Automatic Modulation Classification using LSTMs with an Alternative Input Shape

This paper investigates the impact of input representation and architecture on the performance of Long Short-Term Memory (LSTM) networks for Automatic Modulation Classification (AMC). Building on previous work, we reproduce and extend LSTM-based AMC models using the CSPB.ML.2018 dataset, introducing an alternative input shape that inverts the roles of time steps and features. This approach reduces training time by over 90% compared to conventional LSTMs and improves classification accuracy by up to 13.4% when using IQ inputs, particularly at low Signal-to-Noise Ratios (SNR). However, it requires longer input signal samples to match the performance of conventional models at high SNRs. We evaluate both cartesian (IQ) and polar input representations, demonstrating that IQ inputs consistently outperform polar inputs across all models. Our results highlight the trade-offs between computational efficiency, input length, and classification accuracy, offering insights for optimizing AMC systems in resource-constrained and real-time applications, such as cognitive radio and spectrum monitoring.