Framework for Fish Passage Design and Evaluation: Application for Emerald Shiners in Niagara River
MetadataShow full item record
Anthropogenic modification of natural rivers in the form of dams, dredging, or shoreline alteration can significantly impede river connectivity. Fish passage plays a significant role in mitigating ecological impairment. However, the traditional build-and-test paradigm is cost-ineffective when the observed fish passage efficiency performs below expectations. A systematic approach that integrates design and evaluation of fish passage before installation is suggested here as a means of addressing this challenge. This study proposes a framework that combines fish passage design and pre-installation evaluation and implements it in a fish passage project for emerald shiners (Notropis atherinoides) in the Niagara River near Buffalo, New York. The proposed fish passage differs from other more common structures in that it is based on an unconfined design that allows fish to enter or leave the passage at any point along the installation. There are three components of this framework: (1) computational fluid dynamics (CFD) modeling, (2) physical modeling, and (3) virtual fish modeling. CFD modeling provides a numerical evaluation of performance of the fish passage design. Physical modeling tests the capability of fish passage modular structures to reduce the velocities to a range in which the emerald shiners can swim sustainably and to demonstrate that fish would choose to swim through the structures. The CFD results show that the recommended design achieves the desired lower velocity, and the physical model results also indicate that the proposed unconfined fish passage can serve as a shelter for migrating emerald shiners. Knowledge of emerald shiner movement and associated hydrodynamic experience in complex flows is obtained in the physical modeling through a new comprehensive video data analysis scheme, which integrates large-scale particle velocimetry (LSPTV), a fish detection model, and a background-foreground subtraction method. LSPTV is employed to compute the instantaneous Eulerian flow field, and the fish detection model and background-foreground subtraction method are used to identify fish positions in each video frame. The fish detection model, which is based on deep learning, is shown to have a success rate of over 93% in identifying fish positions in the video frames. Simultaneous measurements of fish movement and local velocities reveal the precise hydrodynamic conditions experienced by fish in complex flow, which provides a set of ‘hydrodynamic cues – decision’ pairs. A virtual fish model is developed based on the observed fish behavior. The virtual fish is formulated by computing the likelihood of a fish choosing one of a set of discrete moving directions at any instant, including holding position, or swimming forward, left, or right. The model consists of a multilayer forward artificial neural network (ANN) and a convolutional neural network (CNN). The ANN takes temporal information such as fish prior swimming speed and direction as inputs, and the CNN takes spatially and temporally varying hydrodynamic cues such as velocity, vorticity, turbulent kinetic energy (TKE), swirl, and strain rate as inputs. The trained hybrid CNN-ANN virtual fish model has an accuracy of just over 60%, when comparing predicted with observed decisions. Also, virtual fish in the physical model show similar trajectory trends to real fish swimming pathways, which suggests this deep learning technique holds promise in developing a non-parametric virtual fish model. A well-trained virtual fish can work as a smart agent in numerical modeling of different fish passage design candidates to optimize performance. The three components of the framework integrate knowledge of hydrodynamics and fish biology systematically, which offers a new approach for fish passage design and evaluation before installation.