Effect of Search Methods on Feature Selection with Hyperspectral Data

Abstract

In this work, we study the effect of different search methods on feature selection using filter and wrapper-based feature selection techniques. Eleven swarm-based search methods (ant search, bat search, bee search, cuckoo Search, elephant search, firefly search, flower search, genetic search, PSO search, and evolutionary search) using Correlation-based Feature Selection (CFS) and support vector machine (SVM) as filter and wrapper methods respectively. Comparison of results in terms of numbers of the selected feature suggests that different search methods output the different number of features with a minimum of 24 and 33 by Best First approach with CFS and SVM wrapper methods respectively whereas maximum number 100 by Elephant Search approach with CFS and 203 by firefly search method with wrapper methods. Comparison in terms of classification accuracy using support vector machine classifier suggests that genetic search method with 100 number of features achieve comparable accuracy to that with full dataset using CFS whereas Bee search method with 168 number of features provides comparable accuracy with wrapper-based feature selection approach. Results from this study suggest that choice of search methods is found to affect both the number of selected features as well as the accuracy achieved by selected features, suggesting the need for further studies using different hyperspectral datasets