Robust spectrogram-based automatic modulation classification via deep CNN-LSTM architecture
Abstract
Automatic modulation classification (AMC) is a crucial task in modern communication systems, enabling the identification of modulation schemes in received signals. This paper presents a proposed model to achieve AMC more accurately by using hybrid techniques, Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) with spectrograms as input features. The system is evaluated under the influence of Additive White Gaussian Noise (AWGN) with Signal-to-Noise Ratios (SNR) set at -5, 5, and 10 dB. The model was trained using the Adam optimizer with a fixed learning rate of 0.00001, mini-batch size of 32, and four epochs. The reported overall accuracies are 75.00%, 87.73%, and 99.68% at − 5, 5, and 10 dB, respectively. CPFSK accuracy is 0% at − 5 and 5 dB and 100% at 10 dB. These results demonstrate strong performance on the evaluated MATLAB/AWGN test configuration, but they do not establish robustness under multipath fading, Doppler, timing offsets, non-Gaussian noise, alternative train/test splits, alternative optimizers or learning-rate schedules, or alternative time–frequency representations. In addition, convergence beyond four epochs, CNN-only/LSTM-only ablations, STFT sensitivity, model parameter count, and inference latency were not experimentally evaluated. Therefore, the reported results should be interpreted as dataset- and channel-specific evidence rather than as a controlled claim of superiority over other AMC approaches.
Related
- Efficient Alzheimer’s disease classification with robust rank pooling
- DSR-VGG: A Lightweight Depthwise Separable Residual VGG Architecture for Nasopharyngeal Histopathological Image Classification
- Adaptive Liquid State-Space Networks: A Novel Neuro-Dynamic Architecture for Tabular Classification with Application to Anemia Diagnosis
- ICE-CAP: ice crystal engineering for cryogel architecture programming. Modulation of material properties with ice-binding polymers.
- Cross-corpus consensus feature selection for language-invariant Parkinson’s disease detection: a dual-stream acoustic and spectrogram-attention framework
BibTeX
@article{obead2026robust,
title = {{Robust spectrogram-based automatic modulation classification via deep CNN-LSTM architecture}},
author = {Ola Fadhil Obead and A. M. El-Assy and Hossam El-Din Moustafa and Hala B. Nafea},
journal = {Scientific Reports},
year = {2026},
month = nov,
doi = {10.1038/s41598-026-74229-5},
url = {https://www.nature.com/articles/s41598-026-74229-5},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Robust spectrogram-based automatic modulation classification via deep CNN-LSTM architecture
AU - Ola Fadhil Obead
AU - A. M. El-Assy
AU - Hossam El-Din Moustafa
AU - Hala B. Nafea
PY - 2026
DA - 2026/11/05
JO - Scientific Reports
DO - 10.1038/s41598-026-74229-5
UR - https://www.nature.com/articles/s41598-026-74229-5
AB - Automatic modulation classification (AMC) is a crucial task in modern communication systems, enabling the identification of modulation schemes in received signals. This paper presents a proposed model to achieve AMC more accurately by using hybrid techniques, Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) with spectrograms as input features. The system is evaluated under the influence of Additive White Gaussian Noise (AWGN) with Signal-to-Noise Ratios (SNR) set at -5, 5, and 10 dB. The model was trained using the Adam optimizer with a fixed learning rate of 0.00001, mini-batch size of 32, and four epochs. The reported overall accuracies are 75.00%, 87.73%, and 99.68% at − 5, 5, and 10 dB, respectively. CPFSK accuracy is 0% at − 5 and 5 dB and 100% at 10 dB. These results demonstrate strong performance on the evaluated MATLAB/AWGN test configuration, but they do not establish robustness under multipath fading, Doppler, timing offsets, non-Gaussian noise, alternative train/test splits, alternative optimizers or learning-rate schedules, or alternative time–frequency representations. In addition, convergence beyond four epochs, CNN-only/LSTM-only ablations, STFT sensitivity, model parameter count, and inference latency were not experimentally evaluated. Therefore, the reported results should be interpreted as dataset- and channel-specific evidence rather than as a controlled claim of superiority over other AMC approaches.
SN - 2045-2322
ER -