论文部分内容阅读
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。
Taking the model maximum likelihood estimation as an example, a semi-supervised classification method combining labeled and unlabeled samples is introduced to solve the small sample problem in remote sensing image classification. A classifier is initialized with a few existing labeled samples, The unlabeled samples are recursively calculated to optimize the classifier until the likelihood function containing all the samples converges to a local maximum. By analyzing the relationship between the classification of remote sensing images and the intrinsic features of images in images, we design two classification experiments under the assumption of different generative models. The results show that the participation of unlabeled samples can greatly improve the accuracy of image classification under the condition of small sample, but the number of two samples should maintain an appropriate ratio. Finally, by comparing with the classification of SVM which has unique advantages in solving small sample classification problems, we find that this method has better potential in the case of small samples.