IEEE - Aerospace and Electronic Systems - July 2021 - 34

ASurveyofArtificial Intelligence Approaches for Target Surveillance With Radar Sensors
Figure 4.
Taxonomy of automatic target recognition and commonly applied AI methods in this area.
deep neural networks excelled in the past decade [61]-
[66]. The core challenge is not only to recognize a target
but also to identify what kind of target has been detected.
Here, we distinguish two fundamentally different types of
classification tasks: Image-like classification for detecting
targets in (inverse) synthetic aperture radar SAR/ISAR or
high resolution range profile (HRRP) data and plot/MicroDoppler/track-based
classification for data originating
from typical radars scanning in range/azimuth/elevation
(cf. Figure 4). Here, we focus on the literature regarding
the classification process itself. For a general overview
of future technological challenges of modern radar systems
to also improve classification see, for example,
[10], [67], [68].
IMAGE-LIKE CLASSIFICATION FOR SAR/ISAR
APPLICATIONS
Image-like classification can be applied to several kinds of
data with the most prominent ones being SAR or ISAR
applications. High resolution range profiles (HRRPs),
however, can also be seen as similar candidates for imagelike
classification algorithms (in one instead oftwo dimensions).
Here, we try to present a few examples out of the
huge body of papers on this topic without trying to be
complete. The core ideas, however, repeat themselves.
Already in 1999 [69] demonstrated the usability of neural
networks for the task of HRRP classification. In [33],
Jurado et al. used simulated HRRPs to identify air-targets
with a 2-layer neural network while Lu et al. [70] present
a " deep learning " approach (with about 3-layers) on
34
satellite HRRPs using a combination of recurrent units
and support vector machines while Ma et al. [39] present
an approach that only uses SVM classification. In contrast,
Lunden [46] uses a convolutional neuronal network of
similar size without a SVM on simulated multistatic radar
data. Further tweaks of SVMs toward improved ATR on
HRRPs are presented in [38] and [71] for multiclass SVM
decision trees while Feixue et al. [42] present a recurrent
neural network applied to profiles in the radial axis.
Already early on Bayes error estimates have been used to
address the problem of not only classifying the data but
also simultaneously providing an error estimate [13].
Note, however, that similar Bayes estimates can be
derived from classifier ensembles for other classifiers like
deep neural networks as well. Finally, the problem has
also been investigated in the bigger context of assisting
human operators [72].
In the SAR/ISAR context, similar to object detection
tasks in the visual domain, there exist two-stage classifiers
first detecting objects before classifying them, for example
[73]-[75], but also single stage detectors directly assessing
whether there is an object and classifying it at the same
time (see Wang et al. [75] for a comparison of the two
kinds of detectors). While modern deep learning methods
often directly use the raw image-like representation, there
is also a huge body of algorithms that build upon features
extracted before (e.g., [73]). The impact of improved SAR
data using polarimetric sensors, for example, has been
sketched by Neumann and Brandfass [76]. The addition of
further radar information due to the usage of polarimetric
radar sensors essentially leads to more discriminating
features that should result in better classification
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
