Unsupervised Sea Clutter Suppression for Maritime Radar Target Detection via Contrastive Twin Squeeze-and-Excitation U-Nets

Maritime surveillance radar plays a vital role in ocean monitoring and navigational safety; however, the detection of small targets is often compromised by complex, nonstationary sea clutter. Existing deep learning–based clutter suppression approaches predominantly rely on supervised training with pixel-level annotations, which are impractical in operational maritime environments due to the non-cooperative nature of targets and the lack of reliable ground truth. To address these limitations, this paper presents Twin SEUNet-CL, an unsupervised framework for sea clutter suppression trained exclusively on unlabeled radar imagery. The proposed architecture employs twin squeeze-and-excitation U-Nets to learn complementary feature representations while exploiting channel-wise attention to enhance target-related responses and suppress clutter. To enable effective label-free discrimination, a contrastive learning strategy with hard-negative mining is introduced using complementary image pairs. Furthermore, a Jensen–Shannon divergence–based mutual information regularization term is incorporated between dense local features and predicted class-probability maps to stabilize the learned representation space. Experimental results on both simulated datasets and real maritime radar measurements show that Twin SEUNet-CL significantly improves detection probability at low false-alarm rates under severe clutter conditions, and exhibits robust sim-to-real transfer when trained on simulated data and evaluated on unseen real rain-clutter measurements without fine-tuning.