A convolutional neural network applied to arctic acoustic recordings to identify soundscape components

Abstract

Underwater acoustic recordings often have contributions from anthropogenic, geophonic, and biophonic sources and deconstructing the sound field into constituent parts can be useful when characterizing the soundscape. This work focuses on the analysis of a yearlong acoustic measurement made on the Chukchi Shelf as part of the Canada Basin Acoustic Propagation Experiment (CANAPE). The data were recorded on a shallow-water array with horizontal and vertical apertures and span time periods including the open water season, marginal ice zone, and complete ice cover. A supervised convolutional neural network is trained on labeled spectra and demonstrates the ability to separate anthropogenic and biologic sources. Details of the model, labeling process, initial results, and lessons learned will be discussed in this talk.