Application of a conditional generative adversarial network to denoising solar observations

The Extreme Ultraviolet Imaging Spectrometer (EIS) on the Hinode spacecraft has substantially advanced our understanding of the Sun’s upper atmosphere. Unfortunately, after being in operation since 2006, the EIS detectors have become noisy, which poses a challenge to data analysis. This paper presents a Conditional Generative Adversarial Network (cGAN) tailored to address the unique noise characteristics inherent in EIS data over the mission. Generative Adversarial Networks are deep learning models that learn to generate realistic data by training a pair of networks in an adversarial process, a mechanism that makes them particularly effective at capturing complex data distributions. Our cGAN model employs a U-Net-based generator and a conditioned discriminator, and it is trained and validated on a synthetic dataset designed to simulate the noise characteristics of EIS observations. The model converges quickly and produces denoised images that closely resemble the ground truth. Application to real EIS observations produces encouraging results, with the model effectively removing noise and largely preserving the spatial and spectral features of the data. When comparing the results of Gaussian fits to the line profiles, however, we find that the model produces only a modest enhancement over the current interpolation method.