Main direction - An effective statistic for hyperspectral image classification

In the hyperspectral image classification literature, mean and covariance are prominently used statistics. The first order statistics, i.e., mean, although easily estimated, can have low discriminative property. However, the discrimination power of the second order statistic, i.e., covariance, is strong. The downside of the estimation of covariance is the large sample size. In this paper, we propose a effective statistic named main direction. The main direction is defined as a direction in which the variation or the distance of the samples becomes the largest. Besides, the estimation of the main direction is stable even when the size of the sample set is very small. Based on the main direction, we propose a modified spectral angle mapper classifier and demonstrate the effectiveness of the modified classifier with experimental results.