# Neural network-based identification for scallops (Pecten maximus) in natural marine habitats

## TL;DR

This study explores using AI to automatically identify and count scallops in underwater videos to improve fisheries management.

## Contribution

The study applies AI to identify naturally buried scallops in marine habitats, unlike previous work that used artificial setups.

## Key findings

- The NetHarn AI model achieved moderate success in identifying scallops with an F1 score of 0.44 and mAP of 0.41.
- Model performance varied with location, camera type, and habitat, highlighting challenges like blurred images.
- The study emphasizes the need for better data collection and larger datasets to improve AI accuracy for marine monitoring.

## Abstract

The Great Atlantic scallop, or King scallop (Pecten maximus), ranks third in value after mackerel and Nephrops in UK fisheries. Its landings have surged over recent decades, making it the UK’s fastest-growing fishery. Scallop stock assessments, crucial for sustainable fisheries management, traditionally rely on fisheries surveys, including underwater imaging and dredge sampling. Data on areas that contain scallops but not fishable using dredges is lacking. Dredge sampling is also potentially destructive. Remote data collection using drop down cameras and towed video are used, but there are few tools available to analyse these data automatically. P. maximus are usually recessed in fine sand and gravel habitats making image identification challenging. This study explores the potential of Artificial Intelligence (AI), specifically the NetHarn model from the VIAME toolkit, to identify and count scallops from underwater video transects. The research utilises diverse video footage from NatureScot, captured with custom camera systems (DDV and miniDDV), providing varied habitat, image quality, and camera specifications. Previous AI studies of this species artificially placed scallops on the seabed and are not representative of natural presentation. This research applies the same AI model to survey images featuring scallops in their natural habitat. Results showed moderate performance of the NetHarn model, achieving an F1 score of 0.44 and a mean Average Precision (mAP) of 0.41 when classifying scallops into three categories: king, queen, and dead. Model performance varied across geographic locations, camera platforms, and habitat types, with challenges including blurred images and mislabelling. The study emphasises the need for improved data acquisition, standardised camera systems, and larger annotated datasets to enhance AI model performance. Despite moderate results, this research highlights AI’s potential for automating estimation of scallop stock abundance and marine habitat monitoring. Future efforts should focus on addressing image quality issues, increasing sample sizes, and optimising data collection for enhanced marine conservation and fisheries management.