An Unsupervised Learning approach to classifying Extended Radio Sources from deep MeerKAT images
2021-10-25, 09:15–09:30, Grand Ballroom

In this talk I present our Machine-learning based approach to analysing image-domain data from the MeerKAT telescope. Images from sensitive radio telescopes such as MeerKAT reveal unforeseen features even in familiar sources; while at the same time the spatial density of sources in these images makes manual classification and analysis unfeasible. I demonstrate that Autoencoders are in particular well-suited for this task as they should isolate key features of different morphologies while not overemphasising the source-to-source variations very common in sensitive radio images. This may also help remove potential biases and lack of repeatability in manual classification. We look at some variations of the autoencoder, including variational autoencoders and the use of pyramidal spatial pooling.

I will show some results from our use of these algorithms on images from the MeerKAT Galaxy Cluster Legacy Survey (MGCLS).

Going forward these algorithms can also be used on future and larger surveys. The number of sources likely to be observed in future surveys, especially with new and upcoming arrays such as the SKA, will far exceed the capacity for manual classification and an automated approach will be necessary.


Understanding and improving machine learning, Big data: How to deal with the 5 Vs (volume, velocity, variety, veracity, value)