Classification of dispersive gunshot calls using a convolutional neural network

Area of Science:

• Marine bioacoustics
• Machine learning applications in ecology
• Signal processing for environmental monitoring

Background:

• Passive acoustic monitoring (PAM) generates vast datasets requiring efficient analysis.
• Identifying specific marine mammal vocalizations, like right whale gunshots, is crucial for population studies.
• Manual screening of acoustic data is time-consuming and labor-intensive.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for automated identification of right whale gunshots.
• To assess the CNN's ability to generalize to unseen datasets from different right whale populations.
• To explore the utility of identified gunshots for acoustic inversion of source and environmental parameters.

Main Methods:

• A CNN was trained using South Atlantic right whale gunshot vocalizations.
• The trained CNN was tested on an independent dataset of North Pacific right whale (NPRW) gunshots.
• The model's performance was evaluated for its accuracy in detecting NPRW gunshots.

Main Results:

• The CNN demonstrated successful generalization, accurately identifying NPRW gunshots.
• Identified gunshots were suitable for inverting source range and environmental parameters.
• The automated approach significantly reduces the time needed for manual data screening.

Conclusions:

• CNNs provide an effective tool for automated detection of marine mammal impulse calls in large acoustic datasets.
• This technology enhances the efficiency of passive acoustic monitoring data analysis.
• Automated gunshot identification facilitates further acoustic research, including environmental and source parameter inversions.