Machine Learning for the Detection and Characterisation of Underwater Munitions in Seafloor Imagery

: (Alternative Format Thesis)

Abstract

Underwater munitions pose a serious risk to the many inhabitants and users of the marine environment. Efforts to monitor and remediate these munitions are complicated by their remote location and the dangers of human intervention. Additionally, the strong attenuation of electromagnetic radiation in water limits the effectiveness of many conventional sensing modalities. Acoustic imaging, and in particular synthetic aperture sonar, coupled with modern autonomous underwater vehicles enables high-resolution imaging of large areas of the sea floor. However, a key challenge remains in enabling an autonomous system to interpret the images to facilitate faster, cheaper surveying and adaptive mission planning.

This thesis explores advances in machine learning to improve the automatic detection and characterisation of underwater munitions. Specifically, it focuses on self-supervised learning, synthetic data augmentation, and sonar-based multi-view analysis. A focus on techniques that do not require human labels was in response to the scarcity and unreliability of labelled training data from subsea environments. These challenges were identified in initial work using multi-view, and multi-modal data to quantify label reliability and accuracy.

The self-supervised method produced an interpretable intermediate result - a controllable re-lighting of the scene. This representation of the 2D image data was extended to explicitly predict 3D geometry by the development of a differentiable sonar simulator. Additionally, a diffusion-based super-resolution model was developed to enhance low-resolution legacy sonar data for use in modern machine learning pipelines. These approaches have been validated on real survey data, demonstrating improved munition detection while mitigating the effects of unreliable human-labelled data.

Together, these findings have the potential to advance autonomous munition remediation, improve large-scale subsea surveying capabilities, and inform future research in self-supervised learning for underwater imaging. The proposed techniques offer a scalable framework for enhancing sonar-based object recognition beyond munition detection, with applications in marine archaeology, environmental monitoring, and offshore infrastructure assessment.

Date of Award	25 Mar 2026
Original language	English
Awarding Institution	University of Bath
Supervisor	Alan Hunter (Supervisor) & Tom Fincham Haines (Supervisor)