Fully constrained target fraction estimation from noise-reduced hyperspectral images

Abstract

Estimating abundance fractions of materials in hyperspectral images is a problem often studied in remote sensing. It can be used in reconnaissance and surveillance applications. The major difficulty of fraction estimation in hyperspectral images is due to the fact that the sampling distance is generally larger than the size of the targets of interest. Under this circumstance, estimation should be carried out at the sub-pixel level. In a linear mixture model, the spectrum of a mixed pixel is represented as a linear combination of the endmember spectra present in the pixel area weighted by fractional area coverage. Accurate fraction estimation requires two constraints (SC) imposed on abundance fractions: the abundance sum-to-one constraint and abundance non-negativity constraint (NC). In this article, we present a new fully constrained least squares computational method to estimate abundance fractions. Another contribution of this article is that the estimation is, unlike many other proposed methods, performed on noise-reduced hyperspectral images instead of original images. Experiments using synthetic data and Airborne Visible Infrared Imaging Spectrometer (AVIRIS) data demonstrate that this fully constrained estimation outperforms the unconstrained and partially constrained least squares methods and that noise reduction can considerably improve the capability of our approach when the noise intensity rises.