Research On Sample Reducation And Combination Of Spatial Information For SVM In Remote Sensing Image Classification

Hyperspectral image analyzing provides the theoretic foundation to the application of hyperpectral image. Based on statistic learning, it uses structure risk minimization instead of experience risk minimization which is widely used in traditional classification. Because of this, the support vector machine has good classification performance under limited samples,non-linear and high dimensional space. However, the relatively low optimizing efficiency restricts its development in realistic world, especially in the circumstances with the high real-time requirement. The Least Squares Support Vector Machine (LS-SVM), modifying the standard support vector machine, uses equality constrains instead of inequality constrains and dramatically reduces the problem solving complexity. But it loses the sparse at the same time which means that the support vector set in the solution contains all the samples. This sacrifices the speed when predicting new samples. The traditional classifier also ignores the spatial relationship between targets and only uses spectral data for training and classification which causes insufficient information extracting.Motivated by all the previous problems, this thesis studies hyperspectral image classification deeply and modifies the training procedure of support vector machine to reduce the classification time and maintain comparable accuracy at the same time. The original classification result is reconsidered based on the information about spatial correlation and it is proved that this can increase the classification accuracy. The main contributions are listed as follows.The research background of hyperspectral image and its value of application are introduced firstly . The development of spectral imaging, the characteristics of hyperspectral data, and previous works about hypersectral image analyzing are also summarized.Next, the simulation results of hypersectral data classification show that the classification method based on support vector machine processes distinct advantages in hypersectral image classification comparing with traditional classification methods. This is obtained by deep understanding of traditional hypersectral image surveillance classification and support vector machine classification method.After analyzing the fundamental of support vector machine, the advantage and disadvantage of LS-SVM and considering the sparse loosing of LS-SVM, this thesis proposes a sample reducing strategy based on Coulomb Force which improves the classification efficiency and maintains or increases classification accuracy at the same time. Experimental results show that this strategy greatly reducing the classification time.Finally, the Markov random field is introduced and this thesis proposes a new classifier model which mixes the spatial information and spectral information. The proposed new model is based on the sample reducing strategy and Markov random field model. This classifier gets the probabilistic classification result from the sample reducing LM-SVM model firstly, and then obtains the probability which represents the spatial relationship from the Markov random field model, and finally calculates the ultimate classification result by taking the product of these two probabilities. It can be seen from the experimental results that the proposed classifier with mixed spatial information proposes higher accuracy then the traditional one using spectral information mining only.