Research On Classification-Oriented Multi/Hyper-Spectral And SAR Images Synergic Processing

Multi/Hyper-spectral images contain not only two-dimensional space information of the ordinary remote sensing images, and have a wealth of spectral information. In particular, the spectral information of hyperspectral images can be formed to an approximately continuous spectral curve. The SAR images mainly reflect the dielectric properties and geometric characteristics of surface materials, resulting in very rich spatial texture details. Especially, Polarimetric SAR images contain much polarized scattering information of surface features. Therefore, the research on the exploitation of the synergy between mult/hyper-spectral and SAR images for land use mapping has important significance and applications.According to the characteristics of each data source and the amount of information per image, in this dissertation, the multi/hyper-spectral and SAR images are divided into two groups for feature extraction and feature selection, in order to make full use of the complementary information of both spectral and polarization scattering. And on this basis, the high accuracy of the classification-oriented multi-level synergic processing is achieved in this thesis:The first group is multi-spectral and SAR images. In accordance with their respective characteristics, this dissertation considers the spectral characteristics of multi-spectral image and the texture features of SAR image, including the local statistical features, the GLCM texture features and fractal dimension feature. And the feature selection method based on genetic algorithm is used on SAR texture features for removing redundant information and much more accurately synergic classification in the next section.The second group is hyper-spectral and polarimetric SAR images, which have much more abundant information. In order to get the best set of spectral features for classification and overcome the "curse of dimensionality" phenomenon, this dissertation studies the double nearsest proportion feature extract method which is especially used for hyper-spectral images. And this method is compared with the traditional principal component analysis and linear discriminant analysis methods. For the Polarimetric SAR images, H/A/αdecomposition, Freeman decomposition, Yamaguchi decomposition and multi-component scattering model based decomposition method is applied for feature extraction. Finally, based on the aboved research on the feature extraction and selection methods of multi/hyper-spectral and SAR images, the feature level synergic classification algorithm based on parallel feature combination and the decision level synergic classification based on fuzzy set theory is studied in this dissertation. Inspired by the above two methods, we propose a new synergic classification method which combine the feature level and decision level method. The precise classification results obtained by the feature level method are the inputs of decision level method, so as to preserve and use the advantages of both levels. Therefore, the ultimate goal of multi-level and high accuracy synergic classification is achieved by this new method.