Generating Space-based SDO/HMI-like Solar Magnetograms from Ground-based Hα Images by Deep Learning

Abstract

Recently, the method of estimating magnetic field through monochromatic images by deep learning has been proposed, demonstrating good morphological similarity but somewhat poor magnetic polarity consistency relative to real observation. In this paper, we propose to estimate magnetic field from Hα images by using a conditional generative adversarial network (cGAN) as the basic framework. The Hα images from the Global Oscillation Network Group are used as the inputs and the line-of-sight magnetograms of the Helioseismic Magnetic Imager (HMI) are used as the targets. First, we train a cGAN model (Model A) with shuffling training data. However, the estimated magnetic polarities are not very consistent with real observations. Second, to improve the accuracy of estimated magnetic polarities, we train a cGAN model (Model B) with the chronological Hα and HMI images, which can implicitly exploit the magnetic polarity constraint of time-series observation to generate more accurate magnetic polarities. We compare the generated magnetograms with the target HMI magnetograms to evaluate the two models. It can be observed that Model B has better magnetic polarity consistency than Model A. To quantitatively measure this consistency, we propose a new metric called pixel-to-pixel polarity accuracy (PPA). With respect to PPA, Model B is superior to Model A. This work gives us an insight that the time-series constraint can be implicitly exploited through organizing training data chronologically, and this conclusion also can be applied to other similar tasks related to time-series data.