Research On Cloud Removal Of Remote Sensing Images Based On Generative Adversarial Network

Remote sensing technology plays an important role in land resource exploration,military defense,and post-disaster reconstruction.However,satellite remote sensing images are susceptible to cloud interference,which reduces the availability of remote sensing images,the traditional cloud removal method only has better results in the task of removing evenly distributed clouds;Generative Adversarial Network is a deep learning model,which has been well applied in removing clouds and fog from remote sensing images.However,the features extracted are single,ignoring the influence of the feature spatial relationship on the image based on the CGAN,resulting in the processed remote sensing image being blurred and cloud shadow remaining.In order to solve the problems,this thesis uses the group normalized SE-Res Block and the group normalized SEDense Block to improve the generator in the Generative Adversarial Network,And used for cloud removal of remote sensing image.The content is summarized as follows:(1)The method that uses group normalization SE-Res Block to improve the residual convolution block,uses the U-Net structure and skips connection to design a new generator.Markov discriminator is used as the discriminator in the SE-Res Gan.Experiments have proved that the SE-Res Gan can remove cloud information in remote sensing images,and it has improved PSNR and SSIM compared with comparison methods.(2)The method that uses group normalized SE-Dense Block to improve the Generative Adversarial Network.This method uses group normalization SE-Dense Block to improve the dense convolution block,uses the U-Net structure to design a new generator,and uses Markov discriminator as the discriminator in the SE-Dense Gan.Experiments have proved that the SE-Dense Gan has a good effect in cloud removal of remote sensing image.Compared with the SE-Res Gan,the SE-Dense Gan model has a certain increase in parameters,and the processed remote sensing image has more ground texture details.In general,this thesis focuses on the task of Generative Adversarial Network to cloud removal of remote sensing image.Corresponding experimental results prove that the improved Generative Adversarial Network has a perfect effect in cloud removal of remote sensing image,and has broad application prospects in this field.