Comparing and validating spectral unmixing algorithms with AVIRIS imagery

Abstract

Spectral unmixing of hyperspectral imagery has been utilized in many application areas to identify and map targets at the subpixel level. Its applicability to forest applications, however, raises concerns because of scene heterogeneity and the multiple scattering between sun radiation and targets of illumination, which may introduce nonlinearity to hyperspectral data and lead to violating the assumption of linearity upon which many spectral unmixing algorithms are based. To investigate the applicability of spectral unmixing to forest spectral mixture analysis, we tested three commonly used spectral unmixing algorithms with hyperspectral imagery of coastal forests acquired by the National Aeronautics and Space Administration (NASA) Airborne Visible / Infrared Imaging Spectrometer (AVIRIS). The algorithms tested were separated based on whether the two constraints on end-member fractions, i.e., non-negativity and sum-to-one, were enforced when solving the unmixing problem. The first algorithm investigated was unconstrained, where both constraints were not imposed. The other two algorithms applied one or both of the constraints. The results of this study indicate that all three algorithms tested were applicable to calculating forest canopy fractions when forest density was high. With the 20 m AVIRIS data employed in this study, the fully constrained unmixing algorithm outperformed the unconstrained and partially constrained algorithms on unmixing accuracy. According to the results derived from the high forest density plots, the average deviations from the ground truth of the unmixed forest fractions were 19 ± 11% by the unconstrained unmixing, 21 ± 13% by the partially constrained unmixing, and 8 ± 7% by the fully constrained unmixing.