IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 19

Image licensed by Ingram Publishing
summarized the work done by many researchers in this field
including high-cited peer-reviewed articles in addition to
some influential conference papers. The overview of this
paper is presented in Figure 1 where in its left side the taxonomy
ofCNN applications to SAR is shown and in its right
side the related topics to both CNN and SAR in order to be
discussed thoroughly are shown. Moreover, corresponding
key words to each item are also given for the ease of the
reader and they will be referred to throughout this article.
SAR is an active imaging sensor installed on a moving
platform such as an aircraft or a satellite and it offers high-resolution
images in all-day and all-weather conditions.Despite the
outstanding capabilities of SAR imagery, applying CNNs to
SAR images is more challenging than optical images. Because
ofthe unique features ofSAR systems, there is a large discrepancy
between SAR and optical images. In other words, radar
illumination direction fundamentally changes how a target
looks due to the self-shadowing phenomenon, noise is unique
to this modality (multiplicative speckle), dynamic range is
larger, etc. In addition, a large-scale SAR annotated dataset for
training is scarce because ofthe high cost ofSAR image acquisition.
As mentioned in Figure 1, Moving and Stationary Target
Acquisition and Recognition (MSTAR) dataset [8], as the
subject for a large body ofpublished work in automatic target
recognition (ATR) application [9] [10], is one of the few publicly
available SAR labeled dataset.
The rest of this article is organized as follows. In the section
" Deep Neural Networks in Computer Vision and Remote
Sensing " we introduce different DL approaches, and then we
concentrate onCNNs by addressing their role in both computer
vision and RS fields and presenting a brief introduction about
their typical structures. In " Applications of CNNs in SAR
Data Analysis, " subareas ofSAR data analysis that are tackled
by CNNs are reviewed. In the section " Related Topics to CNN
and SAR, " both CNN and SAR are thoroughly discussed.
Last, we present " Conclusions. " In addition, challenges and
future research directions are highlighted.
MAY 2022
DEEP NEURAL NETWORKS IN COMPUTER VISION AND RS
We start this section with a short introduction of different
DL approaches in order to highlight their differences and
then we concentrate on CNNs. A few applications to SAR
data for each approach will also be mentioned.
An AE is a symmetrical neural network used for unsupervised
learning from unlabeled data. In an undercomplete
AE, the encoder aims to compress the data into a
reduced dimension and the decoder aims to reconstruct it
back. A stacked autoencoder (SAE) is a deep network
model that makes use of multiple AEs, where the output
of one AE becomes the input of the subsequent one. Convolutional
AEs [11] have also been developed to provide
AEs with spatial characteristics of convolutions. AEs have
been widely used in SAR domain for polarimetric image
classification [12]-[16], SAR change detection [17], [18],
and so forth.
A DBN is formed by stacking two or more individual
unsupervised networks such as restricted Boltzmann
machines (RBMs) blocks [19] in such a way that each RBM
layer can be connected to both previous layers as well as subsequent
ones. In other words, a single RBM, which is an
undirected model with a visible layer and a single hidden
layer, has a limited power to extract features. Therefore, two
or more RBMs are typically stacked together to form DBNs.
It is noteworthy to mention that DBNs have been applied to
SARimage classification [20]-[23].
An RNN is typically employed for the analysis of nonstationary
signals. In an RNN, the input and output data can
have different lengths. In the context of SAR, it has been
used for image classification [24], as an inverse solver for
image formation problems in passive SAR [25], for agricultural
land cover mapping in satellite radar images [26] and
similarly for winter vegetation quality mapping [27].
CNNs have made obsolete the paradigm of handcrafted
feature extraction in the conventional computer
IEEE A&E SYSTEMS MAGAZINE
19
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
