WCS-Net: A Wavelet-Guided Spatial–Frequency Fusion Network for Detection-Oriented Sea Clutter Suppression in Marine Radar Remote Sensing

Maritime target detection from marine radar range–Doppler (R–D) maps is strongly affected by sea clutter with non-Gaussian, nonstationary, and multiscale characteristics. Existing deep restoration networks mainly operate in the spatial domain and often lack explicit modeling of frequency-dependent clutter structures and detection-oriented target preservation. To address these issues, this article proposes a wavelet-guided spatial–frequency fusion network, termed WCS-Net, for end-to-end sea clutter suppression. WCS-Net contains three main components. First, a wavelet transform-based frequency information extraction module employs a Symlet-4 discrete wavelet transform–inverse Symlet-4 wavelet transform U-shaped structure with energy-normalized subband attention to separate low-frequency clutter envelopes from high-frequency target-related responses. Second, a multiscale texture feature extraction module captures local clutter textures using lightweight multiscale R–D blocks with confidence-weighted residual updates. Third, a spatial–frequency alignment representation module adaptively fuses the two branches through cross-domain interaction descriptors, local-global feature gates, and wavelet-dominant confidence-weighted logit correction. A tail-balanced loss is further designed to jointly constrain global reconstruction, target-neighborhood preservation, and high-amplitude background-tail suppression. Experiments on real measured marine radar data with simulated targets show that the default WCS-Net achieves the best overall detection-oriented performance among representative baseline methods on the evaluated datasets. With 1.2685 M parameters and 13.3491 GFLOPs for a 320 x 320 R–D map, WCS-Net provides an effective and efficient framework for detection-oriented sea clutter suppression.