Analyzing the Gaussian ML classifier for limited training samples

The Gaussian ML classifier is one of the most widely used classifiers for remotely sensed data since it is easy to implement and relatively fast. However, as the dimension of hyperspectral images significantly increases, the performance of the Gaussian ML classifier suffers when training samples are not enough, mainly due to inaccurate estimation of covariance matrices. In this paper, we provide thorough performance analyses of the Gaussian ML classifier in terms of the number of training samples. In particular, we analyze how decision boundaries which the Gaussian ML classifier defines vary when limited training samples are available. In order to quantify variations of decision boundaries, we introduce two distance measures. Experimental results show that there is a significant variation in covariance and mean estimation, which subsequently produces noticeably different decision boundaries.