Presenter
Dr. Jianyuan Li, China Institute of Water Resources and Hydropower Research
Co-author(s)
Dr. Chunna Liu, China Institute of Water Resources and Hydropower Research, Beijing
Dr. Xiaochun Lu, College of Hydraulic and Environmental Engineering, China Three Gorges University
Mr. Bilang Wu, China Institute of Water Resources and Hydropower Research, Beijing
Dr. Rui Li, China Institute of Water Resources and Hydropower Research, Beijing
Dr. Yi Liu, China Institute of Water Resources and Hydropower Research, Beijing
Sub-theme
5. Establishing Sustainable Water Infrastructures
Topic
5-1. Exploitation and development of Nature-based Solutions in water engineering and technology
Body
The construction of dams can lead to a loss of vertical connectivity in rivers, affecting the genetic exchange of upstream and downstream populations and reducing river biodiversity, with the most direct negative impact on migratory fish. To mitigate the impact of dam construction on aquatic life, hydroelectric projects have gradually begun to build fish passage projects as links to dams to assist fish migration and spawning. Fish passage monitoring is required during the operation of the fish passage to verify the effectiveness of the fish passage, thus enabling the scientific management of the fish passage and promoting the conservation of fish resources. As traditional overfishing monitoring is difficult to meet the accuracy and efficiency requirements of overfishing monitoring, an improved model based on YOLOv5s is proposed and applied to a hydropower station fish passage site. Firstly, to address the problem of blurred underwater images and difficult target detection, the Efficient-RepGFPN is proposed as the Neck network in YOLOv5s to enhance feature information decoding, which improves the detection capability of the model for targets; secondly, to address the problem of little image information, the Efficient Channel Attention (ECA) attention mechanism as the Bottleneck of the C3 structure of the backbone feature extraction network, which reduces the computational parameters and improves the detection accuracy; then, to address the problem of poor target localization and non-convergence of regression in detection, this paper uses AlphaIOU Loss as the model loss function to optimize the overall performance of the model. Experiments based on a complex water body dataset actually collected showed that the algorithm in this paper mAP@0.5为91.9% has a single image processing time of 10.4ms, which is a 4.7 percentage point improvement in accuracy and 0.4ms in processing speed over YOLOv5s under the same conditions. In summary, the algorithm model in this paper has a high accuracy on the basis of ensuring fast detection and is more suitable for fish passage overfishing monitoring under complex water bodies, which can provide a better alternative or complementary approach for fish passage overfishing monitoring.