Feature Subspace Detection for Hyperspectral Images Classification using Segmented Principal Component Analysis and F-score

Remotely sensed images provide unseen characteristics of the earth's surface as they are composed of hundreds of spectral bands of the electromagnetic spectrum. Both spatial and spectral information of hyperspectral image make it possible to categories vegetation and recognize earth's minerals and materials. But analysis of hyperspectral data suffers from curse of dimensionality due to the huge number of spectral bands or features. Not all features contain useful information and additionally, there is redundant information in some features. This paper proposes a model for detecting effective feature subspace from original hyperspectral data using both Segmented Principal Component Analysis and F-score methods. Depending on the correlation of the spectral bands, the data cube is partitioned into subgroups. Then principal component transform is performed on each subgroup. Finally, the most informative feature subspace is selected by considering discriminative characteristics of the features using F-score method. Two real hyperspectral images are used in the implementation of the proposed model. The classification accuracy of the proposed approach shows the superiority over other studied methods.