IEEE - Aerospace and Electronic Systems - November 2019 - 50

Deep Learning for Classiﬁcation of Mini-UAVs Using Micro-Doppler Spectrograms in Cognitive Radar
Table 2.

Deep Neural Network Test Cases for Classiﬁcation of Mini-UAVs Using Micro-Doppler Spectrograms
Case

Training
data

Test data

Neural
network

Sampling
rate

Integration
time

Accuracy

Simulated

CNN

96 kHz

10.7 ms

91.7%

Simulated

Measured

CNN

96 kHz

10.7 ms

36.7%

Measured

CNN

96 kHz

10.7 ms

82.8%

Measured

CNN

24 kHz

10.7 ms

87.1%

Measured

CNN

24 kHz

2.7 ms

97.7%

Measured

LSTM-RNN

24 kHz

2.7 ms

89.0%

CLASSIFICATION EXPERIMENTS AND RESULTS
To investigate the accuracy of different types of deep neural networks for the classification of mini-UAVs, several
test cases were defined, see Table 2. Test case A involves
the use of a CNN with a high sampling rate (96 kHz) of
the radar signal and a medium coherent integration time
(10.7 ms) in the Doppler filtering and simulated training
and test data. Test case B gives an idea if the simulated
spectrograms can be used to classify measured spectrograms from real mini-UAVs.
Test case C uses measured spectrograms for training
and testing. Test case D also uses measured spectrograms
for training and testing, but employs a reduced sampling
rate by decimating the radar signal with a factor 4. Test
case E uses a low sampling rate and a short integration
time (2.7 ms). Finally, test case F uses a long short-term
memory (LSTM) recurrent neural network (RNN) instead
of a CNN as a classifier [14].
Table 3 shows the configuration of the CNN used in the
first three test cases (A, B, and C). All convolutional and
fully connected layers (except for layer 13 which is linear)
use a rectified linear unit as a nonlinear activation function.
The convolutional and max-pool layers of the CNNs in test
cases D and E have smaller sizes than the configuration in
Table 3 due to the reduced sampling rate and integration
time. The CNN is trained with a categorical cross-entropy
loss function and the ADAM optimizer [5].
The LSTM RNN in test case F does not use a spectrogram as an input, but operates on an input sequence of
Doppler spectra. The LSTM-RNN configuration has four
layers and a hidden state size of 256 for the LSTM cells.
A comparison of the mean classification accuracy in
Table 2 shows that good results are achieved except for
the case where simulated spectrograms are used for training a CNN and measured spectrograms are used during
testing. This is perhaps not surprising due to the lack of
fidelity in the target models, which do not include
50

shadowing, multiple reflections, diffraction, and material
properties. More advanced RCS modeling techniques
using a hybrid finite element method and method of
moments approach may improve the fidelity of the microDoppler spectrograms at the expense of computational
cost [13].
A second conclusion is that the decimation of the sampling rate by a factor of 4 improves the mean classification
accuracy for the measured spectrograms considerably.
This is due to the fact that the part of the spectrogram that
does not contain any relevant information is removed by

Table 3.

CNN Conﬁguration
Layer

Type

Size

Kernel

Filters

Input

1024 Â 177

Convolution

1024 Â 177

5Â5

Convolution

1024 Â 177

5Â5

Max-pool

512 Â 88

2Â2

Convolution

512 Â 88

5Â5

Convolution

512 Â 88

5Â5

Max-pool

256 Â 44

2Â2

Convolution

256 Â 44

5Â5

Convolution

256 Â 44

5Â5

Max-pool

128 Â 22

2Â2

Flatten

1 Â 180224

Fully con.

1 Â 512

Fully con.

1Â5

Soft-max

1Â5

IEEE A&E SYSTEMS MAGAZINE

NOVEMBER 2019

IEEE - Aerospace and Electronic Systems - November 2019

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - November 2019