Round goby [Neogobius melanostomus (Pallas, 1814)], gudgeon (Gobio gobio L.) and bullhead (Cottus gobio L.) show distinct swimming patterns in a vertical slot fish pass

• Published in: Frontiers in Environmental Science
• Main Authors: Joschka Wiegleb, Philipp E. Hirsch, Frank Seidel, Georg Rauter, Patricia Burkhardt-Holm
• Format: Article
• Language: English
• Published: Frontiers Media S.A. 2023-03-01
• Subjects: video tracking; convolutional neural network; machine learning (ML); hydrodynamics; fish pass
• Online Access: https://www.frontiersin.org/articles/10.3389/fenvs.2023.1156248/full

The vertical slots of fish passes represent bottlenecks that must be passed by every fish migrating upstream. The hydraulics in fish passes are well investigated but less is known about the small scale behaviour of fish while passing the vertical slot. Understanding the species-specific swimming behaviour during the passage could allow for creation of future fish passes with hydraulics adapted to the swimming requirements of desired target species. We recorded the swimming trajectories of three fish species as point coordinates per video frame using cameras. Then, two common machine learning algorithms were used to identify species characteristic swimming patterns in the trajectories. A Random Forest model trained on 21 trajectory features revealed that water discharge, the spatial trajectory position, and the trajectory length were most distinct trajectory features among species. The model identified the species with a mean F1 score of 0.86 ± 0.08 SD for round goby [Neogobius melanostomus (Pallas, 1814)], 0.81 ± 0.12 SD for gudgeon (Gobio L.), and 0.58 ± 0.20 SD for bullhead (Cottus gobio L.). A Convolutional Neural Network achieved a mean F1 score of 0.89 ± 0.03 SD for round goby, 0.76 ± 0.05 SD for gudgeon, and 0.67 ± 0.02 SD for bullhead if exclusively trained on the point coordinates of the swimming trajectories. These results demonstrate that fish species exhibit distinct swimming patterns when passing through a vertical slot, and how these patterns can be used for species identification using machine learning algorithms. Because round goby achieved the highest F1 scores, we conclude that round goby showed the most characteristic swimming trajectories among the species tested. Future fish passage research should account for the individual swimming patterns of the fish in these bottleneck flow fields and on adapting the flow to the individual swimming patterns of the target fish. Flow conditions being supportive for swimming patterns of the desired fish could have the potential to improve the river connectivity and thereby support the aquatic biodiversity.