【摘要】 针对高光谱遥感图像分类中标记样本难获取的问题,提出了一种基于同质区和迁移学习的新型半监督分类方法。首先对高光谱图像进行分割得到高纯度的同质分割斑块,获取大量扩展训练样本。并在此基础上引入迁移学习,将扩展训练样本作为源域,剩余未标记样本作为目标域,实现多次迁移,从而减少同一幅图像上各地物的分布差异,并保留其各自的内部属性。实验结果表明,该方法是一种有效的高光谱图像半监督分类方法。更多还原

【Abstract】 In the field of hyperspectral remote sensing images classification, it is difficult to acquire the labeled samples. In this paper, a new semisupervised classification based on homogeneous region and transfer learning is presented to exploit and utilize as much information as possible to realize the classification task. Firstly the hyperspectral image is segmented to high-purity homogenous patches, and a large number of extended training samples are obtained from these patches. Then, the transfer learning is introduced to realize multiple transfer by using the extended training sample as the source domain, and the remaining unlabeled samples are considered to be the target domain. In this way, the distribution differences of land features on the same image are reduced and their internal attributes are retained. Experimental results with real data indicate that it is an effective semisupervised classification method for hyperspectral image.更多还原