Answers are grounded in this paper and cite their sources.
Passive Acoustic Monitoring has become a cornerstone of marine mammal conservation, enabling continuous, non-invasive monitoring through the detection and analysis of vocalizations. As the focus shifts from retrospective analyses to real-time monitoring, classification systems must operate within the strict computational constraints of ultra-low-power embedded devices, generalize across diverse acoustic environments, remain robust to anthropogenic noise from vessel traffic, and classify distinct beluga sound types to enable advanced ecological analyses such as abundance estimation and behavioral inference. We address these challenges by developing a classification model for the endangered beluga whale population of the St. Lawrence Estuary. Our lightweight, quantized convolutional neural network classifies beluga vocalizations under natural field conditions, including overlapping calls and variable background noise. The model achieves a macro-averaged F1 score of 0.85 with a memory footprint of under 512 kilobytes, satisfying the constraints of edge-deployable microcontrollers. Using labeled examples with and without vessel noise, we directly quantified the impact of anthropogenic interference, revealing a drop to an F1 score of 0.75 in the presence of boat noise. Training on data from four distinct sites, we demonstrate that a single multi-site model generalizes effectively across all locations. In a leave-one-site-out evaluation, the full-precision model maintains strong performance on unseen sites with an average F1 drop of only 0.05, while the quantized model shows a larger gap of 0.10, highlighting the increased sensitivity of compressed models to distribution shifts.
No takes yet. Share an insight, caveat, or question.
Fernandez et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: