ENHANCED MULTI-LABEL CLASSIFICATION OF HETEROGENEOUS UNDERWATER SOUNDSCAPES BY CONVOLUTIONAL NEURAL NETWORKS USING BAYESIAN DEEP LEARNING

Abstract

The classification of underwater soundscapes is a challenging task for humans as well as machine learning systems. This is largely due to the heterogenous nature of these soundscapes, especially in coastal zones close to human settlements, where multiple ships and other man-made and natural sound sources are often present simultaneously. This thesis proposes a Bayesian deep learning approach that can accurately classify multiple ships simultaneously present in the vicinity of a sensor (multi-label classification) while also providing an uncertainty measurement for the classification. This is achieved by assuming a Bayesian formulation of standard convolutional neural network architectures to not only assign multi-labels per inference but also to provide per inference uncertainty. The best performing Bayesian architecture on the multi-label task achieves a weighted F1 score of 0.84, where each prediction is accompanied by a measurement of uncertainty that is used to further enhance the understanding of model predictions. Ships, submarines, and unmanned underwater vehicles can use this classification system to aid in the identification, tracking, and/or targeting of contacts to help maintain safety of navigation, to aid in the real-time interdiction of illicit activities (such as drug or human smuggling and covert vessel transits), and to provide port security monitoring while uncertainty filters can help sonar operators prioritize contacts for further analysis.