Multi-Feature Based Low Rank And Sparse Representation For Hyperspectral Image Classification

Difference exists among the different spectral bands in the hyperspectral image,and it can be used for classification and target detection.Hyperspectral remote sensing technology plays an important role in environmental dectction,military,precision agriculture and mineralogy.So its classification is the premise of image understanding.With the development of remote sensing technology,spectral resolution and spatial resolution continue to improve,which increases the information and brings some problems.For example,how to effectivly extract information for classification;how to achieve the effective fusion between different features and so on.In order to solve these problems,this paper proposes a classification method based on sparse representation,low rank representation and extreme learning machine to realize hyperspectral image classification.The experiments are carried out on three hyperspectral datasets including Indian Pines,Salinas Scene and University of Pavia,and satisfactory classification accuracy is obtained.The main work of this paper is summarized as follows:1.Robust joint sparse representation based on multi-feature is proposed for hyperspectral image classification.Due to the complex content and background of hyperspectral images,the differences between classes and within classes vary greatly,and the combination of multiple features often helps to solve these problems.The traditional method is based on sparse representation principle,since a neighborhood can not guarantee that the same thing,and the differences between the characteristics,the use of SOMP solution will be more harsh.However,the use of sparse solution of the idea,in the choice of dictionary atoms,not only the use of common to find a common atom,while using their own characteristics to find a more suitable for their own atoms.This avoids the fact that elements in a neighbor do not belong to the same class and avoid the difference between features.2.A classification method of low rank sparse representation based on multi-feature association is proposed.First,after the superpixel segmentation,and the multi-feature extraction,a super-pixel block is selected under each feature,the pixel block in the respective dictionary is used to get a representation of the coefficient,low rank and sparse constraints on the cofficients can make full use of global and local information to achieve the fusion of features.3.ELM based on multi-task deep learning is proposed to classify hyperspectral images.Combined with the idea of deep learning,a deep model is designed for multi-feature learning.First,the characteristics of the extracted features are learned separately,then the learned features are combined as a new feature throught ELM.At the same time,in order to improve the separability of class,linear discriminant analysis(LDA)is used to study the discriminant information,and finally the classification is realized.