Spectral Subspace Approach in Top Ranked Band Selection for Hyperspectral Image Classification

Abstract

In this paper, spectral subspace decomposition technique is incorporated in the band selection procedure of hyperspectral image data. The ranking of bands based on the information content is indicated by entropy measure and the selection based on class seperability is indicated by spectral divergence. Spectral subspace decomposition is employed to ensure the selection of top ranked bands dispersed across the entire bandwidth without accretion in one region due to high correlation. The ranking of bands after decomposition of spectral space into 3 subspaces such as Visible, Near Infra Red and Short Wave Infra Red has provided better fraction of bands from all the three subspaces. The merit of spectral subspace decomposition is tested through wetland classification in Muthupet lagoon of South East Indian coastline. The incorporation of the subspace at the time of ranking has provided improved classification accuracy with minimal loss of spectral information than the one without decomposition. Here, the overall accuracy of 5 top ranked bands without subspace decomposition is 81.7% due to accretion in NIR region whereas with spectral subspace the accuracy is increased to 91.3% due to the bands from all three subspace region.