Dimensionality reduction of hyperspectral images is essential for reduction of computational complexity and faster analysis. A novel method for band reduction has been proposed here, which has been adapted from the genetic algorithm (GA) along with spatial clustering. Spatial clustering generates overall signature variation present in a particular scene and in turn removes huge redundancy present in the raster data set. GA is applied on the clustered signatures to extract the reduced set of bands that is computed to be the "fittest" i.e., those bands that provide the most discriminating information in a hyperspectral image. This has been computed by taking the sum of Kullback-Leibler divergences (KLD) between consecutive selected bands. A higher KLD value amongst adjacent selected band implies higher divergence in value. The selected band-set image has been classified and the accuracy indices are evaluated respectively. The proposed method shows high performance on the basis of classification accuracy and efficient execution while comparing with two other state-of-the-art methods.

No abstract