Deep learning-based seabird detection in fisheries for seabird protection

Area of Science:

• Marine Biology
• Computer Vision
• Conservation Technology

Background:

• New Zealand has a high diversity of seabirds, earning it the title 'seabird capital'.
• Commercial fishing operations in New Zealand lead to accidental seabird bycatch, posing a threat to population sustainability.
• Existing methods for monitoring seabird bycatch are often manual and labor-intensive.

Purpose of the Study:

• To develop an automated system for detecting seabird interactions with commercial fishing vessels.
• To improve the accuracy and efficiency of seabird bycatch monitoring.
• To address the limitations of previous seabird detection models evaluated in controlled settings.

Main Methods:

• Gathering, annotating, and preprocessing a novel image dataset of seabird-fisher interactions.
• Utilizing transfer learning with established YOLO (You Only Look Once) benchmark models.
• Performing hyperparameter tuning on top-performing YOLO models to optimize detection accuracy.

Main Results:

• The developed YOLO model achieved high performance metrics: mAP@50 of 0.9926 and mAP@50-95 of 0.9147 on test data.
• The model demonstrated effectiveness in unconstrained, real-world marine scenarios.
• Automated detection significantly reduces the need for manual footage review.

Conclusions:

• The developed YOLO model offers a robust solution for automated seabird detection in commercial fishing.
• This technology can accurately quantify seabird-vessel interactions, aiding conservation strategies.
• The study bridges the gap between controlled environment models and real-world maritime applications for seabird protection.