Cloud Removal of Full-Disk Solar H\(\alpha \) Images Based on RPix2PixHD

Abstract

Clouds in the sky can significantly affect full-disk observations of the Sun. In cloud-covered full-disk H\(\alpha \) images, certain solar features become obscured, posing challenges for further solar research. Obtaining both cloud-covered and corresponding cloud-free images is often challenging, resulting in poor alignment of image pairs in the dataset, which adversely affects the performance of cloud removal models. We use RPix2PixHD, a novel network designed to translate cloud-covered images into cloud-free ones while mitigating the effects of misaligned data on the model. RPix2PixHD comprises two main components, Pix2PixHD and RegNet. Pix2PixHD includes a multiresolution generator and a multiscale discriminator. The generator takes cloud-covered images as input to produce cloud-free images. RegNet computes a deformation field using the generated cloud-free images and the ground truth cloud-free images. This deformation field is then used to resample the generated cloud-free images, resulting in registered images. The correction loss is calculated based on these registered images and utilized for training the generator, thereby enhancing the model’s cloud removal effectiveness. We conducted cloud removal experiments on full-disk H\(\alpha \) images obtained from the Huairou Solar Observing Station (HSOS). The experimental results demonstrate that RPix2PixHD effectively removes clouds from cloud-covered solar H\(\alpha \) images, successfully restoring solar feature details and outperforming comparative methods.