Research On Sparse Learning Based Semi-supervised Hyperspectral Image Classification Methods

Benefited from their high spectral resolution,hyperspectral images can distinguish similar objects.Hyperspectral image classification(HIC)is a typical method of hyperspectral image analysis and processing,which has attracted more and more attention in recent years.In general,a large amount of labeled samples are needed to obtain satisfactory results of HIC.However,the cost of labeling hyperspectral pixel is high,so it is an urgent problem to precise HIC with few labeling samples.In order to better solve such problems,this paper proposes two HIC models under the tensor framework,and the main research contents are as follows:(1)This paper proposes the semi-supervised tensor sparse coding model(SS-TSCM).In this model,two kinds of regularization terms are established under the tensor framework,which are denoted as sparse regularization term and semi-supervised regularization term respectively.The sparse regularization term guarantees the spatial neighboring pixels share the same sparse pattern under the dictionary,and the semi-supervised regularization term can effectively utilize the information provided by a large number of unlabeled samples.Then the two regularization terms are introduced into the sparse coding model,which can not only well retain the spatial-spectral structure of hyperspectral image,but also make use of the spatial-spectral information provided by a large number of unlabeled samples.The performance of our proposed SS-TSCM is evaluated on three real hyperspectral datasets.(2)This paper proposes the semi-supervised kernel tensor sparse coding model(SS-KTSCM).Based on the research of SS-TSCM,this model maps the spectral features to the high-dimensional feature space under the tensor framework to make them linearly separable.Thus,it can solve the problem of linear inseparable of spectral features,which is caused by complex scattering and atmospheric influence in the process of hyperspectral image acquisition.In addition,the kernel method is used to convert the inner product operation in the high-dimensional feature space directly to compute the kernel of spectral features,which greatly simplifies the operation.The performance of our proposed SS-KTSCM is evaluated on three real hyperspectral datasets.