Lightweight Domain-Adaptive Modulation Classification Under SNR and Channel Mismatch

Reliable automatic modulation classification (AMC) degrades under mismatched noise and propagation conditions. This letter presents a lightweight and interpretable domain-adaptive AMC framework that jointly addresses both SNR variation and channel mismatch. Modality-specific feature extractors operate on raw IQ samples and spectral representations, while a domain-alignment module combines adversarial domain confusion with covariance matching to learn domain-invariant features. Experiments on RadioML2016.10a show improved robustness under SNR mismatch, and tests on Rayleigh–Rician fading datasets across five carrier frequencies demonstrate strong cross-channel generalization, achieving up to 86.8% accuracy. The proposed models require fewer than 0.75MB of parameters and support sub-millisecond inference, supporting deployment-ready intelligent radio systems.