Multilabel Classification of Heterogeneous Underwater Soundscapes With Bayesian Deep Learning
Brandon Beckler, Andrew Pfau, Marko Orescanin, Sabrina Atchley, Nicholas Villemez, John E. Joseph, Christopher W. Miller, Tetyana Margolina
Abstract
Underwater soundscapes of coastal zones close to human settlements are heterogeneous in nature. Multiple ships and biological sources are often simultaneously present in the passive sonar vicinity. The classification of such heterogeneous underwater soundscapes is a challenging task for humans as well as machine learning systems. In this article, a Bayesian deep learning approach is proposed that can accurately classify multiple ships simultaneously present near the sensor (multilabel classification) and provide uncertainty in the classification. This is achieved by assuming a Bayesian formulation of standard convolutional neural network architecture to not only assign multilabels per inference but also to provide per inference uncertainty. By utilizing almost 3500 h of passive sonar data (spanning more than a year of sensor deployment) labeled through automated fusion with automatic identification system information, both multiclass and multilabel classification tasks of ship-generated noise are addressed. The best performing Bayesian architecture on the multilabel task achieves a weighted <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$F^{1}$</tex-math></inline-formula> score of 0.84, where each prediction is accompanied by a measurement of uncertainty, which is used to further enhance the understanding of model predictions.