IEEE - Aerospace and Electronic Systems - July 2021 - 35

Wrabel et al.
performance. Unfortunately, each of the cited papers uses
its own kind of dataset and/or research focus such that a
final summary is quite difficult (Kovacs et al. [73] use
ISAR data gathered from a passive radar, Lu and Li [74]
created their own dataset out of six TerraSAR-X stripmap
mode SAR images, and Wang et al. [75] also created and
published their own dataset out of images from the Chinese
Gaofen-3 and Sentinel-1 SAR images). Depending
on the task different approaches are certainly worth trying
while our own research suggests that given enough data
and advanced augmentation techniques deep neural networks
perform the best.
Besides ATR applications there are also publications
on SAR-optical data fusion. Schmitt et al. [77], for example,
present a dataset for this task and give a nice overview
of recent work using pixel to pixel neural networks in
terms of generative adversarial networks (GAN) like the
well-known pix2pix network [78] or pseudo-Siamese networks
[79] to facilitate the task. Wang and Patel [80] present
an example of a GAN to generate high-quality visual
images from SAR data. A domain adaptation approach
transferring knowledge from the optical to the SAR
domain has been described in [81]. For this task, more
work needs to be done but the cited papers should give a
good starting point.
MICRO-DOPPLER/TRACK-BASED CLASSIFICATION
Depending on the radar type and the time frame in which
the classification must be completed, simple attributes like
velocity, heading, radar cross section, etc., might suffice
while fast results on single plots might require sophisticated
methods analyzing the Micro-Doppler spectrum.
Here, we will first report findings of the instantaneous
classifiers mostly based on the Micro-Doppler spectrum
and will then present work based on multiple plots/detections/tracks
of radar targets that might also combine the
Micro-Doppler-based classification results with the trackbased
information.
Micro-Doppler-based ATR is often done in multiple
steps. First, the spectrum is typically normalized with
respect to response amplitude or reference speed. Second,
features are extracted that are subsequently fed to a classifier,
see e.g., [82], [83]. Typical features are the main lobe
width, the number of lobes, Cepstrum coefficients [82],
[84], in-phase, and quadrature components like in Jouny
et al. [83], target patterns (e.g., " sawtooth " patterns in the
Cepstrum, or symmetry in the spectrum [85], [86]), or harmonic/subharmonic
analysis as described in Neumann and
Brosch [27] and Neumann and Senkowski [84]. Some of
those features like Cepstrum coefficients can also be automatically
tuned using machine learning techniques in
order to form better discriminative features; see e.g., Erol
et al. [87]. An interesting variant is presented by
JULY 2021
McConaghy et al. [88], in which the audio signal presented
to a tank operator is directly used to perform a classification
typically conducted by the operator. Here, the
preprocessing optimized for the human operator (the audio
signal) is used with a linear autoregressive classifier and
compared to a nonlinear predictor combined with a RBF
neural network. This demonstrates that even with features
optimized for human operators a successful classification
with neural networks is possible (with even higher accuracy
to be expected for an end-to-end optimized system,
of course). Our own studies showed that the best performance
can be achieved using deep neural networks (given
enough data and advanced augmentation techniques).
Depending on the use case and the hardware requirements,
it might still make sense to employ " traditional " classifiers
like decision trees or even nearest neighbor classifiers.
In contrast to the feature based micro-Doppler spectrum
based classification, the spectrum can be directly fed to the
neural network. An early contribution is, for example, Chakrabarti
et al. [89]. Back then, it was already recognized that
the addition of noise as a simple data augmentation significantly
improves results. This finding has recently been confirmedin
our ownstudies,inwhich we foundthatadvanced
data augmentation going way beyond the simple addition of
noise is a vital component to employ really deep networks
with several tens of layers; see Neumann and Brosch [27].
We could show that because of the high number of parameters
in deep neural networks, it is important to modify the
spectrum every time it is presented to the network. We
applied techniques adding noise and emulating slower and
faster target objects by modifying the time signal.
In contrast to Micro-Doppler-based ATR, track-based
ATR uses a longer time-period over multiple target hits to
recognize what kind of object has been detected. As in
modern Hensoldt radar systems, this can be combined
with Micro-Doppler-based ATR (see also Martinez [90])
or used standalone depending on the radar capabilities and
requirements. A multitude of variants has been investigated
ranging from alpha-beta-filters combined with random
forests applied to the trajectory [21], to feature-based
classification derived from the laws of rigid body movement;
see e.g., Garg and Singh [16]. Depending on the use
case, simple classifiers were found to be more robust than
more advanced ones in certain conditions [91], [92].
Mohajerin et al. [93] also use a simple neural network (a
multilayer perceptron) to distinguish simulated drone trajectories
from real air-traffic. A combination of MicroDoppler
and track based classification is presented by
Martinez [90]. Here, a constant false alarm rate filter is
combined with feature-based micro-Doppler classification
(comparing SVM, Naive Bayes, and kNN classifiers) and
joint probabilistic data association (PDA) filters for tracking.
Again the findings suggest that it is still very task-,
data-, and target-platform-dependent, but the cited contributions
should provide a good starting point.
IEEE A&E SYSTEMS MAGAZINE
35
IEEE - Aerospace and Electronic Systems - July 2021

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