Endmember bundles: a new approach to incorporating endmember variability into spectral mixture analysis

Abstract

Accuracy of vegetation cover fractions, computed with spectral mixture analysis, may be compromised by variation in canopy structure and biochemistry when a single endmember represents top-of-canopy reflectance. In this article, endmember variability is incorporated into mixture analysis by representing each endmember by a set or bundle of spectra, each of which could reasonably be the reflectance of an instance of the endmember. Endmember bundles are constructed from the data itself by an extension to a previously described method of manually deriving endmembers from remotely sensed data. Applied to remotely sensed images, bundle unmixing produces maximum and minimum fraction images bounding the correct cover fractions and specifying error due to endmember variability. In this article, endmember bundles and bounding fraction images were created for an airborne visible/infrared imaging spectrometer (AVIRIS) subscene simulated with a canopy radiative transfer/geometric-optical model. Variation in endmember reflectance was achieved using ranges of parameter values including leaf area index (LAI) and tissue optical properties observed In a North Texas savanna. The subscene's spatial pattern was based on a 1992 Landsat Thematic Mapper image of the study region. Bounding fraction images bracketed the cover fractions of the simulated data for 98% of the pixels for soil, 97% for senescent grass and 93% for trees. Averages of bounding images estimated fractional coverage used in the simulation with an average error of /spl les/0.05, a significant improvement over previous methods with important implications for regional-scale research on vegetation extent and dynamics.