In statistical analysis of remotely sensed data, means and variances of each classes are used as the basis of statistical similarity determination. Therefore, the overall accuracy of classification is affected by the training results. It is assumed that the ideal distributions of pixel values follow normal distributions, but practically they have some aggregations and biases. non anomalies of distribution can affect the classification results greatly as well as the variances of training results. In this study, relationships between the inferential variances of the training sets and the distributions of pixel values are examined. and the resulting changes of classification results are studied. Furthermore, the training method which minimizes the effect of underestimation of variances is proposed.