I Can See Clearly Now: A Deep Learning Method to Remove Clouds from Solar Images

Even in the sunny spots where solar telescopes are stationed, clouds are a fact of life. Rejecting cloud-obscured solar images means losing a lot of data, but new machine-learning methods show that we don’t have to cast off cloudy images after all.

Contending with Clouds

Space-based solar telescopes have the luxury of perfect “weather” every day, but ground-based solar observatories have to contend with clouds. A study undertaken at Udaipur Solar Observatory in India, for example, found that the Sun’s disk was unimpeded by clouds only 63% of the time. If clouds can’t be removed reliably from images taken the remaining 37% of the time, that means losing a ton of information about the Sun. Is there a way to salvage these images?

So far, existing methods of handling cloud-contaminated images fall short, either failing to remove the clouds fully or failing to reconstruct the obscured solar features underneath. Now, a study led by Zhenhong Shang (Kunming University of Science and Technology) has shown how a new machine learning method rescues clouded-out images of the Sun’s disk.

Machine Learning Method

The model developed by Shang’s team handles the two critical aspects of cloud removal: 1) identifying and characterizing cloudy images and 2) removing the clouds and recovering the features underneath.

First, the team’s method reports not only whether an image is cloudy but also how cloudy it is. This is an improvement over previous methods that returned only a yes or no to the question of cloud cover. Previous methods that searched only for variations in brightness across the solar disk could also be fooled by widespread cloud cover, whereas the new method also takes in information about the median brightness of the disk, identifying images that are nearly completely cloud covered.

For the second step of the procedure, Shang’s team used data from the Global Oscillation Network Group (GONG), six solar telescopes that maintain eyes on the Sun every hour of every day. The team selected mildly, moderately, and severely cloud-covered images from this dataset. To train the machine learning model to remove clouds, they used 760 pairs of clear and cloudy images from the GONG telescopes. When cloud cover is intermittent, these clear–cloudy pairs are constructed from images taken by the same telescope. When cloud cover is persistent, a cloudy image from one observatory is paired with a clear image from another observatory, taken just seconds later. This dataset was also augmented by resizing and rotating some of the images, making sure it can handle even more cloud-cover scenarios.

Clear Skies Ahead

After testing and validating their method against a separate set of GONG images, Shang’s team compared the model’s performance to that of existing methods. On nearly every quantitative and qualitative metric, the new machine learning method outperformed existing methods. The new method showed particular improvements in recovering details of the solar surface under dense cloud cover.

The team also demonstrated the method’s performance on data from other ground-based solar observatories, showing its generalizability and readiness to tackle the challenge of cloud removal.

Citation

“Cloud Removal in Full-Disk Solar Images Using Deep Learning,” Zhenhong Shang et al 2025 ApJS 276 56. doi:10.3847/1538-4365/ad93ca