Research On Hyperspectral Image Classification Algorithms Based On Low Rank Subspace Representation

Hyperspectral imaging technology records cube data of ‘continuous bands’,‘numerous bands’ and nanoscale spectral resolution,which is widely used in earth observation,land cover analysis,mineral exploration,environmental monitoring and intelligent agriculture etc.,by the high spectral resolution and the ability of discriminating different land cover types with subtle differences.The corresponding intelligent analysis processing techniques mainly include fusion,unmixing,target detection,feature extraction,classification and so on.How to make full use of the rich spatial and spectral information in hyperspectral data and perform detailed classification of ground object is always research hotspot in hyperspectral remote sensing image processing.Traditional machine learning methods have been widely used in hyperspectral image classification,there is still some improvement room for classification accuracy.However,there are a few problems which make it difficult to accurately discriminate various classes of land covers,e.g.,“Hough effect” caused by high dimensionality,information redundancy,sample nonequilibrium,and “identical characteristics” caused by insufficient spatial resolution.In addition,labeling hyperspectral images is costly,laborious and time-consuming.To address aforementioned issues,based on the analysis and summary of hyperspectral supervised classification and unsupervised classification,around the spatial-spectral representation learning,this dissertation studies the supervised classification under small samples and high accuracy unsupervised classification under the framework of efficient representation learning,to fully explore the intrinsic correlation in hyperspectral data.The main contributions of this dissertation are as follows:(1)A novel combined kernel sparse multinomial logistic regression and TV-L1 optimization method is proposed.The traditional two-stage classification methods of hyperspectral image directly measure the initial classification probability by L2 norm without further analysis of the characteristics of the classification error,which makes the classification performance need to be further improved.Firstly,the kernel sparse multinomial logistic regression method is used to explore the spectral information of hyperspectral data,and obtain the initial classification probability map.Then,the L1 norm is adopted to measure the heavy-tailed property of the classification error,and the spatial neighborhood Markov property of the classification probability is measured by total variation(TV)regularization term.The TV-L1 optimization model is established,which can be solved quickly by Alternating Direction Method of Multipliers(ADMM).The experimental results show that the proposed method can effectively combine the spatial-spectral structure information and Markov property of neighborhood,eliminate the classification errors caused by insufficient training samples,and improve the robustness and classification accuracy.(2)A novel spectral-spatial bilateral modulated low rank subspace clustering(SSLRSC)algorithm is presented for hyperspectral images clustering.In order to classify the hyperspectral remote sensing images without any label information,a novel spectral-spatial bilateral modulated low rank subspace clustering algorithm for hyperspectral images is proposed.Based on the assumption that the pixels from the same ground-cover distributed in the same independent subspace,the low rank representation is used to describe the subspace structure of the hyperspectral data,and obtains a low rank representation matrix.Considering that low rank representation can only describe the global structure of data,spectral weighted low rank subspace clustering(SWLRSC)is adopted so that each pixel can be represented by the most correlative hyperspectral pixel in its subspace.Finally,the spatial domain bilateral weighting is adopted to maintain the spatial smoothness of the representation coefficients between adjacent pixels.The experimental results show that the proposed method improves the clustering accuracy of hyperspectral images effectively.(3)A novel low rank subspace clustering method based on spatial-spectral hypergraph regularization for hyperspectral image is proposed.The representation features obtained from the optimized low rank subspace self-representation model mine the global information of the data,but ignore the local structural information.Inspired by the idea of manifold learning,the graph regularized low rank subspace clustering was proposed to explore the intrinsic structure of data by the globality of low rank constraint and locality of graph.However,the structure of simple graph only contains the relationship between two pixels and lacks the exploration of more complex high-order relationship of multiple pixels.Then,based on the theory of hypergraph regularized low rank subspace clustering,this dissertation divides spatially homogeneous pixels into the same spatial region through super-pixel segmentation,and then constructs the hyperedge of the spatial-spectral hypergraph and calculates the weight of the hyperedge.Compared with the previous spatial-spectral hypergraphs with fixed spatial windows,the proposed hypergraph construction method effectively utilizes the structure information of homogenous spatial regions in arbitrary shapes,and simultaneously explore the complex high-order spectral similarity information of hyperspectral data.By fusing the above-mentioned hypergraph regularization and low rank representation,more robust representation features are learned with joint optimization learning to construct similarity matrix,thereby obtaining better clustering results.Experimental results on real hyperspectral data sets verify the effectiveness of the proposed method.(4)We propose a unified low rank subspace clustering method with dynamic hypergraph for hyperspectral image.Among the existing non-dynamic optimization hypergraph regularization methods,the hypergraph is generally preconstructed from the original data by fixed distance measurement rather than dynamic optimization learning,which limits the role of hypergraph regularization in subspace representation learning.In order to address these issues,this dissertation proposes a dynamic hypergraph regularized low-rank subspace clustering algorithm,in which the hypergraph is updated by dynamic learning of subspace features,and at the same time it is used to constrain the local structure information of subspace features.In addition,the existing clustering algorithms have two steps: learn new representation features;construct the similarity matrix with representation features,and then the clustering results were obtained by spectral clustering or K-means.However,K-means algorithm is very sensitive to the initialization of the cluster center,which makes the clustering result very unstable.In this dissertation,a unified optimization model for simultaneously learning dynamic hypergraphs and discrete clustering labels is constructed by introducing a rotation matrix,in which the subspace features and hypergraphs can be updated adaptively according to the clustering results,with good self-learning ability and stable clustering results.The experimental results on real hyperspectral images show that the proposed methods achieve the best comprehensive performance.