Based On The SVM Kernel Function And Parameter Selection Of Hyperspectral Image Classification

Hyperspectral remote sensing plays a very important role in the fields such as agriculture, forestry, geology, Marine, and military as the development front of remote sensing technology. The classification of remote sensing image in Hyperspectral image information process is a very important research direction, where the classification accuracy is the foundation for a variety of applications. Hyperspectral image has a number of multi-bands, high resolution and large amount of data, etc. Using traditional image classification for hyperspectral image classification is difficult to obtain good effect. Taking into account the good generalization of support vector machines in small samples, nonlinearity and high dimension feature space, and considering the particularity of hyper-spectral remote sensing image and the problems existing in the classification process, this thesis deeply studies the kernel function of support vector machine and their parameters selection for hyperspectral remote sensing image classification. The main content of the thesis and contributions are as follows:(1) We introduce the background of remote sensing research, summarize the development and research direction of support vector machine. Analysis the formation of hyperspectral image, this thesis expounds the characteristics of hyperspectral data and the process of classification, discuss the important steps for the necessary in classification.(2) This thesis introduces the principle of support vector machine, and analyzes the support vector machine in linear separability, generalized linear separability, and nonlinear model designed methods, and the role of nuclear technology in solving nonlinear problems are pointed out.(3) In view of the kernel function and its parameters of the support vector machine, their influence on the effect of support vector machine classification are analyzed.Through the experiment, it proves that the selection of parameters affect the effects of the support vector machine classification. we also introduce the grid search method and the basic particle swarm algorithm.. Finally, we put forward an improved parameter optimization method based on particle swarm optimization algorithm, and a contrast experiment was carried out.(4) In this thesis, we conduct the numerical experiments for the proposed algorithm. Firstly,we examine the classification of different kernel functions. We apply different kernel functions of support vector machine to hyperspectral image classification, finally we analysis out the accuracy of the different kernel functions in hyperspectral image classification is not significant. Secondly, we use different parameter optimization methods to classify. Results show that the improved basic particle swarm method increase the classification accuracy and the effectiveness of the method was verified.The innovation of this thesis lies in for the basic particle swarm algorithm setting the dynamical adjustment factor to avoid the local optimal solutions and to search for the global optimal value.