Deep Learning for Automated Detection and Identification of Migrating American Eel Anguilla rostrata from Imaging Sonar Data

Abstract

Adult American eels (Anguilla rostrata) are vulnerable to hydropower turbine mortality during outmigration from growth habitat in inland waters to the ocean where they spawn. Imaging sonar is a reliable and proven technology for monitoring of fish passage and migration; however, there is no efficient automated method for eel detection. We designed a deep learning model for automated detection of adult American eels from sonar data. The method employs convolution neural network (CNN) to distinguish between 14 images of eels and non-eel objects. Prior to image classification with CNN, background subtraction and wavelet denoising were applied to enhance sonar images. The CNN model was first trained and tested on data obtained from a laboratory experiment, which yielded overall accuracies of >98% for image-based classification. Then, the model was trained and tested on field data that were obtained near the Iroquois Dam located on the St. Lawrence River; the accuracy achieved was commensurate with that of human experts.

Automatic detection, identification and counting of anguilliform fish using in situ acoustic camera data: development of a cross-camera morphological analysis approach

“…(12) obtained a higher performance (recall = 91%, precision = 96%) for 5-m wide DIDSON videos, that mainly increase the image resolution; fragmented targets were likely not a serious problem for their dataset. (14) also obtained a similar good performance (recall = 84%), but it was based on a few individuals of American eels, with no other fish species in their dataset. Moreover, both studies used data from a single acoustic camera model at a single monitoring site.…”

“…Multiple automatic or semiautomatic methods have been developed to detect and describe fish using acoustic camera datasets, listed and described in the review of (6). Among the studies quoted by (6), a few authors focused on the distinction of species of interest, such as anguilliform fish, from other species (12)(13)(14). Indeed, an operator can easily distinguish the particular body shape and swimming undulation of anguilliform fish from those of most other fish species (15)(16)(17)(18).…”

“…The distance of the fish from the acoustic camera, as well as its body length and orientation may decrease its detectability, because it may lead to the image of its body into distinct disconnected fragments (13,24). Among recent studies, the deep-learning models of (25) and (14) obtained a recall of 85% when identifying American eels using ARIS sonar recorded under stable flow conditions. The authors recommended that future studies investigate the method’s detection ability over long-term monitoring periods.…”

Monitoring fish using imaging sonar: Capacity, challenges and future perspective

“…The sonar images also contain a large amount of noise, such as speckle noise and intensive reflectance from solid non‐target objects. The mutual interference of the reflected acoustic waves, the reflectance of entrained air or small debris, and ambient environment noise result in a great number of speckle noises (Dos Santos et al, 2017; Zang et al, 2021). In addition, the high‐density objects such as rocks generate strong echoes, which may obscure the sound reflected by fish (Viehman & Zydlewski, 2015).…”

“…For this reason, data conversion is needed for the following processing. For example, the raw data of DIDSON is a 2‐dimensional acoustic array that has to be converted to the polar coordinates (Jing et al, 2019; Zang et al, 2021). As the data are dispersed, interpolation is often used between beams to produce a smoother image (Jing et al, 2019; Kupilik & Petersen, 2014a; Yu et al, 2016).…”