The following section discusses the basic spectral features of green vegetation and describes how the vegetation indices using broad-band remote sensing data have been developed. This section will also show that imaging spectroscopy, or at least remote sensing that uses narrow spectral bands, improves the measurement of vegetation cover. A further discussion of vegetation indices from a slightly different angle is provided in Appendix B.

Work on the spectral properties of plants began as early as 1913 when Willstätter and Stoll proved that light entering leaves is critically reflected internally at the cell walls where the refractive index changes from that of water (1.33) to that of air (1.00). This leads to highly efficient scattering of all wavelengths of light. Gates et al. (1965) performed some of the earliest work which treated the spectral properties of plants from the ultraviolet through the thermal infrared. Table 8 summarizes the absorptions features of some kay plant materials. Figure 54 shows the typical reflectance spectrum of green leaves in the spectral range of the visible and near- to mid-infrared (0.4 mm-2.5 mm). Plant pigments such as chlorophyll strongly absorb light in the visible, and the liquid water in plant leaves absorbs much of the light at wavelengths longer than 1.4 mm. This contrasts with strong reflectance in the near-infrared in the range from 0.75 mm through about 1.4 mm, wavelengths at which plant materials are relatively transparent.

Table 8: Absorption Features of Some Key Plant Materials (data from Gates, 1980 and Elvidge, 1990)
Material	Wavelength (m mm)
a-carotene	0.42-0.5
Arabinogalactan	0.99, 1.21, 1.45, 1.55, 1.74, 1.93, 2.10, 2.28, 2.32, 2.50
B-glucan (a hemicullulose)	1.45, 1.77, 1.93 2.09, 2.33, 2.50
Carnauba wax	0.93, 1.04, 1.21, 1.39, 1.41, 1.54, 1.73, 1.82, 1.93, 2.01, 2.05, 2.14, 2.31, 2.35, 2.39, 2.43
Cellulose	1.48, 1.93, 2.10, 2.28, 2.34, 2.48
Chlorophyll-a	0.38-0.45, 0.675
Chlorophyll-b	0.41-0.47, 0.61
D-ribulose 1,5-diphosphate carboxylase	1.50, 1.68, 1.74, 1.94, 2.05, 2.17, 2.29, 2.47
Humic Acid	0.4-0.7, 1.92, 2.30, 2.34
Lignin	1.45, 1.68, 1.93, 2.04-2.14, 2.27, 2.33, 2.38, 2.50
Lutein (a xanthophyll)	0.4-0.5
Pectin (apple)	1.44, 1.72, 1.92, 2.09, 2.24, 2.36, 2.48
Pectin (citrus)	0.98, 1.19, 1.44, 1.56, 1.68, 1.73, 1.78, 1.93, 2.08, 2.25, 2.32, 2.36, 2.48
Protochlorophyll	0.41-0.47,0.58
Starch	0.99, 1.22, 1.45, 1.56, 1.70, 1.77, 1.93, 2.10, 2.32, 2.48
Tannic Acid	0.99, 1.12, 1.46, 1.66, 1.93, 2.13, 2.26, 2.32, 2.50
Xylan (a hemicellulose)	1.21, 1.45, 1.72, 1.79, 1.93, 2.09, 2.26, 2.33, 2.50

Figure 54: Typical green vegetation spectrum. Left plot is an enlargement of the boxed region. Notice how the pigment absorption in the visible is actually caused by a triplet of pigment absorptions. The strong reflectance from 0.8-1.2 mm is caused by strong scattering from cell walls. Liquid water in the leaf is the most important absorber beyond 1.0 mm, but several plant materials have significant absorptions between 2.02.5 mm. Table 8 provides more information about the absorption locations of plant materials.

Figure 55: Green vegetation and soil spectra for the range 0.4-2.5 mm. Notice how the change in reflectance between the visible and the near-infrared is much smaller for the soil than for the green vegetation.

Much attention in the remote sensing of green vegetation is focused on the strong reflectance contrast between the visible and the near infrared (NIR) which forms a strong step in the spectrum of green vegetation which is often referred to as the “red edge.” In figure 55 the spectrum of a soil is plotted for comparison with the green leaf spectrum. Notice that while there is some difference in the red and near-infrared reflectances of the soil (with the near-infrared being slightly higher) the difference is much smaller that that for plants.

