XLIII SIMPÓSIO BRASILEIRO DE TELECOMUNICAÇÕES E PROCESSAMENTO DE SINAIS - SBrT 2025, 29 DE SETEMBRO A 2 DE OUTUBRO–11 DE 2025, NATAL, RN
Deep Learning Models Applied in Automatic
Modulation Classification of Radar Signals
Pedro de Figueiredo Abissamra, Sarah Negreiros de Carvalho Leite, Renato Machado, and Dimas Irion Alves
Abstract— Artificial Intelligence in Electronic Warfare has in [7] proposed using an AlexNet to classify twelve radar
gained prominence, particularly for Automatic Modulation Clas- modulations. The network accomplished the task with 97.58%
sification tasks. Deep learning methods have demonstrated ro- accuracy at an SNR of -6 dB. In another early study, a
bustness and high accuracy in addressing this challenge. This
study proposed and tested Long Short-Term Memory and Convo- coordinate attention model was introduced to a ShuffleNet,
lutional Neural Network architectures for Automatic Modulation which could also distinguish between twelve classes with
Classification in radar signals. The LSTM model achieved 90% 98.14% accuracy at an SNR of -8 dB [8].
accuracy in classifying eleven modulation types at -2.66 dB SNR, Recurrent Neural Networks (RNN) have been proposed for
while the CNN model reached the same accuracy at 1.50 dB different applications [9], [10]. The study in [9] used RNNs
SNR. Although the LSTM outperformed the CNN, it required
higher computational resources and longer latency. to classify, denoise, and deinterleave pulse streams, while [10]
focused on classifying radar emitters. Additionally, the authors
Keywords— AMC, Convolutional Neural Networks, Long in [4] developed a novel method that combined a shallow
Short-Term Memory, Radar Signals.
Convolutional Neural Network (CNN) with a bidirectional
Long Short-Term Memory (Bi-LSTM) network to address the
I. I NTRODUCTION modulation classification task for radar signals. This combined
There has been a growing demand for automation in Elec- approach achieved an impressive 95% accuracy in classifying
tronic Warfare (EW) systems. To address this need, integrating eight modulation types under -10 dB Signal-to-Noise Ratio
Artificial Intelligence (AI) into EW systems has emerged as a (SNR) conditions. More recently, an LSTM network capable
solution [1]. AI’s ability to provide effective decision support, of classifying twenty-three distinct types of radar modulations
manage large volumes of data, and enhance decision-making was introduced, achieving 90% accuracy at -2 dB SNR [11].
processes allows for improved self-control, self-regulation, and These advancements highlight the growing prevalence of
self-actuation in these military systems [2]. deep learning techniques in electronic warfare, particularly
Automatic Modulation Classification (AMC) or Automatic for radar signal classification. This paper aims to evaluate the
Modulation Recognition (AMR) consists of identifying the performance of LSTM and CNN architectures for automatic
modulation scheme of a received signal without any prior modulation classification (AMC) of radar signals.
information. In a non-cooperative environment, such as in This paper is organized as follows. Section II describes
EW, this becomes a critical challenge, which makes the AMC the dataset, the deep learning architectures used, and how
crucial for achieving tactical superiority, especially when the performance evaluation was executed. Section III presents
dealing with the AMC of radar signals [3]. The challenge the results obtained for each tested model, as well as a
of AMC for radar signals is becoming an urgent problem discussion about those results. Section IV concludes the paper
in electronic countermeasure systems. As the electromagnetic and presents future work to be accomplished.
environment on the battlefield becomes increasingly complex,
the parameters of radar signals are also evolving accordingly. II. M ETHODOLOGY
Conventional techniques are more likely to have poorer perfor- This work applied two deep learning models, namely a Long
mance and higher computational complexity, especially under Short-Term Memory and a Convolutional Neural Network, in
low signal-to-noise ratio (SNR) conditions [4]. Due to the the Automatic Modulation Classification task. For each model,
high accuracy and robustness capabilities of deep learning a baseline architecture was chosen as a reference to verify the
techniques, they have begun to be widely used to solve AMR performance of the proposed architectures. The dataset used
tasks [5]. was the DeepRadar2022, which will be described next.
