Feature Selection for Classification of Hyperspectral Remotely Sensed data using NSGA-II

This paper summarizes the implementation and performance of Nondominated Sorting Genetic algorithm (NSGA-II) [2] for feature selection of remotely sensed hyperspectral imagery. Two step processes have been followed. In first step, a feature subset is selected with optimum spectral and texture information content resulting in a smaller space to be searched in the next step. In the second step, a single objective search algorithm is used to obtain a final smaller subset out of the features already selected (with optimum information content), which have best separability between the classes. Classes are obtained by classifying the subset bands using maximum likelihood classification algorithm. Method of spectral and textural information evaluation of images, genotypic representation of our algorithm, classification methodology and separability criteria between classes have been discussed. Also discussed is the reason for choice of NSGA-II and a strategy to extract optimum results from it.