A deep learning-based model for automatic identification of mesopelagic organisms from in-trawl cameras

Area of Science:

• Marine Biology
• Deep-Sea Ecology
• Artificial Intelligence in Fisheries

Background:

• Mesopelagic organisms are crucial for ocean carbon transport and food webs.
• Acoustic methods detect mesopelagic aggregations but struggle with species identification.
• Trawl catches are used for validation but have limitations like species escapement.

Purpose of the Study:

• To develop and evaluate a deep learning model for identifying mesopelagic species from in-trawl images.
• To assess the performance of the model under different lighting conditions (white and red light).
• To improve the validation of acoustic data for mesopelagic populations.

Main Methods:

• Trained a deep learning object detection model (YOLOv11s) on in-trawl images.
• Used images collected under white and red light with varying gain settings.
• Identified seven mesopelagic groups and larger pelagic fishes.

Main Results:

• The model achieved high accuracy (weighted mean average precision ~0.95) with white light.
• Red light also allowed effective detection (weighted mean average precision ~0.77).
• Lanternfish, silvery lightfish, and barracudina were detected with high precision (>0.89).

Conclusions:

• Deep learning enables accurate identification of mesopelagic species from in-trawl camera data.
• This approach facilitates rapid, depth-stratified data collection and analysis.
• It provides a valuable tool for studying fragile species often lost in trawls.