Incorporating Negative Sample Training for Ship Detection Based on Deep Learning.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Ship detection in high-resolution optical satellite imagery is crucial for maritime surveillance and rescue.
• Detecting ships near land is challenging due to complex backgrounds and false alarms from terrestrial objects.
• Existing methods often require land masking, which can introduce segmentation errors and require parameter tuning.

Purpose of the Study:

• To develop an improved ship detection method for high-resolution satellite images that eliminates the need for land masking.
• To reduce false alarms in ship detection, particularly in coastal and terrestrial areas.
• To enhance the efficiency and accuracy of ship detection algorithms.

Main Methods:

• Utilized the Faster Region-based Convolutional Neural Network (Faster R-CNN) architecture.
• Proposed a novel training strategy incorporating terrestrial images as negative samples without manual annotation.
• Evaluated the method using satellite imagery from Gaofen-1 (GF-1), Gaofen-2 (GF-2), and Jilin-1 (JL-1) satellites.

Main Results:

• The proposed method significantly reduced false alarms in terrestrial areas compared to traditional approaches.
• Achieved substantial improvements in detection performance, with an absolute increment of 70% in F1-measure for images with large landmasses (e.g., GF-1).
• Demonstrated superior performance in complex harbor scenarios (e.g., JL-1 images), showing a 42.5% absolute increment in F1-measure.

Conclusions:

• The negative sample training strategy for Faster R-CNN is effective for accurate ship detection without land masking.
• This approach simplifies the detection process, reduces computational complexity, and saves time.
• The method offers a robust and efficient solution for ship detection in diverse satellite imaging conditions.