Statistical classification of odontocete clicks

Abstract

To the best of our knowledge, all odontocetes produce some kind of click like vocalisation, which is used primarily for echolocation but may also play a role in social communication. Characteristics of these echolocation pulses range from the broad band but relatively low frequency clicks of sperm whales to the ultrasonic, narrow-band clicks of harbour porpoise. Although these clicks are often easily detected, it can be difficult to classify them to species, thereby hampering efforts to monitor and study odontocetes using passive acoustics. Candidate clicks from three species were detected using a simple energy trigger, operating in the frequency band of interest. The clicks were then identified to species using two different statistical classifiers to separate beaked whale vocalisations from those of other odontocete sounds. In the first, a number of parameters (peak frequency, mean frequency, sweep frequency, click duration, width of principal spectral peak and the relative energy in different frequency bands) were calculated and a tree classifier was used to separate clicks of different species. In the second, the spectral energy in 32 relatively coarse energy bands 1.5 kHz wide were used as input to a multivariate classifier. Both classifiers were trained and tested using data provided to the 3rd International Workshop on Detection and Classification of Marine Mammals using Passive Acoustics in order to assess the classifiers performance with Blainville's beaked whales, short-finned pilot whales and Risso's dolphin clicks. The methods were also applied to survey data collected using a towed hydrophone deployed from a sailing research vessel in the Bahamas. Some of the towed hydrophone data were collected over the US Navy's AUTEC range where independent confirmation of beaked whale vocal activity was available from bottom-mounted hydrophones.