A Transfer Learning and Time-Frequency Fusion Framework for Few-Shot Modulation Classification in LEO Satellite Communications

In response to the challenge of low classification accuracy in small sample scenarios encountered by automatic modulation classification (AMC) in low earth orbit (LEO) satellite communication systems, this paper proposes a multiscale encoder framework based on transfer learning. First, by combining continuous wavelet transform to extract time-frequency features and fusing signal I/Q data, the feature robustness under large Doppler effects is enhanced. Then the model adopts a source domain pre-training and target domain fine-tuning strategy, designs a data adaptation module to solve the difference in cross-domain data length, and combines local convolutional encoders with global attention mechanisms to achieve multi-scale feature fusion. Experiments show that under limited training samples, the classification accuracy of the model is significantly better than the benchmark models, especially in distinguishing QAM, PSK and APSK signals, which verifies the effectiveness of transfer learning and multi-scale feature fusion.