Research On Hyperspectral Image Classification Algorithm Based On Collaborative Representation And Manfold Learning

The development of remote sensing imaging technology increasingly promotes the breadth and depth of hyperspectral image application,and the classification of hyperspectral image is closely related to its application.In the research of hyperspectral image classification,many influential techniques have been proposed,among which the collaborative representation based on dictionary learning is an important aspect.In this paper,based on collaborative representation and manifold learning,combining with the characteristics of hyperspectral image data and focusing on the mining of spatial-spectral information,the classification algorithm of hyperspectral image is studied and two improved hyperspectral image classification algorithms are proposed.(1)Collaborative representation is an important tool in the research of hyperspectral image classification.However,in the learning process of hyperspectral image classification,the correlation algorithm of collaborative representation does not well describe the characteristics of hyperspectral image,such as the localization of pixels and label information,etc.,so its performance is limited.For this reason,we propose a divergence kernel collaborative representation technique and a feature acquisition classification method using spatial-spectral fusion(IKCRC).In order to describe the localization and label information of pixels effectively,the proposed method constructs a new divergence kernel collaborative representation model and the corresponding calculation model.In the proposed algorithm formula,we introduce kernel mapping to improve the classification ability,and use the initial feature extraction of spatial-spectral fusion in the calculation process to make the algorithm simple and efficient.Comparison experiments with standard hyperspectral image datasets show that our proposed method,IKCRC,is more effective in improving classification accuracy.(2)In the field of hyperspectral image classification,collaborative representation has the advantage of high efficiency and quickness.However,due to the limitation of dictionary structure in the learning process of hyperspectral image classification,the collaborative and its related algorithms are difficult to describe the manifold information of hyperspectral image.Sparse representation will be affected by the same spectral difference between the dictionary atoms in the neighborhood and the test samples,which will seriously reduce the classification performance.According to the classification characteristics of hyperspectral images,a locally linear embedding kernel collaborative representation(LLEKCR)algorithm is proposed in this paper.The proposed method combines the manifold learning method to construct a new kernel collaborative representation model,and extracts discriminant spectral information and spatial information between spatial-spectral feature learning.In the classification process,the correlation coefficient and classification error between the feature dictionary and the test sample are integrated into the decision making.Comparison experiments with standard hyperspectral image datasets show that the proposed method is more effective in improving the classification accuracy.