中文摘要：

针对图的半监督聚类算法（Semi-Supervised Graph-Based Clustering,SSGC）中出现的对先验信息利用不充分、不足以应对数据异构、计算耗时大等问题,本文提出一种基于半监督复合核的图聚类算法,并应用于高光谱图像。该算法首先通过引入半监督学习方法对径向基函数（Radial Basis Function,RBF）进行了改进,以充分利用少量的标记样本和无标记样本;其次将RBF核与光谱角核进行融合,构造复合核权重矩阵。在权重矩阵的构造过程中,K-近邻方法的引入也简化了计算过程。在Indian Pine和Botswana高光谱数据集上的实验结果表明,相对于SSGC算法,本文算法不仅实现了更高的分类正确率,其总体分类精度提升1%～4%,而且有效提升了运算速度。

英文摘要：

A semi-supervised graph-based clustering method is presented with composite kernel for the hyperspectral images,mainly to solve the problems existed in an algorithm called Semi-Supervised Graph-Based Clustering（SSGC） and improve its performance.As for the realization,it firstly reforms the Radial Basis Function（RBF） by adopting semi-supervised approach,to exploit the wealth of unlabeled samples in the image.Then,it incorporates the spectral angle kernel with RBF kernel,and constructs a composite kernel.At last,the use of K-Nearest Neighbor（KNN） method while constructing the weight matrix has greatly simplified the calculation.Experimental result in Indian Pine and Botswana hyperspectral data demonstrates that this algorithm can not only get higher classification accuracy（1%～4% higher than SSGC,10%～20% higher than K-means and Fuzzy C-Means（FCM）,but effectively improve operation speed compared with SSGC.