Statistical learning for uncertainty-aware active sonar methodology

Research Themes

Sensor Signal Processing 

Multi-agent Systems and Data Intelligence 

Aim

Develop statistical learning methodology for active sonar that provide robust detection and classification across noisy underwater environments while producing calibrated uncertainty estimates to support reliable defence and security decision-making. 

Objectives

The project is divided into an initial exploration stage, followed by one of three possible research avenues, with various possible extensions.  

  • Exploration. Through a review of existing literature, available data sources, and exploratory numerical experiments, investigate the most promising route for maximising the potential of statistical learning methodology in classifying active sonar data while retaining useful quantification of uncertainty. 
  • Methodology development. Enable transfer across frequency bands, sensor configurations, environments, and target classes, including adaptation from broad pretraining to specific operational detection tasks. Depending on the exploration stage, this may involve  
  • the development of a foundation model for active sonar data which can be fine-tuned to specific domains,  
  • a transfer learning model to exploit knowledge from one active sonar domain with large amounts of available data (such as high-frequency side scan sonar) to one with little labelled data (low frequency sonar),  
  • an active learning approach in which the model works only on the target domain data, first clustering it and building a statistical classifier based on human prioritisation of the detected clusters. 
  • Enable practically useful and useable uncertainty quantification, e.g. through Bayesian deep learning approaches or conformal prediction, focusing on uncertainty-aware classification and detection methods in high-clutter situations so that model outputs provide operationally meaningful confidence estimates and prediction sets. 

Description

Active sonar is a key sensing technology for underwater defence and security, but reliable detection and classification remain challenging as acoustic returns are strongly affected by propagation conditions, reverberation, environmental variability, target aspect, and low signal-to-noise ratios. Labelled active sonar data are limited and often highly scenario-specific, making it difficult to train models that generalise reliably to new environments or sensor configurations. 

Existing machine learning research in the acoustics domain focusses largely on passive monitoring due to the financial costs associated with both employing active sonar and labelling collected data.  

This project will investigate statistical learning approaches for active sonar classification. The student will develop machine learning methods that learn transferable representations from large collections of underwater acoustic data, using self-supervised, weakly supervised, or simulation-assisted training where appropriate; the approach will be determined based on initial experiments, a literature review, and available data sources.  

A central focus will be uncertainty quantification. The project will study methods such as Bayesian neural networks, deep ensembles, conformal prediction, calibration techniques, and probabilistic classifiers to ensure that the system outputs not only a prediction, but also an operationally meaningful confidence measure.  

The probabilistic outputs of the methodology will be used in a downstream task in a practically useful way, such as flagging uncertain detections to a human operator. 

 

Closing date: 
Apply now

Principal Supervisor

Eligibility

A strong mathematical background and programming experience, preferably in Python, are desirable. Prior experience in machine learning, signal processing, Bayesian statistics, acoustics, or underwater sensing would be advantageous but is not essential.

Minimum entry qualification - an Honours degree at 2:1 or above (or International equivalent) in a relevant science or engineering discipline, possibly supported by an MSc Degree. 

Funding

This position is fully funded. 

Informal Enquiries