Unsupervised clustering of biological sounds in a Hawaiian coral reef

Manually labeling datasets for supervised machine learning are often time-consuming, expensive, and subjective, particularly in the ocean acoustic domain where a typical human labeler may not be familiar with the signals of interest. Unsupervised learning can be used to automatically cluster signals in large acoustic datasets with no manual labels, which can greatly expedite tasks related to detection, classification, and soundscape characterization. Here, a convolutional autoencoder was built for the purposes of automatically clustering low frequency (>700 Hz) biological sounds in a Hawaiian coral reef. More than 500,000 detections were clustered into five classes—including humpback whale calls, damselfish courtship displays, parrotfish feeding sounds, and two unidentified fish vocalizations. Unique diel, lunar, and seasonal trends are observed for all five classes, demonstrating the potential for unsupervised algorithms to illuminate the environmental factors influencing biological soundscapes.