Hyperspectral Imagery Classification Methods Based On Sparse Representation And Spatial Information

Hyperspectral remote sensing technology,as a kind of advanced technology,has been widely used in many fields,including agricultural production,mineral mapping,target detection,disaster warning,military reconnaissance and urban planning.The hyperspectral images we collected have very high spectral resolution and spatial resolution,the characteristics of hyperspectral data are the large amount of data and redundancy,high dimensionality and there exists a strong correlation between the bands,thereby,it can provides rich information which also brings challenges for image interpretation.Classification plays a very important role in hyperspectral image processing,how to make use of the rich image information efficiently and ensure the interpretation accuracy in the meanwhile have gained increasing attention recently.Meanwhile,compared to the traditional classification methods,sparse representation classification method has shown its great potential in data processing field,and it quickly becomes a research hotspot.Based on the actual projects background,this paper has been written in the framework of hyperspectral image classification theory and methods,and it has focused on the two directions which are how to more effectively use the sparse representation and how to more fully mine the information in the hyperspectral image,and expanded the related researchs and discussions on how to effectively combine the spatial information in the sparse representation,aiming to eliminate the sporadic misclassified pixels often exist in the classification map and achieve high accuracy classification.Main works of this paper are shown as the following:1.In order to have a better expression of hyperspectral images,and use the spatial neighborhood information eliminate the sporadic misclassified pixels often exist in the classification map,a new hyperspectral image classification method based on sparse feature and neighborhood homogeneity is proposed in this paper.In order to make the image well expressed and also greatly facilitate the subsequent processing based on the sparse representation,the method uses the sparse features of the image which contain only a small amount of non-zero elements as the classification features.Then the extracted sparse features for classification with Support Vector Machine(SVM)are used,which can obtain the initial classification results based on sparse features.To improve the classification accuracy,the spatial correlation between neighboring pixels in the image is considered by using the neighborhood homogeneity.Theoretical analysis and simulation results illustrate the good classification performance of the proposed method.2.In order to make full use of sparse feature produced by the sparse representation,while through the Markov Random Field(MRF)effectively combine the contextual information of the hyperspectral image,a new hyperspectral imagery classification method based on sparse feature and Markov Random Field is presented in this paper.A dictionary learning process is first employed to find a set of basis vectors,which is capturing high-level semantics of the hyperspectral image data,and then the sparse features of the data can be extracted through the sparse representation,so the implicit information in the data is made explicit.Using the sparse features for the probabilistic SVM and it get every pixel's probability estimates for the individual classes.Meanwhile,the spatial information based on the MRF is incorporated in the classification steps by modifying the form of a probabilistic discriminative function via adding a term of contextual correlation,which can achieve the final classification results.The experimental results verify that the effectiveness of the proposed classification method.3.Based on the point of using the spatial-spectral feature to give a more comprehensive expression for the hyperspectral image data and more effectively combining the sparse representation classifier to improve the classification performance,a new hyperspectral image classification method based on spatial-spectral features and sparse representation is raised in this paper.In the classification of hyperspectral data,when just using spectral features while ignoring the image spatial information has a shortcoming that better classification performance could not be achieved.In order to avoid the above shortcomings it first combines the spatial and the spectral features,which are extracted from the original image to form the spatial-spectral features to effectively represent the hyperspectral image,then constructs the training set based on the spatial-spectral features and gives the sparse representation for the hyperspectral image,and finally determines the class label by calculating the corresponding redundancy of each class of the reconstructed image to complete the classification process.The experimental results have shown that the proposed method can improve the classification performance by effectively utilizing the spatial-spectral features.4.Motivated by making full use of the spatial-spectral information of the imagery and considering the neighborhood pixels correlation and the locality to achieve an improved classification performance,meanwhile,fusing the sparse-collaborative representation to combine classification process into a unified framework.A new residual fusion classification method based on the spatial-spectral information and sparse-collaborative representation is put forward in this paper.Firstly,the hyperspectral imagery is represented by the spatial and the spectral features effectively,then,the Joint Sparse Representation Classification(JSRC)and the Locally Joint Collaborative Representation Classification(LJCRC)methods are introduced to process the features,finally the residuals of the different classification methods are fused and the class label can be determined by calculating the corresponding fused residual.Experimental results show that:compared with the state-of-the-art classification methods,the proposed method achieves better classification performance.