Monocular burst swimming detection in Pacific bluefin tuna (Thunnus orientalis) using deep learning enhanced by ellipse fitting

Abstract

High mortality rates caused by burst swimming represent a significant challenge in the aquaculture of Pacific bluefin tuna (PBT, Thunnus orientalis). PBT are highly sensitive to external stimuli, often responding with sudden, high-speed swimming that results in collisions with tank walls or entanglement in nets, leading to fatal injuries. Previous mitigation efforts, such as artificial light control, have achieved limited success in preventing burst swimming. To address the issue of burst swimming and better understand its causes, a robust and automated detection method is essential. This study develops an automated approach for detecting burst swimming in PBT using deep learning and simple computer vision techniques. Unlike modalities such as sonar or advanced sensors, video-based methods offer non-invasive, continuous monitoring of fish behavior and detailed spatial data. The required setup, consisting of a standard video camera and computer, is both cost-effective and accessible. The proposed method employs advanced segmentation and object tracking techniques, integrating ellipse fitting to estimate the body height of PBT. By leveraging these data, the method calculates the real-scale three-dimensional velocity from two-dimensional video inputs, enabling precise detection of burst swimming events. Evaluated under real aquaculture conditions, the method achieves an F1 score of 75.9%, significantly surpassing the 54.4% of the conventional approach. This research highlights the potential of video-based analysis to enhance our understanding of fish behavior and contributes to the development of sustainable aquaculture technologies, ultimately improving PBT survival rates and production efficiency.