Self-representation Based Subspace Clustering For Hyperspectral Remote Sensing Imagery

Hyperspectral remote sensing can acquire images for the interested area at the nanoscale spectral resolution,with characteristics of image-spectrum merging and continuous spectrum.It is capable of sub-class recognition and attribute analysis of target ground objects with high spectral diagnosis ability,which greatly enhances the refined information extraction ability of ground objects,and is widely used in precise agriculture,environmental monitoring,mineral exploration,military reconnaissance,and many other fields.As an unsupervised information extraction method,clustering is an effective means of ground object recognition when prior information is missing or training samples are difficult to obtain,which has very important theoretical and practical significance.The research progress of hyperspectral image clustering is systematicly reviewed in this paper,and based on this,three problems for hyperspectral image clustering to be sovled are summarized,mainly including how to accurately model the complex internal structure and spectral variability,how to fully exploit the spatial and spectral information of the image,and the nonlinear structure.Aiming at these problems,with introduction of the latest theory in computer version,machine learning,signal processing,data mining,and other fields,combined with the characteristics of hyperspectral images,the self-representation based subspace clustering theory and algorithm is proposed in this paper for hyperspectral remote sensing images,to solve the above problems and improve the clustering accuracy.Through amounts of experiments,the effectiveness of the proposed clustering algorithms is verified.The main research contents of this paper are as follows:(1)Aiming at the problem that traditional clustering methods cannot accurately model the complex internal structure and large spectral variability of hyperspectral images,the sparse subspace clustering algorithm in the computer verison field is introduced to solve the hyperspectral image clustering problem for the first time in this paper.This algorithm effectively models the internal structure and spectral variability of hyperspectral images with the subspace model.On this basis,considering that traditional methods cannot effectively fuse the spatial domain information and the spectral domain information with large differences by directly constructing the spectral-spatial integration model in the image domain,the spectral-spatial collaboration based sparse subspace clustering theory and algorithom is proposed in this paper.This algorithm realizes the transformation of the image feature domain by self-representation learning,to reduce the differences of the features,and incorporates spatial information in the representation domain to construct the integration model,so as to effectively fuse spectral-spatial information of the image.From the perspective of pixels,the spectral-spatial sparse subspace clustering algorithm and the ₂-norm regularized sparse subspace clustering algorithm are proposed in this paper,through constructing the spatial regularization constraints to reduce the representation bias and promote the picewise smoothness of the coefficient matrix,the clusteing accracy can be improved.From the perspective of object,a mass center reweighted object-oriented sparse subspace clustering algorithm is proposed in this paper.Through flexibly modeling the spatial neighbors of the image with various shapes by object,and extracting the representative features of the objects by mass center learning,the object sparse subspace clustering model can be constructed.In this way,the clustering accuracy can be further improved,and the time cost and the memory consumption can be greatly reduced,which further improves the practical application value of the algorithm.The experimental results suggest that the sparse subspace clustering algorithm can better model hyperspectral images and shows great potential.The proposed spatial-spectral collaboration based sparse subspace clustering algorithm can effective improve the clustering accuracy.(2)Aiming at the nonlinear structure problem of hyperspectral images,a kernel sparse subspace clustering algorithm with a spatial max pooling operation is proposed in this paper.On the one hand,hyperspectral pixels are mapped from the original feature space into the higher dimensional kernel space by the kernel function,to approximately transform the linearly inseperable problem into the linearly seperable problem,which effectively reduces the systematic errors of the model and improves the accuracy of self-representation learning.On the other hand,by fully considering of the working mechinasm of sparse subspace clustering,a spatial max pooling operation is utilized to incorporate spatial neighborhood information in the representation domain to generate more discriminant new features for target pixels,thus further improving the clustering accuracy.The experimental results suggest that the proposed algorithm can effectively sovle the nonlinear structure problem of hyperspectral images to some degree,and has obvious advantages over the linear clustering methods.(3)Aiming at the high computational complexity and low efficiency problems of sparse clustering methods,a total variation regularized collaborative representation clustering algorithm with a locally adaptive dictionary is proposed in this paper.Firstly,the collaborative representation model with low computational complexity is introduced to explore its potential in hyperspectral clustering.In addition,through constructing the locally adaptive dictionary,the interference of dissimilar atoms to the self-representation learning of target pixels can be effectively avoid.Then,through constructing the total variation regularization to incorpate spatial information in the representation domain,the discriminant capability of the features can be enhanced,the representation bias can be reduced,and the piecewise smoothness of the coefficient matrix can be promoted.In this way,the clustering accuracy can be improved,and the spatial homogeneity of the clustering result can be guaranteed.The experimental results suggest that collaborative representation can efficiently explore the underlying structure information of hyperspectral images to some degree.The proposed algorithm can obtain clustering accuracy comparable to the sparse methods,and can be a very competitive clustering algorithm for hyperspectral images.