Confronting Acoustic Data Scarcity: A Physics-Informed Approach to Trustworthy Machine Learning in Canada's Arctic

Abstract

This thesis presents a novel, data-centric framework for building trustworthy artificial intelligence for passive acoustic monitoring in the Arctic, a region increasingly challenged by climate change and geopolitical pressures. The research addresses model brittleness in ship-radiated noise classification, framing it as a data problem that can be solved through a systematic, iterative process of data exploration, diagnosis, and augmentation.

A detailed analysis of ship-radiated noise using horizontal line array element data provides a characterization of the complex variability of acoustic signatures. The thesis quantifies the horizontal directionality of radiated noise from individual ships and the broad-scale impact on the ambient soundscape, leveraging a unique data opportunity presented by the COVID-19 pandemic. Using these insights, a human-in-the-loop methodology is developed to diagnose the specific failure modes of a custom deep neural network classifier. This is achieved by visualising how real-world variability, such as source-receiver range and operational state changes, manifests in the model's learned feature space.

The thesis culminates by demonstrating a physics-informed data augmentation strategy as the solution to data scarcity and diagnosed failures. Through the generation of targeted, high-fidelity synthetic data, this approach measurably improves classifier robustness on unseen real-world data, providing a validated methodology for developing reliable automated passive acoustic monitoring systems in complex environments.