Unmanned aerial vehicle (UAV)–assisted communication leverages Automatic Modulation classification (AMC) to enhance wireless transmission efficiency. The classification of multicarrier modulation (MCM) signals is a critical component AMC, enabling robust signal demodulation and adaptive resource allocation. However, existing deep learning-based solutions for MCM signal classification often entail significant computational overhead, posing substantial limitations for real-time deployment in resource-constrained and edge environments. To overcome these limitations, this work presents an AMC framework that utilizes Erlang spectrograms for time-frequency feature extraction and introduces a Squeeze-Excitation Lightweight CNN SELight-CNN, a novel deep neural architecture incorporating Squeeze-and-Excitation mechanisms with L2-regularized convolutions for accurate classification of MCM signals. The proposed SELight-CNN model demonstrates high classification accuracy across 10 multicarrier modulation types under different Signal-to-Noise Ratio (SNR) conditions.
This work presents an AMC framework that utilizes Erlang spectrograms for time-frequency feature extraction and introduces a Squeeze-Excitation Lightweight CNN SELight-CNN, a novel deep neural architecture incorporating Squeeze-and-Excitation mechanisms with L2-regularized convolutions for accurate classification of MCM signals. The proposed SELight-CNN model demonstrates high classification accuracy across 10 multicarrier modulation types under different SNR conditions.
The proposed SELight-CNN consistently outperforms the baseline CNN across all SNR levels, achieving 99% accuracy at −10 dB and reaching 100% from −6 dB onwards. In contrast, the baseline model attains only 95% at −10 dB. Performance is evaluated using recall, precision and F1-score. SELight-CNN stands out by achieving high classification accuracy even in low-SNR conditions, all while keeping the model lightweight and fast. This makes it a practical and efficient choice for real-time use in 5G networks, IoT devices and edge computing systems.
In the front-end, Erlang spectrograms are applied to raw signals from the 5G Waveforms dataset, introducing a novel time-frequency representation for AMC tasks. For classification, the back-end employs SELight-CNN, a lightweight CNN that integrates squeeze-and-excitation (SE) blocks with L2 regularization to effectively learn discriminative features from the Erlang spectrograms of MCM signals. SELight-CNN is benchmarked against a baseline CNN without SE blocks using standard evaluation metrics including F1 score, precision, recall and accuracy, demonstrating improved classification performance, particularly under low SNR conditions.
