Industry-Led Research Provides Accurate, Artificial Intelligence–Based Approach for Analysis of Sensitive Species Passing Through or Around Dams

The Electric Power Research Institute (EPRI) successfully achieved proof of concept for automated identification of silver phase American eels in multibeam sonar data, providing evidence of a new mechanism to more accurately detect and analyze species of concern in the United States. One of the key industrywide challenges to improving fish passage outcomes at hydropower dams across the country is effective monitoring of population, movement, and behavior of different aquatic species. By applying convolutional neural network (a type of artificial intelligence) tools to automate time-consuming processes for eel identification, EPRI was able to achieve greater than 98% accuracy in laboratory-based testing and almost 90% in field-testing results in distinguishing between eels and similar-sized naturally occurring river debris such as sticks. These results are comparable to the classification accuracy achieved by analysts in previous studies and provide a promising application for studies of other fish species in different environments.

The American eel has experienced population declines in eastern North American rivers and is a species of concern in the United States. To address this challenge, EPRI’s Eel Passage Research Center has explored new use cases and analytical methods for sonar-based detection of migrating American eels near hydropower dams. In 2017, the institute conducted a series of field tests at New York’s St. Lawrence River, which not only provided needed monitoring data but also established the Adaptive Resolution Imaging Sonar as a promising tool for eel detection within 20 meters during periods of high river flow. The crux of their most recent WPTO-funded analysis was how to accurately automate eel detections using these sonar data. Convolutional neural networks have recently shown superior automated imaging analysis over a variety of artificial intelligence approaches and are capable of “learning” features from data at varying levels of resolution.

To expand on the 2017 field data, EPRI partnered with PNNL to conduct laboratory tank tests, comparing American eels with other moving objects to demonstrate the validity of convolutional neural networks data analytics methods. Testing results proved to be exceedingly promising: while lab and field-test evaluations yielded high accuracy rates (approximately 98% and 90%, respectively), EPRI found that convolutional neural network algorithms trained on a combination of laboratory and field data achieved 100% classification accuracy for eels versus similar-sized sticks and other objects in the field. Automating sonar-based detection with such high accuracy rates has the potential to save time and reduce costs of monitoring fish passage and, ultimately, fully automate fish passage systems. Key next steps for this project include acquiring more data, modifying tools to identify other species of fish, integrating hydroacoustic software tools, and more.

Given the results of this study, EPRI and PNNL are also conducting outreach with both hardware and software providers to further develop and commercialize the technology. In 2020, PNNL received a Technology Commercialization Fund award from WPTO to establish a sonar image database of eels and other objects and to transition the convolutional neural network algorithms into user-friendly software with a graphical interface. The award will also provide an opportunity for the project team to acquire more field data from partners, including the Technical University of Munich, the U.S. Geological Survey, Hydro Tasmania, and the Great Lakes Fishery Commission. The resulting open-source tools developed throughout the rest of the project will be provided at no cost to the public.

Software tools developed by this project are posted at github.com/xzang/Deep-learning-for-sonar-images.

For additional information, contact Paul Jacobson.