以生成模型最大似然估计为例, 引入结合已标记样本和未标记样本的半监督分类方法来解决遥感影像\n分类中的小样本问题, 应用已有的少量已标记样本初始化一个分类器, 结合大量未标记样本, 通过递归计算的方式\n对分类器进行优化, 直到包含所有样本的似然函数收敛到局部极大值。通过分析遥感影像待分类类别与影像中地物\n类型固有特征之间的关系, 设计两个在不同生成模型假设下的分类实验。结果表明, 未标记样本的参与可在很大程\n度上提高小样本条件下的影像分类精度, 但两种样本的数量应保持一个适当的比例。最后通过与在解决小样本分类\n问题方面有独特优势的SVM 方法的分类比较, 发现在小样本情况下, 本文方法具有更好的应用潜力。

This paper proposes a generative model based semi-supervised learning method of remote sensing image\nclassification, which makes use of both the labeled and unlabeled samples to handle the insufficient labeled training samples\nproblems. We first train an original classifier by the small number of labeled samples alone. Then we re-train it by both the labeled\nand a large amount of unlabeled samples. This process is iterated until the likelihood function of all the samples are converged to\nthe local maxima. Through the designed experiments of the two different mixture models, It is found that the unlabeled samples\nhelp us to get the method to enhance the classification performance to a large extent on condition, which the ratio of the unlabeled\nsamples to the labeled ones must be appropriate. Thus, we have also compared the method by using the state-of-the-art support\nvector machines (SVMs) with the same labeled samples, of which results show that our method works better.