Thick Cloud And Shadow Removal In Multi-temporal Remote Sensing Images Based On Generative Adversarial Network

Remote sensing images are often easily disturbed by the atmospheric environment during the imaging process,resulting in information loss in the areas occluded by thick clouds.Remote sensing data is playing an increasingly important role in industrial applications in my country,and cloud pollution inevitably leads to information loss,which greatly reduces imaging quality,and causes problems for the subsequent applications of remote sensing images in various industries.Therefore,removing remote sensing images with many cloud layers through cloud detection and removing thick clouds from remote sensing images with few cloud layers and still available ground object information are the basic prerequisites for the application of remote sensing images in many fields.The research content of this thesis is to solve the problem of thick cloud cover in remote sensing images,that is,the location of clouds and shadows can be obtained through cloud detection,and the multi-temporal complementary information is used to remove them.Cloud detection is essentially a classification problem,classifying clouds and their shadows into one category to be detected,and other types of objects into one category.Thick cloud removal is essentially the process of missing information reconstruction.The methods of reconstructing missing information in remote sensing images are mainly divided into two types of reconstruction methods based on multi-temporal and multi-spectral.Due to the increasing maturity of deep learning technology,the field of remote sensing image processing has new directions and more advanced results,and its main representative algorithms are mainly based on convolutional neural networks（CNN）and generative adversarial networks（GAN）.The introduction of deep learning into information reconstruction has brought breakthroughs to the research,development,and innovation of remote sensing image cloud detection and thick cloud removal.The main work and innovation of the thesis are:（1）Before removing thick clouds,it is necessary to obtain cloud masks through the cloud detection algorithm.For ordinary convolutional networks,it is easy to misclassify cloud shadows,waterbody information,and mountain information,as well as poor detection of thin clouds and broken clouds.Based on the various advantages of U-Net,a GAM-DUnet cloud detection algorithm is proposed.This method uses U-Net as the main framework,adds dilated convolution to expand the receptive field,and then adds the global attention mechanism GAM to amplify the global cross-dimensional interaction,which can capture more information of different scales and dimensions,reduce information dispersion,and reduce computational complexity and improve model performance.Cloud detection experiments are carried out using different cloud coverage scenes,which verifies the effectiveness of the algorithm for detecting clouds and their shadows in remote sensing images.（2）The lack of spectral information caused by thick clouds and their shadows exists in almost all bands,which increases the difficulty of removing thick clouds based on multispectral methods.Aiming at this problem,this thesis combines a generative adversarial network based on a multi-temporal method to reconstruct the thick clouds and their shadow missing areas in remote sensing images.This method can reduce information loss,improve output resolution,and facilitate the generation of more globally consistent and locally consistent reconstruction results.Using different remote sensing image datasets to conduct simulation experiments and real experiments,it is verified that the method can effectively remove thick clouds and their shadows in remote sensing images.