Research On Cloud Detection Algorithm Of Remote Sensing Image Based On Unet Neural Networ

Remote sensing images often have the problem of cloud occlusion,which leads to the reduction of image utilization and affects the application effect in various industries.Meanwhile,the existence of clouds seriously affects the accuracy of quantitative remote sensing inversion,and can affect the global climate by changing the balance of radiation income and expenditure,and can also affect the atmospheric environment through photochemical reactions,which is of great significance for cloud detection.The features of clouds are complex and diverse,especially the cloud boundaries usually exist in the form of filamentary small fragments,which affects the effectiveness of cloud detection.Deep learning methods can mine the deep features of images to achieve end-to-end cloud detection with high accuracy,but a certain number of high-quality training samples are required.However,due to the limitation of lighting conditions,the sensors in visible and near-infrared bands on board cannot obtain nighttime information,and the current common way of visual interpretation and labeling is not applicable to nighttime,which makes nighttime cloud detection difficult,but cloud detection at night is indispensable for meteorological research.To address the above issues,the main research work of this paper is as follows.(1)A high-quality dataset construction method is proposed.Firstly,the thermal infrared image is formed by selecting the thermal infrared band whose central wavelength is larger than 3.9 microns that is not affected by solar radiation in the data,and then the cloud products corresponding to this image are processed into cloud labels by the high-quality cloud label generation method proposed in this paper to complete the training dataset and the daytime test dataset construction,and the nighttime images lacking visible light and cloud products are verified quantitatively by radar data.In this paper,the radar data and the image cloud detection results are matched in time and space,and the "plurality principle" and "proximity principle" are proposed to realize the correspondence of cloud detection results between the two types of data,and realize the construction of the nighttime test dataset.(2)A cloud detection method based on the improved Swin-Unet is proposed.In this paper,the cloud detection model of improved Swin-Unet,TIR＿CDISUnet,is constructed by using not only the spectral features of thermal infrared images,but also the texture and geometric features of the images for cloud detection,and for the weak texture features,low resolution and unclear edges of thermal infrared images,the model increases the depth of the model network,enhances the network and adjusts the network structure.Performing network enhancement and adjusting the network structure to further improve the model’s feature extraction ability for thermal infrared images and enhance the detection effect.(3)Through the analysis of the experimental results,the accuracy of the test data set in the daytime is 93.49%,which achieves a better effect in the detection of image clouds in different seasons and different scenes,and the accuracy reaches 95.17% in the detection of the more difficult fragmented cloud-rich image scenes,and the accuracy of the test data set at night is 89.59%.the lack of nighttime cloud products.