In a typical classification procedure of high dimensional data, feature extraction is first applied to reduce the dimensionality and a classifier is employed. However, in most feature extraction methods, covariance matrices must be estimated. When training samples are limited, this estimation is inherently biased, thereby generating ineffective features. In this paper, we propose a new feature extraction method for high dimensional hyperspectral data when limited training samples are available. In the proposed method, we construct a feature matrix using available training samples. The proposed method calculates the difference vector feature matrix using weighted difference vectors among the training samples. Experimental results show that the proposed method improves classification accuracy even if the size of training sample is very small.