Semi-supervised Graph Clustering With Spatial-spectral Kernel And Its Application In Hyperspectral Image

Because of its rich information and high resolution, hyperspectral remote sensing image has the ability to accurately describing the ground objects. However, due to its high number of dimensions and large amount of data, hyperspectral data can no longer use the conventional image classification means, then how to effectively excavate useful information to achieve highly accurate classification is already a serious problem.Based on the characteristics of hyperspectral image, this paper respectively studied the graph-based clustering algorithm and spatial-spectral based clustering algorithm, the main contents are:(1)The implementation process of semi-supervised learning, graph-based clustering, kernel clustering and spatial-spectral based clustering algorithm was summarized, as well as some common clustering algorithms and clustering evaluation being introduced.(2)Semi-supervised graph clustering with composite kernel(SSGCK) algorithm was proposed for hyperspectral images. To solve the problem of tedious calculations and inefficient use of label information possibly existed in traditional clustering algorithms, this paper used K-nearest neighbor(KNN) method and semi-supervised learning method during transforming and combining the radial basis function(RBF) kernel and spectral angle kernel, got a neighbor matrix based on composite kernel, and then utilized spectral graph theory for clustering, to better revel the internal structure of the hyperspectral data, and obtained the SSGCK algorithm. The experimental results on Indian Pine and Botswana hyperspectral data sets showed that the SSGCK algorithm obtained a more accurate clustering accuracy, and could significantly reduce computation time.(3) An algorithm used spatial-spectral information was put forward. After introducing spatial knowledge and the means of combining spectral and spatial information, according to the characteristics of the spatial distribution of hyperspectral images, that is, neighbor points for the same, this paper used spatial information to revise the similarity matrix based calculated with spectral information, then got the semi-supervised graph clustering with spatial-spectral kernel(S4GC) algorithm. The experimental results on Indian Pine and Botswana hyperspectral data sets showed that, with respect to SSGCK algorithm, clustering accuracy of S4 GC are greatly enhanced, have better results for feature division on hyperspectral remote sensing images.In summary, this paper studied the graph-based and spatial-spectral based clustering method on hyperspectral remote sensing images and then proposed two new algorithms, the experiments on hyperspectral data sets proved the good effectof the two algorithms.