Authors: Subhamoy Chatterjee (SwRI), Andres Munoz-Jaramillo (SwRI), Maher Dayeh (SwRI & UTSA), Hazel Bain (CIRES & NOAA SWPC), Kim Moreland (UTSA & SwRI)

Solar images in Extreme Ultraviolet (EUV) provide a window to understand the coupling of solar atmosphere to heliosphere and acts as a missing link between solar surface phenomena and the dynamics of solar corona. Having a long baseline of high quality solar EUV images can thus help decrypt the source of space weather events and strengthen the prediction models. We utilize the almost decade-long overlap of SoHO/EIT and SDO/AIA 171 Angstrom full-disc images to create a homogeneous survey. We break the images into smaller non-overlapping patches, perform alignment between input and target, build a Convolutional Neural Network (CNN) to transform low resolution EIT patches to high-resolution AIA patches. Super-resolution being an ill-posed problem, we make use of model-ensemble to estimate the uncertainty in reconstruction. We apply an approximate Bayesian ensembling approach to derive the ensemble of CNNs. We clearly depict the effect of loss function on model performance through different metrics and the improvement in reconstruction with respect to a baseline approach. We also discuss the dependence of uncertainty on training set size, and tuning parameters for ensemble generation. Lastly, we show how the uncertainty adds value when inferring on unseen data that is not well represented by the training set.