Automatic Extraction and Counting of Fish from Underwater Videos Using YOLO-Based Deep Learning Algorithms

In the context of increasing pressure on marine species, effective and automated methodologies for biodiversity monitoring are essential, particularly in sensitive ecosystems such as the Mediterranean Sea. This study presents an integrated deep learning approach for multi-object tracking and species recognition from underwater videos, aiming to automate and improve fish population censuses within Marine areas with a focus in Marine Protected Areas in Sardinia. Various detection models, including DeepFins, DeepEcomar, YOLOv8, and YOLOE, and tracking algorithms such as DeepSort, ByteTrack, and SAMURAI were evaluated. According to the achieved tests the YOLOE model, carefully trained on the Mediterranean-specific SardinIA dataset, demonstrated the best detection performance, while DeepSort proved most effective in maintaining individual identities across complex scenarios. The AI based achieved results compared with traditional visual census methods (underwater visual census, UVC and diver operated video census, DOVC), showing high accuracy in total abundance estimation and good agreement for dominant species. These findings suggest that deep learning techniques offer a promising, scalable solution for marine biodiversity monitoring, although challenges remain in species-level classification. In the present paper the following methodologies and the achieved results are reported.