Abstract:
Passive acoustic monitoring (PAM) enables large-scale, non-invasive observation of marine megafauna, but detecting and temporally localizing whale vocalizations in long, noisy recordings remains a challenge. This work benchmarks two distinct approaches: an unsupervised Radon + Ridge Transform-based model for detecting characteristic frequency sweeps in whale calls, and a weakly supervised multiple instance learning (MIL) model that jointly performs detection and tempo ral localization. The Radon-based model provides interpretable detection without requiring annotations, while the MIL model uses attention mechanisms to identify and localize call seg ments, achieving an F1-score of 82% for detection and 70% for localization. Although the MIL model does not outperform a supervised CNN baseline in detection (83% F1), it uniquely enables localization, offering a more complete solution. These benchmarking results highlight the complementary strengths of interpretable unsupervised methods and flexible weakly supervised models, and underscore the value of modular pipelines for advancing automated whale call analysis.
