Deep learning applied to big astronomical data from SKA and its precursors

Abstract

The SKA, the completion of which is fast approaching, will be a revolutionary radio telescope, not only in terms of its sophisticated design as a telescope, but especially in terms of the ambitious scientific goals enabled by its observations. It will be made up of two complementary interferometers in different continents, with over 130,000 telescopes in Australia and almost 200 in South Africa, which is expected to produce images with unprecedented resolution, sensitivity, and field of view. This is expected to result in the generation of hundreds of petabytes of processed data annually. Currently available solutions for processing this data, most of which rely on manual inspection by trained astronomers, would be intractable for the volume of the data expected. This has necessitated the development of automated tools which can handle these tasks, enabling astronomers to develop knowledge from the information extracted by these tools, rather than raw data from telescopes. This thesis focuses on the application of various state-of-the-art methods to develop automated solutions for two select problems of relevance in astronomical instrument data. The first of these is the classification of gravitational lensing, for which a framework for dealing with classification tasks was developed. This framework contains functionality for loading and pre-processing astronomical data, as well as several CNN-based models for performing the classification task. This tool not only obtains very good results for the dataset used, improving upon the previously reported best performances in other works, but is also very adaptable to similar problems, where an object inside an image needs to be labelled as belonging to one of a number of classes. The second problem is source finding applied to radio astronomy images, for which an Instance Segmentation solution was developed. This solution, based on Mask R-CNN, is capable of detecting, classifying, and producing a pixel-accurate mask for imaging sidelobes, point sources, and radio galaxies. While the performance achieved on sidelobes was less than desired, that for sources and galaxies was satisfactory. No previously existing solutions perform instance segmentation on astronomical data, so it is difficult to have a fair and direct comparison with other solutions. In fact, the dataset for training and evaluating this model was also collected and labelled as part of this work, and will eventually be published. Again, this tool can be very easily adapted to other astronomical problems of a similar nature.