IEEE - Aerospace and Electronic Systems - March 2021 - 69

Kechagias-Stamatis and Aouf
Table 10.

Hybrid Reﬂectivity Attribute-Based Methods
No

Reference

Type

SAR chip
size

Data augmentation

Main features

Zhang et al. [129]

features and deep
learning

128x128

Gabor filters, LBP, LSTM

Lv and Liu [130]

CNN

128x128

ASC-based various
orientations

Kechagias-Stamatis and
Aouf [139]

CNN, SRC

80x80

decision fusion

Ding et al. [141]

ASC, SRC

128x128

feature fusion

Karine et al. [132]

features, SRC

128x128

Gishkori and Mulgrew
[136]

Zernike moments,
SRC

96x96

Ding et al. [137]

azimuth sensitivity,
raw data

128x128

decision fusion

Zhang et al. [138]

Zernike moments,
SRC

70x70

2D slices of the SAR image along the
amplitude direction

Sun et al. [134]

features, raw data

65x65

LC-KSVD learned dictionary

ROI and for each detected keypoint a feature vector is created. The latter ones are converted into a matrix that is
used as a dictionary during the SRC process. Given that
for this method each SAR image provides a 2-D signature
rather than a 1-D, which is the norm for SRC, the authors
employ a multitask SRC, where the reconstruction error
originates from a block of atoms. Sun et al. [134] used the
original SAR image and the SIFT features, where the former representation incorporates intensity information and
the latter gradient information. Both representations are
used jointly [135], while the learned dictionary relies on
the LC-KSVD algorithm [77]. Gishkori and Mulgrew
[136] proposed a similar solution but utilize a set of rotation invariant Pseudo Zernike Moments (PZM) that are
generated from the entire SAR image. Then, the PZM are
employed to create a dictionary, which is used during the
SRC process. Ding et al. [137] suggested a decision-level
fusion of a two-stream SRC pipeline. The first SRC utilizes the SAR imagery, while the second SRC the azimuth
sensitivity image (ASI) which reveals the target sensitivity
at an azimuth angle. Finally, the target identity is defined
by comparing the classification scores of both SRC
streams. Zhang et al. [138] generated multilayer 2-D slices
of the SAR image along with the amplitude direction and
then calculate a set of Zernike moments (ZM) [2] for each
slice. Finally, an SRC scheme is employed exploiting the
ZM feature vector.
In [139], Kechagias-Stamatis and Aouf combined the
strengths of SRC and CNN in a decision fusion scheme.
For the SRC module, a novel adaptive elastic net type of
optimization is proposed that balances the advantages of
l1 À norm and l2 À norm depending on a Gaussian mixture model (GMM) [140] analysis of each SAR image.
MARCH 2021

decision fusion
-

Regarding the CNN module, the authors suggest utilizing
the AlexNet response map originating from a shallow
layer, while target classification is relying on a multiclass
SVM scheme. It should be noted that AlexNet is not
retrained or any transfer learning is applied, but rather the
pretrained weights of AlexNet are employed. Finally, the
target identity is determined based on a decision-level
scheme that adaptively changes its fusion weights. In
[141], Ding et al. combine the concepts of SRC and ASC
by hierarchically fusing the global and local features of a
SAR image. Hence, in the first stage, the Gaussian random
projection [142] features are used as global features that
are then employed for an SRC process. If the decision is
reliable, i.e., the reconstruction error does not exceed a
threshold, then the classification process terminates, while
if not, local features based on the ASC are extracted. The
target identity is defined by the template ASC closest to
the target ASC based on the Hungarian algorithm.

SUMMARY
Tables 4 to 10 give a summary of the reflectivity attributed-based methods. This category of techniques exploits
the raw SAR image without remapping it into a different
data domain.
a) Most recent papers involve deep learning structures
with the vast majority relying on convolutional neural networks.
b) CNN-based architectures incorporate various schemes
ranging from manual or GAN-based data augmentation to CNN distillation strategies, transfer learning
and fully exploiting pretrained state-of-the-art deep
networks.

IEEE A&E SYSTEMS MAGAZINE

IEEE - Aerospace and Electronic Systems - March 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - March 2021