IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 24

A Survey on the Applications of Convolutional Neural Networks for Synthetic Aperture Radar: Recent Advances
Figure 4.
MSTAR dataset including optical and SAR images of ten military vehicles.
SAR-ATR. It was collected and released by the U.S. Air
Force Research Laboratory and Defense Advanced
Research Projects Agency. This dataset has been widely
used by so many studies for training and testing in SARATR
applications. MSTAR dataset, which contains ten
classes of ground military vehicles as shown in Figure 4,
has been collected by an X-band SAR sensor with 0.3 m
0.3 m resolution and full 360 aspect angles coverage.
Most of the target chips have the size of 128 128 pixels.
MSTAR has been used in two different configuration sets:
standard operation conditions (SOCs) and the extended
operation conditions (EOC). In SOC, typically, the depression
angle of 15 is used for training and the depression
angle of17 is used for test. Whereas, in EOC, there exists
more dissimilarity between training and test sets such as
larger depression angle variations and targets' intraclass
variants. Furthermore, OpenSARShip [70] and SAR Ship
Detection Dataset (SSDD) [71] are two other datasets typically
used for marine surveillance and ship detection
applications [72]. These two datasets will be discussed in
" Applications ofCNNs in SAR Data Analysis. "
LAND USE AND LAND COVER CLASSIFICATION
LULC classification is a fundamental application in RS.
SAR systems are generally employed for LULC classification
as they produce unique features that cannot be
obtained with electro-optical (EO) systems [73]. Land use
refers to the purpose of the land (such as agricultural, residential,
etc.), whereas land cover refers to the physical
land type (such as water, ice, etc.). There have been so
many studies for investigating the applicability of CNNs
for LULC classification, such as buildings [74], [75], urban
flood [76], [77] , and so forth.
Considering the input of proposed architectures, Chen
and Tao [78] have addressed the problem of land cover
classification using PolSAR data by exploiting null angle
features derived from the matrix rotation [79] as the input
of their CNN. Zhou et al. [80] proposed a CNN with six
input channels for terrain classification (such as water,
building, grass, etc.) using polarimetric SAR images.
Although monostatic multilook complex POLSAR data
can be represented by the Pauli-based polarimetric coherency
matrix T [81] that is a 3 3 complex matrix, they
24
proposed a new 6-D RV vector representation to feed their
CNN. Liu et al. [82] have also proposed a polarimetric
scattering coding matrix to feed a classifier based on a
fully convolutional network (FCN) for PolSAR land cover
classification. Wang et al. [83] proposed a fixed-feature
size CNN, which makes some reference to Lenet-5 [3], to
classify all pixels in a patch simultaneously using PolSAR
data for land cover classification. The polarimetric data of
all the pixels in a patch are used to generate a matrix as
the input to their CNN.
Considering some modifications to the structure of the
conventional deep CNNs, Ahishali et al. [84] proposed a
compact and adaptive CNN for LULC classification of
single-polarized COSMO-SkyMed and dual-polarized
TerraSAR-X intensity data. Bi et al. [85], [86] integrated
a graph model and CNN for PolSAR land cover classification.
Xie et al. [87] employed convolutional AE with
Wishart classifier, which uses the Wishart distance to measure
similarity, for land cover classification in PolSAR
images. Ren et al. [88] proposed an architecture for SAR
land cover classification by integrating CNN and sparse
AE for unsupervised feature learning.
The Flevoland1 image has been widely used for many
studies related to LULC PolSAR image classification.
Acquired by the AIRSAR platform in 1989, this image
consists of fifteen classes. The result of [80] together with
the Flevoland image are shown in Figure 5.
CNN-BASED REGRESSION FOR PARAMETERS
ESTIMATION
As previously mentioned, regression techniques are
employed to predict continuous and numerical parameters.
In CNN-based regression applications, the classification
layer is removed and the output of the last fully connected
layer is compared with the ground truth using loss functions
that are suitable for regression purposes. Wang
et al., in [89] and [90], used CNN for the regression problem
of ice concentration estimation using SAR images
during melt and freeze-up, respectively. In both regression
1The dataset including the toolbox can be downloaded from:
https://earth.esa.int/web/polsarpro/datasources/sample-datasets.
https://earth.esa.int/documents/653194/658149/AIRSAR_Flevoland
IEEE A&E SYSTEMS MAGAZINE
MAY 2022

https://earth.esa.int/web/polsarpro/datasources/sample-datasets https://earth.esa.int/documents/653194/658149/AIRSAR_Flevoland

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