SupeNetUAV: A Physics-Informed Multiscale Lightweight Network for Automatic Modulation Classification of UAV Wireless Signals

Abstract

Automatic modulation classification of unmanned aerial vehicle wireless signals presents difficulties due to noise, fading, frequency offset and other time-varying channel impairments, and the existing deep learning methods usually exhibit a trade-off between recognition performance and computation complexity. To address this problem, we propose SupeNetUAV, a lightweight network motivated by UAV Unmanned aerial vehicles (UAVs) wireless sensing. A dual-branch stem combines raw in-phase/quadrature samples with signal-derived amplitude and differential-phase descriptors. Parallel depthwise dilated convolutions extract temporal patterns at several scales, while squeeze-and-excitation and residual paths refine the features. A compact head controls the number of channels and temporal bins used for classification. On RML2016.10A and HisarMod2019.1, SupeNetUAV achieves overall accuracies of 61.54% and 76.77% with 39,239 and 145,046 parameters, respectively. An accuracy-first comparison gives both datasets equal weight and considers resource efficiency after joint recognition accuracy. SupeNetUAV ranks first among nine evaluated implementations, with a geometric-mean accuracy of 68.73%, compared with 68.18% for MobileAmcT. Its parameter counts are 93.30% and 75.45% lower than those of MobileAmcT.