Passive acoustic monitoring can inform our understanding of baleen whale behavior by recording and analyzing their vocalizations. Two crucial factors in the analysis of whale calls are their detection and localization. In this study, we first develop a machine learning method to detect Bryde's whale calls observed by ocean-bottom instruments deployed in the Panama basin, and back-project the detected events to determine their localizations. Using previously identified Bryde's whale calls, we apply data augmentation strategies to increase the size of our training dataset to ultimately obtain 890 214 training examples. Using an evaluation dataset, we determine which detection thresholds optimize false positives and negatives and apply these to continuously recorded hydrophone data. The detection resulted in 4514 potential events, of which 899 were recorded by at least three instruments. The waveforms of these events were automatically extracted, cross correlation probability envelopes were computed between hydrophones, and these were finally back-projected onto a 3D grid to obtain final event localizations. For this network, this procedure is shown to be robust to high noise levels, random time errors, and systematic bias introduced by the velocity model. This approach has further advantages, such as being computationally efficient and requiring minimal manual intervention.
Tary et al. (Fri,) studied this question.