Automated classification of tursiops aduncus whistles based on a depth-wise separable convolutional neural network...

Area of Science:

• Marine Biology
• Bioacoustics
• Artificial Intelligence

Background:

• Whistle classification is crucial for understanding cetacean (toothed whale) communication, social behaviors, and habitat usage.
• Tursiops aduncus (bottlenose dolphin) whistles contain vital information about individual and group dynamics.
• Existing classification methods may struggle with feature extraction and imbalanced datasets.

Purpose of the Study:

• To propose a novel deep learning architecture for accurate whistle classification.
• To address the challenge of imbalanced sample categories in whistle datasets.
• To enhance the generalization ability and classification performance for Tursiops aduncus whistles.

Main Methods:

• Development of a depthwise separable convolutional neural network (CNN) for whistle feature extraction, utilizing depthwise and pointwise convolutions.
• Implementation of a random series audio augmentation method with five algorithms to address data imbalance and improve generalization.
• Passive acoustic monitoring over eight months in the Xiamen area to collect Tursiops aduncus whistle data.

Main Results:

• The proposed depthwise separable CNN achieved superior classification performance compared to conventional CNNs, with reduced computational complexity and parameters.
• The random series audio augmentation method effectively improved the model's generalization ability.
• The combined approach achieved a high classification accuracy of 98.53% for Tursiops aduncus whistles.

Conclusions:

• The proposed depthwise separable CNN architecture is effective for classifying Tursiops aduncus whistles.
• The novel audio augmentation technique successfully mitigates issues related to imbalanced datasets.
• This study provides a robust computational tool for advancing bioacoustic research on cetaceans.