Using tropical reef, bird and unrelated sounds for superior transfer learning in marine bioacoustics

Area of Science:

• Ecological acoustics
• Machine learning applications
• Bioacoustics

Background:

• Passive acoustic monitoring (PAM) offers ecological assessment potential but faces high annotation and computing costs.
• Generalizable pretrained networks can reduce these costs, but require large annotated libraries, currently limiting development to data-rich taxa like birds.

Purpose of the Study:

• To identify the optimal pretraining strategy for machine learning models in data-deficient acoustic monitoring domains.
• To evaluate pretraining approaches using tropical reef sound data as a case study.

Main Methods:

• Assembled ReefSet, a 57,000-sound annotated library of tropical reef sounds from 16 datasets.
• Conducted pretraining experiments using bird audio, ReefSet data, unrelated audio, and cross-domain mixtures.
• Evaluated model generalizability on held-out reef datasets using AUC-ROC metrics.

Main Results:

• Pretraining on a large bird audio library (50x ReefSet size) yielded superior generalizability (AUC-ROC 0.881) compared to ReefSet pretraining (0.724) or unrelated audio (0.834).
• Cross-domain mixing of bird, reef, and unrelated audio during pretraining achieved the highest transfer learning performance (AUC-ROC 0.933).
• Developed SurfPerch, an optimized pretrained network for tropical reef PAM data analysis.

Conclusions:

• Cross-domain pretraining is a highly effective strategy for improving machine learning performance in data-deficient acoustic monitoring.
• The SurfPerch network provides a cost-effective foundation for automated analysis of tropical reef and similar PAM datasets.
• This approach significantly reduces annotation and computing burdens, facilitating wider adoption of PAM in ecological research.