Research On Classification Of Hyperspectral Images Based On Support Vector Machine

Hyperspectral remote sensing has been become the front line of remote sensing area, and plays an important role in fields such as military and civil field. High spectral resolution, many bands, narrow band width and larg data amount of hyperspectral image not only bring great research value for human beings, but also bring great challenge for processing them. Methods of multispectral imagery processing are no longer applicable to hyperspectral imagery. It's urgent to find rapid and accurate methods for discovering interested information from the huge data produced by hyperspectral sensors.Hyperspectral remote sensing image classification is one of the key technologies in remote sensing applications field, and its high speed as well as high accuracy algorithm is the precondition of practical applications. Traditional pattern classification methods are based on the principle of experiential risk minimization, and they can achieve the best result only when the number of samples approaches infinity. Unfortunately, in hyperspectral image classification, training samples are usually limited. According to classification features of hyperspectral remote sensing image, this dissertation takes deep study on support vector machine (SVM) and its applications in hyperspectral remote sensing image classification by means of the good generalization of support vector machines in small samples, nonlinearity and high dimension space. Main contributions of this thesis are given as follows:1. It introduces the characteristics of HSI, the development of imaging spectroscopy, as well as of the-state-of-the-art of improving HSI resolution, and explains the background and application value of this research.2. It analyzes the classification methods of both traditional supervised of hyperspectral images and support vector machines. The simulation experiments result of hyperspectral imagery classification shows their classification performance, points out the disadvantages of traditional classification methods while applying to hyper spectral data, and presents the unique advantages of the SVM.3. By deep study on SVM, it analyzes that classification value is deviated by introducing the kernel function into nonlinear SVM. The optimal setting of classification threshold is proposed. This assay proposes a new threshold before classification, and uses a new threshold for classification. The simulation experiment results show that this proposed method can efficiently improve classification accuracy of nonlinear SVM.4. Training algorithm for large-scale support vector machine is an important and active subject in the field of SVM research. After analyzing the nature and difficulties in training SVM, a new reduction strategy is proposed in this paper for training SVM with large-scale sample set with purpose of overcoming its slow training speed. In this strategy, the samples corresponding to non support vectors are solved and removed before training SVM. This method is fast in convergence without classification accuracy loss. The simulation experimental results prove the feasibility and effectiveness of this method.