Research On Hyperspectral Image Classification Algorithm Based On Spatial Spectrum Features And Dictionary Learning

Hyperspectral remote sensing technology,as the main means of earth observation,is widely used in urban planning,ecological protection,military and many other fields.Pixel classification is the basis of hyperspectral image analysis and application and has great research significance.However,hyperspectral images have problems such as noise,information redundancy and complex spatial structure,which bring many challenges to the classification task.In this thesis,the spatial features,spectral features and dictionary learning of hyperspectral images are studied in depth,and the following algorithm is proposed.1.Aiming at the problems of hyperspectral image such as noise,complex spatial structure and complex spectral information,a collaborative representation hyperspectral image classification algorithm based on feature perception is proposed.Firstly,the hyperspectral image is reconstructed by adaptive weighting method.Secondly,spectral bias matrix and spatial bias matrix between pixels were calculated to perceive spectral and spatial features.Finally,the offset matrix is used to constrain the synergistic coefficient in the form of regular term.Experimental results show that,compared with the existing and advanced algorithms,the proposed algorithm can effectively perceive features,smooth noise and eliminate outliers,and has better spatial recognition and classification performance.2.In order to solve the problems that traditional dictionary learning algorithms ignore the locality between samples,overemphasize the orthogonality between dictionary atoms,and destroy the information consistency between dictionary atoms and test samples,a joint collaborative representation algorithm for hyperspectral image classification based on dictionary learning is proposed.Firstly,the algorithm realizes the dictionary manifold structure perception by learning the manifold structure inside the dictionary atom.Secondly,we learn the generalized spatial manifold structure and the generalized spectral manifold structure between the test sample and the dictionary atom,and perceive the spectral and spatial features.Finally,the above structural relations are used as important prior information to constrain the synergistic coefficients by means of closed modeling.The experimental results show that the proposed algorithm can perceive the structural features of images by learning the manifold structure between samples,and has a good spatial recognition ability.It is superior to the comparison algorithm in several important evaluation indexes,and has better classification performance and strong robustness.3.In order to effectively solve the problem of information redundancy of hyperspectral images and improve the classification accuracy of similar samples,a sparse representation hyperspectral image classification algorithm based on composite dictionary learning is proposed.Firstly,the feature learning algorithm is used to learn the space spectrum fusion features of the original image.Secondly,in the feature space,the KNN algorithm is used to construct the local constraint dictionary to improve the discriminability of the dictionary.Then,the multi-feature residual fusion framework is used to make up for the deficiency of single feature in the sample.Finally,the category labels of the test samples are determined according to the principle of minimum error.The experimental results show that the proposed algorithm can effectively reduce the correlation between different samples and improve the discriminant ability of the dictionary by learning the space spectrum features and constructing the local constraint dictionary.Compared with the existing and advanced algorithms,the classification accuracy and classification effect of the proposed algorithm are significantly improved.