Spectral-spatial Unmixing And Classification Methods Of Hyperspectral Imagery

Hyperspectral remote sensing is one of the greatest cutting-edge technologies in re-mote sensing filed. Due to the wealth of radiation, spectral and spatial information, hy-perspectral image has been widely applied in precision agriculture, mineral exploration, military object recognition, resources and the environment survey, disaster assessment and other fields. Therefore, interpreting hyperspectral data plays a critical role on theoretical significance and realistic applications.Unmixing and classification are the fundamental problems for hyperspectral remote sensing. They are also the most important basis of quantitative analysis and follow-up ap-plications of hyperspectral image. Due to the facts of instrument, atmospheric radiation, uneven illumination and the structure of the coverland, pixels in the same class usually have difference spectra, leading to a low accuracy for spectral-only unmixing and classifi-cation methods. Spatial information can fully characterize the feature structure, effectively reducing the "isomer spectrum" effects. And spectral-spatial methods have drawn more and more attention recently. This thesis focuses on the interpretation of hyperspectral im-age, such as unmixing and classification. Several spectral-spatial models and approaches are put forward for unmixing and classification problems. Followings are the primary works and achievements of the dissertation.1. Under the linear mixture model, spectral unmixing can be seen as a sparse re-gression problem by using the library known in advance as the dictionary of endmembers. Using the/1/2 regularizer to enforce the sparsity of the fractional abundances and imposing the ANC and ASC constraints, a fully constraint/1/2 regularized sparse regression model of hyperspectral unmixing is proposed. A reweighted l1 iterative algorithm is introduced to solve the model. The experiments on simulated and real data both show that the/1/2 regularized sparse regression method is effective and accurate on linear hyperspectral un-mixing.2. For supervised hyperspectral classification, a new spectral-spatial method, com-bining sparse representation and MRF-based spatial prior, is proposed under the Bayesian framework, in which the likelihood probability is modeled by the sparse representation method, and the spatial prior is modeled as a Gibbs distribution, which specifies a Markov random field, on the classification labels. The graph-cut algorithm is adopted to solve the proposed method. Real hyperspectral data sets are used to validate our proposed method. Experimental results show that the proposed method outperforms many of the state-of-the-art methods.3. Under the Maximum A posterior (MAP) framework, we propose a supervised classification model which includes a spectral data fidelity term and a spatially adaptive Markov Random Fields (MRF) prior in the hidden field. The data fidelity term is learnt from the sparse multinomial logistic regression (SMLR) classifier, while the spatially adap-tive MRFs prior is modeled by a spatially adaptive total variation (SpATV) regularization. To further improve the classification accuracy, the true labels of the training samples are fixed as an additional constraint on the proposed model. An efficient hyperspectral im-age classification algorithm, named SMLR-SpATV (sparse multinomial logistic regression based spatially adaptive total variation method), is then developed to solve the final pro-posed model using the alternating direction method of multipliers (ADMM). Experimental results on real hyperspectral data sets demonstrate that the proposed approach outperforms many state-of-the-art methods.4. A new multiple-classifier approach combing multinomial logistic regression (MLRsub) and sparse representation for spectral-spatial classification of hyperspectral im-ages (HSI) is proposed. The multiple-classifier approach is based on the decision fusion of full probability distribution by MLRsub and sparse probability distribution by sparse unmixing method separately. Spatial information in the method is exploited by an edge preserving Markov random field (MRF). Experimental results with real hyperspectral data sets indicate that our proposed multiple-classifier leads to better classification performance than the state-of-the-art methods.