Hyperspectral Image Clustering Based On Spectral-Spatial Collaborative Framework

Hyperspectral image clustering is an important part of the remote sensing data analysis.It is one of the common and effective measures for people to extract and use the remote sensing information.In the hyperspectral images,because the spectral information shows the characteristics of high intra-class difference and high inter-class similarity,it is difficult to obtain high accuracy clustering results when only spectral information is used for the hyperspectral image clustering.The hyperspectral images have the characteristics that the spectral changes are large,the spatial structure is complex,the number of labeled samples is small and the amount of data is large.In order to meet the requirements of the clustering accuracy,based on multi-prototype clustering,anchor extraction and sparse subspace clustering,two hyperspectral image clustering algorithms based on spectral-spatial collaborative framework are proposed by using the spectral and spatial information.The main work is as follows:1.When using the method of anchor extraction for spectral clustering of hyperspectral images,the random extraction method is generally used to select the anchors,and then use them to construct the similarity matrix.However,this extraction method may lead to the misalignment of the measurement of similarity relationship due to the large correlation between the anchors.Therefore,the hyperspectral image spectral clustering based on anchor extraction with K-multiple-means is proposed.First,multi-prototype clustering of hyperspectral images is carried out by using the K-multiple-means clustering method.Then,the obtained multiple prototypes are set as the anchors,and the objective function is constructed by combining the spatial information.And the similarity matrix and anchor graph are further constructed under the spectral-spatial cooperative framework.Finally,the spectral analysis on anchor graph is performed to obtain the clustering results.2.When the sparse subspace clustering is performed in the hyperspectral images,there exist the problems of high dimensionality,large amount of data and strong correlation of adjacent samples,so it may be disturbed by redundant information when constructing the similarity matrix,which in turn leads to the decline of the similarity matrix's ability to characterize the similarity relations between samples.And it also causes the problem of excessive storage burden.Therefore,the sparse subspace clustering for hyperspectral image based on anchor extraction is proposed.First,the hyperspectral image is preprocessed by the hierarchical guidance filtering.Then,the obtained spectral-spatial representation is sparse by adding constraints combining spatial information to the objective function under the spectral-spatial cooperative framework.Finally,the anchors are extracted in the sparse coefficient matrix,and the subspace clustering with the spectral-spatial cooperation of the hyperspectral images based on anchor extraction is carried out to obtain the clustering results.This thesis carries out the experiments on three real hyperspectral image data sets named Indian Pines,Pavia University and Salinas to validate the effectiveness of the methods.The results of the experiments reveal that the two hyperspectral image clustering algorithms proposed in this paper can achieve higher clustering accuracy compared with the traditional clustering algorithms.