Hyperspectral Image Classification And Feature Extraction Based On Kernel Methods

Comparing with panchromatic and multispectral image, hyperspectral image which enriches spectrum information can be better on classifying objects on the ground precisely. However, the conventional image analysis methods can not meet the requirements of hyperspectral image applications. Since kernel methods were successfully applied to support vector machine, using kernel functions, people try to extend the ordinary linear methods of feature extraction and classification to nonlinear situation. This paper gives a summary of classification and feature extraction methods of hyperspectral image. Using the successful applications in many fields for references and basing on kernel methods, several researches on classification and feature extraction methods of the hyperspectral image were made. The major works implemented and the goals achieved are listed as follow:1. Firstly, illustrating the hyperspectral RS technologies and several matters which need to be resolved in its application. Then analysis of the methods of hyperspectral image classification and its characterastics was given, and the methods of bands selection and feature extraction were summed up.2. For the hyperspectral image classification based on Support Vector Machine (SVM), several kinds of technology was integrated, including sequential minimal optimization training algorithm, cross-validation grid search parameters selection technology, multi-class classifier decomposition methods, to design multi-class SVM classifier which is more fast and robust. The classification experiments of PHI and Hyperion images show that the accuracy with the chosen methods is satisfying, and reboust.3. For the hyperspectral image classification based on Kernel Fisher Discriminant Analysis (KFDA), we used methods of SVM classifier's parameters selection and multi-class classifier decomposition for references, to improve KFDA's classification performance. Through the experiments of OMIS and AVIRIS images, it can be concluded that, comparing with SVM classifier, KFDA can get almost the same classification accuracy and require less time.4. Bringing Generalized Discriminant Analysis (GDA) into feature extraction of hyperspectral image. Through feature extraction experiment of hyperspectral image, a conclusion can be made that the feature extraction method based on GDA can improve hyperspectral image's classification accuracy.