Automated Classification of Humpback Whale Calls Using Deep Learning: A Comparative Study of Neural Architectures and...

Area of Science:

• Marine bioacoustics
• Computational intelligence
• Signal processing

Background:

• Passive acoustic monitoring (PAM) generates large datasets requiring automated analysis.
• Humpback whale vocalizations are crucial for population studies but challenging to classify.
• Existing classification methods may lack accuracy and robustness.

Purpose of the Study:

• To develop and evaluate an automated humpback whale detection system.
• To compare the performance of different neural network architectures and feature representations.
• To optimize data processing and augmentation for improved detection accuracy.

Main Methods:

• A curated dataset of humpback whale audio was created from public repositories.
• Data augmentation techniques were applied to expand the dataset.
• Multiple neural networks, including MobileNetV2 and custom CNNs, were trained using TensorFlow and Keras.
• Mel spectrograms and Mel-Frequency Cepstral Coefficients (MFCCs) were used as feature representations.

Main Results:

• Mel spectrograms consistently outperformed MFCCs across all models.
• Pre-trained MobileNetV2 with mel spectrograms achieved 99.01% accuracy, 99% precision/recall, and 0.98 MCC.
• Custom CNN with mel spectrograms achieved 98.92% accuracy and a 0.75% false negative rate.
• MFCC-based models showed lower robustness and higher false negative rates.

Conclusions:

• Mel spectrograms are a superior feature representation for humpback whale detection compared to MFCCs.
• The developed neural network models, particularly MobileNetV2, demonstrate high efficacy for automated humpback whale identification.
• This research provides a robust framework for analyzing PAM data and advancing marine mammal monitoring.