Super-resolution of SDO/HMI Magnetogram Using Novel Deep Learning Methods.

Abstract:

Image super-resolution (SR) is a technique of enhancing the resolution of an image where a high-resolution (HR) image is reconstructed from a low-resolution (LR) image. In this study, we apply two novel deep learning models (residual attention model and progressive GAN model) for Solar Dynamics Observatory (SDO)/Helioseismic and Magnetic Imager (HMI) magnetograms. For this, we consider line-of-sight (LOS) magnetograms taken by SDO/HMI as output and their degraded ones with 4 by 4 binning as input. Deep learning networks try to find internal relationships between low-resolution and high-resolution images from the given input and the corresponding output image. We consider SDO/HMI magnetograms from 2014 May to 2014 August for training, from 2014 October to 2014 December for validation, and 2015 January to 2015 March for test. We find that the residual base model generates higher-quality results than the progressive GAN model and the bicubic interpolation in terms of visual aspects and metrics. The results of the residual base model for the test data set are as follows: 48.35 dB for peak signal-to-noise ratio (PSNR), 0.93 for correlation coefficient, 15.65 G for root mean square error (RMSE) and 0.98 for structural similarity (SSIM). We apply this model to a full-resolution SDO/HMI magnetogram and then compare the generated magnetogram with the corresponding Hinode/The Solar Optical Telescope Narrow Brand Filtergrams (NFI) magnetogram. This comparison shows that the generated magnetogram is consistent with the Hinode one with a high correlation (CC: 0.94) and a high similarity (SSIM: 0.93), which are better than the bicubic method.