Research On Feature Extraction And Fusion Methods Of Remote Sensing Images

With the rapid development of micro-electronics and sensor technologies,the image acquisition capacity of remote sensing satellites becomes more and more stronger,which can yield various types of images,such as panchromatic(PAN),multispectral and hyperspectral data with high spatial resolution and/or spectral resolution.Hyperspectral images(HSIs)contain plenty of spatial and spectral signature information,and have a strong ability of ground object detection.They are widely used in the fields of surface mineral detection,environmental monitoring,military target detection and analysis,etc.However,because of the constraints of sensor technology,PAN images take only one spectral band but high spatial resolution,multispectral images(MSIs)have several to more than ten bands but low spatial resolution,while HSIs occupy dozens to hundreds of spectral bands but the lowest spatial resolution,resulting in a large number of mixed pixels.This leads to at least two technical problems:the first is how to extract the spectral signature of the substances and their corresponding spatial distribution accurately from the mixed pixels;the second is how to improve the quality and spatial resolution of HSIs for reducing the mixed pixels.The former amounts to hyperspectral unmixing,and the latter can be solved by the fusion of HSIs and MSIs which is a very effective method.In addition,HSIs and MSIs are all cube data,so the data need to be reshaped into matrices in spectral unmixing and image fusion,which will lead to the loss of some data structure information.Under the framework of spectral unmixing,the intrinsic relationship between spatial and spectral signatures is deeply explored,and a feature extraction method based on spectral unmixing and spatial-spectral regularization is studied by utilizing the minimum volume criterion in convex geometry.On the basis of non-negative matrix factorization and low-rank tensor factorization,this dissertation conducts an in-depth research on hyperspectral and multispectral image fusion,using the regularization methods such as proximal minimum volume,anisotropy total variation,and low-rank sparsity.The main research work of this dissertation include the following aspects:(1)Based on minimum volume belief and vector-total-variation-based smoothness in spectral unmixing,a feature extraction method based on matrix facorization and spatial-spectral regularization is proposed.Firstly,three quadratic regularization expressions based on minimum volume are introduced in this method to fully explore the correlation between spectral and spatial signatures.According to Craig's criterion,the minimum volume sim-plex is a convex geometry spanned by endmember vertices,which should enclose all the pixels in the image.Therefore,the minimum-volume criterion can effectively improve the performance of this method.Secondly,because there are a large number of mixed pixels in hyperspectral images,especially when corrupted by noise,the unmixing performance will be rapidly dropped.Thus,the regularization of spatial-spectral smoothing based on total variation is introduced to improve the anti-noise performance.Finally,structured matrices are constructed to reformulate the unmixed model as a convex problem based on matrix variables.The experimental results show that the proposed feature extraction method can effectively capture the endmember signature and fractional abundance of hyperspectral data.(2)The fusion of hyperspectral and multispectral images is regarded as an alternat-ing dynamical system,in which a novel fusion method is proposed on basis of a proximal minimum-volume regularization in convex geometry.Firstly,the volume of simplex spanned by endmembers is approximately equivalent to the distance sum between the endmember vertices and the centroid of reconstructed data in each iteration.Thus,the correlation be-tween hyperspectral and multispectral subspaces is fully utilized to improve the structural information loss caused by data matrixing.Secondly,the proximal alternating optimization is employed to design efficient algorithm,which decouples a bi-convex fusion problem into two single-variable convex subproblems.Then,the alternating direction method of multi-pliers(ADMM)iterates the primal and dual variables after multivariate separation via the equality constraints for the convergence of the solvers.The experimental results on three datasets of Pavia university,Moffett and Washington DC demonstrate that the fusion algo-rithm can significantly improve the quality and spatial resolution of reconstructed images.(3)Another novel image fusion method is proposed with sparsity and smoothing reg-ularization,which can denoise the pixels in the minimum-volume simplex via anisotropy total variation.First of all,the minimum-volume simplex is equivalent to the square sum of the distances between the endmembers,which is used to regularize the endmember signa-true,so as to alleviate the loss of spatial structure information caused by matrix decomposi-tion.Then,because the anti-noise performance of minimum-volume-based fusion method is weak,i.e.the performance will drop quickly with the increase of noise,the anisotropic total variation is imposed on data fitting terms of coupled non-negative matrix factorization to suppress the influence of noise in the two-dimensional image of each band along the ver-tical and horizontal directions,respectively.Finally,ADMM is used to design an efficient solver,in which tensor operators are used to reduce the dimension of large-scale matrices The experimental results display that this fusion method based on joint spatial-spectral reg-ularization can not only effectively improve the performance of reconstructed images,but also yield good anti-noise performance.(4)To utilize the spatial-spectral structure information more completely,a data fusion method of remote sensing images is proposed based on low-rank tensor decomposition.Firstly,the tensor-based observation model of hyperspectral and multispectral images is constructed on the basis of Tucker tensor decomposition.and then the redundant information of spatial dimensions and the shadow effect of spectral signatures can be eliminated by using the low-rank characteristic of factor matrices.Secondly,the sparse decomposition representation of the target image on three modes is enhanced by the sparsity of core tensor,which contains the weight coefficients of three factor matrices.Thirdly,considering the gradual change of signature vectors in the spectral-dimension factor matrix,total variation is introduced to carry out vertical smoothing to suppress the corruption of noise.Then,a set of efficient solvers is designed by conjugate gradient and ADMM to reduce the computational complexity of the solvers.The simulation results indicate that the proposed algorithm can improve the resolution and performance of super-resolution images,and decrease the noise sensitivity of this fusion method.