Automated marine oil spill detection algorithm based on single-image generative adversarial network and YOLO-v8 under...

Area of Science:

• Environmental Science
• Remote Sensing
• Artificial Intelligence

Background:

• Increasing marine oil spills threaten ecosystems and human safety.
• Accurate and rapid detection is crucial for marine environmental protection.
• Deep learning models for oil spill detection are hindered by limited real-world sample data.

Purpose of the Study:

• To develop an effective marine oil spill detection method for limited sample scenarios.
• To address the challenge of insufficient training data in deep learning models for oil spill identification.
• To improve the efficiency and accuracy of marine oil spill detection using enhanced datasets.

Main Methods:

• Utilized Landsat-8 satellite imagery for a small marine oil spill dataset.
• Employed a single image generative adversarial network (SinGAN) for data augmentation, generating diverse oil spill samples.
• Pre-trained a YOLO-v8 model using transfer learning and trained it on both original and augmented datasets for detection.

Main Results:

• The YOLO-v8 model trained on the augmented dataset showed significant improvements: 12.3% in recall, 6.3% in precision, and 11.3% in average precision.
• The enhanced dataset led to superior performance compared to the model trained on the unexpanded dataset.
• The SinGAN-based data augmentation technique proved beneficial for other object detection algorithms.

Conclusions:

• The proposed YOLO-v8 based marine oil spill detection model demonstrates leading or comparable performance metrics (recall, precision, AP).
• Data augmentation using SinGAN effectively overcomes the limitations of small sample sizes in training deep learning models.
• This approach offers a viable solution for accurate and efficient marine oil spill detection in data-scarce environments.