Quad-Level Inshore Awareness for Automatic Marine Ship Surveillance Using Satellite SAR System

Automatic marine ship surveillance using satellite synthetic aperture radar (SAR) system is receiving widespread attention. However, the inshore scene complicacy exceedingly increases the search difficulty, causing a huge accuracy gap between inshore and offshore. Therefore, we propose quad-level inshore awareness (QLIA) to boost the ship surveillance performance among inshore scenes. QLIA provides comprehensive inshore scene perception at four levels, achieving progressively higher accuracy, that is, input-level inshore awareness (i-LIA), feature-level inshore awareness (f-LIA), anchor-level inshore awareness (a-LIA), and objective-level inshore awareness (o-LIA). i-LIA first devises a saliency map generation module at the original input to ease land interference and focus on ship regions, and then constructs an aggregation excitation collaboration network to comprehensively balance the original information and saliency information for collaborative transcendence. f-LIA designs the wavelet spatial self-attention and ship inshore topology modeling for suppressing land noise and perceiving ship–land topology relationships, respectively, thereby making the extracted features more discriminative. a-LIA adaptively predicts sparse anchor positions through a dynamic anchor selection module, effectively differentiating between ship and land areas, which not only strengthens the representativeness of generated anchors but also improves operational efficiency. Finally, o-LIA proposes a difficulty scene concentration loss during training, which focuses on supervising difficult inshore scene samples. Such biased learning can correct the previous imbalance in learning benefits owing to the imbalance between shore and offshore samples. Satellite SAR collected data from RadarSat-1, TerraSAR-X, and Sentinel-1 is used for experimentation. Results show the validity of i-LIA, f-LIA, a-LIA, and o-LIA, meanwhile showcasing the state-of-the-art performance of QLIA.