Spectral mixture analysis (SMA) uses linear mixture models to provide physical representations of land surface reflectance as continuous fields of spectral endmember abundances. While most spectral mixture analyses are site-specific and use local image endmembers to model abundance of specific materials, the linear mixture model can also provide a physical basis for a more general representation of land surface reflectance as mixtures of generic endmembers. A spectral mixture analysis of a global composite of 30 spectrally diverse Landsat ETM+ subscenes indicates that a wide variety of ETM+ reflectance spectra can be accurately represented as linear combinations of soil and rock substrate, green vegetation, and dark surface reflectance spectra. The global analysis indicates that >98% of ETM+ image spectral variance can be represented within a three-dimensional spectral mixing space and that >90% of the variance can be described with a two-dimensional projection of the mixing space. A wide variety of individual subscene and composite mixing spaces consistently show similar triangular distributions of mixed reflectances bounded by linear mixtures of the substrate, vegetation, and dark surface endmembers. Distinct binary mixing spurs representing ice/snow and submarine/reef reflectance are also linear but separate from the primary triangular mixing space. While the three-endmember substrate, vegetation, dark surface (SVD) linear mixing model does not accommodate some significant characteristics of the three-dimensional mixing space, it is able to represent more than 95% of the 30 million observed ETM+ image spectra with misfits of less than 0.04 reflectance units. Linear binary mixing continua between the dark surface endmember and both the substrate and vegetation endmembers indicate the extent to which shadowing and nonreflective surfaces combine with illuminated substrate and vegetation at subpixet scales to modulate spectrally mixed ETM+ reflectances in a wide variety of land cover types. SVD fractions estimated with a standard set of generic endmember spectra could offer many of the benefits of standard metrics like vegetation indices (e.g. NDVI) and linear transformations (e.g. Kauth-Thomas Tasseled Cap), while retaining the unique benefits of physically based fraction estimates. These benefits include good agreement with ground based measurements and output units corresponding to physical properties of land cover that can be used directly as input to land surface process models. (C) 2004 Published by Elsevier Inc.
859UWTimes Cited:29Cited References Count:44