Class-Probability Based Semi-Supervised Dimensionality Reduction for Hyperspectral Images

Hyperspectral images (HSI)are very useful due to the rich information they contained. However, for the same reason, it is also inconvenient to be analyzed due to its high dimension and also because it contains a lot of redundant information. Therefore, dimensionality reduction (DR)is often an indispensable step for the analysis of HSI. Due to the expensiveness of labeling samples, semi-supervised learning technique that performs DR with only a small amount of labeled samples, has attract more and more attention during the past several years. In this paper, we propose a novel method called class probability semi-supervised DR (CPSDR). Unlike previously semi-supervised DR methods, which only focus on a small number of labeled samples and depend on their local geometry information, our approach also pay much attention on unlabeled samples. Moreover, in our approach, not only local geometry information but also class structure information was exploited. We then combined these two information together to yield a more discriminative scatter matrix. We formulate our problem as an optimization problem and solve it by eigenvalue decomposition. The experimental results on Salinas and PaviaU hyperspectral data suggested that our algorithm achieved state-of-the-art performance.