Hyperspectral Image Classification Based On Semantic Extraction

Hyperspectral images(HSIs)contain invaluable information in both spectral and spatial domains,which provide well support for image processing.Meanwhile,classification is one of the most popular processing techniques in the analysis of HSIs.High spectral resolutions provide useful information for discriminating different materials and objects.There are three main issues for traditional HSI classification methods: Previously,the original feature space may not be the most effective space for classification task.Secondly,traditional methods cannot make fully use of spatial information in HSIs.Finally,the valuable class information which is provided by a few training samples is much ignored.To solve these problems,we propose three semantic extraction methods,which based on Markov random field(MRF)model for HSI classification.Mapping all samples to the semantic space from original feature space with the valuable training samples,and then exploiting a novel MRF model to offer the effective spatial constraint for HSI semantic extraction.Experimental results on three real hyperspectral data sets show that the proposed methods provide satisfactory classification accuracy and robust performance,specially the region uniformity is better than most conventional methods.The main contributions of this thesis are summarized as follows:1.We propose a semantic fusion and extraction method based on multi-feature with spatial constraint(MFS)for HSI classification,which extracts some low level features from the original data and then maps these features to the same semantic space,finally various semantic features and local spatial information are integrated into the Markov random field(MRF)model.This method enhances the discrimination capability of each sample,thus the overall accuracy and the region consistency have been improved.2.A semantic extraction approach based on adaptive local and non-local spatial constraint for HSI classification is proposed in this thesis.The proposed approach exploits local information with a superpixel-based method and extracts redundant spatial information from non-local regions is an extension of MFS.The experimental results show that the proposed approach further enhances the overall accuracy.Meanwhile the robustness of this approach is better than others.3.We develop an object-oriented semantic extraction method with multi-layer segmentation for HSI classification.Based on the object-oriented thinking,a superpixel can be regarded as an object in HSI.The HSI image is segmented into several superpixels with multi-layer segmentation strategy at first,then samples belonging to the same superpixel at last layer should be classified to the same class.Meanwhile two regular terms(neighborhood constraint and brother constraint)are proposed to improve the algorithm performance for the novel MRF model.The proposed method is an object-level classification method,which has better accuracy and region uniformity than traditional pixel-level methods,also reduces the time complexity effectively.