Hyperspectral Image Classification Based On Guided Filtering And Collaborative Representation

Hyperspectral imagery covers hundreds of bands of spectral information from visible light to infrared light.With its high spatial resolution and high spectral resolution,accurate classification of ground objects is possible.Due to the huge data dimension of hyperspectral imagery and the complex and changeable environmental conditions during the capture process,there is a large amount of redundant data in the image,accompanied by the phenomenon that homogeneous pixels have different spectrum and that heterogeneous pixels have same spectrum.The classification accuracy and efficiency of hyperspectral imagery still needs to be improved.Aiming at the characteristics of hyperspectral remote sensing data,this paper is devoted to obtaining robust and low-dimensional features for classification and recognition to overcome the adverse effects of data redundancy and data mixing on the recognition results.The main work content is as follows:(1)In view of the problem that the low classification accuracy caused by phenomenon that homogeneous pixels have different spectrum and that heterogeneous pixels have same spectrum,and the edge pixels are easily confused when combining spatial and spectral information in hyperspectral images,this paper proposes a method based on hierarchical guided filtering and the nearest regularized subspace.The method first uses principal component analysis to obtain the first principal component of the hyperspectral image.Then perform hierarchical guided filtering operation on the hyperspectral image with a guidance image,the first principal component.The edge-preserving characteristic of the guided filtering can effectively protect the spectral feature of heterogeneous pixels at the edge during the filtering process and reduce the difference in spectrum within-class at the local area.Finally,the nearest regularized subspace classifier is applied to classify the hyperspectral image preprocessed by hierarchical guided filtering.Compared with existing methods on three hyperspectral datasets,the experimental results demonstrates that the method proposed in this paper has achieved better classification accuracy and visualization effects.(2)Aiming at the problems that the higher dimensionality of hyperspectral data brings greater time overhead to the classification process and the lower classification accuracy under the condition of the small samples,a method based on dual-scale guided filtering and reconstruction error map filtering optimization proposed in this paper.Considering that single-scale filtering is difficult to accurately capture the features of geographic structure in diverse scales in hyperspectral remote sensing images,the introduced method first implements two variable-scale filtering operations on the hyperspectral image to obtain the robust characteristic of joint spatial and spectral information.Then the dual-scale filtered feature data is fused by principal component analysis to achieve dimensionality reduction and avoid the Hughes phenomenon.Finally,the framework employs the nearest regularized subspace classifier to obtain the reconstruction error map of the image,and the guided filtering is used to optimize the local weight of the error map to further improve the classification accuracy before decision-making.Perform performance evaluation of the proposed algorithm on hyperspectral datasets,the experimental results under the condition of small samples demonstrate that the method proposed in this paper effectively improves the classification accuracy compared with the traditional edge-preserving filtering method,and has higher real-time performance than the representation learning classification method.The paper contains 22 figures,15 tables,and 110 references.