Hyperspectral Imagery Classification Based On Support Vector Machine

Hyperspectral remote sensing has been playing an important role in many fields of both military and civil applications. The outstanding characteristic of hyperspectral imagery is the high spectral resolution at the expense of the high data size and high dimension. The methods of multispectral imagery processing are not applicable to hyperspectral imagery. It's urgent to develop fast and accurate methods to discover interested information from the huge data produced by hyperspectral sensors.Classification is one of the methods for people to obtain information. When traditional classification methods used for hyperspectral imagery classification, the "Hughes Phenomenon" is easily caused because of the limited training samples. Support vector machine is a new general machine learning method based on Statistical Learning Theory, can solve this phenomenon effectively. Based on the properties of hyperspectral imagery data, SVM is applied in classification of hyperspectral imagery in this thesis, the main contents and innovations are as following:1. Analyzes the characteristic of hyperspectral data and the classification process of hyperspectral imagery. Points out the disadvantages of traditional classification methods while apply to hyperspectral data, and presents the advantages of the Support Vector Machine.2. Analyzes various methods of Multiclass Support Vector Machines from the four views as follows: constructing a multiclass classifier by combining several binary classifiers, using hierarchical structure, solving a single optimization problem and using error correcting codes SVM. This paper systemically introduces several typically methods and make a clear comparison between them.3. Presents a secondary classification method. This method modifies the weight coefficients of the sub-class machines, and enhances the divisibility of the classes that are difficult to classify. The experiments results of hyperspectral imagery classification show that the proposed multiclass SVM has its superiority.4. Gives an FSVM method which based on 1-a-1 SVM, and applies it to hyperspectral imagery data classification. The results show that it not only depresses the existence of unclassifiable regions but also gets better classification accuracy than the traditional SVM.