Ship detection using ensemble deep learning techniques from synthetic aperture radar imagery.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Maritime Security

Background:

• Synthetic Aperture Radar (SAR) is crucial for maritime surveillance, aiding in traffic management, piracy detection, and illegal fishing prevention.
• Existing SAR ship detection models struggle with scale variance and uneven ship size distribution, leading to insensitivity to small vessels and increased false alarms.

Purpose of the Study:

• To develop an effective ensemble model for multi-scale ship detection in SAR imagery.
• To address the limitations of existing models in accurately identifying ships of varying sizes.

Main Methods:

• An ensemble model, eYOLO, was developed by integrating YOLOv4 and YOLOv5 architectures.
• Weighted box fusion was employed to combine the outputs from YOLOv4 and YOLOv5.
• A generalized intersection over union loss function was utilized to enhance model generalization and reduce scale sensitivity.

Main Results:

• The eYOLO model demonstrated high performance in multi-scale ship detection.
• Achieved an F1 score of 91.49% and a mean Average Precision (mAP) of 92.00% on an open-source SAR-ship dataset.
• Outperformed existing methods in detecting ships across various scales.

Conclusions:

• The proposed eYOLO model is effective for multi-scale ship detection in SAR imagery.
• The ensemble approach and generalized IoU loss significantly improve detection accuracy and reduce scale sensitivity.
• eYOLO offers a promising solution for enhanced maritime surveillance and security.