关键词: 高光谱影像分类, 图的半监督学习, 主动学习, 空-谱特征

Abstract:

Currently, it is difficult and time-consuming to obtain enough labeled samples for hyperspectral image(HSI) classification, while numerous unlabeled samples can be easily identified but unused. Here, in order to overcome these shortcomings, we proposed a semi-supervised graph-based combined with active learning mechanism approach in this paper. Firstly, we extracted the spatio-spectral feature by reorganizing the spectrum of a pixel with its neighbors, followed by a sorting scheme to make the feature representation to be rotation invariant. Then, the most uncertain samples(namely largest ambiguity samples for classifier) were selected for operator to label with active learning algorithm. Finally, both labeled samples and unlabeled samples were used for semi-supervised classification. The proposed algorithm could guarantee that boundary samples were selected, which would help construct the boundary of a classifier. Simultaneously, even when less labeled samples were available, the proposed method achieved a good classification result by introducing a large number of unlabeled samples. The experimental results on three real hyperspectral images confirmed that the proposed method can obtain higher classification accuracy.

Key words: hyperspectral image classification, semi-supervised graph-based learning, active learning, spatio-spectral feature