Unsupervised-to-Supervised Sea Clutter Suppression via Adversarial Prior and Manifold Idempotent Theory

Marine radar is widely employed in ocean monitoring systems. However, the presence of sea clutter and noise significantly hampers the performance of the marine radar. In this paper, we propose an unsupervised-to-supervised sea clutter suppression and denoise framework. In the unsupervised stage, we devised a two-branch blind spot network tailored for noise suppression under conditions of relatively high signal-to-clutter ratio (SCR) and signal-to-noise ratio (SNR). In the supervised stage, we designed the Global Fourier Block to expand the receptive field and promote information exchange, used the implicit prior-based discriminator to boost the fidelity, and added idempotent constraints to refine the manifold of the decluttered manifold. The unsupervised stage provides sufficiently high-quality pseudo-labels to the supervised stage, effectively addressing the challenge of obtaining clean labels in current supervised deep learning-based (DL-based) sea clutter suppression methods. The experiments demonstrate that our proposed method exhibits excellent adaptability in both single-target and multi-target scenarios, as well as under various numbers of pulses. To validate the widespread applicability of our unsupervised stage, we integrate it into other supervised DL-based clutter suppression methods, and the experiments show a noticeable improvement in the performance of these methods. Our code will be released after possible acceptance.