Address correspondence to:
Dr. Milton O. Smith
Dept. of Geological Sci.
AJ-20, Univ. of Washington
Seattle, WA 98195
A method was tested in the semiarid Owens Valley, California for measuring
sparse vegetation cover using Landsat Thematic Mapper (TM) multispectral
images. Although green vegetation has a characteristic reflectance spectrum
in the visible and near-infrared, using conventional image-processing methods,
it has been difficult to quantify vegetation cover of less than about 40%,
owing to the spectral dominance of the background soils and rocks. Thus
multispectral images have been of limited use in mapping variations in
vegetation cover in arid and semiarid regions. In this study fractions
of vegetation, soils, and shading and shadow within the smallest resolution
elements (30 x 30 m pixels) of the TM images were computed by applying
a mixing model based on laboratory and field reference spectra. Fractions
of vegetation were calculated for each pixel in TM images taken in December
1982 and May 1985, and the results were compared with ground transects.
Despite spatial variations in background soil, temporal differences in
satellite instrument response, and differences in atmospheric and lighting
conditions, the fractions of vegetation computed from each image gave a
spatially consistent measure of the projected vegetation cover. Results
were obtained for a 150-km segment of Owens Valley; they indicate that
the method can facilitate mapping and monitoring sparse vegetation cover
over large regions covered by satellite images.
Ecologists now regard the earth as an integrated ecosystem with fluxes of gases and nutrients among soils, water, biota and the atmosphere (e.g., Walter, 1979; Odum, 1983; Greegor, 1986). New awareness of the integration of processes in the biosphere has increased concern for the fragility and stability of biological communities subjected to a variety of both man-induced and natural climatic changes (Waring et al., 1986). Desert communities are especially sensitive to such changes. Increased human demand for water and predicted climatic changes due to increased atmospheric CO2 have stimulated a need for greater understanding of the ecology of arid lands and the factors that control the abundance and distribution of natural vegetation.
Detailed maps of the present-day vegetation in the world's arid lands would facilitate a variety of ecological studies and would provide a reference against which to measure the effects of future changes; however, such information is not generally available. Although extensive field studies of desert vegetation in western North America were made in the past (e.g., Kearney et al., 1914; Shantz and Piemeisel, 1940; Billings, 1945; 1949), such efforts are too expensive and labor-intensive to attempt today. Elsewhere in the world's arid lands there exist few detailed maps of vegetation, because of cost factors and logistical difficulties in conducting ground surveys in remote areas. In the place of extensive field work there presently is an emphasis on modeling of large-scale biophysical processes. Models of ecological phenomena, such as regional desertification, however, still must be validated by ground observations especially when studying processes of global climatic and land-use changes.
Remote-sensing techniques have the potential to overcome the manpower and fiscal restrictions that now limit large-scale ecological surveys, and recent studies have demonstrated their utility in assessing vegetation distribution (e.g., Justice et al., 1985; Tucker et al., 1985; 1986). Visible and near-infrared multispectral images are the most useful data currently available to examine vegetation patterns and corresponding ecological processes at regional and global scales. Multispectral data are available from earth-observing satellites that date back to 1972, and current satellites provide an expanding data base.
There are two main challenges in applying the vast multispectral image data base to ecological mapping: 1) relating spectral data to the conventional ground-based measurements that are used to characterize vegetation communities, and 2) bridging from the scale of local field measurements to regional and global scales.
1) The first challenge arises because ecologists measuring the properties of vegetation in the field typically do not have to be concerned about reflected solar radiation; however, to make use of remote sensing data it is essential to understand the correspondence between scene radiance as recorded in multispectral images and field parameters such as vegetation cover, species, community type, and developmental stage. The radiance measured by a multispectral image also is influenced by various factors that are unrelated to the materials on the ground, such as instrumental sensitivity and drift, viewing and illumination geometry, atmospheric backscatter and absorption, and the geometric orientation of surface elements (including topography) within the scene. When these factors have been taken into account, it is possible to interpret the light reflected from a surface in terms of the materials and their mixtures.
2) Satellite images cover spatial scales from tens of meters to hundreds
of kilometers. One challenge is to connect observations of vegetation in
the field to the measurements made by the smallest resolution elements
of the multispectral images (30 x 30 m in the images used in this study).
This requires understanding the reflective properties of the vegetation
at the field scale. The next challenge is to extend ecological observations
from local areas to larger regions based on the properties measured by
the images. If this can be done, the images can become a unique vehicle
for exploring regional patterns of vegetation abundance and character,
providing new ecological insights.
