Deep learning models used to detect fish movement over resistivity counters

Abstract

Diadromous fish are one of the most threatened groups of fish species, being subject to pressures from freshwater, estuarine and marine environments. Of these fish, Atlantic salmon is the most economically important and increasingly threatened. To assess salmonid (Atlantic salmon and sea trout) stocks, resistivity counters have been widely used. However, verification of data from the counters can be challenging due to miscounts, misidentification and biases in human verification of fish counts. We applied deep learning models to identify diadromous fish using continuous electrical resistivity data from resistivity fish counters. Our models were tested on three rivers (Frome, Fowey and Test in the South and South-West of England) and compared with a minimum of one year's manually validated data. We detected fish signals from background noise with an F1-score of 99%, large from small fish (≥30 cm) with a precision of 95%, and an increase of >38% small and large fish waveforms. The F1-score for salmonids was 92%, and a significantly greater proportion (>173%) of upstream-moving large salmonids (≥30 cm) were detected compared to manual methods. To date, abundance estimates for resistivity counters have only been applied to salmonids because of labour-intensive waveform identification. Using deep learning methods, we quantified salmonids and other diadromous fish with varying accuracies. Our method can be applied to resistivity counters to detect diadromous fish globally, reducing human bias and improving detection accuracy.