Research On Key Technologies Of Hyperspectral Remote Sensing Imagery

Over the past few decades,the rich hybrid spatial-spectral information of hyperspectral image has greatly improved the perceptual ability of remote sensing image,which makes the hyperspectral remote sensing widely used in in various fields,such as space remote sensing,precision agriculture and military application.In the applications of hyperspectral remote sensing,it is necessary to exploit and analyze hyperspectral remote sensing data after we use the spectrometer to obtain it,and the processing flow of hyperspectral data mainly includes three stages: hyperspectral data preprocessing,hyperspectral data analysis and hyperspectral data application.Hyperspectral data preprocessing and hyperspectral data analysis are the premise and guarantee of hyperspectral data application,which is the key to expanding the breadth and depth of hyperspectral data application.In this dissertation,we focuses on the study of three key techniques of hyperspectral data preprocessing and hyperspectral data analysis,i.e.denosing in hyperspectral data preprocessing,unmixing and classification in hyperspectral data analysis.The main works and contributions of this dissertation are as follows.1.In order to remove the mixed noise in hyperspectral image,we propose a new hyperspectral image denoising method based on superpixel segmentation and low rank matrix factorization named DBSCAN-LRMF.First,the principal component analysis is adopted to obtain the base image of hyperspectral data,then superpixel segmentation technique DBSCAN is adopted on the the base image to segment the hyperspectral image into different homogeneous regions,finally the low rank matrix factorization?LRMF?is used to remove the mixed noise in different homogeneous regions.Extensive experiments have demonstrated that the proposed DBSCAN-LRMF can make full use of the spatial-spectral information of hyperspectral image,remove the mixed noise efficiently,and maintain the texture and details of the hyperspectral image.2.In order to solve the hyperspectral unmixing problem based on the generalized bilinear model,we propose a new nonlinear hyperspectral unmixing method based on bound projected optimal gradient method?BPOGM?.We transform the unmixing of generalized bilinear model into least square problem under the bound constraint,which can be solved by the Nesterov's optimal gradient method,and it is named as BPOGM.The BPOGM can achieve the optimal convergence rate of O?1/?k²??,with k denoting the number of iteration in BPOGM,finally we apply the BPOGM to solve the unmixing of generalized bilinear model under the alternating least squares framework.Extensive experiments have demonstrated that the proposed BPOGM can overcome the following problems: the price to pay with the Bayesian algorithm is its high computational complexity,the semi-NMF algorithm is sensitive to initialization and the GDA is a pixel-wise algorithm,which hinders us from applying to large hyperspectral images.3.In order to solve the sparse unmixing problem in hyperspectral image,we propose a new sparse unmixing method based on noise level estimation under the sparse regression framework?SU-NLE?.First,the noise in each band is estimated based on the multiple regression theory,then the noise weighting matrix can be obtained by the estimated noise,finally the noise weighting matrix is integrated to the sparse regression unmixing framework.Extensive experiments have demonstrated that the proposed SU-NLE can alleviate the impact of the noise levels at different bands,thus can improve the accuracy of unmixing based on sparse regression.4.In order to solve the classification problem in hyperspectral image,we propose a new robust sparse classification method named RSRC,which can take the sparse representation residuals into consideration.Besides,to make full use of spatial information of hyperspectral data,we extend the RSRC to the joint robust sparsity model named JRSRC.Extensive experiments have demonstrated that the proposed RSRC and JRSRC can obtain better classification performance than OMP and SOMP when using the Indian Pines and the Pavia University data sets,respectively,which demonstrates that it helps to improve the classification performance by taking the sparse representation residuals into consideration.Moreover,the classification performances of JRSRC are obvious better than these of RSRC,which indicates that ultilizing the spatial information of HSI can further improve the classification performance.5.In order to solve the classification problem in hyperspectral image,we propose a new robust sparse classification method based on spatial filtering and l_2,1 norm named SFL.First,the spatial filtering is adopted on the hyperspectral image by taking the average of neighborhood,then the l_2,1 norm regularization is used to extract relevant training samples among the whole training data set with joint sparsity,and the l_2,1 norm loss function is adopted to make it robust for outliers,finally the non-negative constraint is added to the sparse representation matrix motivated by hyperspectral unmixing.Extensive experiments have demonstrated that the proposed SFL can improve the classification performance and cost less time than the pixel-wise sparse regression based classification methods.