Improvements to the effectiveness of supervised training procedures

Abstract

Two simple methods are proposed to improve the effectiveness of the supervised training in an interactive environment. An iterative procedure for cleaning training data is introduced to provide more appropriate class statistics (in terms of normality criterion) for classification. A binary graph representation of the training category set—called a separability tree—makes the checking of training statistics significantly easier. Some examples are presented with Landsat Thematic Mapper data concerning the use of both techniques.