Research On Band Selection Method Of Hyperspectral Image Based On Low Rank Representation

Hyperspectral remote sensing as a new type of remote sensing technology characterized by the integration of image and spectrumhas been one of the important technological breakthroughs of earth observation technology in recent years,which is widely used in the fields of modern military,mineral exploration,precision agriculture,environmental monitoring.A great deal of band data in hyperspectral remote sensing can provide rich information for researcher to detect ground objects,which is extremely pivotal for the following ground object classification and target recognition.However,the application of hyperspectral images suffers from the information redundancy problem and a new challenge to the difficulty of data processing brought by hundreds of continuous and subdivided spectral bands.How to select the significant band information from abundant row hyperspectral wave range data to maintain the effective classification information and improve the speed of data processing has become a new focus in the field of hyperspectral remote sensing classification gradually.To address the above problems,the band selection of hyperspectral classification data has been researched in this thesis by utilizing the low rank representation model based on the inherent spatial and spectral low rank characteristics of hyperspectral images,which realizes the dimension reduction of hyperspectral images.Specifically,the main contributions of this paper are as follows:(1)A low rank representation band selection(LRRBS)method based on low rank representation and graph structure has been proposed.First,the low rank representation model of hyperspectral image is constructed by using the low rank coefficients to represent the structure of the image,and then the spectral clustering algorithm is adopted to divide the band into several subsets,at last,the band selection criteria is designed to select the most salient band in the subsets(2)A new hyperspectral band selection method based super pixel segmentation constraint low rank representation band selection(SSCLRRBS)is proposed in the fourth chapter.The superpixel representation module performs regular optimization on the model to improve the representation of the low-rank model of the hyperspectral images,which utilizes Pearson coefficient measure to select the corresponding significant bands from the original data.Moreover,to verify the performance of the above band selection methods,this paper applies the selected bands to the classification applications of three classic hyperspectral data sets.The experimental results illustrate that the LRRBS method has a better classification effect on the three datasets,and the SSCLRRBS method constrained by superpixel constraints generates the great classification performance of the low-rank representation model.In addition,the experiments have also shown that the two band selection methods proposed in this paper have better performance compared with other state-of-the-art band selection methods.