Source Detection and Characterisation with Machine Learning

: in Next Generation Radio Continuum Survey Data

Abstract

The field of radio astronomy is on the brink of groundbreaking scientific advances with the construction of the Square Kilometre Array (SKA) telescope. Reaching science goals with current radio telescopes, which have modest data products compared to those promised by the SKA, is inhibited by a lack of accurate and automated source-finding techniques. These techniques are required to extract sources and their parameters from the data in order to constrain evolutionary models of active and star forming galaxies. In this thesis we present a new machine learning (ML) driven source-finding tool for next generation radio surveys that performs fast source extraction on a range of source morphologies at large dynamic ranges with minimal parameter tuning and post processing. We have investigated dimensionality reduction ML techniques for the application of radio image segmentation in next generation survey data, including convolutional and variational autoencoders. These methods were successfully applied to simulated radio continuum data, but they were unable to manage the dynamic range of sources in large radio images. We then developed a novel source-finding method, ContinUNet, powered by an ML segmentation algorithm, U-Net, that has proven highly effective and efficient when tested on SKA precursor data sets. Our model was trained and tested on simulated data and proved comparable to the state-of-the-art source-finding methods in terms of recovery of the source population and their characteristics. ContinUNet was then tested on MeerKAT data without retraining and was able to extract point sources and extended sources with equal ease; processing a 1.6 square degree field in seconds on a supercomputer and minutes on a personal laptop. We were able to associate components of extended sources without manual intervention due to the powerful inference capabilities learnt within the network. These advances make ContinUNet a promising tool for enabling science in the upcoming SKA era.

Date of Award	1 Oct 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	Mark Birkinshaw (Supervisor), Natasha Maddox (Supervisor) & Ben J Maughan (Supervisor)