Automatic Object Detection and Categorisation in Deep Astronomical Imaging Surveys Using Unsupervised Machine Learning

I present an unsupervised machine learning technique that automatically segments and labels galaxies in astronomical imaging surveys using only pixel data. Distinct from previous unsupervised machine learning approaches used in astronomy the technique uses no pre-selection or pre-filtering of target galaxy type to identify galaxies that are similar. I demonstrate the technique on the Hubble Space Telescope (HST) Frontier Fields. By training the algorithm using galaxies from one field (Abell 2744) and applying the result to another (MACS0416.1-2403), I show how the algorithm can cleanly separate early and late type galaxies without any form of pre-directed training for what an ‘early’ or ‘late’ type galaxy is. I present the results of testing the technique for generalisation and to identify its optimal configuration. I then apply the technique to the HST Cosmic Assembly Near-infrared Deep Extragalactic Legacy Survey (CANDELS) fields, creating a catalogue of 60000 labelled galaxies, grouped by their similarity. I show how the automatically identified groups contain galaxies with similar morphological (and photometric) type. I compare the catalogue to human-classifications from the Galaxy Zoo: CANDELS project. Although there is not a direct mapping, I demonstrate a good level of concordance between them. I publicly release the catalogue and a corresponding visual catalogue and galaxy similarity search facility at www.galaxyml.uk. I show how the technique can be used to identify rarer objects and present lensed galaxy candidates from the CANDELS imaging. Finally, I consider how the technique can be improved and applied to future surveys to identify transient objects.