Figure 56: In this scatter diagram, NIR (0.755 mm) and red (0.674 mm) reflectances (normalized to 1000) are plotted for each pixel from an AVIRIS image. The lower right boundary of this sort of plot is taken to be formed by pixels containing only bare soil, and this boundary is referred to as the soil line. The “tip” opposite the soil line, which has high NIR reflectance and low red reflectance, is taken to be where pixels completely covered with vegetation plot on this diagram. All pixels covered by a mixture of bare soil and vegetation will plot between these two extremes. This sort of figure is sometimes called a tasseled-cap, because of its shape.

One consequence of the distinctive green vegetation spectrum is shown in figure 56. The lower right edge of the data point cloud forms a boundary which is generally identified as the soil line. There is also a tip to the graph opposite the soil line due to a strong NIR reflectance and a relatively weak red reflectance which is characteristic of green vegetation, and this point is defined as the green vegetation point. Points partly covered by green plants and partly covered by soil will plot between the bare soil line and the vegetation point. This figure was referred to by Kauth and Thomas (1976) as the “tasseled cap-agram” and is commonly known as the tasseled cap.

There have been two general approaches taken to develop indices for measuring green vegetation cover based on the characteristics of the tasseled cap. The first approach is to measure the distance between where the pixel plots in the tasseled cap plot from the soil line. The soil line is used because it is generally easier to find than the 100% vegetation point. The perpendicular vegetation index (PVI) of Richardson and Wiegand (1977) assumes that the perpendicular distance of the pixel from the soil line is linearly related to the vegetation cover. This index is calculated as follows:

PVI NIR red = - sin a (NIR) cos a (red) (9)

where NIR is the near-infrared reflectance, red is the red reflectance and a is the angle between the soil line and the near-infrared axis. This means that the isovegetation lines (lines of equal vegetation) would all be parallel to the soil line. A special case of this is the vegetation index (VI) mentioned by Lillesand and Kiefer (1987) which has more recently been christened the difference vegetation index (DVI) by Richardson and Everitt (1992):

VI = DVI = NIR - red. (10)

This case occurs when the soil line has a slope of 1.0.

The next possibility is to assume that the isovegetation lines all intersect at a single point. As the first approximation, Jordan (1969) developed the ratio vegetation index:

	NIR
RVI = -------. (11)
	red 

       NIR-red   RVI-1
NDVI = ------- = ----- . (12)
       NIR+red   RVI+1

         NIR     1		
IPVI = ------- = - (NDVI+1). (13)
       NIR+red   2

Huete (1988) suggested a new vegetation index which was designed to minimize the effect of the soil background, which he called the soil-adjusted vegetation index (SAVI). This vegetation index takes the form:

        NIR-red
SAVI = ---------(1+L) . (14)
       NIR+red+L

Qi et al. (1994a) developed a vegetation index which is basically a version of SAVI where the L-factor is dynamically adjusted using the image data. They referred to this index as the Modified Soil Adjusted Vegetation Index or MSAVI. The factor L is given by the following expression:

L = 1 - (2*slope*NDVI*WDVI) (15)

WDVI = NIR - (slope*red).

MSAVI2 = 1/2 * ((2*(NIR+1)) - (((2*NIR)+1)2 - 8(NIR-red))1/2). (16)

One important difficulty which has been encountered in using the vegetation indices which attempt to minimize the effect of a changing soil background is an increase in the sensitivity to variations in the atmosphere (Leprieur et al., 1994; Qi et al., 1994b). There have been several approaches in the development of vegetation indices which are less sensitive to the atmosphere, such as the Atmospherically Resistant Vegetation Index (ARVI) of Kaufman and Tanré (1992) and the Global Environmental Monitoring Index (GEMI) of Pinty and Verstraete (1991). Chehbouni (1994, personal communication) has data demonstrating that GEMI is highly sensitive to soil noise. Qi et al. (1994b) demonstrated that soil noise caused GEMI to violently break down at low vegetation covers, and that all of the vegetation indices designed to minimize the effect of the atmosphere have increased sensitivity to the soil, which makes these indices completely unsuitable for arid regions.

There are still specialists trying to develop vegetation indices using two band combinations other than the NIR-red. Pickup et al. (1993) proposed an index similar to PVI which made use of MSS bands 4 and 5. Although Pickup et al. (1993) claimed that it was highly effective at detecting both dry and green vegetation, this band combination would be sensitive to the iron oxide absorption feature present in many soils. The detection of dry and green vegetation with this index, which is called PD54, is probably due to the fact that vegetation cover would obscure this feature, but this feature can also be highly variable in many soils.

