The Bering-Chukchi-Beaufort (BCB) bowhead whale population produces a rich diversity of call types during their annual fall migration. While many calls are recognizable as frequency-modulated (FM) sweeps, a significant portion are difficult to categorize and are often lumped into a generic “complex” category. In this study we employ a range of deep learning approaches to enhance detection and classification of bowhead whale vocalizations. Our analysis uses a manually-curated database of 2 million call detections from 2008–2014 off Alaska’s North Slope, alongside several millionadditional calls from an automated database. Beyond using convolutional neural network methods to improve detection, we utilize an encoder-decoder architecture to create a embedding space upon which we apply unsupervised and supervised algorithms to cluster or classify complex calls into discrete classes. The effort is rendered more challenging by the frequent occurrence and spectral diversity of seismic airgun signals that are overlap at the same bandwidth. Work supported by the North Pacific Research Board.
Thode et al. (Wed,) studied this question.