Research On Key Issue Of Intelligent Analysis For Hyperspectral Image Based On Support Vector Machine (SVM)

It is a complicated and systematic project for processing and analysis of the remote sensing(RS) image, the different data is subject to different theory of analysis and processing. The hyperspectral RS image has always been one of the important directions of the RS technology research and development, because it has sophisticated information holdings, it has been used in a lot of field and industry generally. As the rapid development of hyperspectral sensors, the spectral resolution and temporal resolution of RS image has greatly improved, which raises the amount of longitudinal data (spectral dimension) greatly. Because of special nature of the data, The new research topics is proposed for information extraction and processing,the traditional image processing methods have not been fully applied to the analysis of hyperspectral RS image recognition. So, it is urgent to to study the appropriate classification strategy for the effective analysis and identification of hyperspectral RS image.The RS image analysis is a kind of pattern recognition, can be categorized into the area of machine learning, SVM is a statistical learning theory developed from mathematical statistic theory, it is able to solve such problems as the structure selection problem, local minimum points, and so on. In this paper,Considering the characteristics of hyperspectral RS images and the theoretical performance of SVM, it is studied how the SVM used in hyperspectral RS image classification in which the key issues are discussed and improved.Firstly. SVM ideology is elaborated in detail and their theoretical properties is studied. It is obtained through experiments that the traditional SVM possesses many advantages existing in analysis of the hyperspectral RS image. It can solve the machine learning problems with small sample, can improve the generalization performance, can solve the high-dimensional problems and nonlinear problems, it also can avoid neural network structure selection and local minimum problems. Meanwhile, There is also lots of weak points.such as an alarming amount of calculation, the optimal parameters settings is very difficult.In this paper, for high dimensional features of hyperspectral RS images, The kernel PCA(KPCA) is used for feature extraction, which is the combination of kernel function and classic PCA, in which the kernel theory of SVM can be achieved with an effective interface with the SVM classification following. In KPCA, physical spectral information would be lost or distorted, to make up for the shortcoming, this paper introduced the theory of fractal dimension to extract fractal dimension feature values of spectral feature curve. Then feature conversion and extraction carry on with the joint KPCA and fractal dimension theory. The results of the experiments show that the new feature extraction algorithm is more effective, especially in the SVM classification. Then, It is studied and analyzed for the parameters of SVM classification model in the paper, and is discussed for the essence of parameter and its effect for the SVM classification. Firstly, it is to illustrate the importance of parameter settings through experiments, the SVM classification model without punishment coefficient is proposed of combined with NPA calculation principle, in which it is selected for the Radial Basis Function (RBF). The sectional directed binary tree optimization strategy is designed to set the optimal kernel parameter. Both of the designs realize the effective SVM model parameters setting. The tests show that the effective parameter settings make SVM performance improve greatly.SVM theory unit is developed based on bi-classification problems, when category is more than two, the multi-classification need to be converted into bi-classification problem. The multi-classification strategy will affects the speed and accuracy of SVM classification greatly. In this paper, a new algorithm is drawn that the SVM classification is carried on by the order of complexity of each category, in which the NPA calculation of bi-classification problems is used combined with "1 V m" multi-classification strategy. Firstly the indicators of the comprehensive complexity of each category(index) are calculated. The classification order is determined according to the index. Using NPA SVM bi-classification, separate the high complexity of category data firstly from the others, which avoids the category with high complexity involved in the calculation repeatly. The test shows new strategy improvs the efficiency of SVM multi-classification significantly.Finally, this article discusses the effectiveness of the design of SVM hyperspectral RS image classification framework, highlighting the innovative features, it also proposes some problems which need further improvement, which wish to inspire future research helpfully.