Research On Spectral-spatial Dimensionality Reduction Method For Hyperspectral Images Based On Superpixel

Hyperspectral remote sensing is a milestone in the development of remote sensing technology.With the rapid development of computer science,sensor technology and communication technology,huge amount of hyperspectral images(hyperspectral data)are stored.With hundreds of wavebands and high spatial resolution,hyperspectral images can provide rich information for accurate feature identification.In recent years,hyperspectral remote sensing technology has been successfully applied in many fields such as geological survey,urban planning,fine agriculture,and vegetation monitoring.Hyperspectral images with high-dimensional features usually contain a large amount of redundant information,which not only increases the computational cost of later data processing,but also easily leads to the Hughes phenomenon.Therefore,in order to solve this problem,it is necessary to carry out the research on dimensionality reduction of hyperspectral images.Because hyperspectral images not only have rich spectral information,but also have high spatial resolution,which means that hyperspectral data also contain rich local spatial structure information,so how to incorporate these local spatial information in the process of dimensionality reduction calculation and improve the quality of the reduced dimensional data is a very worthy research problem.Based on superpixel and principal component analysis methods,this thesis proposes an effective spectral–spatial dimensionality reduction method for hyperspectral images,with the goal of improving the quality of dimensionality reduction data by incorporating local spatial information provided by the image in the process of dimensionality reduction.The method mainly consists of the following steps:(1)using an entropy rate-based superpixel segmentation algorithm,the given hyperspectral image is cut into a large number of homogeneous regions with adaptive shapes and sizes;(2)by improving the local modularity function in complex networks,the over-segmentation results are merged to obtain the merged superpixel map;(3)On each merged superpixel,the dimensionality is reduced using the principal component analysis algorithm and to the same dimensionality;(4)Integrating the dimensionality reduction results on each superpixel to generate the overall dimensionality reduction matrix.The contribution of the proposed method is to increase the role of local spatial structure information in the dimensionality reduction calculation by superpixel merging,and weaken the dependence of the dimensionality reduction results on the superpixel segmentation scale.To demonstrate the dimensionality reduction effect of the proposed method,extensive experiments were conducted on three commonly used hyperspectral benchmark datasets from Indian Pines,Salinas and Pavia University.The test results and comparison results on three hyperspectral images show that the classification accuracy of the images using the dimensionality reduction data obtained by the proposed method is significantly improved and outperforms several other competing methods.This indicates that the proposed method can effectively remove redundancy in hyperspectral data by enhancing the utilization of spatial information in the dimensionality reduction calculation.