Modified Co-Training With Spectral and Spatial Views for Semisupervised Hyperspectral Image Classification

Hyperspectral images are characterized by limited labeled samples, large number of spectral channels, and existence of noise and redundancy. Supervised hyperspectral image classification is difficult due to the unbalance between the high dimensionality of the data and the limited labeled training samples available in real analysis scenarios. The collection of labeled samples is generally hard, expensive, and time-consuming, whereas unlabeled samples can be obtained much easier. This observation has fostered the idea of adopting semisupervised learning techniques in hyperspectral image classification. In this paper, a semisupervised method based on a modified co-training process with spectral and spatial views is proposed for hyperspectral image classification. The original spectral features and the 2-D Gabor features extracted from spatial domains are adopted as two distinct views for co-training, which considers both the spectral and spatial information. Then, a modified co-training process with a new sample selection scheme is presented, which can effectively improve the co-training performance, especially when there are extremely limited labeled samples available. Experiments carried out on two real hyperspectral images show the superiority of the proposed semisupervised method with the modified co-training process over the corresponding supervised techniques, the semisupervised method with the conventional co-training version, and the semisupervised graph-based method.