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

## TL;DR

This paper introduces a deep learning model that automatically identifies mesopelagic organisms from in-trawl camera images, improving data collection for ocean research.

## Contribution

A novel deep learning model (YOLO11s) is developed for automated identification of mesopelagic species from in-trawl camera images.

## Key findings

- The model achieved high precision (~0.95) for white-light images and still performed well (~0.77) under red-light conditions.
- Lanternfish, silvery lightfish, and barracudina were detected with high accuracy (average precision > 0.89).
- Lower precision for krill and pelagic shrimp was improved with higher resolution and expanded training data.

## Abstract

Mesopelagic organisms play an important role in the ocean’s carbon transport and food webs and have been regarded as a potential harvestable resource. Their extensive aggregations in the upper thousand meters of the water column are frequently detected acoustically as deep scattering layers. However, extracting species and length composition from acoustics alone is challenging. Trawl catches, commonly used for ground-truthing acoustic data, suffer from size- and species-specific escapement and are spatially integrated along the trawl path. In-trawl cameras offer records at a finer spatial scale and are unaffected by mesh selectivity in the codend. Hence, integrating optical systems into trawling operations can enhance the validation of acoustic data without increasing sampling time. In this study, we trained a deep learning-based object detection model (YOLO11s) to automate the identification of seven mesopelagic groups common in the North Atlantic Ocean (lanternfish, silvery lightfish, barracudina, krill, pelagic shrimp, gelatinous zooplankton, and squid) along with a group of larger pelagic fishes from in-trawl images collected under white, and red-light with two gain settings. The model generally performed better on white-light images (weighted mean average precision ~ 0.95). However, using red light did not greatly reduce the model’s ability to detect mesopelagic organisms (weighted mean average precision ~ 0.77). The model performed especially well at detecting lanternfish, silvery lightfish and barracudina (average precision > 0.89). Object classes with average precision values under 0.80 (e.g., pelagic shrimp, krill) benefited from increasing the image resolution and expanding the training dataset. Our study demonstrates that employing the latest machine learning algorithms enables the detection of small-sized mesopelagic species from in-trawl camera images, allowing for rapid extraction of depth-stratified data and records of fragile species that are typically lost in the codend meshes.