Denoising odontocete echolocation clicks using a hybrid model with convolutional neural network and long short-term...

Area of Science:

• Marine Bioacoustics
• Signal Processing
• Deep Learning

Background:

• Ocean noise significantly degrades the quality of recorded odontocete echolocation clicks.
• Accurate recording of echolocation clicks is vital for studying toothed whale behavior and communication.
• Existing denoising methods struggle with the complexities of underwater acoustic environments.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning model for denoising odontocete echolocation clicks.
• To address the challenge of limited noise-free training data for deep learning models.
• To determine the optimal network architecture for effective echolocation click denoising.

Main Methods:

• A hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) model was proposed.
• Echolocation clicks were artificially corrupted with ocean noise for model training.
• Data augmentation and transfer learning, including Gabor function-based simulated clicks, were used to overcome data scarcity.
• Network parameters were fine-tuned using real odontocete echolocation click data.

Main Results:

• The hybrid CNN-LSTM model demonstrated significant effectiveness in denoising both narrowband high-frequency and broadband echolocation clicks.
• Experimental evaluations using synthetic data confirmed the model's denoising capabilities.
• Hybrid models with one convolutional layer and multiple LSTM layers showed superior performance.

Conclusions:

• The proposed hybrid CNN-LSTM model offers a robust solution for removing ocean noise from odontocete echolocation clicks.
• The recommended architecture (one convolutional layer, multiple LSTM layers) is suitable for various echolocation click types.
• This approach can improve the quality of acoustic monitoring data for marine mammal research.