A Dual-Polarimetric SAR Ship Detection Dataset and a Memory-Augmented Autoencoder-Based Detection Method.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Multipolarized SAR imagery is increasingly used for object detection.
• Existing SAR ship datasets often lack dual-polarization data, limiting the use of polarization characteristics.

Purpose of the Study:

• To introduce a novel dual-polarimetric SAR dataset for ship detection (DSSDD).
• To establish baseline performance for existing detectors on this new dataset.
• To propose a weakly supervised method for improved ship detection in dual-polarimetric SAR images.

Main Methods:

• Construction of DSSDD from 50 Sentinel-1 dual-polarimetric SAR images, creating 1236 slices with fused VV and VH polarization information.
• Labeling ships with rotatable bounding boxes (RBox) and horizontal bounding boxes (BBox).
• Implementation of R³Det and Yolo-v4 detectors and a proposed weakly supervised anomaly detection method using advanced memory-augmented autoencoder (MemAE).

Main Results:

• DSSDD provides 8-bit pseudo-color and 16-bit complex data.
• Baseline performance established for R³Det and Yolo-v4 on DSSDD.
• The proposed MemAE method effectively reduced false alarms and demonstrated high efficiency and performance, comparable to supervised methods.

Conclusions:

• DSSDD is a valuable resource for dual-polarimetric SAR ship detection research.
• The weakly supervised MemAE method offers a promising, efficient, and less annotation-intensive approach for ship detection in dual-polarimetric SAR data.