Spectral Gaps Extrapolation for Stepped-Frequency SAR via Generative Adversarial Networks

Ultra-wideband (UWB) radars achieve high-resolution imaging along with penetration capability thanks to the use of an ultra-wide spectrum ranging from hundreds of MHz to a few GHz. However, this spectrum is also shared by many other communication systems. Severe interference requires the systems to hop over these crowded frequency bands or even bypass sensing in certain prohibited bands. To recover this missing spectral information, we propose a generative adversarial network, called SARGAN, that learns the relationship between the full-bandwidth original and the observed missing band signals by observing these training pairs. We employ a simple sparse model to generate the training data and then test SARGAN performance on real-world data that it has never observed before. Initial results show that this approach is promising in tackling the challenging missing band problem. Our generative network can interpolate/extrapolate up to 90% of the missing spectrum - significantly outperformed recent state-of-the-art sparse-recovery/low-rank methods.