A YOLOv8-Based Visual Approach for Dolphin Whistle Call Detection

Recognizing dolphin whistle signals holds significant importance for understanding the health of marine ecosystem and investigating cetacean communication mechanisms and social behaviors. Passive acoustic monitoring (PAM), a non-invasive method, has been widely adopted for underwater acoustic signal acquisition. However, detecting and annotating dolphin whistle calls is challenging due to limited training data and the relatively low signal-to-noise ratio in marine environments. This study proposes a novel visual approach to detecting dolphin whistle calls using a pretrained YOLOv8 object detection framework. Experimental results demonstrate robust performance in simultaneous vocalization classification and temporal localization, achieving a mean classification accuracy of 0.924 and temporal detection intersection-over-union (IoU) of 0.946. Notably, our framework successfully integrates multi-category vocalization analysis with precise temporal resolution, representing a significantly advancement over previous single-task approaches in cetacean acoustics signal detection. This work introduces an automated solution for detecting and annotating vocalization signals of marine mammal animals, with potential applications in endangered species conservation and the study of dolphin communication, syntax, and social behavior.