Different deep learning techniques have been used in the
context of AMC for radar signals. One example is the ap- A. DeepRadar2022
plication of a Convolutional Neural Network (CNN) Le-Net-5
to classify eight types of radar modulations, achieving 96.52% DeepRadar2022 is a dataset created by [11]. It consists of
accuracy under a -2 dB SNR condition [6]. The work presented modulated In-Phase and Quadrature (IQ) radar signals sampled
at 100MHz. It has 21 modulation classes plus a noise and a
Pedro de Figueiredo Abissamra, Sarah Negreiros de Carvalho Leite, non-modulated class, totaling 23 balanced classes. All data
Renato Machado, and Dimas Irion Alves are with the Electronic Warfare was created using Matlab, and signals with different SNRs
Laboratory, Aeronautics Institute of Technology, São José dos Campos-SP.
E-mails: [Link].101808@[Link], [Link]@[Link], {rmachado, were created for each class. The SNR varies from -12 dB to
dimasirion}@[Link]. 20 dB with a step of 2 dB, summing up 782000 different
XLIII SIMPÓSIO BRASILEIRO DE TELECOMUNICAÇÕES E PROCESSAMENTO DE SINAIS - SBrT 2025, 29 DE SETEMBRO A 2 DE OUTUBRO–11 DE 2025, NATAL, RN
signals with a (1024 × 2) shape. They are already separated three. A max pooling layer is used between each convolution
into 60% for training and 20% for test and validation. The full layer with a pool size of two. These layers are followed by
dataset can be found in [12]. two fully connected layers with 128 neurons and the output
In this paper, due to hardware limitations, a reduced dataset layer with 11 neurons. The ReLU activation function is used
with 11 classes was used: Linear Frequency Modulation, in all convolution layers. The dense layers employ a SELU
Frequency Modulation with Costas Code, Binary Phase-Shift activation function, whereas the output layer uses the Softmax
Keying, Phase Modulation with Barker code, Phase Mod- activation function. The input layer has a (1024 × 2) shape,
ulation with Frank code, Phase Modulation with P1 code, and the total number of parameters of CNN 1 is 158.155.
Phase Modulation with P2 code, Phase Modulation with P3 The proposed CNN architecture (CNN 2) consists of six
code, Phase Modulation with P4 code, Non-Modulation and one-dimensional convolutional layers, each with a distinct
Complex White Gaussian Noise (noise). The total number of number of filters. In the first layer, there are 512 filters, and
training data was 224400; and for test and validation, this as the next layer arises, the number of filters is divided by
amount was 74800. two. This way, in the sixth layer, there are only 32 filters.
As discussed in CNN 1, the kernel size is three, and a max
B. Long Short-Term Memory pooling layer with a pool size of two is used between each
The LSTM baseline architecture (referred to here as LSTM convolutional layer. After the first and second max pooling
1) was proposed by [11]. It consists of three stacked LSTM layers, a dropout of 20% was inserted. Four fully connected
layers, a dense layer, and the classification layer beside the layers were used. In the first layer, there are 256 neurons,
input layer. Each LSTM layer comprises 128 cells; the fully while in the others, this number decays by a factor of two
connected (FC) layer uses a softmax activation function with until 32 neurons in the last layer before the output, which
11 output neurons, and the input layer has a (1024 × 2) shape. has 11 neurons. A dropout rate of 20% was also applied after
The total number of parameters is 331.659. the first and second FC layers. The activation functions used
The proposed LSTM architecture (referred to here as LSTM were the same as in CNN 1. A description of this network is
2) is similar to the baseline architecture. The only difference presented in Figure 2. The total number of parameters of the
is that this new one has four stacked LSTM layers with 64 new network is 713.579.
cells each. The input and the dense layer are the same as the
previous, as shown in Figure 1. This new network has a total
of 116.939 parameters.
Fig. 1. Layers description of the proposed LSTM architecture (LSTM 2).
The training parameters were the same for both architec-
tures. A total of 500 epochs were used with a batch size of 512
samples. The optimizer used was Adaptive Moment Estimation Fig. 2. Layers description of the proposed CNN architecture (CNN 2).
(ADAM) with a cyclical learning rate varying from 1.10−7 to
1.10−3 . These networks were trained on a personal computer
equipped with an AMD Ryzen 7 3700x CPU, 32 GB RAM,
and a 12 GB RTX3060 Nvidia GPU. The testing hardware The same training parameters were used for both architec-
was a server equipped with 104 x Intel Xeon Gold 5320 CPU tures. A total of 80 epochs were used with a batch size of 128
and 1 TB of RAM. samples. The optimizer was ADAM with a cyclical learning
rate varying from 1.10−7 to 1.10−3 . An early stopping monitor
was also used with a patience value of 5 epochs, and the
C. Convolutional Neural Network monitored parameter was the accuracy of the validation set.
The CNN baseline architecture (referred to here as CNN 1) The training and testing hardware was a server equipped with
was proposed by [13]. It consists of seven one-dimensional 104 x Intel Xeon Gold 5320 CPU and 1 TB RAM (the same
convolution layers with 64 filters each and a kernel size of used for testing the LSTM).
