Classification of sounds from Pacific white-sided dolphins using a convolutional neural network and a method to reduce false-positive detections.

An automatic detector for identifying the clicks and pulsed calls of Pacific white-sided dolphins (Lagenorhynchus obliquidens) was developed using a convolutional neural network architecture for passive acoustic monitoring, particularly in the areas surrounding the Mutsu and Funka Bays in Japan. Recordings were made at one site in each bay during the spring and early summer in both 2022 and 2023. The data exhibited different soundscapes, as broadband pulses, possibly attributed to snapping shrimp, were found far more frequently in Mutsu Bay than in Funka Bay. The developed detector showed a precision, recall, and accuracy of 0.94-0.95, 0.94, and 0.98, respectively, for both call types. Furthermore, considering the social and gregarious characteristics of the investigated species, an additional selection criterion using a two-process model was proposed to eliminate hours with few positive images. The selection criterion could remove 58%-84% of false-positive images, 0%-0.5% of true-positive images for clicks, 32%-96% of false-positives, and 6%-33% of true-positives for pulsed calls during the periods that were manually inspected. Processing using a combination of the detector and selection criterion can be applied to passive acoustic monitoring around these bays to reveal the migration patterns of Lagenorhynchus obliquidens.