Abstract:
Modern and future radio surveys performed with increasingly powerful instruments, such as the 64-antenna MeerKAT interfereometer and eventually the Square Kilometre Array (SKA), will catalogue upwards of hundreds of thousands to millions of radio sources. This can make classification of source morphology and searching for specific source classes extremely challenging. MeerKAT excels at imaging large-scale and faint emission features due to its high sensitivity and excellent imaging quality, allowing for many exotic, scientifically rich radio objects to be identified for the first time. However, finding them is a problem, especially using manual classification. Moreover, MeerKAT’s moderate angular resolution (~ 5 arcsec) means that a typical field is crowded with many sources, including many point-like sources. An automated approach to classification is therefore required. The aim of this project is to isolate the most morphologically unusual or exotic sources. The approach explored in this project is the use of autoencoders, neural networks that encode an input into some latent space and then attempt to reconstruct the input from the code form. We test this on the MeerKAT Galaxy Cluster Legacy Survey, comprising of 115 galaxy clusters at 1.28 GHz with µJy/beam sensitivity. A subset of these are manually classified and used to train numerous configurations of autoencoder algorithms, including ensembles of autoencoders, and test the algorithms’ performance in isolating potentially interesting sources. It is found that the autoencoders significantly reduce the work required to locate potentially interesting sources.
