Blind source separation in complex marine soundscapes: An unsupervised two-stage clustering based on non-negative...

Area of Science:

• Ecoacoustics
• Bioacoustics
• Signal Processing

Background:

• Soundscape monitoring is crucial for biodiversity assessment using acoustic signals.
• Overlapping sound sources in complex environments challenge traditional analysis methods.
• Effective source separation is needed for accurate ecoacoustic analysis.

Purpose of the Study:

• To develop an unsupervised blind source separation algorithm for complex soundscapes.
• To improve the performance of biodiversity assessment through enhanced acoustic signal separation.
• To provide a practical approach for fine-grained source separation.

Main Methods:

• Proposed an unsupervised blind source separation algorithm utilizing nonnegative matrix factorization (NMF).
• Implemented a two-stage coarse-to-fine clustering strategy: NMF-based coarse clustering followed by hierarchical clustering-guided K-means.
• Evaluated performance on simulated and real-world data using separation quality, detection metrics, and a composite score.

Main Results:

• The proposed two-stage method significantly outperformed one-stage clustering in source separation quality.
• Demonstrated superior performance on both simulated and real-world acoustic data.
• Successfully recovered a greater number of individual sound source components compared to existing methods.

Conclusions:

• The developed algorithm offers a practical and effective solution for fine-grained source separation in complex soundscapes.
• This approach supports more quantitative and accurate ecoacoustic analysis for biodiversity monitoring.
• The method shows robustness across varying levels of acoustic mixture complexity.