Research On Feature Extraction For Remote Sensing Imagery Based On Spectral Information And Spatial Information

Remote sensing imagery is information product of non-contact, remote detection technology, which contains kinds of spectral information and spatial information. And it is widely applied in military, agriculture, environmental pollution, geological disasters, and other fields. With the development of remote sensing imaging technology, the processing technology of remote sensing imagery obtains more and more attention. In general, the data of remote sensing imagery is large and redundancy, therefore, how to extract effective information for application becomes a problem to be solved.In this thesis, feature extraction methods for remote sensing imagery based on spectral information and spatial information are studied. The main research contents are as follows:On the one hand, sparse feature extraction methods for hyperspectral imagery are studied. As to the weakness of Sparse Principle Component Analysis (SPCA) method and Sparse Discriminant Analysis (SDA) method, a feature extraction method combining spectral information and spatial information for hyperspectral imagery is proposed, which is named for Sparse Tensor Discriminant Analysis (STDA). And K-Nearest Neighbor (KNN) is used for classification of hyperspectral imagery. The experiment results with Pavia University dataset and Indian Pines dataset show that the classification precision of hyperspectral imagery is improved by extracting the spatial feature on the basis of spectral feature.On the other hand, spatial texture feature extraction methods for SAR remote sensing imagery are studied. Due to the fact that Local Binary Patterns (LBP) can not describe the texture feature in a wide range of areas, Gabor and Three-Patch Local Binary Patterns (TPLBP) are applied to extract spatial texture feature for SAR remote sensing imagery. And Kernel Extreme Learning Machine (KELM) is used for SAR target recognition. The experiment results with MSTAR dataset indicate that the edge feature information between the SAR target and its background are extracted effectively by Gabor, and TPLBP has better ability of texture feature description than LBP. Therefore, the recognition precision of SAR remote sensing imagery is improved.