XLIII SIMPÓSIO BRASILEIRO DE TELECOMUNICAÇÕES E PROCESSAMENTO DE SINAIS - SBrT 2025, 29 DE SETEMBRO A 2 DE OUTUBRO–11 DE 2025, NATAL, RN
TABLE I
D. Performance Verification
AVERAGE PERFORMANCE TEST RESULTS OF THE LSTM ARCHITECTURES
To gain a more comprehensive understanding of the mod-
els’ and architecture’s performance, both the baseline and Accuracy Latency/
SNR SNR
the proposed networks were trained and tested. The LSTM Model Epochs Modulations General
(%) Sample (ms)
(dB) (dB)
network underwent three training and testing cycles due to
LSTM
its high computational demands and limited hardware avail- 89.67 500 29.23 1.33 -2.58
1
ability. For the CNN network, five training and testing cycles LSTM
were performed. After each cycle, all performance parameters 2
89.33 500 21.33 0.83 -2.66
(explained in Section III) were recorded, and their average
values were calculated.
Once all tests were completed, a statistical analysis was
conducted solely for the CNN models, given the limited Figure 3 presents the LSTM 2 general accuracy as a function
observation quantity for the LSTM network, to ensure that of the SNR in one of the tests done. The general SNR that
the proposed CNN network performed better than the base- gives a 90% accuracy is -2.50 dB. Figure 4 presents the
line method. The statistical test employed was the one-way SNR per modulation obtained for the same architecture in the
ANOVA, with its approval criteria of a p-value of 0.05 same test. It can be seen that the minimum SNR at which
or lower. In other words, if the calculated p-value of the the network can maintain 90% accuracy is 0 dB, and the
performance parameters observations was equal to or lower modulation that limits this condition is the phase modulation
than 0.05, an improvement with statistical significance could with P1 code.
be noticed. To ensure that this evaluation method would be
valid for these data (which respected a normal distribution),
the Kolmogorov-Smirnov test was previously executed for all
parameters in the evaluation, except for the number of epochs
and latency per sample, which can vary depending on the
computer in use.
III. R ESULTS AND D ISCUSSION
The performance of all architectures was verified by their
accuracy on the test set, number of epochs to train the model,
latency (for a batch of 32 samples), and minimum (SNR) for
a 90% accuracy. The SNR evaluation was done using two
approaches. The first approach outputs the average accuracy
(general accuracy) performance considering the assessment of
all modulations, while the second considers each modulation Fig. 3. General accuracy as a function of SNR of the proposed LSTM
separately. These approaches were tested for the LSTM and architecture.
CNN architectures, as described and discussed throughout this
section.
A. LSTM
The average results obtained for the LSTM 1 as well as
for the LSTM 2 are presented in Table I. Observing these
results, it is noticeable that, in general, the performance of the
proposed model is similar to that of the baseline model, except
for the latency per sample parameter, which is 27.03% lower
in LSTM 2. This is a consequence of the reduced number
of parameters compared to LSTM 1. However, although the
total number of parameters of LSTM 2 is reduced by 65%
compared to LSTM 1, the latency per sample does not follow
this proportion. The probable reason for that is the addition of
the fourth layer in the network.
Figure 3 presents the LSTM 2 general accuracy as a function Fig. 4. Modulation accuracy as a function of SNR of the proposed LSTM
of the SNR in one of the tests. The general SNR that gives architecture.
a 90% accuracy is -2.50 dB. Figure 4 presents the SNR per
modulation obtained for the same architecture in the same test. From Figure 5, it is possible to identify that the LSTM has
It can be seen that the minimum SNR required for the network more difficulty in classifying phase modulations, especially
to maintain 90% accuracy is 0 dB, and the modulation that with Frank and P codes. In contrast, the LSTM excelled in the
limits this condition is the phase modulation with the P1 code. classification task for chirp, 2PSK, noise, and non-modulated
XLIII SIMPÓSIO BRASILEIRO DE TELECOMUNICAÇÕES E PROCESSAMENTO DE SINAIS - SBrT 2025, 29 DE SETEMBRO A 2 DE OUTUBRO–11 DE 2025, NATAL, RN
signals, achieving 90% accuracy under a -9 dB condition. presented an improved performance compared to CNN 1 when
observing the accuracy and the general SNR.
