Abstract

We study feature selection and classification for a land-cover mapping task from airborne high resolution AeS-1 data in radar X- and P-Band. The studied feature selection methods are the well-established sequential addition of features to an initially empty feature set and genetic algorithm search, both based on statistical distance measures, and exhaustive evaluation based on actual classifier performance. It was observed, that different criteria and search strategies come up with different subsets of features. We present results of combined classifications derived from classifiers trained on different subsets of features. The considered combination strategies are product, sum, maximum and majority rules. Combination turned out to bring significant improvement. The task of discriminating two forest classes, three classes of agricultural area, two classes of built-up area and a specific class devoted to radar imaging ambiguities for a test site located in Switzerland provided a satisfying result for machine classification from radar data.

Citations