Sparse representation-based classification of mysticete calls.

Area of Science:

• Bioacoustics
• Machine Learning
• Signal Processing

Background:

• Mysticete (baleen whale) vocalizations are crucial for understanding whale behavior and population dynamics.
• Automatic classification of whale calls is challenging due to complex acoustic environments and overlapping signals.
• Existing methods may require extensive feature engineering or retraining for new call types.

Purpose of the Study:

• To develop an automatic, noise-robust classification method for mysticete calls.
• To create a modular classifier that can be easily updated with new call types.
• To evaluate the method's performance on diverse mysticete call datasets.

Main Methods:

• Utilizes sparse representations and dictionary-based modeling to define whale call subspaces.
• Employs a classifier that rejects signals outside the learned linear subspaces as noise.
• Achieves noise rejection without requiring feature learning, enhancing modularity and ease of design.

Main Results:

• Achieved an average recall of 96.4% across five tested mysticete call types.
• Successfully rejected 93.3% of persistent and transient noise samples.
• Demonstrated high performance on Antarctic blue whale Z-calls, pygmy blue whale calls, fin whale 20 Hz calls, and North-Pacific blue whale D-calls.

Conclusions:

• The proposed sparse representation-based classifier offers an effective and robust solution for automatic mysticete call identification.
• The method's modularity allows for flexible adaptation to new whale call species or types.
• The classifier demonstrates significant potential for bioacoustic monitoring and research applications.