kfas

SEARCH FOR ARTICLE

kfas, vol. 57, no. 6, pp.743-751, December, 2024 DOI. https://doi.org/10.5657/KFAS.2024.0743

Research on Automatic Classification of Dolphin Whistle in Underwater

Acoustic Data Using Convolution Neural Network (CNN) Model

윤영글1·김한수1·김범규1,2·조성호1·강돈혁1·김선효1*
1한국해양과학기술원 해양력강화·방위연구부, 2부경대학교 인공지능융합학과

ABSTRACT

With the advancement of underwater acoustic observation technology and the development of high-performance equipment, passive acoustic monitoring (PAM) has become a widely used method for monitoring ecosystem cetaceans. As the amount of high-resolution acoustic data increases exponentially, researchers continuously strive to automate the classification of marine mammal sounds and improve the speed of acoustic data analysis. This study developed a method for automatically classifying the dolphin whistles of common dolphins, false killer whales, and bottlenose dolphins using a convolutional neural network model. Whistle signals were analyzed using three spectrogram images: Original, median filter, and edge detection. The F1-Scores analysis results for the three data types were 84.40%, 84.65%, and 80.29%, confirming the possibility of automating whistle classification for the three dolphin species. In Case 2 of the F1-Scores, the common dolphin scored 90.90%, the false killer whale scored 80.00%, and the bottlenose dolphin scored 85.71%, indicating that the common dolphin showed relatively high automatic detection. This study contributes to developing an automatic detection and classification technique for cetacean species in South Korea. Long-term underwater acoustic measurement data acquired through PAM can be utilized more effectively in the future.
Keyword

Deep learning, Convolution neural network (CNN), Underwater acoustics, Whistle, Automatic classification