Rapid detection of fish calls within diverse coral reef soundscapes using a convolutional neural network

Abstract

The quantity of passive acoustic data collected in marine environments is rapidly expanding; however, the software developments required to meaningfully process large volumes of soundscape data have lagged behind. A significant bottleneck in the analysis of biological patterns in soundscape datasets is the human effort required to identify and annotate individual acoustic events, such as diverse and abundant fish sounds. This paper addresses this problem by training a YOLOv5 convolutional neural network (CNN) to automate the detection of tonal and pulsed fish calls in spectrogram data from five tropical coral reefs in the U.S. Virgin Islands, building from over 22 h of annotated data with 55 015 fish calls. The network identified fish calls with a mean average precision of up to 0.633, while processing data over 25× faster than it is recorded. We compare the CNN to human annotators on five datasets, including three used for training and two untrained reefs. CNN-detected call rates reflected baseline reef fish and coral cover observations; and both expected biological (e.g., crepuscular choruses) and novel call patterns were identified. Given the importance of reef-fish communities, their bioacoustic patterns, and the impending biodiversity crisis, these results provide a vital and scalable means to assess reef community health.