中文摘要：

提出一种利用属类概率距离构图的半监督学习算法，并利用高光谱图像数据集进行了性能测试。首先，该算法利用基于分类的稀疏表达方法来预估未标记样本的属类概率向量；然后，利用这个概率向量对描述数据相似性的距离函数进行改造，改造后的距离函数能有效扩大异类样本点之间的距离，在新的距离函数的度量下，每个样本点的邻域中可包含更多同类的样本点；最后，将该距离函数应用于半监督学习线性邻域传播算法标签传播算法中。在Hyperi on和AVIRIS高光谱遥感图像上进行试验，结果表明，相比于传统的基于图的半监督学习算法，该算法能有效提高高光谱遥感图像分类精度。

英文摘要：

A class‐probability distance based semi‐supervised learning method is proposed for hyperspectral remote sensing image classification .In the method ,sparse representation based classification （SRC） is adopted for estimating the class‐probability of unlabeled sample .Then a distance metric that describes the data simil arity is developed based on the estimated cl ass‐probability .With this new distance metric ,the distance between samples of different classes is enlarged effectively ,and the neighbors of each sample can contain more samples belonging to the same class .Finally ,this distance metric is applied to linear neighborhood propagation and l abel propagation algorithms .Experimental results using Hyperion and AVIRIS hyperspectral remote sensing images show that the approach outperforms the existing semi‐supervised learning methods in terms of classification accuracy .