Deep embedded clustering of coral reef bioacoustics

Area of Science:

• Marine bioacoustics
• Machine learning
• Signal processing

Background:

• Coral reef soundscapes contain complex acoustic signals from various marine species.
• Distinguishing between different biological sound sources, like fish and whales, is crucial for ecological monitoring.
• Automated methods are needed to analyze large volumes of underwater acoustic data.

Purpose of the Study:

• To apply deep clustering techniques to unlabeled coral reef soundscape data.
• To differentiate between fish pulse calls and whale song segments using automated signal classification.
• To compare the performance of deep embedded clustering (DEC) with Gaussian mixture models (GMM) and conventional clustering methods.

Main Methods:

• Deep embedded clustering (DEC) was used to learn latent features from signal power spectrograms.
• Spectral and temporal features were extracted and clustered using Gaussian mixture models (GMM) and conventional clustering.
• Simulated datasets with varying signal parameters (bandwidth, duration, SNR) were used for initial testing.
• Real-world acoustic data from a coral reef near Hawaii was analyzed.

Main Results:

• DEC and GMM achieved high accuracy in distinguishing fish and whale signals on simulated data.
• Conventional clustering methods showed low accuracy, especially with overlapping or unequal clusters.
• DEC features yielded the highest classification accuracy of 77.5% on a manually labeled real-world dataset.
• Both DEC and GMM successfully identified distinct clusters for fish, whale, and overlapping signals.

Conclusions:

• Deep embedded clustering (DEC) is a promising method for automated classification of bioacoustic signals in complex soundscapes.
• DEC offers superior performance compared to conventional clustering for analyzing underwater acoustic data.
• This approach can aid in the ecological assessment and monitoring of marine life through sound analysis.