Adaptive local feature based classification for multispectral data

Abstract

In this paper, a new adaptive feature selection based supervised classification technique in which features are selected locally rather than globally as in Principal Component Analysis (PCA) and Minimum Component Analysis (MCA) is presented. Classification techniques based on such global parameters tends to degrade because all classes are projected along the principal component direction for PCA and minimum component direction for MCA. All the classes are projected along these directions under the assumption that separability is uniform for all, which is not always true. The new Adaptive Feature Selection classification technique overcomes this disadvantage by selecting features based on the local information of the classes instead of global information. In addition, a minimum likelihood decision rule is employed instead of maximum likelihood decision rule. Good performance of our technique can be seen from the experimental results on the Kennedy Space Center (KSC) TM images.