Automatic detection of dolphin whistles and clicks based on entropy approach

Abstract

Abstract Long-term monitoring of cetacean vocalizations allows for the exploration of their occurrence, seasonality and abundance. However, accurate automatic detection of vocalizations from vast acoustic datasets containing diverse sound sources remains a challenge. In this study, we propose the permutation entropy (H) and the sample entropy (SE) as metrics for the unattended detection of whistles and clicks. We tested the detection performance of whistles and clicks in various scenarios commonly occurring in marine habitats, including dense snapping shrimps, vessel engine noise and overlapping whistles and clicks. The use of the entropy metrics resulted in detection accuracy of over 95%. In particular, H outcomes correctly detected whistles even if associated with snapping shrimps or engine noise, while SE was a reliable indicator for clicks and robust to vessel noise. These algorithms do not require prior training in vocalization and are computationally fast. The advancement of metrics such as those presented here, will enable non-invasive and cost-effective assessment of cetacean population dynamics and health and may inform future conservation management.