Soundscape Characterization Using Autoencoders and Unsupervised Learning.

Area of Science:

• Ecology
• Bioacoustics
• Machine Learning

Background:

• Passive acoustic monitoring (PAM) using acoustic recorder units (ARUs) is valuable for detecting landscape changes and biodiversity patterns.
• Current PAM methods often rely on supervised approaches, which are limited by large data volumes and the need for prior dataset knowledge.

Purpose of the Study:

• To propose and evaluate a non-supervised framework utilizing autoencoders for extracting soundscape features from ARU data.
• To demonstrate the utility of unsupervised learning for analyzing large-scale acoustic datasets and identifying ecological patterns.

Main Methods:

• Developed a non-supervised framework employing autoencoders to extract soundscape features.
• Applied the framework to acoustic data from Colombian landscapes collected by 31 audiomoth recorders.
• Generated soundscape clusters using autoencoder features and represented them with prototype spectrograms via centroid features and the neural network's decoder.

Main Results:

• The autoencoder framework successfully identified significant soundscape patterns, including recurring and intense sound types across various frequency ranges.
• Analysis provided insights into the distribution and temporal dynamics of sound compositions within the study area.
• The method enabled pinpointing crucial sound sources and enhancing understanding of the acoustic environment.

Conclusions:

• Unsupervised algorithms, particularly autoencoders, offer a promising alternative for soundscape analysis and environmental monitoring.
• This approach addresses challenges associated with large acoustic datasets and the limitations of supervised methods.
• The findings support the broader application of machine learning in ecological soundscape research.