SPECTRAL UNMIXING OF VEGETATION, SOIL AND DRY CARBON IN ARID REGIONS: COMPARING MULTISPECTRAL AND HYPERSPECTRAL OBSERVATIONS Gregory P. Asner 1 and Kathleen B. Heidebrecht 1 1. Introduction Remote sensing of vegetation cover and condition is critically needed to understand the impacts of land use and climate variability in arid and semi-arid regions. However, remote sensing of vegetation change in these environments is difficult for several reasons. First, individual plant canopies are typically small and do not reach the spatial scale of typical Landsat-like satellite image pixels. Second, the phenological status and subsequent dry carbon (or non-photosynthetic) fraction of plant canopies varies dramatically in both space and time throughout arid and semi-arid regions. Detection of only the “green” part of the vegetation using a metric such as the normalized difference vegetation index (NDVI) thus yields limited information on the presence and condition of plants in these ecosystems. Monitoring of both photosynthetic vegetation (PV) and non-photosynthetic vegetation (NPV) is needed to understand a range of ecosystem characteristics including vegetation presence, cover and abundance, physiological and biogeochemical functioning, drought severity, fire fuel load, disturbance events and recovery from disturbance. Many approaches have been devised to analyze PV, NPV and bare soil cover in arid and semi-arid regions. A wide variety of studies have attempted to correlate vegetation indices (e.g., NDVI) to the fractional coverage of PV and bare soil (e.g., Duncan et al. 1993, Carlson and Ripley 1997). The typical spectral regions used to detect PV – the visible and near-infrared wavelengths (0.4-1.3 um) – do not easily separate the individual contribution of NPV and bare soil to the measurement (van Leeuwen and Huete 1996, Asner 1998, Roberts et al. 1998, Asner et al. 2000). More recently, spectral mixture analysis (SMA) was developed to decompose image pixels into constituent PV, NPV and bare soil covers. Many SMA efforts have now been applied in arid and semi-arid ecosystems using airborne and spaceborne sensors (e.g., Smith et al. 1990, Sohn and McCoy 1997, Wessman et al. 1997, Elmore et al. 2000). Most SMA approaches assume that image pixels contain endmember cover fractions that are linearly summed: ρ(λ) pixel = Σ [C e • ρ(λ) e ] = [C pv • ρ(λ) pv + C soil • ρ(λ) soil + C npv • ρ(λ) npv ] (1) where ρ(λ) e is the reflectance of each land-cover endmember (e) at wavelength λ. The sub-pixel cover fraction (C e ) of each land-cover endmember may be PV, NPV, bare soil or other constituents. Solving for the sub-pixel cover fractions (C e ) therefore requires that the observations (in this case, reflectance or ρ(λ) pixel ) contain enough information to solve a set of linear equations, each of the form of equation (1) but at a different wavelength (λ). The selection of reflectance endmembers (ρ(λ) e ) for equation (1) is also critical to the accurate estimation of the sub-pixel cover fractions (C e ). These endmembers are usually selected either from the image data (e.g., Wessman et al. 1997) or from spectral libraries built from field surveys (e.g., Roberts et al. 1998). Each approach has distinct advantages and disadvantages. Image-based endmembers are ideal because they are drawn from the population of data points to be analyzed, which increases the likelihood that image pixels will be decomposed using endmembers that actually exist in the area. However, selection of image endmembers often requires the availability of pixels comprised purely of each dominant cover type. Pure image pixels are rarely available in images of ecosystems, especially in arid and semi-arid regions. A very unique method for addressing this issue has been developed by Bateson and Curtiss (1996) and Bateson et al. (2000). Nonetheless, no automated, fully objective methods have been developed for dealing with sub-pixel heterogeneity in image endmember selection. The alternative approach of spectral endmember libraries has its advantages and problems as well. The major advantages are that endmembers can be readily collected from large field-based surveys and that the quality and interpretation of the endmembers are easy to control. The potential problems in using spectral libraries relate to endmember generality and scalability. Spectral endmembers collected in one area may not be applicable to another 1 Carnegie Institution, Stanford, California, 94305. 1