Two Effective Feature Selection Criteria for Multispectral Remote Sensing

In an earlier study, Swain et al. reported on two statistical separability measures which for multiclass feature selection were shown experimentally to be more reliable than divergence. However, the empirical results of that study together with the best theoretical results in the literature left open some practical questions regarding the quantitative characterization of these separability measures. This paper is concerned with an empirical study aimed at answering such questions. It has been possible to further substantiate that the Jeffreys-Matusita Distance and a saturating transform of divergence are effective feature selection criteria for remote sensing applications. In fact, an explanation as to why this should be the case has now been made apparent.