Feature selection for ERS-1/2 InSAR classification: high dimensionality case

Abstract

A systematic way of selection and assessment of the performance of a large number of texture features extracted from spaceborne interferometric SAR data and classified with different types of classifiers is presented. Multi-seasonal ERS-1 and ERS-2 SAR data of the Czech Republic is used to classify into four different land-cover classes. A multistage search method in the space of all possible feature subsets taken from local statistics, fractal analysis and co-occurrence matrices is proposed and tested. In the early stages of the method, features are ranked according to its discriminatory power measured by a ranking coefficient based on subset performance measured by Jeffreis-Matusita-distance. Best ranked features are chosen and a new set is formed and evaluated using the hold-out method employing maximum-likelihood, nearest neighbor and multilayer perceptron classifiers.