The appearance of vegetation varies with the scale of observation. At the large scale of field observations a shrub is seen to consist of different components, including sunlit and shadowed leaves and branches and stems, plus a substrate of litter and soil. These various scene components are characterized by different reflectance spectra. At a smaller scale the shrub components are not spatially resolved, although individual shrubs may be. For example, Figure 2 is a smaller-scale oblique view of a scene of desert scrub in Owens Valley, within our study area. In Figure 2, a single picture element or pixel of ~ 30 x 30 m, typical of Landsat Thematic Mapper (TM) images, has been schematically delineated. This smallest element of a satellite image encompasses shrubs of several species, the shadows cast by shrubs and boulders, shading due to local variations in incidence angle, and the unshadowed soil visible between shrubs. At the small scale of satellite images, components of the individual shrubs--leaves, branches and litter--are not spatially resolved, although they contribute spectrally, even at the subpixel scale. Even at the scale of Figure 2, it is not possible to determine the species of individual shrubs accurately, and this also is true at the scale of most satellite images. The interpretation of the scene may vary with changes in spatial scale, although the materials on the ground remain invariant.
Vegetation indices, reviewed by Jackson (1983), Tueller and Oleson (1989), and others, generally are based on ratios of the radiance in red and infrared spectral bands, chosen to maximize the reflectance contrasts between vegetation and other materials. Some indices have been used to estimate leaf area index (LAI) and other vegetation parameters (e.g., Choudhury, 1988; Elvidge and Lyon, 1985; Huete et al., 1985; Huete, 1986; Tucker, 1979; Jackson, 1983; Kauth and Thomas, 1976). Most commonly used are the vegetation ratio index (VRI), the normalized difference vegetation index (NDVI), and the perpendicular vegetation index (PVI) (Richardson and Wiegand, 1977). Statistical methods of data classification include maximum-likelihood, clustering, and discriminant analysis (e.g., Haralick and Fu, 1983) and methods based on principal-components analysis (e.g., Crist and Cicone, 1984). The objective of image classification is to link image spectra to dominant scene components or to characteristic mixtures of components. It is assumed that spectrally similar data will describe thematically similar elements within a scene. It is also assumed that for each pixel there is a dominant scene component, or at least a unique and identifiable suite of components that are present in distinctive proportions.
Vegetation-index and classification techniques have been primarily used to map vegetation in agricultural or forest lands, where the argument can be made that there is a single scene component or class represented in at least some pixels. However, in sparsely vegetated areas this is rarely the case, and index and classification techniques have been shown to perform less well (Ustin et al., 1986a; Tueller and Oleson, 1989; Tucker and Miller, 1977; Huete et al., 1984; Elvidge and Lyon, 1985; Heilman and Boyd, 1986; Huete et al., 1985). In desert scrub environments, for example, thematic classes correspond to characteristic suites or mixtures of components that occur in preferred proportions and at certain illumination geometries (e.g., shaded or sun-facing slopes). The radiance recorded in an image pixel may be mixed from both soil and vegetation, but because soil and vegetation can theoretically mix in any proportion there exists, in theory, an infinite number of classes, even for a scene containing but two components. Clusters of radiance values are thus typically indistinct, and otherwise distinct clusters may overlap because of illumination differences in rugged terrain. Thus rules must be applied to designate thematic classes for many scenes.
Spectral mixture analysis transforms radiance data into fractions of a few dominant endmember spectra which correspond to scene components (Adams and Adams, 1984; Adams et al., 1986; Smith et al., 1985). "Fraction images" depict the mixing proportions of these endmember spectra and thus, via calibration to field data, the mixing proportions of the scene components. Mixture analysis differs significantly from statistical classification in a number of ways, perhaps most significantly in the small number of endmembers compared to the potentially large number of thematic classes required to describe a scene. Mixture analysis separates the spectral contributions of these intrinsic scene components from shadow and other effects of illumination. This approach is particularly useful for measuring vegetation cover, especially in desert regions where the proportions of vegetation and soil may vary significantly over small distances.
The application of spectral mixture analysis presented in this paper differs from previous discussions of spectral mixing by Horwitz et al. (1975), Jackson (1983), Conel and Alley (1984), Huete et al. (1985), and Pech et al. (1986) in that the approach is directed at using a simple mixture model to link reflectances measured by field and/or laboratory instruments with image relative-radiance measurements acquired by satellite. This is done by referencing to the known spectra of materials and their mixtures on the ground. Although previous studies have discussed the importance of spectral mixtures, they do not provide the methodology to determine the combined atmosphere and instrument calibration at the time of image acquisition, to remove variations in lighting geometry caused by topography and other factors, and to separate spectral mixtures.
In this article we discuss the use of spectral mixture analysis of Landsat Thematic Mapper (TM) multispectral satellite images to estimate vegetation abundance in deserts. In particular, we explore the construction, calibration, and significance of vegetation, soil, and related fraction images of semiarid Owens Valley, California. The method, however, is not limited to any one imaging system, and has been applied to such diverse data as Viking Lander images of Mars (Adams et al., 1986) and the 224-channel Advanced Visible Infrared Imaging Spectrometer (AVIRIS) images of terrestrial scenes (Smith et al., 1988a).