The CNN 2 result for general accuracy as a function of the
SNR is shown in Figure 6. As we can see, the SNR that gives
a 90% accuracy is 1 dB. The accuracy per modulation as a
function of the SNR for the same test is presented in Figure
7. As a result, the minimum SNR at which the network can
maintain 90% accuracy is approximately 7 dB. The modulation
that limits this condition is the phase modulation with the
P4 code. Figure 8 presents a confusion matrix obtained for
this CNN architecture. The CNN also has more difficulty
classifying phase modulations, especially with Frank and P
codes. Similarly to the LSTM, the CNN also excelled in the
classification task for chirp, noise, and non-modulated signals,
classifying them with 90% accuracy under a -7.5 dB condition.
Fig. 5. Confusion matrix showing the misclassification errors of the proposed
LSTM architecture.
B. CNN
The average results obtained for each performance param-
eter for both CNN models tested are presented in Table II.
Analyzing it, it can be seen that the proposed network (CNN
2) was able to overcome 1,8% the baseline network in terms
Fig. 6. General accuracy as a function of SNR of the proposed CNN
of accuracy. Consequently, the minimum SNR (general and architecture.
per modulation) parameter is also reduced.
Comparing both SNR parameters, a 19% reduction (1.6
dB) was observed when examining the modulations separately,
and a 35% reduction (0.70 dB) was noted when considering
all modulations together. Nevertheless, when the number of
epochs and latency per sample is observed, it is noticed that
CNN 2 is heavier than CNN 1 once the number of epochs
needed to train the network is almost twice that of CNN 1,
and the latency parameter is 2.88 times higher than that of
CNN 1.
TABLE II
AVERAGE PERFORMANCE TEST RESULTS OF THE CNN ARCHITECTURES
SNR SNR
Accuracy Latency/
Model Epochs Modulations General
(%) Sample (ms)
(dB) (dB)
CNN Fig. 7. Modulation accuracy as a function of SNR of the proposed CNN
82.00 27.80 0.39 8.60 2.30 architecture.
1
CNN
83.80 56.20 1.04 7.00 1.50
2
C. Models Comparison
The statistical test results were 0.0013, 0.0602, and 0.0139 From the previous results, both models have demonstrated
for accuracy, SNR per modulation, and general SNR, re- difficulty classifying phase modulations, especially in distin-
spectively. As mentioned in Section II, the p-value approval guishing between similar modulations, such as those with P
criterion for this work is 0.05. Therefore, looking rigorously code. Similarly to the LSTM, the CNN can easily classify
at the p-values presented, it is possible to verify that CNN 2 noise, chirps, and non-modulated signals. LSTM was also
XLIII SIMPÓSIO BRASILEIRO DE TELECOMUNICAÇÕES E PROCESSAMENTO DE SINAIS - SBrT 2025, 29 DE SETEMBRO A 2 DE OUTUBRO–11 DE 2025, NATAL, RN
if processing time is a crucial demand, the CNN can be a
good option, albeit at the expense of slightly compromising
accuracy capability.
ACKNOWLEDGMENT
The authors thank the Electronic Warfare Laboratory of
the Aeronautics Institute of Technology and the LOM Project
for providing the facilities to carry out this work. This work
was partially sponsored by SIATT, FINEP (under grant no.
01.22.0581.00), National Institute of Science and Technology
(INCT-Signals) sponsored by Brazil’s National Council for
Scientific and Technological Development (CNPq) under grant
no. 406517/2022-3, Ministry of Science, Technology and
Innovation (MCTI), and by the São Paulo Research Founda-
tion (FAPESP) under grant no. 20/09838-0 (BI0S—Brazilian
Institute of Data Science).
R EFERENCES
Fig. 8. Confusion matrix showing the misclassification errors of the proposed
CNN architecture. [1] P. Sharma, K. K. Sarma, and N. E. Mastorakis, “Artificial intelligence
aided electronic warfare systems- recent trends and evolving applica-
tions,” IEEE Access, vol. 8, pp. 224 761–224 780, 2020.
