Research On Cloud Detection And Removal Method Of Optical Remote Sensing Image Based On Deep Learning

Optical satellite remote sensing images have the characteristics of strong Optical satellite remote sensing images have strong intuitiveness and rich surface information.With the improvement of imaging resolution of optical remote sensing satellites,their applications are becoming more and more extensive in disaster prevention and mitigation,resource survey,agricultural investigation,and military fields.However,due to imaging mechanism limitations,optical remote sensing images are easily disturbed by clouds.According to research data published by the International Satellite Cloud Climatology Project(ISCCP),more than 60% of the earth’s surface is covered by clouds on average each year,resulting in widespread cloud coverage in optical remote sensing images.Thick clouds block surface information,forming shadow areas on the ground;thin clouds weaken surface radiation,causing image distortion and blurring.Clouds destroy the overall consistency of remote sensing images and severely limit their use.Therefore,how to effectively remove cloud interference in remote sensing images,restore the original true features of remote sensing images,and enhance their usability is currently a research hotspot in photogrammetry and remote sensing fields.This study focuses on the existing problems of insufficient detection,incomplete removal,loss of surface information,large color differences,and difficulty in processing thin clouds,thick clouds,and shadows simultaneously in existing cloud detection and removal methods.Based on the sufficient detection of clouds and shadows in remote sensing images,and following the research idea of "detectionfusion-reconstruction",this study proposes a remote sensing image cloud detection method based on multi-scale feature alignment fusion network(MSFAF-Net)and a cloud removal method based on image fusion-feature learning reconstruction(IF-FLR)using deep learning.Cloud detection is achieved by constructing a multi-scale feature alignment fusion network,and cloud removal is realized by constructing a multi-scale feature fusion and reconstruction network(MSFFIR-GAN).Taking Sentinel-2 satellite images as an example and comparing with the weighted linear regression(WLR)method,the results show that our proposed method can effectively remove clouds and shadows in remote sensing images,restore the original surface information features of cloud areas,and ensure the overall and local consistency of cloud removal results.The main work is as follows:(1)Cloud detection in optical remote sensing images.This paper designs the MSFAF-Net cloud detection network,which uses multiple feature extraction modules to obtain feature information at different depths of the image.The network also uses the atrous spatial pyramid pooling(ASPP)to obtain feature information at different scales of the image.Additionally,the network uses the feature alignment structure to align the features of different levels,correct the deviations between feature maps,and alleviate the problem of feature mismatch,which results in higher detection accuracy and spatial resolution.Compared with UNet,FCN8 s,and Deeplab V3 networks,and using Landsat8 38-cloud,GF1-WHU,and Sentinel-2 Cloud Mask datasets.The average intersection over union(MIo U)of our proposed method is superior to the comparison networks,including UNet,FCN8 s,and Deeplab V3.For the Landsat8 38-cloud,GF1-WHU,and Sentinel-2 Cloud Mask datasets,our method improves the MIo U by 0.92%,2.61%,and 2.80%;1.34%,1.08%,and 0.86%;and 9.45%,4.08%,and 5.32%,respectively.The effectiveness and wider applicability of MSFAF-Net for cloud detection in optical remote sensing images were verified.(2)Cloud Removal of Optical Remote Sensing Images.Following the research approach of "detection-fusion-reconstruction," the IF-FLR cloud removal method is proposed.Based on the cloud mask obtained through the MSFAF-Net network,the problem of cloud and shadow occlusion of ground objects is addressed by fusing with a reference image to supplement the missing surface information.Furthermore,based on the generative adversarial network,the MSFFIR-GAN image reconstruction network is designed,which uses multiple feature extraction modules in the generator to obtain image reconstruction features at different levels.The up-sampling feature fusion module is used to integrate features of different depths step by step,to reconstruct the input image.A hierarchical Markov discriminator is designed to make the reconstructed image have clear details and texture information,which makes the reconstructed results closer to the target image.(3)Feasibility verification and analysis of the method.Two sets of Sentinel-2images were selected as the research objects.Based on the IF-FLR cloud removal method,a complete experiment was designed and carried out from cloud detection to cloud removal.The reconstruction results were comprehensively evaluated from the aspects of visual effect,spectral characteristics,and reconstruction quality.The model was used to reconstruct the fused image,achieving the removal of clouds in remote sensing images.The experiment results show that the IF-FLR cloud removal method can effectively remove clouds under various land cover scenes,while maintaining the overall consistency of the cloud removal result image,to eliminate or weaken the impact of clouds on remote sensing images and restore the original surface information features.In summary,this paper provides a complete solution from detection to removal,and experimental results demonstrate that the proposed approach is effective in removing thin and thick clouds as well as shadows in remote sensing images while preserving the original surface information features of the cloud-covered regions.This method improves the usability of remote sensing images and provides a valuable reference for cloud detection and removal in remote sensing applications.