Research On The Technology Of Single Multispectral Image Cloud Detection

Remote sensing satel ite is an important means for people to detect and perceive the earth,and the satel ite images obtained by optical remote sensing satel ite can visual y reflect all kinds of information on the earth’s surface,and provide information support and guarantee for national census,urban planning,agricultural survey and environmental protection.But not all of the data collected by satel ites meets the requirements,and one of the main reasons is cloud cover.In the pre-processing stage of satel ite ground application system,it is necessary to perform the surface quality test by calculate the cloud content of each scene image.In the practical application of satel ite image,it is necessary to precisely remove the cloud-containing area of the image,so accurate detection and identification of clouds on the image has become a necessary process for optical remote sensing image processing.The characteristic of most optical remote sensing satel ites is wide photographic range and fast updating speed,but due to the limited number of satel ite bands and narrow spectral range,as well as the complex changes of cloud shape,texture and brightness and the influence of surface highlights,there are always certain problems in the automatic cloud detection of satel ite image products.Based on the above background,this paper focuses on the problem of rapid and accurate cloud detection in single remote sensing images with less spectral information.The main tasks include:1.To solve the problem of lack of high-resolution image cloud detection data set,an objectoriented cloud mask extraction method was designed and used for data set construction.This paper studies image information extraction through object-oriented technology,and constructs the cloud detection knowledge base by analyzing the characteristics of thin cloud,thick cloud,bare ground,snow and ice on the image.Multi-scale segmentation of TH-1 satel ite multispectral images with different underlying surfaces was performed on the e Cognition platform to obtain image objects.Then,the object was classified by the nearest neighbor taxonomy to obtain the initial detection results,and then the results were modified by combining with manual classification editing.Finally,the cloud detection data set was constructed according to the data set fabrication standards.2.Aiming at the problem that traditional methods are difficult to distinguish cloud and snow effectively,the semantic segmentation technology based on deep learning is studied and applied to the detection of cloud and snow in remote sensing images.The cloud detection model was obtained through training experiments of deep learning semantic segmentation network such as FCN,Seg Net and Deeplab using cloud detection data set on Tensor Flow deep learning platform.The experimental results show that the data set constructed in this paper can be used for the training of deep learning cloud detection model,which can distinguish clouds,snow,ice and bare ground in the image.3.A single multispectral image cloud detection method based on deep learning was designed,and the TH-1 satel ite multispectral image cloud detection system was designed and implemented based on Py Qt5 development platform.The system integrates such functional modules as data set making,cloud detection model training based on deep learning,cloud detection of single multispectral image and detection result display,etc.,and realizes the automation and visualization of model training and effect display on Linux system.The experiment shows that the cloud detection method based on deep learning has a faster detection speed and higher accuracy than the object-oriented method,and it has the ability to distinguish clouds and snow on the image without the influence of thin clouds.