Deep Regression for Imaging Solar Magnetograms Using Pyramid Generative Adversarial Networks

Rasha Alshehhi; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 204-205

Monitoring a large active region in the farside of the Sun is important for space weather forecasting. However, direct imaging of the farside is currently not available and usually physicists rely on seismic holography to infer farside magnetograms. On other hand, mapping between holography and magnetic images is non-trivial. In this work, Generative Adversarial Network (GAN) is used; which consists of a pyramid of modified pixel2pixel architectures to capture internal distributions at different scales with higher quality. Generative model is trained and evaluated using frontside of Solar Dynamic Observatory (SDO): Atmospheric Imaging Assembly (AIA) and Helioseismic and Magnetic Imager (HMI) magnetograms. Farside solar magnetograms from Extreme UltraViolet Imager (EUVI) farside data is also generated. The generative model successfully generates frontside solar magnetograms and outperforms state-of-the art method. It also help to monitor the magnetic changes from farside to frontside using generated solar magnetograms.