Fish migration monitoring from audio detection with CNNs

Abstract

The monitoring of migratory fish is essential to evaluate the state of the fish population in freshwater and follow its evolution. During spawning in rivers, some species of alosa produce a characteristic splash sound, called “bull”, that enables to perceive their presence. Stakeholders involved in the rehabilitation of freshwater ecosystems rely on staff to aurally count the bulls during spring nights and then estimate the alosa population in different sites. In order to reduce the human costs and expand the scope of study, we propose a deep learning approach for audio event detection from recordings made from the river banks. Two different models of Convolutional Neural Networks (CNNs), namely AlexNet and VGG-16, have been tested. Encouraging results enable us to aim for a semi-automatized and production oriented implementation.