Soil and vegetation cover estimation for global imaging spectroscopy using spectral mixture analysis

The Earth surface Mineral dust source InvesTigation (EMIT) is a visible-to-shortwave infrared imaging spectrometer currently aboard the International Space Station. Derivations of fractional cover from spectral unmixing algorithms have provided insights into various ecosystem functions. In the case of EMIT, they will be used by multiple global Earth systems models to constrain the sign of dust-related radiative forcing. This study aims to evaluate the efficacy of different approaches for estimating fractional cover and quantifying the corresponding uncertainty, and serves as a model to encapsulate the true error budget for EMIT. We simulated surface reflectance from a spectral library compiled from various drylands to generate millions of candidate spectra made up of different random fractions of nonphotosynthetic vegetation (NPV), green vegetation (GV), and soil. Simulated spectra were used as-is but we also tested the impact of atmospheric conditions/surface reflectance retrieval by using them to calculate top-of-atmosphere radiance then using the current EMIT surface reflectance retrieval algorithm to estimate apparent surface reflectance. We tested approaches to unmixing these simulated spectra using multiple strategies for dealing with spectrum brightness, within-class spectral variability, and library selection. We also incorporated a Monte Carlo approach to stabilize fractional cover retrievals and quantify uncertainty. The best spectral unmixing approaches produced mean absolute error < 0.10 for NPV and soil and < 0.06 for GV with uncertainties ≤ ± 0.02 for all classes. We named this innovative approach EndMember Combination Monte Carlo, E(MC) 2 , unmixing and found that our fractional cover retrievals are insensitive to atmospheric residuals in the surface reflectance data. • We present novel methods for the selection of endmembers for spectral unmixing. • We present a new method for spectral unmixing, E(MC) 2 . • We assess the error and uncertainty of unmixing algorithms using simulated data. • We find excellent performance and uncertainty from spectral unmixing algorithms.