[2] T. Singh and A. Gulhane, “8 key military applications for artificial
able to classify 2PSK modulations effortlessly. Comparing intelligence in 2018,” [Link]
the results obtained from the two models tested, it is evident applications-for-artificial-intelligence-in-2018, 2018, accessed: Apr. 2,
2025.
that the LSTM outperforms the CNN in terms of accuracy, [3] G. Song, G. Jeon, and D. Yoon, “Deep learning-based automatic
making it capable of classifying modulations under worse modulation classification for composite modulated radar signal using
SNR conditions. Nonetheless, this improvement incurs a high time-frequency image,” in 2024 34th International Telecommunication
Networks and Applications Conference, ITNAC 2024. Institute of
computational cost to train the model and a high latency time Electrical and Electronics Engineers Inc., 2024.
after training, compared to the CNN model. The choice of [4] S. Wei, Q. Qu, X. Zeng, J. Liang, J. Shi, and X. Zhang, “Self-
using an LSTM or a CNN will depend on the context of the attention bi-lstm networks for radar signal modulation recognition,”
IEEE Transactions on Microwave Theory and Techniques, vol. 69, pp.
application. On the one hand, if a more precise classification 5160–5172, 11 2021.
is needed and processing time is not a critical problem, the [5] B. E. Altuntas, O. Aksu, M. Y. Celik, and M. H. Durak, “Deep learning
LSTM can be an appropriate model. On the other hand, if based automatic modulation recognition using gelu activation function,”
in 4th International Conference on Emerging Smart Technologies and
processing time is crucial and the application permits slightly Applications, eSmarTA 2024. Institute of Electrical and Electronics
neglecting accuracy, the CNN might be a good approach. Engineers Inc., 2024.
[6] G. Ruan and W. Wu, “Automatic recognition of radar signal types based
on convolutional neural network,” in 2019 International Conference on
IV. C ONCLUSIONS Image and Video Processing, and Artificial Intelligence, R. Su, Ed.
This paper proposes an LSTM and a CNN architecture for SPIE, 11 2019, p. 124.
[7] D. Li, R. Yang, X. Li, and S. Zhu, “Radar signal modulation recognition
the Automatic Modulation Classification task applied to radar based on deep joint learning,” IEEE Access, vol. 8, pp. 48 515–48 528,
signals. The performance of the proposed architectures was 2020.
compared to that of the baseline architectures. The LSTM [8] Y. Wen, X. Wang, G. Xu, and F. Wang, “Lpi radar signals intra-pulse
modulation recognition based on ca-shufflenet,” in Proceedings - 2022
architecture can reduce latency time by 27%, while maintain- 14th International Conference on Signal Processing Systems, ICSPS
ing similar accuracy and SNR performance to the baseline 2022. Institute of Electrical and Electronics Engineers Inc., 2022, pp.
architecture. Regarding CNN, the proposed architecture could 144–149.
[9] Z. M. Liu and P. S. Yu, “Classification, denoising, and deinterleaving
improve accuracy and SNR performance, but at the cost of of pulse streams with recurrent neural networks,” IEEE Transactions on
higher latency time. Aerospace and Electronic Systems, vol. 55, pp. 1624–1639, 8 2019.
LSTM has proven to be a powerful approach to this [10] X. Li, Z. Liu, Z. Huang, and W. Liu, “Radar emitter classification with
attention-based multi-rnns,” IEEE Communications Letters, vol. 24, pp.
problem, although it demands high computational power for 2000–2004, 9 2020.
training and presents high latency time in the test stage. [11] V. Clerico, J. González-López, G. Agam, and J. Grajal, “Lstm
Convolutional Neural Networks have been demonstrated to framework for classification of radar and communications signals,” in
2023 IEEE Radar Conference (RadarConf23). IEEE, 5 2023, pp. 1–6.
be an efficient method for the AMC task. Besides having an [Online]. Available: [Link]
accuracy slightly lower than that of the LSTM architectures [12] Kilian, “Deepradar: A radar signal dataset for deep learning,”
tested, the CNN architectures had a latency time during the [Link] 2022, accessed: Apr.
2, 2025.
testing stage that was at least 95% lower than that of the faster [13] T. J. O’Shea, T. Roy, and T. C. Clancy, “Over
LSTM architecture. The choice of using an LSTM or a CNN the air deep learning based radio signal classification,”
model must be made according to the application’s needs. 12 2017. [Online]. Available: [Link]
[Link]
If it requires higher accuracy and processing time is not a
concern, LSTM might be an excellent model. Alternatively,