Unsupervised band selection method based on improved N-FINDR algorithm for spectral unmixing

Abstract

Hyperspectral imagery (HSI) has high spectral dimensionality which presents a serious challenge to HSI processing, and so reduction of dimensionality is necessary. Band selection (BS) is one of the categories of dimensionality reduction methods. Existing BS methods have expensive cost, need prior information or only cater for classification. In order to get an efficient and unsupervised BS method for spectral unmixing, two aspects work are done. First, original N-FINDR algorithm is greatly improved by substituting volume calculation for distance test. Second, the improved N-FINDR algorithm is used to construct an unsupervised BS method for spectral unmixing. Both theory and experiments prove that the new unsupervised BS method is very effective