The Research Of Classification Techniques For Hyperspectral Remote Sensing Image Data

During the last 20 years, hyperspectral remote sensing has been playing an important role in many fields of both military and civil applications. It's urgent to develop fast and accurate methods to discover interested information from the huge data produced by hyperspectral sensors. Based on the properties of hyperspectral image data, effective classification algorithms for extracting most information of groundcover types from hyperspectral image data are studied in this thesis.The special properties of hyperspectral image data are first analyzed with its challenging influence on traditional classification methods. It's showed that hyperspectral image data can be properly modeled using the Gaussian mixture model, and can obtain a even better match to the model after linear projections.According to the characteristics of hyperspectral image, unsupervised classification techniques of hyperspectral image data are analyzed, which can be used for segmenting hyperspectral image, or as an important unsupervised step for the following supervised classification and unmixing processing. Model-based classification method is applied to effectively discover data structure and probability distribution, and dimensionality reduction, or feature extraction method is introduced to solve the high dimensionality problem. Based on the interactional relationship between dimensionality reduction and classification, feature extraction and classification are properly combined in two different ways, one by classification after dimensionality reduction, and the other by realizing dimensionality reduction and classification simultaneously. The critical problem in unsupervised classification of class number determination and intrinsic dimensionality derivation was discussed, and several selection rules are proposed accordingly. In detail, an unsupervised classification method combining Principal Component Analysis (PCA) linear transform and Gaussian Mixture Model (GMM) is developed, with the model derived by EM algorithm; a new model selection principle for data after PCA transform, PMDL, is proposed, and the class number is determined by selection of optimal component number of mixture according to the principle. Based on mixture of Probabilistic Principal Component Analysis (PPCA) model, dimensionality reduction and classification are simultaneously realized embodying the ideology of subspace methods of pattern recognition. The unifying realization of the seeming independent steps can provide good understanding of data structure, and the intrinsic dimensionality and the precise dimensionality reducedrepresentation of data can be obtained. The model selection principle of determining effective number of dimensionality reduction for different clusters is proposed.Supervised classification algorithms with limited training samples for hyperspectral image data are then studied, including how to apply the powerful supervised classification method for high dimensional data, SVM, to classify hyperspectral image data; the problems of kernel function selection and parameter determination are analyzed. The algorithm combining Gabor filtering and SVM classifier is proposed, and integrating spatial and spectral information contained in hyperspectral image data increases classification performance.