Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Removing cloud shadows from ground-based solar imagery

Amal Chaoui, Jay Paul Morgan, Adeline Paiement, Jean Aboudarham

TL;DR

This work tackles cloud-shadow removal in ground-based solar imagery using a U-Net backbone trained under fully supervised and conditional GAN frameworks. It investigates three output strategies (direct cleaning, shadow-mask division, and residual-mask subtraction) and finds the residual-mask subtraction to be the most stable and effective. The approach is evaluated on Ca-II and H-alpha modalities with real and synthetically generated cloud textures, achieving state-of-the-art performance against non-learning baselines across PSNR, SSIM, RMSE, and FID metrics, while showing generalization to streaked clouds. The findings offer practical guidance on selecting training setups and highlight directions for future improvements in filament restoration and downstream analyses.

Abstract

The study and prediction of space weather entails the analysis of solar images showing structures of the Sun's atmosphere. When imaged from the Earth's ground, images may be polluted by terrestrial clouds which hinder the detection of solar structures. We propose a new method to remove cloud shadows, based on a U-Net architecture, and compare classical supervision with conditional GAN. We evaluate our method on two different imaging modalities, using both real images and a new dataset of synthetic clouds. Quantitative assessments are obtained through image quality indices (RMSE, PSNR, SSIM, and FID). We demonstrate improved results with regards to the traditional cloud removal technique and a sparse coding baseline, on different cloud types and textures.

Removing cloud shadows from ground-based solar imagery

TL;DR

This work tackles cloud-shadow removal in ground-based solar imagery using a U-Net backbone trained under fully supervised and conditional GAN frameworks. It investigates three output strategies (direct cleaning, shadow-mask division, and residual-mask subtraction) and finds the residual-mask subtraction to be the most stable and effective. The approach is evaluated on Ca-II and H-alpha modalities with real and synthetically generated cloud textures, achieving state-of-the-art performance against non-learning baselines across PSNR, SSIM, RMSE, and FID metrics, while showing generalization to streaked clouds. The findings offer practical guidance on selecting training setups and highlight directions for future improvements in filament restoration and downstream analyses.

Abstract

The study and prediction of space weather entails the analysis of solar images showing structures of the Sun's atmosphere. When imaged from the Earth's ground, images may be polluted by terrestrial clouds which hinder the detection of solar structures. We propose a new method to remove cloud shadows, based on a U-Net architecture, and compare classical supervision with conditional GAN. We evaluate our method on two different imaging modalities, using both real images and a new dataset of synthetic clouds. Quantitative assessments are obtained through image quality indices (RMSE, PSNR, SSIM, and FID). We demonstrate improved results with regards to the traditional cloud removal technique and a sparse coding baseline, on different cloud types and textures.
Paper Structure (7 sections, 5 equations, 5 figures, 2 tables)

This paper contains 7 sections, 5 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction and previous works
  2. Data
  3. Methodology
  4. Results
  5. Choice of DNN output
  6. Comparative evaluation of the methods
  7. Conclusion

Figures (5)

  • Figure 1: Overview of cloud removal from solar imagery. A cloud mask is predicted and subtracted from the original image, to prepare for a subsequent detection of solar structures (here ARs in red boxes). Our proposed method only deals with the cloud removal stage.
  • Figure 2: Examples of real (top) and synthetic (bottom) cloud contamination in Ca-II images.
  • Figure 3: Loss curves when training the proposed models using Eqs. \ref{['eq:division']} and \ref{['eq:subtraction']}.
  • Figure 4: Comparative illustration of the three DNN outputs on some Ca-II images with synthetic clouds. We display the results of a random run of each network, since all runs produced roughly the same results.
  • Figure 5: Examples of removal of real and synthetic clouds from Ca-II and H-$\alpha$ images by compared algorithms. Images with real clouds (Real Ca-II, Real H-$\alpha$, Ca-II AR, H-$\alpha$ Filament) do not have a clean target counterpart, as explained in Section \ref{['sec:data']}. Therefore, for these columns, the target row is left blank.