Monitoring of a Nearshore Small Dolphin Species Using Passive Acoustic Platforms and Supervised Machine Learning Techniques

Abstract

Passive acoustic monitoring (PAM) is increasingly being adopted as a non-invasive method for the assessment of ocean ecological dynamics. PAM is an important sampling approach for acquiring critical information about marine mammals, especially in areas where data are lacking and where evaluations of threats for vulnerable populations are required. The Indo-Pacific humpback dolphin (IPHD, Sousa chinensis) is a coastal species which inhabits tropical and warm-temperate waters from the eastern Indian Ocean throughout Southeast Asia to central China. A new population of this species was recently discovered in waters southwest of Hainan Island, China. An array of passive acoustic platforms was deployed at depths of 10-20 m (the preferred habitat of humpback dolphins), across sites covering more than 100 km of coastline. In this study, we explored whether the acoustic data recorded by the array could be used to classify IPHD echolocation clicks, with the aim of investigating the spatiotemporal patterns of distribution and acoustic behavior of this species. A number of supervised machine learning algorithms were trained to automatically classify echolocation clicks from the different types of short-broadband pulses recorded. The best performance was reported by a cubic support vector machine (Cubic SVM), which was applied to 19,215 5-min recordings (similar to 4.2 TB), collected over a period of 75 days at six locations. Subsequently, using spectrogram visualization and audio listening, human operators confirmed the presence of clicks within the selected files. Additionally, other dolphin vocalizations (including whistles, buzzes, and burst pulses) and different sound sources (soniferous fishes, snapping shrimps, human activities) were also reported. The detection range of IPHD clicks was estimated using a transmission loss (IL) model and the performance of the trained classifier was compared with data synchronously collected by an acoustic data logger (A-tag). This study demonstrates that the distribution and habitat use of a coastal and resident dolphin species can be monitored over a large spatiotemporal scale, using an array of passive acoustic platforms and a data analysis protocol that includes both machine learning techniques and spectrogram inspection.