Discriminating and classifying odontocete echolocation clicks in the Hawaiian Islands using machine learning methods

Area of Science:

• Marine biology
• Bioacoustics
• Machine learning

Background:

• Passive acoustic monitoring (PAM) is a non-invasive method for studying marine mammals.
• Large datasets of toothed whale echolocation clicks from Hawai'i were collected between 2008-2019.
• Previous analyses were limited by dataset size and classification complexity.

Purpose of the Study:

• To develop and apply a machine learning toolkit for classifying toothed whale echolocation clicks.
• To enable multi-year, multi-species analyses of acoustic data.
• To improve the efficiency of processing large passive acoustic monitoring datasets.

Main Methods:

• Utilized unsupervised clustering and human-mediated analysis to identify echolocation click types.
• Developed a neural network classifier trained on 5-minute data segments.
• Attributed a new click type to rough-toothed dolphins (Steno bredanensis) using auxiliary data.

Main Results:

• Distilled ten unique echolocation click types attributable to regional odontocetes.
• Achieved high accuracy (>96%) and recall (>75%) for most species.
• Identified variable precision for false killer whales (Pseudorca crassidens) and one short-finned pilot whale (Globicephala macrorhynchus) call class.

Conclusions:

• Machine learning is highly effective for analyzing large passive acoustic monitoring datasets.
• The developed classifier and timeseries facilitate spatiotemporal analyses of toothed whales.
• This approach can enhance global multi-species PAM data processing for echolocation clicks.