Cloud Detection Algorithm Research For Optical Satellite Lmage

The optical satellite image has advantages of high-resolution, rich information etc.,and is used widely. A major problem inherent is that, the possibly existing clouds mayblock ground objects and even make the image data of no use. This not only increasesthe expenses and time of image transmission and processing, but also greatly reducesthe use of optical satellite images. Thus cloud detection becomes one of the primarytasks and one key technology of optical satellite image processing. In practice, it isdifficult to character the physical and imagery properties of natural clouds, which makescloud detection a major difficulty in applications. In this thesis, we focus on solving thisproblem by using the popular machine learning method, support vector machine (SVM),and the cartoon texture decomposition. The results have important theoreticalsignificance and application value.In this thesis, we first summarize the present development of optical satellite imageand cloud detection. Then we describe the support vector machine (SVM) learningmethod and cartoon-texture decomposition of images. Analysis of cartoon-texturedecomposition results of some optical satellite images show that, clouds of large scalescan be effectively separated in the cartoon component, while other small-scale details inthe image can be stored in the texture component. Therefore, the cartoon-texturedecomposition of the optical satellite image can effectively eliminate the interference ofsmall-scale details on cloud detection. Based on this, we proposed a new clouddetection algorithm incorporating SVM and cartoon-texture decomposition. First, theimage to be tested is decomposed into the structure component and the texturecomponent. Then cloud is detected in the structure component. Finally, under the guideof this preliminary detection result, a further cloud detection is conducted on theoriginal image by using SVM. Experimental results show that this method caneffectively improve the detection rate and reduce the false alarming rate.