Deep Machine Learning Techniques for the Detection and Classification of Sperm Whale Bioacoustics

Area of Science:

• Marine Biology
• Bioacoustics
• Computational Science

Background:

• Sperm whale bioacoustics is crucial for understanding their behavior and communication.
• Traditional methods for analyzing sperm whale vocalizations are often labor-intensive and limited in scope.
• Machine learning (ML) offers novel approaches to process and interpret complex acoustic data.

Purpose of the Study:

• To implement ML techniques, specifically neural networks, to analyze sperm whale (Physeter macrocephalus) bioacoustics.
• To develop and evaluate ML models for classifying echolocation clicks, coda types, vocal clans, and identifying individual whales.
• To assess the accuracy and feasibility of using ML for advanced bioacoustic analysis.

Main Methods:

• Convolutional Neural Networks (CNNs) were used to develop an echolocation click detector.
• Recurrent neural networks, including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), were trained for classification tasks.
• Datasets from Dominica and the Eastern Tropical Pacific (ETP) were utilized for training and validation.

Main Results:

• The CNN-based click detector achieved 99.5% accuracy in classifying sperm whale spectrograms.
• Coda type classification accuracy reached 97.5% (Dominica) and 93.6% (ETP).
• Vocal clan classification achieved 95.3% (Dominica) and 93.1% (ETP) accuracy, with 99.4% accuracy for individual whale identification.

Conclusions:

• ML techniques, particularly CNNs and recurrent neural networks, are highly effective for analyzing sperm whale bioacoustics.
• These models demonstrate significant accuracy in classifying clicks, codas, vocal clans, and identifying individuals.
• The study validates the use of neural networks for extracting detailed information from whale vocalizations, paving the way for future research.