A Combined Framework for Dimensionality Reduction of Hyperspectral Images using Feature Selection and Feature Extraction

Due to the presence of large number of bands, computation overhead for classification of hyperspectral images becomes very high. Proper selection of subset of bands is necessary for reducing dimensionality. In this article while reducing dimension, discriminating capabilities of both extracted and selected features are exploited which, in turn, improves the performance of classification. Maximum margin criterion is used for extracting features while correlation based method is used to obtain subset of features. Two different approaches are considered for combining extracted and selected features. In the first method, combination of two types of features is done exhaustively where the numbers of each type of features may differ; while in the other they are kept the same. Thereafter, on this combined set of features, two popular feature selection algorithms namely, sequential backward selection (SBS) and ReliefF, are employed to reduce dimensionality Experiments have been conducted on three hyperspectral image datasets. Results show that, with such reduced combined features, classification accuracy of hyperspectral images is better in contrast to working with the set of extracted/ selected features having the same cardinality.