Adaptive Superpixel-Level CFAR Detector for SAR Inshore Dense Ship Detection

Ship monitoring is an important application of synthetic aperture radar (SAR). The constant false alarm rate (CFAR) methods are commonly used for ship detection. However, CFAR detectors usually face challenges for inshore dense ship detection. Due to the significant mixture of ship candidates and sea clutters within the clutter window, the detection threshold may be overestimated leading to many missed detections. To mitigate this issue, a superpixel-level CFAR detector is proposed. The main contribution contains two aspects. First, a labeling procedure is established for pure clutter superpixels and mixture superpixels discrimination in terms of unsupervised clustering. Second, a nonlocal topology strategy is proposed to adaptively determine a sufficient number of pure clutter superpixels for detection threshold estimation. In this vein, an adaptive superpixel-level CFAR approach is constructed and validated with Radarsat-2, Sentinel-1, and AIRSARShip-1 data sets. Comparison studies demonstrate the superiority of the proposed method. Compared with a traditional CFAR detector and two recent superpixel methods, the proposed method achieves clearly better performance for inshore dense ship regions.

Automated summary

Ship monitoring using synthetic aperture radar (SAR) is an important application, with constant false alarm rate (CFAR) methods commonly used for ship detection. However, challenges arise in dense ship detection near shorelines. A new superpixel-level CFAR detector is proposed, incorporating a labeling procedure for discriminating between pure clutter superpixels and mixture superpixels, as well as a nonlocal topology strategy for adaptive threshold estimation. The proposed method outperforms traditional CFAR detectors and recent superpixel methods in detecting inshore dense ship regions.