Jackson (1983) showed how a set of n bands could be used to construct a perpendicular vegetation index. The basic procedure is to pick two or more soil points to define a soil line in n-space and then to use Gram-Schmidt orthogonalization to find the “greenness” line which passes through the 100% vegetation cover point and is perpendicular to the soil line. Two of the indices, both known as the Green Vegetation Index, have seen considerable use. A four-band GVI, for use with MSS data, was developed by Kauth and Thomas (1976), and a six-band version was developed for TM data by Crist and Cicone (1984).

An extremely important point to realize about the soil line used in many of these vegetation indices is that the soil line will be different for different areas, and the soil line will vary for different NIR and red bandpasses. Table 9 gives the slope and intercept for the soil line calculated from AVIRIS data for different bandpasses. The clear implication is that the only truly valid way of making use of a vegetation index which uses a soil line is to compute the soil line for each image. If a good calibration is available, calculating the soil line for each target for each instrument once might suffice. Of course, even the assumption that all of the bare soil spectra in a single image form a line may also be inaccurate.

Table 9: Red-NIR Soil Line Parameters for AVIRIS Data Sampled at Different Bandpasses
Instrument Simulated	Red Bandpass mm	NIR Bandpass mm	Slope	Intercept
MSS	0.6-0.7	0.8-1.1	0.9034	52.95
TM	0.63-0.68	0.8-0.9	0.7939	71.39
AVIRIS	0.674	0.755	0.8863	55.00

Elvidge and Chen (1995) found that SAVI and PVI consistently provided better estimates of LAI and percent green cover than did NDVI or RVI. They also found that there was a steady improvement in all of these vegetation indices as narrower and narrower bands were used for the near-infrared and red reflectances, with SAVI being the best index at the very narrowest bandwidth. The advantage of narrow bands for use with vegetation indices provides additional arguments for the use of high spectral resolution remote sensing.

The advent of imaging spectroscopy in remote sensing opens up the possibility of using the entire spectrum of the target pixel to measure vegetation cover. Elvidge and Mouat (1988) and Elvidge and Chen (1993) showed that trace quantities of green vegetation could be detected in data acquired by the JPL AVIRIS instrument. The red edge was found to be a persistent feature down to less than 5% green cover. However, these detections were made in areas with a relatively uniform soil background (Elvidge and Chen, 1993).

169 Elvidge and Chen (1995) investigated the feasibility of a vegetation index based on high spectral resolution data. The first of these indices used the first derivative of the spectra. The first derivative vegetation index (1st DGVI) was calculated by integrating the first derivative of the spectra from 0.626 mm to 0.795 mm as shown in figure 57. The next vegetation index tested by Elvidge and Chen (1995) used the integral of the absolute value of the second derivation of the target spectrum over the same range as used for the 1st DGVI. This second derivative green vegetation index 2nd DGVI was found to be superior to SAVI and PVI at estimating LAI and percent green cover, while the 1st DGVI was inferior to SAVI and PVI. They found that they could produce a vegetation index superior to 2nd DGVI by integrating the departures in the first derivative spectrum from a baseline given by the first derivative of the reflectance at 0.625 mm. This local baseline 1st DGVI had the most linear relationship with LAI and percent green cover of any of the vegetation indices tested. They found that there was no gain in using a local baseline for the 2nd DGVI. The local baseline correction to the 1st DGVI seems to eliminate the overall slope observed in the spectrum of nearly all materials from the ultraviolet to the near-infrared, and since this would cause a DC shift in the first derivative, it should be absent in the second derivative.

Figure 57: The spectrum of a green leaf (top) and its first and second derivatives (middle and bottom). The shaded regions on the first and second derivative spectra is the area used for the derivative-based vegetation indices of Elvidge and Chen (1995).

It should be noted that all of these vegetation indices are fundamentally based on empirical measurements and correlation between the vegetation index and vegetation cover found with on-site investigation. The adjustment performed to produce SAVI was justified by Heute (1988) by using basic radiative transfer, but it was initially constructed from experimental data. There has been no success at relating vegetation index values to plant cover on the basis of actual soil and plant spectra included in a rigorous radiative scattering model. Results like those of Elvidge and Chen (1995) show that different soil backgrounds give different vegetation index values for the same cover of a given plant type. It is likely that different plant types may also give different vegetation index values for the same plant cover.

All of these vegetation indices are dependent on the presence of the red edge feature. The red edge is due to the distinctive spectral properties of green leaves which contain pigments which strongly absorb visible wavelengths while remaining highly transparent in the near-infrared, and also due to strong scattering from leaf cell walls as photons try to travel from the water-rich cells into the air-filled intercellular spaces. Clearly, the ideal target plant for these indices is one with large leaves with high water content and high pigmentation, and little soil or stem structure visible in the overhead view.