Classification of large acoustic datasets using machine learning and crowdsourcing: application to whale calls

Area of Science:

• Marine biology
• Bioacoustics
• Animal communication

Background:

• Vocalizations are crucial for marine mammal communication, enabling information transfer over various distances.
• The diversity of cetacean calls presents challenges in classification and regional identification.
• The Whale FM project utilizes advanced methods to study whale vocal behavior.

Purpose of the Study:

• To investigate the effectiveness of computer and citizen science methods in categorizing killer whale and pilot whale sounds.
• To determine if distinct acoustic repertoires exist between whale populations from different geographic locations.
• To assess the accuracy of automated versus human analysis in identifying whale populations by sound.

Main Methods:

• Audio recordings were collected from killer whales and pilot whales using sensors near Norway, Iceland, and the Bahamas.
• Computer algorithms were employed for automated analysis of whale vocalizations.
• Citizen scientists participated in the Whale FM project to analyze sound data.

Main Results:

• Automated analysis accurately separated Icelandic and Norwegian killer whales based on their calls.
• Pilot whales were successfully differentiated into Norwegian long-finned and Bahamas short-finned groups by computer analysis.
• Citizen science analysis also distinguished whale locations but with slightly lower accuracy than computer methods.

Conclusions:

• Geographic location influences the acoustic repertoires of killer whales and pilot whales.
• Computer-based analysis demonstrates high efficacy in identifying regional differences in whale vocalizations.
• Citizen science offers a viable, albeit less precise, complementary approach to bioacoustic analysis.