Research On Hyperspectral Dimensionality Reduction And Classification Based On Spatial-Spectral Sparse Structure Learning

Hyperspectral remote sensing technology is a significant progress of the earth observation in last century.The spectral resolution of remote sensing images is raised to nanoscale for the first time,which provides more abundant spectral information for classifying different kinds of ground object precisely.However,due to the mutual restriction of sensor space and spectral resolution as well as the complexity and diversity of the distribution of ground objects in nature,the automatic interpretation of hyperspectral remote sensing images still has a lot of problems: 1)There are a large amount of redundant information and high correlation in the fine spectral,which makes the information representation and the feature extraction of land objects get into trouble;2)As a result of the mixed pixel problems that are general to the hyperspectral remote sensing images,it's common that the different bodies have the same spectrum and the same bodies have the different spectrum.Meanwhile,the classification accuracy can't satisfy the requirement of practical applications.3)The accuracy and intelligent level of the land objects classification are insufficient because the statistical property of the intra-class and inter-class is sophisticated and the real samples are lacking.Thus,attribute reduction(feature selection and extraction)of the hyperspectral remote sensing images and the intelligent classification of ground objects have become the key issues that need to be solved urgently in remote sensing domain.Based on the National Basic Research Program(973 Program)of China(No.2013CB329402)and the National Natural Science Foundation of China(No.91438103),this dissertation researches the hyperspectral remote sensing images dimensionality reduction and classification by drawing lessons from the sparse perception and cognitive mechanism in the process of biological cognition and the theories of the sparse representation learning,structural learning and semi-supervised learning etc.The main content and the innovation of this research are as follows:1)A Structure Regularied sparse Band Selection(SRe BS)via learned pairwise agreement was proposed.Discriminative features usually have sparse characteristics.Constructing a coding cost function consisting of pairwised agreement and responsibility matrix,the sparse band subset selection problem is modeled as the least cost problem under the sparse constraint of the responsibility matrix.First,using non-negative low rank representation to obtain a accurate spectral pairwise agreement,and then refine it by the relative position relationship between the bands.Second,a manifold regularization term is introduced into the coding cost function to ensure the smoothness of the optimal band subset in the local spatial area.A multi-multiplier alternation optimization method is designed for the new objective function,and the optimal feature subset reflecting the feature of ground features is optimized and solved.Experimental results show that the algorithm can effectively reduce the band redundancy and improve the classification accuracy.2)A sparse feature learning algorithm based on spatial-spectral compressed tensor coding is proposed.The inherent high-order tensor structure of spectral data,which takes into account its tensor characteristics in data processing,helps to improve the processing performance.On the other hand,band selection in the original data space has a clear physical meaning,but the classification performance is generally lower than that based on feature transformation.Aiming at the above problems,a feature learning method based on sparse coding of kernel tensors is proposed.In the high-dimensional kernel tensor space,a feature subset with higher discriminative power is selected.First,the three-dimensional tensor spectral data is mapped to a higher-dimensional tensor kernel space by the tensor kernel mapping trick,and the feature selection is modeled as a sparse tensor encoding problem in the kernel space.Furthermore,by introducing the spatial local to construct the spatial-spectral tensor,and the sparse band selection model of cooperative coding of space-spectrum compression tensors is constructed.A Tensor Multiple Measure Vector(TMMV)optimization algorithm is designed accordingly.Experiments on synthetic and real hyperspectral data show that the proposed method has better classification accuracy than the traditional method.3)A Superpixel Tensor Sparse Coding(STSC)based Hyperspectral Image Classification(HIC)method is proposed.Hyperspectral data is not only a set of spectral data,but also image data with specific spatial semantics.In order to obtain more accurate classification performance,a classification method based on sparse coding of structural superpixel tensor by combining the tensor sparse coding model with spatial semantics.Firstly,a superpixel segmentation method based on Hierarchical Spatial Affinity Propagation(HSAP)method is designed to obtain local semantic information of image data.Secondly,superpixel tensor is constructed for each superpixel and its corresponding neighbor superpixels,and a superpixel tensor joint sparse coding model under the labelled sample is established to obtain superpixel class labels through the least encoding errors.Meanwhile,in order to refine the classification results,sparse coding classification of the boundary pixels is introduced.Fi-nally,the pixel-superpixel label information is integrated using the principle of majority voting to obtain the final label.The experimental results on the hyperspectral data show that the classification method based on sparse coding of superpixel tensors can effectively describe the spatial semantic information,remove the pixel-wised classification errors.4)A semi-supervised subspace learning algorithm based on low-rank structure is designed.In addition to the sparse properties in spectral space,high-order tensor space,and tensor kernel space,hyperspectral data also has structural sparsity,i.e.low rank property.Considering a small number of sample conditions,in order to make full use of the un labelled samples,a low-rank manifold is established based on the low-rank representation,which effectively describes the global structural sparsity of the data.In the supervised subspace learning model,the objective function including the low rank manifold regularization and the maximum discriminative spatial-spectral margin is constructed,and a semi-supervised subspace learning method based on the low rank structural regularity is proposed.The proposed method automatically determines the optimal dimensionality of the subspace and can achieve higher classification performance under a small number of labeled samples.5)A semi-supervised HIC method with dual geometric low-rank learning was proposed.In order to explore the geometric structural characteristics of hyperspectral data in the spectral space and the local spatial domain,a dual geometric structure learning model based on low-rank representation of global spectral structure learning and local spatial structure learning was constructed.Then the local spatial Geometry regularized Laplician Low Rank Representation(GLap LRR)method was constructed to learn the geometric similarity between the samples.And then construct the regularization term of the manifold between samples according to the learned GLap LRR.Finally,a semi-supervised Support Vector Machine(SVM)classification framework based on the GLap LRR graph regularization term was designed.The experimental results show that the proposed dual-geometric low-rank learning model can effectively characterize the spatial-spectral structure of hyperspectral data and improve the classification accuracy.