Hyperspectral Remote Sensing Image Classification Algorithms Based On Spatial-spectral Tensor

Hyperspectral remote sensing image has a large number of spectral dimensions and a lot of redundant information,and it is difficult to effectively and completely represent hyperspectral remote sensing image information only with spectral vectors,which leads to the limited classification ability of traditional classification methods for complex ground objects.In view of the above problem,this paper proposes two spatial-spectral tensor based hyperspectral remote sensing image classification algorithms under the framework of tensor theory.The spatial spectral tensor,composing of spectral vectors of the central pixel and its spatial neighborhood pixels in hyperspectral remote sensing image,can more completely represent the spatial and spectral information of the pixel,which is conducive to the accurate classification of ground objects.The kernel functions can only deal with the vector data in traditional Support Vector Machine(SVM)classification models.When the input data is tensor,it can only be used after vectorization,resulting in the loss of internal structure information of tensor data.Combining Tensor Algebra(TB)with Singular Value Decomposition(SVD),this paper proposes a Tensor based Radial Basis Function(Tensor-RBF)algorithm based on classical Radial Basis Function(RBF).The SVM multi-classes classification model is designed to classify the hyperspectral remote sensing image by combining one-against-one(OAO)parallel strategy.Aiming at the low-rank and sparse properties of spatial-spectral tensors,a Tensor based Sparse Representation(Tensor-SR)algorithm is proposed based on the traditional vector based Sparse Representation(SR)algorithm and TB theory in this paper.The sparse representation tensors are extracted,which have lower dimension and complete internal structure information of spatial-spectral tensor,as the basic unit of subsequent classification.The Support Tensor Machine(STM)multi-classes classification model,which directly acts on sparse representation tensors,is designed to classify the hyperspectral remote sensing image.To qualitatively and quantitatively analyze the effectiveness and feasibility of the proposed algorithms in the process of hyperspectral remote sensing image classification.The contrast experiments of hyperspectral remote sensing images from different regions acquired by different sensors are carried out by using the proposed algorithms and contrast algorithms in this paper.User accuracies,product accuracies,overall accuracies and kappa coefficient of all the classes are calculated by the confusion matrix.The experimental results show that the spatial-spectral tensor can represent the pixel information of hyperspectral remote sensing image more completely than a single spectral vector.In addition,the two classification algorithms proposed in this paper can more effectively classify ground object hyperspectral remote sensing images,and the hyperspectral remote sensing image classification based on spatial spectral tensor framework has considerable research potential.