基于未标签信息主动学习算法的高光谱影像分类

Hyperspectral remote sensing image classification based on active learning algorithm with unlabeled information

针对高光谱遥感影像分类中,传统的主动学习算法仅利用已标签数据训练样本,大量未标签数据被忽视的问题,提出一种结合未标签信息的主动学习算法。首先,通过K近邻一致性原则、前后预测一致性原则和主动学习算法信息量评估3重筛选得到预测标签可信度高并具备一定信息量的未标签样本;然后,将其预测标签当作真实标签加入到标签样本集中;最后,训练得到更优质的分类模型。实验结果表明,与被动学习算法和传统的主动学习算法相比,所提算法能够在同等标记的代价下获得更高的分类精度,同时具有更好的参数敏感性。

In hyperspectral remote sensing image classification, the traditional active learning algorithms only use labeled data for training sample, massive unlabeled data is ignored. In order to solve the problem, a new active learning algorithm combined with unlabeled information was proposed. Firstly, by realizing triple screening of K neighbor consistency principle, predict consistency principle, and information evaluation of active learning, the unlabeled sample with a certain amount of information and highly reliable prediction label was obtained. Then, the prediction label was added to the label sample set as real label. Finally, an optimized classification model was produced by training the sample. The experimental results show that, compared with the passive learning algorithms and the traditional active learning algorithms, the proposed algorithm can obtain higher classification accuracy under the precondition of the same manual labeling cost and get better parameter sensitivity.