| description abstract | Fish passage design could greatly benefit from knowledge of the movement and path choices of fish in complex, turbulent environments, but empirical information required to construct and validate agent-based models is generally not available. To address this issue, an image-based data analysis scheme is proposed to measure turbulent flow and fish motion simultaneously using a small keystone species to exemplify the procedures. Using the same video images collected in a laboratory channel, large-scale particle tracking velocimetry is employed to compute the instantaneous Eulerian flow field, and a fish detection model and background-foreground subtraction method are used to identify fish positions. The fish detection model uses deep learning with an architecture based on a faster regional convolutional neural network (faster R-CNN). This methodology and architecture are shown to have a success rate greater than 93% in identifying fish positions. Data are then presented to illustrate the utility of this technique as it relates to agent-based models for fish movement. These results show that when velocities surrounding a swimming fish are spatially averaged, the values derived from time-averaged velocities are markedly different when compared to those derived from instantaneous velocities, as evidenced by relatively large root-mean-square errors. A swimming fatigue analysis also shows that, in general, the fish would swim for relatively short durations when experiencing fatigue in contrast to swim trial data. The experimental and data analysis methods proposed here offer a new approach to examine fish-flow interactions in complex hydrodynamic environments and provide a potentially important tool for assessing and evaluating fish passage design and river restoration projects. | |
