The Research Of Classification Of Hyperspectral Imagery Based On RVM

Hyperspectral imagery which can distinguish different ground objects and have tinyspectral difference by virtue of high spectral resolution is a new type of remote sensing data.Hyperspectral imagery also it doesn’t need any prior knowledge of the target spectralsignature. Thus it is very practicable in real scenes. Nowadays it becomes a hotspot in thefield of classification, and attracts many specialists and scholars’ attention. Based onanalysing of structure and characteristics of hyperspectral imagery, and applying some imageprocessing techniques, the dissertation does the following researches in order to solve thedifficulties in hyperspectral imagery classification, such as the lack of classification accuracy,the classification efficiency is not good, and easy to produce the Hughes phenomenon, do theresearch in three parts:First, based on the mathematical theory and the short of support vector machine (SVM)hyperspectral image classification algorithm, the relevance vector machine (RVM) ofhyperspectral image classification is introduced to solve the classification problem which thealgorithm is using the relevance vector machine learning theory combines with thehyperspectral image classification. This algorithm has some advantages such as the test timeis short than SVM, the number of vectors is less and the predict output follows the probabilityproperties. The experimental results show that even the RVM hyperspectral imageclassification has some advantages, but the disadvantages is also obvious including the traintime is quite long and the accuracy of classification lower than SVM.Second, according to the disadvantages of the relevance vector machine algorithm andalso combines with the probability model calculation characteristic, the variational therelevance vector(VRVM) hyperspectral image classification algorithm has been given. Thisalgorithm leads into a new distribution in traditional probability model, making the highercomplexity of the computational can be approximately divided into the sum of two simplelogarithms form. The experimental results show that VRVM hyperspectral imageclassification algorithm has the same classification accuracy with RVM and the RVs are alsoless than SVM same as the RVM, and the training time and test time are reduced.Finally, point to the disadvantages of the relevance vector machine hyperspectral imagescombine with the characteristics and the application of the kernel function, the two improved methods is introduced to solve the disadvantages:1、Using the wavelet kernel function andthe kernel principal component analysis to preprocess the Hyperspectral image data. Theexperimental results show that this algorithm is different with the traditional principalcomponent analysis at it can effectively increase the distance between the different featurescategory, and increase the ratio of the distance in the same classes and the different classes,finally using the preprocessing date to the relevance vector machine hyperspectral imageclassification, the accuracy of the classification has been increased.2、Using the waveletkernel function to replace the traditional kernel function in RVM, the new method isintroduced to improve the accuracy of the classification and reduce the train time.