditional softmax
classifier and has been also applied to the SAR-ATR task
in [49]. It has also been noted by some studies that training
using the LM-softmax can be unstable [50]-[53].
Some studies have focused on the compression and
efficiency of CNNs for SAR-ATR tasks. A possible problem
of lightweight CNNs is that in case of a new dataset
they may lose their accuracies (in comparison with original
CNNs) due to underfitting and decreased model capacity.
Shao et al. [54] exploited depthwise convolution,
which is a kind ofmodel compression technique, to reduce
the CNN parameters and accelerate the calculation for 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