Classification of odontocete echolocation clicks using convolutional neural network.

Area of Science:

• Bioacoustics
• Machine Learning
• Marine Mammal Science

Background:

• Odontocete echolocation clicks are crucial for species identification.
• Manual classification of these clicks is time-consuming and requires expertise.
• Automated methods are needed for large-scale passive acoustic monitoring (PAM).

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for automatic classification of odontocete echolocation clicks.
• To assess the CNN's performance using raw time-domain audio signals.
• To explore strategies for improving classification accuracy with increasing data.

Main Methods:

• A six-layer CNN architecture was designed, including convolutional, fully connected, and softmax layers.
• Rectified linear units were used as activation functions.
• Species prediction was performed on groups of clicks (m) using majority vote and maximum posterior strategies.

Main Results:

• The CNN successfully modeled odontocete species directly from the raw time signal of echolocation clicks.
• Classification accuracy improved as the number of clicks (m) per group increased.
• The proposed method demonstrated effective classification performance on two independent datasets.

Conclusions:

• The developed CNN provides an effective tool for automated odontocete species classification from echolocation clicks.
• The method shows promise for enhancing passive acoustic monitoring efforts for delphinid species.
• This approach can facilitate future research on odontocete bioacoustics and population studies.