Automatic detection of fish sounds: A comparison of traditional machine learning with deep learning

Many species of fish produce sounds that can be used to monitor them non-intrusively and could complement traditional monitoring techniques. However, the manual annotation of fish sounds in acoustic recordings remains time-intensive, limiting the use of passive acoustics as a viable monitoring tool. This study compares two automated approaches for detecting fish sounds: Random Forest (RF) and Convolutional Neural Networks (CNN). Both algorithms were trained on 21,950 manually labeled fish and non-fish sounds recorded between 2014 and 2019 in the Strait of Georgia, British Columbia, Canada. Performance calculated on data from the Strait of Georgia, Barkley Sound, and the Port of Miami showed that the CNN performed up to 1.9 times better than the RF (F-score: 0.82 versus 0.43) and was in some cases able to find more faint fish sounds than the analyst. Noise analysis in the 20–1000 Hz frequency band shows that the CNN is still reliable in noise levels greater than 130 dB re 1 μPa in the Port of Miami but becomes less reliable in Barkley Sound past 100 dB re 1 μPa due to mooring noise. We show that the proposed approach can make passive acoustics viable for monitoring fish in a variety of environments.