Automatic modulation classification (AMC) is a critical function in cognitive radio, software- defined radio (SDR), and modern wireless communication systems, enabling dynamic spec trum access and adaptive receiver operation. While deep learning approaches achieve high classification accuracy, their computational demands limit deployment in real-time and resource-constrained environments such as software-defined radios and edge devices. This work proposes a parameter-efficient CNN–BiLSTM hybrid architecture that combines lightweight convolutional spatial feature extraction with bidirectional temporal modeling for robust AMC across varying channel conditions. Evaluated on the RadioML 2016.10a benchmark across −20 to +18 dB SNR, the proposed model achieves 56.77% overall accuracy while using only 246,027 parameters—a 69% reduction compared to a standard CNN baseline (800,397 parameters). SNR-wise analysis reveals that the hybrid architecture consistently outperforms convolution-only models across all evaluated regimes, with the most pronounced gains at high SNR (>10 dB), achieving 84.59% accuracy compared to 66.97% for the base line CNN—a 17.62 percentage point improvement. Comprehensive efficiency metrics show a 56% reduction in computational cost (0.08 vs. 0.18 GFLOPs) and inference latency of 1.0 ms, making the architecture suitable for edge deployment. Ablation studies confirm the complementary nature of spatial and temporal processing, with convolutional feature extraction contributing the primary accuracy foundation and BiLSTM providing targeted gains in clean-channel conditions. This work addresses the critical gap between classification performance and computational constraints in practical wireless systems.
Bharat Paudel (Sat,) studied this question.