Multi-Level Spatiotemporal Framework for Automatic Modulation Classification

Abstract

Automatic modulation classification (AMC) has been a critical task in non-cooperative communication systems, yet the application of deep learning to this domain faces challenges. In this letter, we propose a multi-level spatiotemporal framework for AMC. The model integrates a series of fundamental deep learning components, including the Mamba for efficient long-sequence temporal modeling, dynamic soft-threshold denoising for adaptive noise suppression, and multi-channel dynamically dense Transformer to overcome the performance saturation in deep stacking. Enhanced with data augmentation inspired by chromosome variation mechanisms, the framework effectively captures the joint time-frequency characteristics of I/Q signals. Experiments on public datasets (i.e., RadioML 2016.10a and RML22) demonstrate the model’s strong AMC performance, achieving classification accuracies of 65.17% and 75.80% overall, and reaching >98.9% at SNR >10 dB. The proposed method provides a structured and extensible framework for AMC.