Detection and localization of Bryde's whale calls using machine learning and probabilistic back-projection

Area of Science:

• Marine Biology
• Bioacoustics
• Machine Learning

Background:

• Passive acoustic monitoring (PAM) is vital for understanding baleen whale behavior through vocalization analysis.
• Accurate detection and localization of whale calls are critical challenges in PAM data interpretation.

Purpose of the Study:

• To develop and validate a machine learning (ML) method for detecting and localizing Bryde's whale calls using ocean-bottom hydrophones.
• To assess the robustness and efficiency of the ML-based localization technique in real-world acoustic conditions.

Main Methods:

• A novel ML model was trained on augmented Bryde's whale call data (890,214 examples).
• Detection thresholds were optimized to balance false positives and negatives.
• Acoustic events were localized by back-projecting cross-correlation data onto a 3D grid.

Main Results:

• The ML method successfully detected 4514 potential whale call events, with 899 localized using data from at least three hydrophones.
• The localization procedure demonstrated robustness against high noise, time errors, and velocity model bias.
• The system proved computationally efficient with minimal need for manual intervention.

Conclusions:

• The developed ML-based passive acoustic monitoring system provides an effective and robust method for detecting and localizing Bryde's whale vocalizations.
• This approach offers significant advantages for marine mammal research, enabling efficient and accurate behavioral analysis in challenging acoustic environments.