Thumbnail Image

Machine learning for automated sonar monitoring of outmigrating American eel behavior

Abundance of American eel (Anguilla rostrata) is low compared to historical levels, which makes it a species of management concern. Adult American eel are vulnerable to hydropower turbine mortality during outmigration from inland waters to their spawning ground in the Sargasso Sea. Morphological and behavioral characteristics of the species make it particularly challenging to provide safe downstream passage at hydropower projects. Furthermore, outmigration of adult eels is episodic and protracted, typically extending over a period of several months or more each year. Consequently, design, optimization, and operation of downstream passage facilities that are economical and biologically effective requires knowledge of eel behavior during their downstream migration. Site-specific information requirements include knowledge of when eels are approaching hydropower facilities; pathways of approach and passage; and near-field behavioral responses to facility structures including intakes, guidance structures, and bypasses. The EPRI Eel Passage Research Center investigated three sonar technologies for observing the abundance and behavior of outmigrating adult eels. Among the three sonar types tested, ARIS multibeam imaging sonar holds the most promise for correctly identifying eels at up to 16-20 meters range. While multi-beam imaging sonar can provide the needed data, the protracted nature of adult eel outmigration and the rate of sonar data production make automated data analysis essential for cost-effective monitoring. Our ongoing project has the goal of developing machine-based detection of American eel from ARIS sonar data. The project will: (1) extract training and validation data from ARIS sonar records collected at the Iroquois Water Control Dam on the St. Lawrence River, USA/Canada; (2) utilize two-dimensional wavelet transform analysis to filter noise and increase contrast of the sonar images; (3) apply convolutional neural network analysis to classify the images into three types of objects – background, eel, and moving, non-eel target; (4) quantify the performance of the analysis system.