Introduction
The aquaculture industry has grown substantially over the past 25 years, with production increasing by 200% and global fish trade by 300% since 1995 (FAO, 2021). As the fastest-growing food production sector, aquaculture requires efficient and sustainable practices that support both productivity and animal welfare. Key anatomical features such as the pectoral fins, tail, and skin are critical for fish locomotion, posture, and health, yet are highly vulnerable to damage caused by handling, high stocking densities, and underlying health issues. Such impairments can adversely impact welfare, feeding efficiency, and growth. Despite their importance, the condition of these features remains largely under-monitored due to the reliance on manual assessment methods, which are time-consuming, labor-intensive, and potentially harmful to fish. This work aims to develop and validate non-invasive computer vision models for automatically monitoring fish morphometric traits—including fins, tails, and skin—and for classifying their condition as ‘good’ or ‘bad.’ The objective is to provide an efficient and reliable tool to improve health monitoring and welfare management in aquaculture systems.
Materials and methods
Two YOLO architectures—YOLOv8 and YOLOv12—were trained to detect three morphometric characteristics (tail, pectoral fin, and red skin spots) in two fish species: European seabass and gilthead seabream. Data were collected using a stereo vision system (1080p resolution) with a patented software (Voskakis et al., 2021) in circular polyester cages (40 m diameter, 9 m depth) at HCMR’s pilot farm in Souda Bay, Crete (certified: GR94FISH0001). Images were annotated into “good” and “bad” classes for tails and fins, and a single class for red skin spots. Training datasets included 2,383 fin images (gilthead seabream), 1,414 fin images (European seabass), 5,041 tail images (gilthead seabream), 4,276 tail images (European seabass), and 584 skin images with red spots for both species. Models were trained using Python on an NVIDIA RTX 4080 GPU (batch size: 12, image size: 640×640), with early stopping (patience = 30) to prevent overfitting. Extensive data augmentation strategies for underwater object detection were tested. A total of over 125 training runs were conducted as datasets were incrementally expanded.
Results
YOLOv8 and YOLOv12 both performed well in detecting morphometric features across species, with YOLOv12 showing superior computational efficiency and less overfitting. Transitioning from YOLOv8 Large to YOLOv12 Small model led to a 78.7% reduction in parameters (43.7M → 9.3M) (Jocher et al., 2021; Tian et al., 2025) without loss in detection accuracy. Despite its smaller size, YOLOv12 Small matched or outperformed YOLOv8 Large. In terms of detection performance, detection of red skin spots achieved the highest accuracy, with a mean Average Precision at IoU 0.5 (mAP50) of 0.89. Tail condition classification in gilthead seabream reached mAP50 = 0.80, while in European seabass, it achieved 0.68 (Figure 1). For pectoral fin classification, the seabream model reached mAP50 = 0.73, whereas the seabass model performed better, achieving mAP50 = 0.79.
Conclusion
Overall, detection accuracy varied across species, morphometric traits but also from differences in image quality due variable environmental conditions. Red skin spots were the most accurately detected feature, while tail condition in European seabass showed the greatest need for further improvement. Both YOLO models performed reliably; YOLOv12 demonstrated advantages in computational efficiency. Future steps include to test the system on large scale aquaculture farms by estimating the percentage of damaged morphometric features.
Acknowledgement
This work has received funding from the European Union’s Horizon 2020 Research and innovation programme under grant agreement 101084204 (CURE4AQUA).
References
FAO. 2021 COFI Declaration for Sustainable Fisheries and Aquaculture. FAO, 2021. doi: 10.4060/cb3767en.
Dimitris Voskakis, Alexandros Makris and Nikos Papandroulakis. Deep learning based fish length estimation. An application for the Mediterranean aquaculture. OCEANS 2021: San Diego – Porto, San Diego, CA, USA, 2021. pp. 1-5, doi: 10.23919/OCEANS44145.2021.9705813.
Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8. https://docs.ultralytics.com/models/yolov8/, 2023. YOLOv8 Large: 43.7M parameters, 165.2B FLOPs.
Yunjie Tian, Qixiang Ye, and David Doermann. Yolov12: Attention-centric real-time object detectors. https://docs.ultralytics.com/models/yolo12/, 2025. YOLOv12 Small: 9.3M parameters, 21.4G FLOPs.