Using Machine Learning Techniques to Detect, Classify and Separate Strong Gravitational Lensing Systems from Astronomical Images

Davies, Andrew (2022). Using Machine Learning Techniques to Detect, Classify and Separate Strong Gravitational Lensing Systems from Astronomical Images. PhD thesis The Open University.



This thesis introduces novel methods to detect or discover strong gravitational lenses in future large surveys, with a particular focus on machine learning techniques. Strong Gravitational lenses are rare objects and difficult to recognise, with only a few thou- sand being detected and confirmed. Purpose built machine learning architectures have already been shown to produce state-of-the-art results for a diverse set of problems, including successes in astronomy. For certain simple to define tasks, computational methods are able to learn patterns and features to improve their ability to perform such tasks. A Convolutional Neural Network (CNN) that can be used to identify images con- taining lensing systems has been developed with the aim to achieve the accuracy and speed required to detect gravitational lenses in images. CNNs have already been used in astronomy for star-galaxy classification. The CNN in this thesis has been trained and tested on Euclid-like and KiDS-like simulations from the Euclid Strong Lensing Group, successfully classifying 77% of lenses, with an area under the Receiver Operating Char- acteristic (ROC) curve of up to 0.96. The CNN also classifies lenses in COSMOS HST F814W-band images. After convolution to the Euclid resolution, the CNN can recover most systems that are identifiable by eye. Computational methods are only a feasible solution if they are sufficiently reliable and accurate. A Zooniverse project, Euclid - Challenge the machines, was set up to determine the relative strengths of using Convo- lutional Neural Networks and human volunteers at classifying images, and to determine whether or not computational methods could be used to find strong lensing candidates. This work showed that human volunteers using visual inspection are less successful at classifying images containing strong gravitational lenses than a trained CNN. How- ever by having multiple classifications for each image, human inspection can produce a much more precise (or purer) sample of lenses than the trained CNNs. A method for deblending multiple objects in images using their relative flux ratios and then using Gaussian Mixture Models to classify as distinct objects has been developed with a view to future improvements on lens detection and characterisation. This method can be used to separate multiple objects in an image, and has been shown to pick out new objects where labels are missing or to classify previously distinct objects as a single object in the Abell 2744 catalogue.

Viewing alternatives

Download history


Public Attention

Altmetrics from Altmetric

Number of Citations

Citations from Dimensions

Item Actions