Detection and Classification of Vocalizations for the Study of Marine Mammal Distributions in the Chukchi Sea

Abstract

Automatic detection and classification of marine mammal vocalizations were performed on a large acoustic dataset collected almost continuously between July 2007 and October 2009 in the Alaskan Chukchi Sea. The purpose of this work was to determine spatial and temporal distributions of marine mammals over a wide area of the Chukchi Sea and to characterize ambient and anthropogenic noise. The acoustic data were obtained from multiple consecutive deployments of between 8 and 44 underwater acoustic recorders sampling at 16 kHz. Median filter and split-window normalizer detection processors were implemented to effectively detect vocalization events. The classification of calls by species was found to be more difficult due to a wide range of vocalization types produced by at least nine species; vocalizations were identified from bowheads, belugas, gray whales, fin whales, killer whales, walruses, bearded and ribbon seals, and arctic cod. Many of these species produced multiple call types and some call types evolved seasonally. Several classification approaches were implemented, and their performances were quantified by comparing classifier outputs with the results from manual classification analyses of a subset of the data. This presentation discusses the classification approaches implemented and the performance evaluations of the classifiers for selected species.