Generative Domain Adaptation for Pixel-Level RFI Segmentation in Ku-Band Satellite Spectrograms

Radio frequency interference (RFI) segmentation in Ku-band satellite communications remains challenging because of weak, non-stationary interference characteristics and the scarcity of pixel-level annotated empirical data. To address this limitation, this study proposes a synthetic-to-real deep learning framework in which four parameterized RFI morphologies—narrowband, broadband, impulsive, and frequency-varying—are superimposed onto empirical Ku-band spectrogram backgrounds acquired from a 12-m ground-station platform. A conditional Generative Adversarial Network (cGAN) is then employed for domain adaptation to reduce the synthetic-to-real gap by harmonizing the hybrid spectrograms with empirical thermal noise characteristics. The refined spectrograms and their exact binary masks are subsequently used to train a U-Net model for pixel-level segmentation. Quantitative evaluation on held-out empirical-background hybrid test data shows that the proposed framework achieves an Intersection over Union (IoU) of 0.849 and an F1-score of 0.918, outperforming traditional threshold-based methods and unrefined learning baselines. Additional qualitative validation on naturally observed empirical Ku-band RFI recordings further supports the practical applicability of the proposed framework beyond controlled hybrid test data. These results indicate that generative domain adaptation provides a practical and scalable alternative to manual labeling for automated RFI monitoring in operational Ku-band environments.