IEEE - Aerospace and Electronic Systems - July 2021 - 32

ASurveyofArtificial Intelligence Approaches for Target Surveillance With Radar Sensors
Figure 3.
Illustration of the investigated selection modes transforming radar return measurements (cells) into input vectors processable by MLPs
according to [32].
The applied classifiers for clutter reduction are all
based on MLPs applying a backpropagation learning algorithm.
The number of neurons in the input layer results
from the chosen number of radar return measurements
(cells) that are selected either in a 1-D mode fashion, i.e.,
horizontally [49] or vertically [50], or a 2-D mode fashion,
i.e., horizontally and vertically from a square of cells
(cf. Figure 3) [32]. The MLP output layer consists of one
node that provides either a value of1 or 0 for target or clutter,
respectively. This output is generated for the middle
element of the selected radar measurements referred to as
cell under test. Indeed, transforming images to allow the
processing with MLPs was not uncommon at that time.
Today, the application of convolutional neural networks
would be preferable for this image processing task [60].
Considering the selection of the horizontal [49] or vertical
[50] 1-D mode, the application of the neural network
based approaches resulted in an average clutter power
reduction of 11 or 10 dB, respectively, and target power
increases of1 dB for both. As a result, SCR improvements
of 12 or 11 dB could be achieved. Both approaches lead to
similar results even though the input and hidden layer
compositions of the applied MLPs were different. Their
similar performances can be explained by the symmetrical
32
nature of radar coverage [32]. For comparison of the performances,
a cell averaging constant false alarm rate (CACFAR)
system served as a competing artifact. The CACFAR
system reduces sea clutter by determining a local
detection threshold through estimating the average clutter
level of the cells located next to the investigated cell [53].
While the CA-CFAR system could only achieve average
SCR improvements of 1.2 or 1.4 dB for 1-D or 2-D mode,
respectively, already the 1-D neural network approaches
considerably outperformed the CA-CFAR system with
approximately 12 dB SCR improvements [32].
The main advancement ofVicen-Bueno et al. [32] in this
context is the so-called Rhombus mode. The Rhombus mode
considers the combination ofboth 1-D modes leading to a 2D
selection mode. This mode outperformed the 1-D modes
by approximately 5 dB, achieving average SCR improvements
of more than 17 dB. Besides, all of the approaches
showed a satisfying robustness regarding varying conditions.
However, due to the higher complexity, the Rhombus mode
requires four times higher computational costs compared to
1-D modes. Thus, if no sufficient computational resources
are available, the application ofa 1-Dmode is preferable.
A more recent approach investigating image-based sea
clutter suppression is presented by Callaghan et al. [26]. In
IEEE A&E SYSTEMS MAGAZINE
JULY 2021

IEEE - Aerospace and Electronic Systems - July 2021

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