K-CloudGAN : Weakly Supervised Cloud Removal towards Automated Satellite Surveillance

Wu, Xiaofan and Li, Changtsun and Atkinson, Peter M. and Zhang, Ce and Angelov, Plamen P. and Jiang, Richard (2025) K-CloudGAN : Weakly Supervised Cloud Removal towards Automated Satellite Surveillance. In: IGARSS 2025 - 2025 IEEE International Geoscience and Remote Sensing Symposium :. IGARSS 2025 - 2025 IEEE International Geoscience and Remote Sensing Symposium . IEEE, pp. 7390-7393. ISBN 9798331508111

Text (Cloud_Removal)
Cloud_Removal.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (8MB)

Abstract

Cloud occlusion significantly impairs the quality and usability of remote sensing imagery. This work tackles the problem in two stages: cloud segmentation and cloud removal. For segmentation, we employ transfer learning with three advanced models—U-Net, DeepLabV3+, and PSPNet—achieving approximately 90% accuracy on real-world datasets, thereby demonstrating the effectiveness of transfer learning for this task. For cloud removal, we introduce K-CloudGAN, a weakly supervised inpainting framework that integrates K-means-based pseudo-labeling with SN-PatchGAN and transfer learning. This design reduces reliance on large-scale labeled datasets while enhancing restoration quality. Compared to existing approaches like CloudGAN, our method achieves a 10% improvement in terms of SSIM and PSNR metrics, confirming its superior performance in restoring cloud-obscured satellite imagery.