Part I

A wide variety of methods exist to derive measures of vegetation cover from remotely sensed data. These methods range widely in complexity, sophistication, and accuracy. The discrepancies are important for researchers to consider when incorporating or generating vegetation data for larger studies. Since the accuracy disparity between techniques is well documented, it should concern researchers who rely on simple linear vegetation indexes for accurate vegetation cover estimates. It is important to determine precisely how different the vegetation measurements will be between a simple, easy method, as compared to a more sophisticated, yet perhaps less accessible one.

I have chosen to test two commonly used methods on either side of this divide to determine when and how researchers should use certain types of methodologies in deriving vegetation estimates from remote sensing data. Arguably the most widely used of these techniques is the Normalized Difference Vegetation Index (NDVI). It is a simple operation, needing little time, expertise, or processing capacity, which uses two bands of data to generate an index of relative vegetation abundance. Spectral Mixture Analysis (SMA) is a more sophisticated approach, using all bands of data in an image to generate percent cover of various specified ground covers within each pixel. Unfortunately field data was not available to assess the absolute accuracy of either of these techniques. However, a good evaluation is possible based on data comparisons, detailed image interpretation, reviews of possible error sources, and reviews of the wealth of literature on both these techniques.

Normalized Difference Vegetation Index

The dominant method for vegetation change detection using remotely sensed data is through vegetation indexes (Deering & Haas, 1980). Vegetation indexes are algorithms aimed at simplifying data from multiple reflectance bands to a single value correlating to physical vegetation parameters (such as biomass, productivity, leaf area index, or percent vegetation ground cover) (Tucker, 1979). These vegetation indexes are based on the well-documented unique spectral characteristics of healthy green vegetation over the visible to infrared wavelengths.

As illustrated in Figure 3, healthy green vegetation generally reflects very little solar energy in the visible wavelengths (0.4-0.7 um), with a sharp increase in reflectance in the near-infrared wavelength region (0.7-1.1 um). This “red edge” is unique to vegetation as a surface material. Dead or senescent vegetation and soil generally reflect relatively greater amounts of energy in the visible wavelengths and less in the near-infrared.[2] This unique spectral property of green vegetation is used in various indexes ranging in complexity from applying correlation coefficients to brightness values of a near-infrared band, to multi-band ratioing combined with complex algorithms (Jensen, 1996). Arguably the most successful and commonly used of these techniques is the Normalized Difference Vegetation Index (NDVI).

NDVI is the traditional vegetation index used by researchers for extracting vegetation abundance from remotely sensed data (Tucker, 1979). It divides the difference between reflectance values in the visible red and near-infrared wavelengths by the overall reflectance in those wavelengths to give an estimate of green vegetation abundance (Tucker, 1979). In essence, the algorithm isolates the dramatic increase in reflectance over the visible red to near infrared wavelengths, and normalizes it by dividing by the overall brightness of each pixel in those wavelengths. Specifically NDVI is:

NDVI = IR-red

IR+red

where the values in either band have been converted from raw DN values to reflectance of solar electromagnetic radiation.[3] The result of this algorithm is a single band data set, ranging from -1 to 1, with values corresponding to photosynthetic vegetation abundance.[4]

NDVI has been used extensively to measure vegetation cover characteristics on a broad-scale worldwide, and has been incorporated into many large-scale forest and crop assessment studies (Peterson et al., 1987; Asrar et al., 1984; Bausch, 1993; Benedetti & Rossini, 1993; Hatfield et al., 1985; Wanjura & Hatfield, 1987). It is used to provide weekly vegetation maps, monitor crops over large regions, monitor vegetation change in much of the tropics, and estimate biomass. Specifically, for example, Shih (1994) used it to monitor agricultural areas in the Everglades, Dejong (1994) used it in a model of soil erosion, Wood (1993) used NDVI to help monitor water and energy fluxes for a climate model, and Dymond et al. (1992) used NDVI to estimate rangeland degradation.

Despite this wide use, some well-documented accuracy limitations exist. The limitations of vegetation indexes emanate from the fact that relationships between vegetation abundance and electromagnetic reflectance values in complex forest structures (and areas with high vegetation abundance) are many times nonlinear, whereas vegetation indexes are simple linear algorithms. Therefore, because of increased mutual shadowing in mature stands, aging forests may show a decrease in NDVI while actual biomass increases. Consequently, once vegetation indexes reach a threshold level they no longer accurately correlate to actual vegetation abundance (see Saturation Effects in Figure 4) (Bégué, 1993; Chance, 1981; Waller et al., 1981; Wanjura & Hatfield, 1987; Wiegand et al., 1991).

