DSAMNet: A Signal-to-Noise Ratio Adaptive Multibranch Framework for Automatic Modulation Classification

Automatic modulation classification (AMC) is a key technique in cognitive radio and non-cooperative communication systems. However, existing deep learning-based AMC methods show limited robustness across varying signal-to-noise ratio (SNR) conditions. When the channel environment changes, feature distributions tend to shift, which leads to performance degradation. This issue becomes more severe under low SNR conditions and limits classification accuracy in practical scenarios. To address this problem, we propose a deep SNR-adaptive modulation network (DSAMNet). The framework integrates multi-domain representation learning with SNR-aware dynamic feature recalibration. Specifically, a parallel multi-branch architecture is constructed to extract complementary features from the in-phase/quadrature (I/Q), amplitude–phase (AP), spectrogram (Spec), and phase-unwrapped instantaneous frequency (IF) domains. A lightweight blind SNR estimator is introduced to provide global channel-quality information. Based on the estimated SNR, a feature-wise linear modulation mechanism (SNR-FiLM) is employed to perform dynamic feature recalibration across different branches. In addition, a content-aware soft gating strategy is designed to combine local feature responses with global SNR information, enabling adaptive multi-domain feature fusion. Experiments on the RadioML2016.10a and RadioML2016.10b datasets show that the proposed method outperforms existing approaches in overall classification performance. It achieves accuracies of 65.11% and 67.77%, respectively. The method also shows improved robustness under low SNR conditions, which demonstrates its effectiveness across varying SNR conditions.