Remote Sensing Cloud Detection Algorithm Based On Unet

With the rapid development of remote sensing technology,the resolution of remote sensing image and the number of bands have been continuously improved,but the cloud detection task in remote sensing image still faces many challenges.The time of remote sensing satellite shooting,Angle,height,cloud thickness and ground object background all have an impact on the cloud detection results of remote sensing image.At present,remote sensing cloud detection is widely used in agriculture,urban management,military exercises and many other fields.Therefore,the research on remote sensing cloud detection algorithm has a very high practical significance.At present,semantic segmentation method based on deep learning has been used in remote sensing cloud detection.As a semantic segmentation method of CNN architecture,UNET is simple in construction and can obtain features of different scales to a certain extent.However,the initial model encoder is composed of convolutional stack,which cannot extract the deep information of remote sensing images and has poor detection effect on thin clouds.In addition,when responding to the input of non-RGB three-channel remote sensing images,the encoder does not optimize the corresponding channel feature extraction of non-RGB three-channel remote sensing images.When extracting the remote sensing semantic information with large scale changes,UNET cannot obtain good feature extraction effect,and the detection effect of similar background clouds such as ice,snow and desert is poor.The detection effect of UNET is not good in the scene where cloud distribution is complex and spatial information quality is high.This dissertation studies the three problems of UNET in the field of remote sensing cloud detection,and proposes a new remote sensing cloud detection algorithm based on UNET.The validity of the proposed model is verified on the public38-Cloud remote sensing cloud detection data set.Major research contributions are summarized as follows:(1)To solve the problem that the existing remote sensing cloud detection algorithms cannot fully extract the deep semantic features of remote sensing images and the detection effect of thin clouds is poor,the UNET model RUNET combined with RESNET is proposed.By comparing RESNET models of different depths,the most efficient RESNET-50 was selected as the backbone network for subsequent model modifications.On the basis of UNET,Runet’s Jaccard index increased by 1.8,and F1-score increased by 1.1.It can extract deeper semantic information of remote sensing.(2)Most of the existing remote sensing cloud detection models fail to make full use of spatial characteristics,multi-channel characteristics and adaptive generation of up-sampling,leading to poor detection effect of similar background clouds such as ice,snow and desert.We proposed MCNET to use Epres Net-50 combined with CARAFE to obtain multi-scale spatial information,pay attention to inter-channel relationship information,and obtain high-quality up-sampling results adaptively.We proposed that MCNET could improve Jaccard by 5.0 and F1-score by 3.3 on the basis of Runet.And the high quality detection images are obtained.(3)As cloud distribution is complex and cloud detection requires high quality of spatial information,the detection model cannot obtain high quality of spatial information.CDMCNET combined with DCM is proposed.On the basis of MCNET,DCM module is added to obtain adaptive information of different scales,and Coord Conv is added to add richer spatial scale information for the input image.On the basis of MCNET,the Jaccard and F1-score of CDMCNET are increased by 0.2 and 0.1,and better results are obtained.