A comparison of methods for the automatic classification of marine mammal vocalizations in the Arctic

JASCO Applied Sciences provided acoustic data collection services to Shell Offshore Incorporated in support of their explorations of the Chukchi and Beaufort Seas during the summer of 2007. A total of 37 ocean bottom hydrophones data recording systems were deployed resulting in a data set in excess of 5 TB. This data equates to almost 5 years of continuous sound recording. This campaign provided information on the migration routes of the marine mammals, which include bowhead, beluga, humpback, gray whales and walruses. Given the large amount of data, manual analysis of the recordings was not feasible. A method to detect and classify the marine mammal vocalizations automatically in a reasonable amount of time had to be developed. The processing is structured in several steps, (1.) the detection of energy events, (2.) the feature extraction, and (3.) the classification into a species variety. This paper focuses on combining the Gaussian mixed models (GMM) algorithm for classification with attributes taken from two different feature extraction algorithms: cepstral coefficients and wavelets. Combinations of these different algorithms are compared using classification operating characteristic (COC) curves for each species tested. This paper compares the performance of these algorithms and their parameters against a large training data set.