IEEE - Aerospace and Electronic Systems - July 2021 - 30

ASurveyofArtificial Intelligence Approaches for Target Surveillance With Radar Sensors
Figure 2.
Scheme of the data processing chain used for clutter identification.
clutter but identified by the operator as suspects or targets
have been added to the training set, labeled as targets. By
this means, the target plots which are difficult to distinguish
from clutter enhance the training set in terms
of diversity.
Further on, the data were processed according to the
scheme displayed in Figure 2. Before classification, a
preprocessing procedure is performed starting with the
removal of ground clutter near the airport based on a
ground clutter map, which is followed by the application
of a CFAR system and a plot extractor. When using
single radar plots, the plot extraction is directly followed
by the feature selection [51], whereas an additional
tracking step to assign the plots to each other is
required when two plots from consecutive scans are
considered [31].
In [51], the features were selected following two different
methods. The first set of features comprises second
and third order moments that result from the moments of
inertia. In contrast, the second set of features consists of
the six most prominent features which characterize the
plot shape. Those are determined by considering the features
an expert radar operator uses to visually distinguish
between snow clutter and target. In [31], the feature selection
was performed based on the latter method, but as two
consecutive plots instead of single plots are investigated,
the feature vector is composed of 13 components, resulting
from the 6 components of each plot as well as one
component representing a measure of the distance
between them.
Further on, classification performances were assessed
by conducting a 10-fold cross-validation with a Bayes
classifier, serving as baseline for comparison, as well as
MLPs and radial basis function (RBF) networks [31],
[51]. Both networks included only one hidden layer.
Thereby, the special characteristic of RBF networks consists
of the composition and activation of the hidden layer:
Each hidden neuron contains a center vector and is activated
by the radial basis function based on the distance
between input and center vector [57], [58].
The results of [31] revealed varying performances
of the evaluated classifiers depending on the considered
class: For the correct classification rate of clutter,
the RBF network showed the best result with 98.3%
followed by the Bayesian classifier (97.8%) and the
MLP (96.6%) with a slightly worse performance. However,
the MLP reached the highest correct classification
30
rate of targets with 95.1% compared to the RBF network
with 88.4% and Bayes, whose classification rate
was considerably lower with 78.5%. Regardless of the
number of hidden layer neurons, the MLPs applied in
[31] showed higher classification rates for targets as
well as for clutter compared to the Bayes classifier
with a correct classification rate of 93.1% for targets
and 91.3% for clutter. The best performance showed a
MLP with 11 hidden neurons considering two plots
from consecutive scans leading to correct classification
rates of 98.6% for clutter and 98.8% for targets. Therefore,
the MLP accomplished to outperform the approaches
from Pierucci and L. Bocchi [51] that
considered only single radar plots as input. Overall,
their results reveal that all of the neural networks
could outperform the Bayes classifiers, while the best
results for the MLP with 11 hidden neurons using two
consecutive plots as input demonstrate a reliable capability
for distinguishing snow clutter from targets [31].
Summing up, the comparison of the classification performances
of different authors indicates that neural network
based approaches outperform Bayes classifiers for
identifying clutter and reveal promising classification
accuracies for application in practice.
AI-BASED CLUTTER IDENTIFICATION IN A MARINE
ENVIRONMENT
One of the main differences among the AI-based clutter
identification approaches in a marine environment is the
kind of input data used for classification. Whereas some
researchers focus on the application of AI techniques
using extracted features in the form of parameters [20],
[55], others use radar images as input [26], [32], [49],
[50], [52], [56].
One of the earliest approaches based on features in
the form of parameters was presented by Bouvier et al.
[55]. In order to distinguish four clutter types (including
land, calm sea, rough sea, and composite clutter), they
applied a neural network classifier on clutter data collected
by a S-band radar located on Mediterranean seashore.
This 2-D (azimuth and range) surveillance radar
was operated under different conditions regarding the
sea swell structure (calm and rough sea) and the seashore
structure (flat or cliff seashore) in order to collect sea or
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
