Detecting ringed seal vocalizations using deep learning

Ringed seals (Pusa hispida) in the Arctic face habitat loss from climate change, making monitoring critical for conservation. Passive acoustic monitoring provides essential data; however, large data volumes require automated analysis. We developed a ResNet-based detector for ringed seal barks in passive acoustic recordings. The detector achieved F1 scores above 0.90 on development data and held-out datasets from different locations and recording periods. In continuous recordings, it maintained high recall (>0.90), with lower precision (∼0.5) reflecting a trade-off between detecting most vocalizations and false positives from acoustically similar non-target sounds. The open-source detector supports large-scale bioacoustic monitoring.