Detecting, classifying, and counting blue whale calls with Siamese neural networks

Area of Science:

• Marine Biology
• Bioacoustics
• Artificial Intelligence

Background:

• Blue whale populations exhibit distinct acoustic signatures.
• Accurate population assessment is crucial for effective conservation strategies.
• Analyzing large acoustic datasets for blue whale vocalizations presents significant challenges.

Purpose of the Study:

• To develop and evaluate a deep learning model for detecting, classifying, and counting blue whale calls from acoustic recordings.
• To compare the performance of Siamese neural networks (SNN) against convolutional neural networks (CNN) for blue whale call analysis.
• To enhance the assessment of blue whale population status through improved acoustic monitoring.

Main Methods:

• Utilized 350 hours of underwater hydrophone recordings from the Indian Ocean.
• Developed a deep learning model using Siamese neural networks (SNN) for acoustic signature analysis.
• Employed manual annotations for training and validation of the SNN model to identify four blue whale acoustic song types.

Main Results:

• Siamese neural networks (SNN) demonstrated superior performance compared to convolutional neural networks (CNN).
• SNN achieved a 2% accuracy improvement in blue whale population classification.
• SNN provided 1.7%-6.4% greater accuracy in call count estimation for each population and learned ordinal relationships in call counts.

Conclusions:

• Siamese neural networks (SNN) are a robust and effective method for automatically analyzing large acoustic datasets for blue whale calls.
• The developed SNN model significantly enhances the ability to detect, classify, and count blue whale vocalizations.
• This advancement offers improved tools for monitoring blue whale populations and informing conservation efforts.