Data-dependent semi-supervised hyperspectral image classification

Hyperspectral imagery provides more powerful information than multispectral remote sensing data. However, when hyperspectral data is used for classification task, the high-dimension features often lead to ill-conditioned problems, such as the Hughes phenomenon. To tackle this problem, various supervised dimensional reduction methods are proposed. However, these methods only exploit the labeled training data and ignore the huge unlabelled data. To utilize the unlabelled data space structure information in dimension reduction, a method is proposed as Data-dependent semi-supervised (DDSS). The proposed method exploits the space structure of labeled data and unlabelled data jointly to reduce the dimensionality of the image cures. Experimental results show that this method significantly outperforms the state-of-the-art dimension reduction methods for classification and denoising.