Unsupervised remote sensing image thin cloud removal method based on contrastive learning

• DOI: 10.1049/ipr2.13067
• Publication Date: 2024-03-11T04:01:15Z
• Authors: Zhan Cong Tan; Xiao Feng Du; Wang Man; Xiao Zhu Xie; Gui Song Wang; Qin Nie
• Keywords: Discriminator

Abstract Cloud removal algorithm is a crucial step of remote sensing image preprocessing. The current mainstream cloud algorithms are implemented based on deep learning, and most them supervised. A large number data pairs required for training to achieve removal. However, real with/without datasets difficult obtain in the world, models obtained by synthetic often need generalize better natural scenes. And existing unsupervised thin methods Cycle‐GAN framework with considerable model complexity unstable not an excellent solution problem lack paired datasets. Based this, this paper, authors propose method contrastive learning—GAN‐UD. It network consisting frequency‐spatial attention generator discriminator. In addition, introduce local loss global content constrain generated images ensure that cloud‐free consistent input terms content. Experimental results show proposed paper can still effectively remove clouds from without datasets, outperforms methods, achieves comparable performance supervised methods.