Tensor Analysis Based Feature Extraction And Classification Of Hyperspectral Remote Sensing Image

There are tens or even hundreds of bands in spectral dimension of hyperspectral remote sensing images usually.On the one hand,the abundant spectral information makes it possible for us to carry out land-cover analysis.On the other hand,so much spectral information also brings huge storage and computing costs,as well as the curse of dimension and so on.There is large redundancy among different bands of hyperspectral remote sensing images.Information of different bands may interfere and conflict each other.At the same time,the mixed pixels in hyperspectral remote sensing images may lead to the phenomenon that the same class samples have different spectral image and the different classes samples have the same spectral image,which can reduce the representation of original spectral information in hyperspectral remote sensing images.So,how to extract feature of hyperspectral remote sensing data,as well as to enhance discriminative between different classes,is a very important problem in hyperspectral remote sensing domain.Low rank representation,sparse representation,graph model,discriminant analysis and other methods have been widely used in feature extract of hyperspectral domain and obtain good performance.However,the classical model of methods mentioned above mostly is based on vector samples of hyperspectral remote sensing images.We have to convert 3-dimension cube into 2-dimension matrix firstly when we use these representation methods,which may destroy spatial nearest neighbor information in hyperspectral remote sensing images.Therefore,we focus on tensor samples representation in hyperspectral remote sensing images and extend the low rank representation,sparse representation,graph model and discriminant analysis ways vector-based to tensor-based representation space in this paper.Based on this,we perform hyperspectral remote sensing images feature extract methods.The work of this paper is as follows:(1)Extend low rank representation and graph model to tensor space.Using tensor samples represent hyperspectral remote sensing data,which can keep the original spatial nearest neighbor information.In framework of tensor-based graph,we use the Kronecker product of tensor factor matrix in each mode represent the similarity metric between tensor samples.Compared with conventional construct graph methods,which method can make full use of the similarity in each mode of tensor data,thus way can produce stronger representation.In the end,we employ Tucker decomposition based on graph model achieve feature extract of hyperspectral remote sensing images.(2)Extend sparse representation and discriminant analysis to tensor space.Taking advantage of tensor samples represent original hyperspectral data,so that we can keep spatial nearest neighbor structure information in origin data.To make full use of sparse representation and discriminant analysis of tensor samples,we combine them into an unity framework,and exploit Lagrange multiplier method solve the object function,after obtain factor matrix in each mode,under Tucker decomposition framework,realize feature extract of origin hyperspectral data.(3)Extract Gabor feature,morphology feature of hyperspectral remote sensing images,and combine them with origin spectral feature by low-rank representation,achieve multi-feature low rank representation of hyperspectral remote sensing images.At the same time,to exploit manifold information in hyperspectral remote sensing images,extend local preserve projection method to tensor space and perform tensor local preserve projection model.After acquiring the most optimal factor matrix on each mode of tensor hyperspectral images based on this model,under Tucker decomposition framework extract hyperspectral images feature finally.