Solar Feature Recognition with Deep Learning

Background

Many of the identifiable features on the Sun are the result of the magnetic fields that permeate the solar atmosphere. Magnetic flux across the solar surface manifests at varying scales both spatially and temporally. Many of these magnetic features can be highly dynamic and couple between layers of the solar atmosphere, channelling energy to higher regions. These features can be difficult to categorise due to their constantly evolving structures. Likewise, various forms of non-magnetic brightenings can manifest in ways similar to magnetic features. This can mean it is challenging to disentangle the various features present in an image, as well as identifying the source of those features (e.g. magnetic fields, vortex flows, transient phenomena).

Project Description

The sheer volume of imaging data available of the Sun makes it a challenge to study small-scale features of interest across the solar atmosphere. Feature detection algorithms exist but are not optimised for the volume of data currently available, especially with regards to newer facilities such as the 4-m Daniel K. Inouye Solar Telescope (DKIST).  Furthermore, feature detection algorithms are often limited to a specific type of feature and are therefore highly constrained and not adaptable. Deep learning techniques such as Convolutional Neural Networks are a possible solution to these issues in object recognition and classification. This project will look to build a deep learning application to identify and track features in solar imaging data at multiple wavelengths. The project will work with data from DKIST as well as other facilities to classify images. The outputs of the image classification model will be utilised to analyse the physical properties of the various classified objects for a better understanding of the complex evolutionary processes displayed at varying spatial scales and heights in the solar atmosphere.

For more information contact Dr Peter Keys (p.keys@qub.ac.uk)