Automated classification of frequency-modulated bowhead whale calls using contour tracing and image segmentation methods.

Abstract

In 2007, directional autonomous seafloor acoustic recorders (DASARs) were deployed at 35 locations over a 280 km swath in the Beaufort Sea in order to monitor potential changes in bowhead whale (Balaena mysticetus) locations and/or acoustic activity during seismic exploration activities. The large amount of acoustic data generated motivated the development of computer-aided methods to assist in detecting and classifying bowhead whale calls. In the classification stage, bowhead whale calls were divided into six categories: (1) upsweeps, (2) downsweeps, (3) constant calls, (4) u-shaped and (5) n-shaped undulated calls, and (6) complex calls. The frequency range, duration, and fine structure of individual calls vary considerably, creating difficulties when using match-filtering or spectrogram correlation methods. A manually reviewed test data set was assembled, containing examples from each call category, arranged by signal-to-noise ratio. The data set was then used to test several methods (based on image segmentation techniques) for extracting relevant parameters from the signal for subsequent classification. An optimization procedure was then used to generate receiver operating characteristic curves and thus determine appropriate decision boundaries for optimum statistical classifiers. [Work supported by Shell Exploration and Production Company.]