Studies have also shown that background soil color affects NDVI, especially in heterogeneous scenes (Bausch, 1993; Huete, 1985; Huete, 1988; Huete et al., 1991; Major et al., 1990; Oi et al., 1993; Todd et al., 1998). Because the difference between the bands is divided by the overall brightness of the two bands, extreme variations in background soil brightness can cause NDVI values to be artificially high or low. In theory, pixels with dark soil backgrounds, such as the basaltic soil patches found in much of the northern scene, have a lower overall brightness. Therefore NDVI values would be artificially higher in these areas, as the difference between the visible and near infrared would be divided by less. Similarly, bright soil backgrounds would raise the overall brightness levels and therefore vegetation values derived through NDVI would be artificially lower than areas with similar abundance that have dark soil backgrounds. This background soil effect is additionally complicated by multiple scattering effects between vegetation and soil (Bausch, 1993; Huete, 1985; Huete 1988; Oi et al., 1993; Waller et al., 1981).

While NDVI has been shown to correlate reasonably well in medium to low vegetation abundance with various ecological parameters (such as leaf area index or green leaf biomass) (Anderson et al., 1993; Deering, & Haas, 1980; Hardisky et al., 1984), the literature suggests that in certain environments specific types of changes in vegetation may not be accurately depicted by NDVI (Asrar, 1984; Peterson et al., 1987; Waller et al., 1981).

Spectral Mixture Analysis

Many other means of quantifying vegetation cover from remotely sensed data have been developed. One of the most promising is Spectral Mixture Analysis (SMA), a more sophisticated technique that utilizes all available bands of data to separate each pixel into fractions of specified land covers (Huete, 1986). The conceptual model used to develop SMA is that most pixels in scenes are mixtures of a few specific ground covers (or endmembers), especially in arid areas and mixed land use environments. If pure spectra of spectrally distinct primary land cover materials (i.e. vegetation, water and soil) in a scene can be found, a data set can be converted to fractions of each pre-defined land cover for each pixel (Adams, 1993; Huete, 1986; Smith et al., 1990 a&b).

Linear mixture modeling assumes that endmembers (pre-defined primary land covers) are arranged in spatially distinct areas in each pixel and can therefore be extracted through the application of specific algorithms.[5] Each pixel is modeled as a spatial mixture of endmember spectra to determine the physical abundance of land cover types in each pixel area. Specifically, the following equations are solved for each pixel, over each band in all scenes of the study:

DN=Sum(F_i*DN_i +E_i)

SumF_i=1

RMSE=SqRoot of SumE_i²

DN is the brightness value of a given pixel for a specific wavelength, or band. F_i is the fractional abundance of a particular endmember. DN_i is the intensity of the image endmember at each particular wavelength or band. E_i is the error of the fit for each particular band. There is one equation for each band and provided that the number of endmembers is less than the number of bands, this system of equations can be solved using a least squared inversion.[6] The sum of these endmember fractions is constrained to equal 1. Therefore, the fraction of cover for each endmember image should be between 0 and 1. The third equation is the total root mean square error for all four bands (Elmore et al., 2000).

The results of SMA are the percent coverage of each defined ground cover material, or endmember, in each pixel. This method had the advantage of deriving not only vegetation data, but land cover fractions for all the endmembers used, as well. Additionally, the data is generated into a physically-based measure, and therefore easily integrated into studies as measures of percent live cover rather than an indexed, relative measure. Elmore et al. (2000) found percent live cover estimates using Spectral Mixture Modeling to be accurate within 4.0% and change in percent live cover to have a precision of 3.8%.[7] Specific examples of research using SMA include monitoring rural land use changes and their effects on soil characteristics in Europe (Sommer et al., 1998), creating heat budget models (Moriyama & Takebayashi, 1999), estimating erosion (Metternicht & Fermont, 1998), analyzing changes in Arctic Sea ice (Piwowar et al., 1998), and detecting fire and grazing pattern (Wassman, 1997).

Despite the utility and accuracy of this method, limitations to SMA exist. In addition to being limited in the total number of possible endmembers, SMA is limited by the type of endmembers that can be used, especially when using multispectral data (Adams, et al., 1993; Roberts et al., 1993). Endmembers must be spectrally distinct from one another and generally account for the dominant land cover and spectral characteristics of the scene. Therefore, without hyperspectral[8] data, it is virtually impossible to use this method to generate fractions of different types of photosynthetic vegetation. Generally researchers are limited to vegetation, soil, sand, and shade when using multispectral satellite imagery. Additionally, this process requires a good deal of processing capabilities and expertise in order to find pure spectra of appropriate endmembers in the scene. However, it is important to emphasize the advantage of additional datasets of endmember surface cover that are generated using SMA, as these can be used to extract more information from scenes than vegetation index information alone.

Back to Index

For more information please contact Laurel_Firestone@yahoo.com