Dimensionality reduction of hyperspectral data: band selection using curve fitting

Abstract

Hyperspectral sensors offer narrow spectral bandwidth facilitating better discrimination of various ground materials. However, high spectral resolutions of these sensors result in larger data volumes, and thus pose computation challenges. The increased computational complexity limit the use of hyperspectral data, where applications demands moderate accuracies but economy of processing and execution time. Also the high dimensionality of the feature space adversely affect classification accuracies when the number of training samples is limited – a consequence of Hughes’ effect. A reduction in the number of dimensions lead to the Hughes effect, thus improving classification accuracies. Dimensionality reduction can be accomplished by: (i) feature selection, that is, selection of sub-optimal subset of the original set of features and (ii) feature extraction, that is, projection of the original feature space into a lower dimensional subspace that preserves most of Information. In this contribution, we propose a novel method of feature section by identifying and selecting the optimal bands based on spectral decorrelation using a local curve fitting technique. The technique is implemented on the Hyperion data of a study area from Western India. The results shows that the proposed algorithm is efficient and effective in preserving the useful original information for better classification with reduced data size and dimension.