Fine-scale classification of odontocete echolocation clicks using deep learning

Abstract

All toothed whale species produce echolocation clicks, which can be used for passive acoustic monitoring. However in biodiverse habitats, such as off the coast of southern California, it is common for acoustic encounters with multiple species to overlap in time. This creates a challenge for acoustic classification in the context of long term monitoring. Machine learning methods may facilitate classification at the level of individual detections. A generic echolocation click detector was applied to passive acoustic recordings from the 2013 DCL dataset, which has manual labels of five odontocete species encounters available as a ground truth. We evaluated two different methods for training deep neural networks to classify the detections to species, using either individual click waveforms or clustered sets of similar echolocation clicks. Results show that deep neural networks have the potential to accurately classify clicks to species at fine scales, allowing for improved handling of complex acoustic environments and multi-species encounters. We suggest that click-level classification can facilitate more advanced quantitative analysis of passive acoustic recordings, and that it can be done efficiently using deep learning.