Detecting Spinner Dolphin (Stenella longirostris) Clicks In Noisy And Low Sampling Rate Hydrophone Recordings

Abstract

Development of automated detection algorithms for cetacean vocalizations is important to facilitate marine mammal research. This thesis focuses on click train detection in cases in which sampling rates are too low to capture the full bandwidth of the clicks, and in which impulsive noise confounds current detection methods. We develop an algorithm to detect/classify odontocete click trains based on the regular timing of clicks; the method relies on the slowly- varying nature of Inter-Click Intervals (ICIs) within a click train. The algorithm is refined and evaluated using simulated data. It is motivated and applied to recordings of spinner dolphins collected in Hawaii. Performance is quantified using receiver-operating and precision-recall curves for both simulated and real data. While the method shows promise (including the ability to separate multiple clicking animals) for click trains with stable ICI and in relatively low-noise conditions, the performance on the spinner dolphin dataset is marginal.