Thin And Thick Cloud Removal Methods For Optical Remote Sensing Imagery Based On Generative Adversarial Networks

High-resolution optical remote sensing images are widely used in important fields such as land and resource investigation,geological resource exploration,environmental monitoring and military reconnaissance.However,the presence of clouds causes some information of images to be contaminated,which seriously affects the imaging quality of image data and makes the data utilization rate decrease,limiting its application potential.According to the thickness of the cloud layer and whether the feature information is completely obscured,the cloud removal research can be divided into thin cloud removal and thick cloud removal.In recent years,deep learning methods have received much attention in the field of remote sensing cloud removal due to their powerful feature mapping capability,but they are still unable to effectively remove thin and thick clouds in complex texture scenes.In this paper,the relevant research problems in the field of thin cloud removal and thick cloud removal using deep learning methods are studied and discussed and analyzed.The main works and innovations of this paper are as follows:(1)For the case of non-uniformly distributed thin cloud contamination in high-resolution optical remote sensing images,this paper proposes a generative adversarial network and physical modeldriven thin cloud removal method for remote sensing.The method uses a generative adversarial network to train adversarial on thin cloud images and cloud-free images,and learns the potential nonlinear mapping relationship between thin cloud images and cloud-free images.Meanwhile,the atmospheric physical transport model is introduced into the thin cloud removal method,and the end-to-end network module is used to extract the transmittance and atmospheric light information of real clouds,so as to remove the non-uniformly distributed thin clouds from remote sensing images.After the evaluation of simulated and real experiments,this method improves the quality of thin cloud removal in recovered images compared with other thin cloud removal methods.(2)For high-resolution optical images covered by thick clouds,it is difficult for the previous methods to effectively reconstruct thick clouds in complex texture regions because the feature information is almost completely lost.In this paper,we propose a texture complexity-guided and self-stepping learning remote sensing thick cloud removal method,which uses texture complexity to guide the generation of training samples,and utilizes a self-stepping learning training strategy to train progressively according to the texture complexity of the image from easy to difficult to improve the stability and generalization performance of model training,so as to achieve the purpose of thick cloud removal in complex texture regions.Finally,we combine the Structural Similarity(SSIM)and Mean Square Error(MSE)loss functions to further improve the visual quality of the reconstructed images.The simulated and real experiments demonstrate that the method is significantly better than other methods,the reconstructed textures are more detailed,and the thick cloud occlusions in complex texture scenes can be effectively removed.