Answers are grounded in this paper and cite their sources.
Complex sea clutter remains a major challenge for radar target detection because of its non-Gaussian statistics, heavy-tailed fluctuations, and pronounced nonstationary behavior, which severely degrade the reliability of weak-target detection. Existing deep learning methods usually perform clutter suppression and target localization within a single-stage framework, where the two objectives are optimized simultaneously in a shared feature space. However, these two objectives are inherently conflicting in clutter-dominated radar observations: aggressive clutter suppression may oversmooth weak target responses, whereas excessive preservation of local target-like structures may retain clutter spikes and increase false alarms. To address this problem, we propose a two-stage framework that explicitly decouples clutter suppression from target localization. In the first stage, a sequential clutter representation network (SCR-Net) treats radar echoes as sequential representations and suppresses structured sea clutter by modeling its long-range spatial dependencies. In the second stage, a target-aware spatial refinement network (TSR-Net) takes the channelwise concatenation of the original radar image and the clutter-suppressed output as input, and focuses on target-oriented spatial refinement and localization. By separating global clutter modeling from local target localization, the proposed framework alleviates the representation conflict between background attenuation and weak target preservation. Extensive experiments on both physically realistic simulated datasets and real radar measurements demonstrate that the proposed method achieves stronger clutter suppression, higher detection reliability, and superior computational efficiency compared with representative traditional detectors and deep learning baselines. These results indicate that task decoupling provides an effective and practical solution for robust radar target detection in complex sea clutter environments.
No takes yet. Share an insight, caveat, or question.
Yi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: