IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 30, NO.

2, MARCH 1992 26 1

Atmospherically Resistant Vegetation

Index (ARVI) for EOS-MODIS
Yoram J. Kaufman and Didier Tanr6

Abstract- Atmospherically resistant vegetation index (ARVI) plotted (based on data from Bowker et al. [4]), showing the
is proposed and developed to be used for remote sensing of veg- sensitivity of the red and near-IR reflectances to vegetation.
etation from the Earth Observing System (EOS) MODIS sensor. The NDVI was shown to be related to canopy photosynthesis
The same index can be used for remote sensing from Landsat
TM, and the EOS-HIRIS sensor. The index takes advantage of [5], and NDVI integrated throughout the growing season is
the presence of the blue channel (0.47k0.01 pm) in the MODIS sensitive to primary productivity [6], to the rainfall in semiarid
sensor, in addition to the red (0.6620.025 pm) and the near IR regions [7], and to seasonal and latitudinal variations of the
(0.86520.02 pm) channels that compose the present normalized atmospheric CO2 concentration [8]. As such it was applied to
difference vegetation index (NDVI). The resistance of the ARVI to regional [9]-[ 111 and global vegetation problems.
atmospheric effects (in comparison to the NDVI) is accomplished
by a self-correction process for the atmospheric effect on the red The success of the NDVI to monitor vegetation variations on
channel, using the difference in the radiance between the blue a large scale, despite the degradation of the AVHRR calibration
and the red channels to correct the radiance in the red channel. in these two bands [12], [13], and despite the presence of
Simulations using radiative transfer computations on arithmetic atmospheric effects [14]-[16] is due to the normalization
and natural surface spectra, for various atmospheric conditions, involved in its definition. The normalization reduces the effect
show that ARM has a similar dynamic range to the NDVI,
but is, on average, four times less sensitive to atmospheric effects of degradation of the satellite calibration from 10-30% for
than the NDVI. The improvement is much better for vegetated a single channel to 0 4 % for the normalized index [12],
surfaces than for soils. It is much better for moderate to small size [13]. The effects of the angular dependence of the surface
aerosol particles (e.g., continental, urban, or smoke aerosol) than bidirectional reflectance and of the atmospheric effects, is also
for large particle size (e.g., maritime aerosol or dust). Due to a reduced significantly in the normalized index [16]-[18]. The
fortunate coincidence, the same optimal value of the parameter 7
that defines the weighting of the blue band radiance in the ARVI effect of scattering and absorption by atmospheric aerosol
definition, is found for vegetated areas with small to moderate and gases (mainly water vapor) and by undetected clouds is
aerosol particle size (e.g., anthropogenic aerosol and smoke) and reduced significantly in the compositing of the NDVI from
for arid regions with large particle size (e.g., dust). Therefore, a several consecutive images, choosing a value that corresponds
single combination of the blue and the red channels in the ARVI to the maximum vegetation index for each pixel [18]-[19].
may be used in all or most remote sensing applications. Due to the
excellent atmospheric resistance of the ARVI, it is expected that Despite the inherent “resistance” of the vegetation index
remote sensing from MODIS of the vegetation index over most to calibration and atmospheric effects, any future advances
land surfaces will include molecular and ozone correction with no in the quantitative application of the vegetation index to
hrther need for aerosol correction, except for dust conditions, remote sensing of the biosphere and to studies of the bio-
like in the Sahel.
sphere-atmosphere interactions, would require further reduc-
tion of the atmospheric and sensor effects on the NDVI. This
I. INTRODUCTION further reduction can be achieved in the EOS-MODIS era by
The Normalized Difference Vegetation Index (NDVI) de- several possible approaches:
rived from meteorological satellite data is used to detect Reduction of the effect of gaseous absorption by choos-
changes in global vegetation [l], [2]. It is computed from ing narrow channels in atmospheric windows [20], [21],
the difference between the near-IR, LNIR,and the red, Lred, so that there will be no dependence of the NDVI on
radiances reflected from the surface and transmitted through water vapor.
the atmosphere: High spectral stability and calibration accuracy of the
MODIS channels [20].
Development of methods for remote sensing of the
This difference is sensitive to the presence of vegetation, since atmospheric aerosol from satellite imagery [21]-[24]
green vegetation usually decreases the signal in the red due and application to atmospheric corrections of remotely
to chlorophyl absorption and increases the signal in the near sensed data [22], [25].
IR [3]. In Fig. 1, typical spectra of vegetation and soils are Redefinition of the NDVI so it will remain sensitive to
variations in green vegetation, while being resistant to
Manuscript received September 26, 1991; revised November 1, 1991.
The authors are with the NASNGoddard Space Flight Center, Greenbelt, atmospheric effects.
MD 20771. The term “resistant to atmospheric effects” refers to signif-
D. Tanr6 is on leave from Laboratoire d’Optique Atmospherique, Universite
de Sciences et Techniques de Lille, Villeneuve d’Ascq, France. icantly smaller variations in the value of ARVI for a given
IEEE Log Number 9105384. variation in the atmospheric opacity, than that of the NDVI.
0196-2892/92$03.00 0 1992 IEEE

~.
262 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 30, NO. 2, MARCH 1992

0.7

y 0.6
Z
--8 - YOUNG WHEAT
5
?!
0.5
0.4
--*-
BURROAKLEAF
0.40
0.18
UI
U
-.C BURNED FOREST
a 0.3
w
0 0.20
3 0.2 0 9 QUARTZ BEACH SAND
a
3 0.1
=.Y. DRY PEDOCALSOIL
0 0.00 0.00
0.4 0.5 0.6 0.7 0.8 0.9 0.4 0.8 1.2 0.4 0.8 1.2
WAVELENGTH em) WAVELENGTH em)
Fig. 1. (a) Surface reflectance spectra for several vegetation types, thin black lines, and several soil types, thick gray lines (data from Bowker et al. [4]),
showing the relation between the reflectance in the blue band and the reflectance in the red band. The MODIS land channels in this spectral region are also
shown (shaded areas). (b) Surface spectral reflectance for dense alfalfa - solid line [37] and the upward spectral normalized radiance for two atmospheric
models with moderate aerosol loading (aerosol optical thickness of 0.25 at 0.56 pm) for low (dashed line) and high (dotted line) absorption, after Fraser
and Kaufman [31]. Computations are performed for Angstrom exponent a=1.0. The effect of atmospheric scattering is strong for short wavelengths, and
the effect of atmospheric absorption for longer wavelengths. (c) The normalized atmospheric path radiance p o , and the total upward normalized radiance
p* for constant surface reflectance of p = 0.05 as a function of wavelength (after Kaufman, [25]).

The latest approach is the purpose of the present paper. It is L(A) = Lo(A) + P ( A ) [ F d ( A ) T ( A ) / 7 d / [ 1 - S(A)P(A)I (2)
envisioned that in the EOS era, remote sensing of vegetation
will be performed by a combination of atmospheric correction where, L,(A) is the atmospheric path radiance, due to solar
for the well accounted for, molecular scattering and ozone radiation that was scattered by the atmosphere to the sensor
absorption [IS] followed by application of a vegetation index without being reflected from the earth surface, F d ( A ) is the
that is significantly less dependent on the atmospheric aerosol spectral downward flux, T (A) is the total transmission of solar
scattering. This approach is expected to reduce the atmospheric radiation reflected from the surface, through the atmosphere
contamination so significantly, that for remote sensing of most to the sensor, and s(A) is the backscattering coefficient of
of the globe no additional correction for the atmospheric the upward radiation by the atmosphere. The coupling term
effects on the vegetation index will be able to further reduce [l - s(A)p(A)] is usually small [31] and for simplicity will
the errors. Detailed atmospheric correction for the aerosol be ignored in this discussion (the detailed computations in
effect will be still needed for highly contaminated atmospheric this paper, do account for it). The radiances L ( A ) , L,(A)
conditions such as arid and semiarid regions [23], [26], tropical are normalized to reflectance units and referred to as -
regions during biomass burning [27], [28], and summertime normalized radiance:
in the eastern U.S. or Europe, contaminated by anthropogenic
emissions [15], [29]. Note that atmospheric corrections will be
still needed for other remote sensing applications, e.g., remote
sensing of the bidirectional reflection of the surface, surface
albedo, spectral classification, etc. [ 181. (3’)
In Section I1 the conceptual approach to the new vegetation
index, ARVI, is developed. In Section I11 the resistance of the where p0 is the cosine of the solar zenith angle and F, is the
new vegetation index to atmospheric effects is demonstrated extraterrestrial solar flux. Therefore, (2) is reduced to:
for several natural surface covers and compared to the original
vegetation index. A sensitivity study of the ARVI is developed P*(X) = Po(A) + P(A)f(A) (4)
in Section IV, and optimization for remote sensing of soils
and vegetation is given in Section V. Relation of the ARVI to The atmospheric effect on the detected signal p*(A) is com-
other vegetation indexes are discussed in Section VI. Section posed from the normalized path radiance po (A) that increases
VI1 concludes this paper. the detected signal and from the transmission f ( A ) that de-
creases the detected signal. For low surface reflectance ( p <
O.l), e.g., for all green vegetation and many soil types in the
red channel, the net atmospheric effect is usually an increase in
11. CONCEPTUAL APPROACH TO the detected signal. For higher surface reflectance ( p > 0.15),
ATMOSPHERICALLY RESISTANTNDVI e.g., for most surfaces in the near IR, the net effect is a mixed
The conceptual approach for the development of the atmo- one. Increase or decrease of the signal, in this case, depends
spherically resistant NDVI is based on the spectral character- more on the characteristics of the atmospheric aerosol (e.g.,
istics of vegetation and soils (Fig. l(a)) and on the spectral the ratio of scattering to absorption) rather than their amount
characteristics of the atmospheric effect (Figs. l(b) and (c)). or total opacity [31]. As a result, the atmospheric effect on the
The spectral upward radiance at the top of the atmosphere red channel is much larger than on the near IR. Therefore, the
L(A) is related to the spectral surface reflectance p ( A ) by [30]: effort to redefine the NDVI, to be more resistant to atmospheric
KAUFMAN AND TANRE: ATMOSPHERICALLY RESISTANT VEGETATION INDEX 263

effects, is directed toward reduction of the effect on the red analytical expressions (5-8). In the rest of the paper we shall
channel. demonstrate that a single value of y can reduce significantly
For MODIS, molecular scattering and ozone absorption the atmospheric effects and we shall optimize this value.
take place mainly in atmospheric layers that are above the Measurements show that the aerosol radiative characteris-
aerosol layer. Therefore, it is possible to correct the normalized tics needed to describe the scattering phase function can be
radiances for molecular scattering and ozone absorption, thus modeled by separating the different continental aerosol types
reducing the normalized radiances observed from the satellite into two types: dust storm aerosol and other aerosol types
to just above the aerosol layer. We performed this correction [33]. For large dust particles the aerosol optical thickness
assuming that the apparent reflectivity of the earth as observed and the path radiance are wavelength independent. Therefore,
from above the haze layer (and including the haze effect) is Po-b M Po-r, resulting in y + cc in (8). As a result, the
Lambertian. This simplification was found to introduce some ARVI is not expected to be much better than the NDVI for
errors when applied to radiances over the ocean [32], and large dust particles. But for other aerosol types the aerosol
methods for its correction were suggested [32]. In the rest optical thickness, T,, is wavelength dependent -7, M CA-",
of the text the normalized radiances are considered to be where a=1-2 [33-351 resulting in Po-b > po-r. Therefore, it
corrected for the molecular scattering and absorption. can be expected that a single value of y will reduce the value
In order to make the ARVI less dependent on variations of of Po-& significantly below the value of po-r.
the atmospheric effect, it will be defined using, in addition to The self-correction process described in (7), for a proper
the red channel, the blue channel, in a self-correcting approach. value of y,reduces to a minimum the effect of the normalized
Similar to (l),but written for the normalized radiance: path radiance Po-rb on the vegetation index. But it also
involves the reflectance of the surface in the blue and the red
bands, as well as atmospheric attenuation introduced in A p r b .
where For zero aerosol optical thickness, r, = 0, A p r b introduces a
shift in the value of the ARVI relative to the corresponding
Prb = P: - y ( d - P:) (5/) value of the NDVI (see next section for examples), by replac-
ing p r with P r b . For a nonzero r,, A p r b introduces a small
The subscripts r and b denote the red and the blue channels, atmospheric effect that is not present in the NDVI of
respectively. Since the quantities p t , p: and pkIR are already
corrected for molecular scattering and gaseous absorption, the 6Prb = ' Y [ p b ( l - f b ) - p r ( 1 - f r ) ] (9)
only contamination due to the atmosphere is from aerosol
scattering and absorption (we do not discuss the effect of which, for example, for moderate atmospheric conditions over
undetected clouds [19] in this content). The reason for the grass is 6 P r b = -0.007. This atmospheric effect is accounted
self-correcting characteristics of the new index, ARVI, can for, together with the residual atmospheric effect on the path
be explained, by comparing the red normalized radiance, radiance, in the computation of the optimum value of y in
p:, used in the original NDVI, with the combined red-blue Section V. For pure vegetation, the difference between the
normalized radiance, p:b that is resistant to the atmospheric reflectance in the blue and the reflectance in the red is small
effects. Substituting p* from (4), and using the definition of (Fig. l(a)). Therefore, the value of the difference A p r b is
prb5/: small, and the corresponding atmospheric contamination is.
small. For soils, the difference is larger, and the contamination
is expected to be larger.
This qualitative discussion describes the principle of the "at-
mospherically resistant" new vegetation index, ARVI. Quanti-
tative simulations follows in the next sections.
Po-rb = Po-r(1 + 7) - y p o - b and A P r b = P r f r - P b f b

The main atmospheric effect in the new vegetation index, 111. DEMONSTRATION
OF THE NEW VEGETATION INDEX
ARVI, is introduced through the normalized path radiance In this section the resistance of the ARVI to atmospheric
Po-rb. Since the value of y is not specified as yet, we can
effects is compared with the atmospheric dependence of the
choose it so that Po-rb is minimal: Po-rb << po-r or even original NDVI. The indexes are applied to three vegetation
equal to zero. For a fixed aerosol model, we can choose y covers: grass [4], forest [36], and alfalfa [37], and two types
to be:
of soils, [4] and [37], as a function of the fraction of the
= Po-r/(Po-b - Po-,) * Po-rb =0 (8)
surface covered by the vegetation. The resultant reflectance is
a weighted average of the soil and vegetation reflectance (for
But since the relative value of pOpr to Po-b depend on the each vegetation and soil type) with the fraction of vegetation
aerosol characteristics, y will depend on the aerosol type. cover serving as the weighing factor. The simulation of the
Therefore, the degree of resistance of ARVI to atmospheric atmospheric effects is performed using the 5s radiative code
effects depends on the success in finding a single value [38J for the three spectral bands that compose the ARVI index:
of y (or a simple scheme to compute y) that reduces the 0.47+0.01 pm, 0.66+.0.025 pm, and 0.86520.02 pm. Since
atmospheric effect significantly on a global scale. In the atmospheric effects are slowly varying with the wavelength,
next few paragraphs we shall discuss it further using the the single value of the wavelength represents the average over
264 IEEE TRANSACTIONS ON GEOSCIENCEAND REMOTE SENSING, VOL. 30, NO. 2, MARCH 1992

the spectral band. The 5s code, that was shown to be very 1 ~ " ' " " " " " " " " ~
accurate for low to moderate aerosol content [38], is applied
0.8
to compute the normalized radiance detected by the satellite 3
n
for three atmospheric aerosol models: continental aerosol with z 0.6
visibility of 25 km and 10 km, and maritime aerosol with z
visibility of 25 km. As mentioned in the previous section, 0
!-
0.4
before computing the vegetation index, the radiances were a
corrected for molecular scattering by subtracting the path
kl 0.2

radiance for T, = 0 and dividing by f ( A ) for T, = 0.

E> 0
In Fig. 2, the vegetation indexes, ARVI and NDVI, are
-0.2
plotted for the three vegetation types as a function of the
fraction of the surface covered by vegetation. The compu-
tations of ARVI are done for y = 1, which will be shown 3 0.8
later to be the optimum single value of y in remote sensing n
applications. The indexes are plotted for the three atmospheric $ 0.6
models as well as for an atmosphere without aerosol loading, 0
where the vegetation indexes are computed directly from 2 0.4
the actual surface reflectances. The differences in the values kl
of the vegetation indexes between these models, shows the E> 0.2
magnitude of the atmospheric effect. The original NDVI varies
as a function of the atmospheric conditions by up to SNDVI 0
1 ~ " ' " " " " " " " " ~
= 0.2. The new index varies only within SARVI = 0.05. The
atmospheric effect on the NDVI and its improvement by the 3 0.8
ARVI index, are larger for densely vegetated surface than for n
bare soil. z
z 0.6
0
I-
IV. SENSITIVITY
STUDY 0.4
t;
Q

The effect of the atmospheric aerosol on the NDVI is defined

by the spectral dependence of the normalized atmospheric 8 0.2
path radiance po(A) and transmission function f ( A ) (4). Its
magnitude depends mainly on the ratio of the atmospheric path 0 0.2 0.4 0.6 0.8 1
radiance to the surface reflectance. The spectral dependence of FRACTION OF VEGETATION
po(A) and f ( A ) are determined by the aerosol type, size and
chemical composition. The magnitude of p o ( A ) and f(A) is Fig. 2. The original vegetation index, NDVI (dashed lines) and the new
vegetation index, ARVI (solids lines), as a function of the fraction of the
mainly determined by the aerosol content (or optical thickness) surface covered by vegetation for forest, alfalfa, and grass. Computations
~51. were performed using the 5s radiative code [38], and surface reflectances
In order to simplify the computations in the sensitivity were taken from Bowker et al. [4], Deering, [37], and Williams et al. [36].
Solar zenith angle is 60' and nadir view.
study, the atmospheric effect is described by simple analytical
expressions for p,,(A) and f ( A ) ,
with an Angstrom exponent a [34] within the range a = 0 - 2.
p 0 ( 4 = w ~ ( X ) T ~ ( A ) PA() ~/ (~~ P O P , ) (10) a = 0 corresponds to large particles such as the Saharan
and dust over arid regions [23], [26], (Y = 2 corresponds to
f ( A ) = 1 - b(~)wo(A).r,(X)(l/PO+ l / P u ) (11) small particles such as smoke particles resulting from fires
[28], [39]. Anthropogenic aerosol fall in between these values
where (,.A) is the spectral aerosol optical thickness, U,( A)
of a [35]. The radiative parameters: phase function, single
is the aerosol single scattering albedo, p ( 0 , A) the aerosol
scattering albedo, and asymmetry parameter are assumed to
phase function and b ( A ) can be expressed by the asymmetry
factor g(A) by: b ( A ) = (1 - g(A))/2 [38]. po and p, are, be spectrally independent and the only variable is the optical
respectively, the cosines of the solar zenith angle 00 and the thickness, T ~ ( X ) , with a dependence T ~ ( X ) = CA-". In
viewing zenith angle 8., Equations (10) and (11) are computed the application and optimization section (Sections I1 and V)
using a linearized single scattering approximation [31], [38]. the complete radiative 5s code of TanrC et al. [38] is used
The resistance of the ARVI to atmospheric effects depends with the appropriate spectral radiative properties of several
on the surface type and on the radiative properties of the aerosol models. The similarity between the results of this
atmosphere. The sensitivity study is based on three types of section and the full radiative computations given in Section
natural surfaces: boreal forest [36], grass [4], and bare soil V, show that, as expected, the present simplifications do not
[37]. The surface reflectance in the three spectral bands are alter substantially the results. Herein, the radiative parameters
reported in Table I. For the atmospheric component, the type are taken from a continental aerosol model, p(0,A) = 0.20,
of aerosols is accounted for by a power law size distribution w 0 ( X ) = 1.0 and g(A) = 2/3 [40] and standard or nominal
KAUFMAN AND TANR~:
ATMOSPHERICALLY RESISTANT VEGETATIONINDEX 265

TABLE I
REFLECTANCES
OF TYPICAL
SURFACES IN THE THREE
BANDS
Surface Cover Rbhe Rred RNIR
Soil [37] 0.110 0.190 0.243
Grass [4] 0.012 0.052 0.660
Forest [36] 0.010 0.016 0.210

atmospheric conditions corresponding to a = 1.3 ( a was

y = 0.0
computed using the data of Shettle and Fenn [41]). The view -II
and illumination angles are 60 = 60" and 6 = 0". -0.05 F

The sensitivity of the ARVI to concentration of atmospheric

aerosol can be expressed as a derivative, q,

77 = S(ARVI)/G.r, (12)
0.15 F GRASS
y = 2.0 A
The analytical expressions (10) and (11) allows us to compute
-
h

7 and to find the optimum value of y (defined in (5)) for which

z
the derivative, 7,is minimum. The sensitivity of the optimum 0.00
value to atmospheric and surface conditions is addressed in a
z
this section.
-0.15

A. Sensitivity to Aerosol Size Distribution

-0.30 I ' ' ' I " ' I " ' I I ' ' I ' ' ' I
For the three selected surfaces, the sensitivity of ARVI to
0.60
atmospheric variation, 7 = S(ARVI)/Sr,, is computed as a
function of a for four values of y: 0.0, 0.5, 1.0, and 2.0. y =
0 is the value for which the ARVI is equal to the NDVI. The
results are plotted as a function of the Angstrom exponent in h

2
0.20
0-40
0.00
1
Fig. 3.
The ARVI for bare soils (Fig. 3) has a weak sensitivity F -0.20
n
[q z f0.051 to aerosol optical thickness. For the moderate -0.40
values of soil reflectance (see Table I) it is less sensitive to -0.60
the presence of a scattering layer because of compensation
between the increase in the normalized path radiance, po(X),
and the decrease in the normalized flux f ( A ) (in (4)) [MI,
-0.80 L
-1.00 t '
0.00
"

0.40
I "

0.80
'
1.20
I

1.60
" ' I I ' ' I " -' -1

2.00
[31]. For instance, a variation in the aerosol optical thickness
of AT, = 0.2, results in an error of S(NDV1) M 0.01 in the ANGSTROM EXPONENT a
NDVI. Use of ARVI results in a systematic improvement for y
= 0.5. For y = 1.0 ARVI gives the same sensitivity [q = 0.051 Fig. 3. Sensitivity of the new vegetation index, ARVI, to variations in the
aerosol optical thickness 7,: 6(ARVI)/67,, as a function of the Angstrom
as NDVI for nominal conditions (a: = 1.30), e.g., continental exponent a , for four values of y: 0.0, 0.5, 1.0, and 2.0. = 0 is the value
aerosol, and better results for larger particles (a = 0.0). y = 2.0 for which ARVI is equal to the NDVI. Results are given for soil, grass, and
results in a large overcorrection for any value of a resulting forest. The exponent a is used here to describe the aerosol type (cy = 0 for
large dust particles, cy = 0.2 for maritime aerosol, CI = 1.3 for continental
in a sensitivity to the aerosol effect that is larger than for the aerosol, and cy = 2 for smoke particles).
original NDVI. For grass, due to the lower surface reflectance
in the red channel, NDVI shows a larger sensitivity, around
q = -0.23 and ARVI gives better results for any values of a overcorrection of the atmospheric effect, resulting in errors
and y 2 0.5. Forest, with even lower surface reflectance in the that are similar or even larger than the errors in the original
red band, presents the strongest sensitivity, 7 z -0.75 which. NDVI (y = 0.0). A value of y of 1.0 seems more appropriate
corresponds to a large error in the NDVI of S(NDV1) = 0.15 if information on the aerosol type is not available.
for an aerosol optical thickness of 0.2. As for grass, ARVI
provides for the forest better results for any value of a and
y 2 0.5, with best results obtained for y between 1.0 and 2.0. B. Relation Between y, a, and Vegetation Fraction
These results confirm the importance of aerosol scattering In Fig. 4 we examine the sensitivity of the new vegetation
effects on vegetation index for densely vegetated areas [18]. index to variation in the fraction of vegetation cover (for grass)
ARVI shows substantially less sensitivity to the aerosol content and in the Angstrom coefficient, a, for four values of y.
than NDVI. Except for bare soil, y = 1.0 and y = 2.0 reduces Contour lines of equal S(ARV1) for r, = 0.1 are plotted in
substantially the sensitivity of ARVI to atmospheric effects the parameterization space with two dimensions: vegetation
in respect to the sensitivity of the NDVI. For small particles fraction and a, for four value of y: 0 (no correction), 0.5,
( a > 1.5) a high value of y, y = 2.0, may result in an 1.0, and 2.0. This parameterization space covers all the basic
266 IEEE TRANSACTIONS ON GEOSCIENCEAND REMOTE SENSING, VOL. 30, NO. 2, MARCH 1992

ro 0.5 4 4 0 1 0

0.00 0.20 0.40 0.60

VEGETATION FRACTION

lines of equal change in ARVI - 6(ARVI) are plotted for

0.80

conditions of remote sensing, varying the surface properties

by changing the vegetation cover (though only for grass and
one type of soil) and changing the size distribution, and the
1.00
8

4
Bts
I
B
$I

4U
2.00

1.60-/--

0.00
0.00
- 0.20
-
VEGETATION FRACTION

Y = 2.0

0.40 0.60 0.80

VEGETATION FRACTION
&to015

- -0005
-0- 4.010
1

1.00

Fig. 4. Sensitivity of the new vegetation index, ARVI, to variation in the fraction of vegetation cover (for grass) and in the Angstrom coefficient, a. Contour
= 0.1 in the parameterization space with dimensions of the vegetation fraction and a. Results are
given for four value of y: 0.0 (no correction), 0.5, 1.0 (the optimum value), and 2.0. This parameterization space covers, in a simplified form, all the possible
conditions of remote sensing, varying the surface properties by changing the vegetation cover and changing the size distribution,and the corresponding optical
characteristics of the aerosol layer. Note that continental aerosol is characterized by a = 1.3 and maritime aerosol by a = 0.2.

C. Sensitivity to the Relation Between the Reflectance

in the Blue and the Red Channel
In Fig. 5 the relation between the sensitivity of ARVI to the
corresponding optical characteristics of the aerosol layer. Note atmospheric effect and the values of the surface reflectance
that for y = 1.0, the isoline that corresponds to zero error in the red (0.66e0.025 pm) and the blue (0.47 kO.01 pm)
covers the “main grounds” in this parameterization space. It channels are displayed. A scatter diagram of 40 different
starts from low vegetation cover (e.g., arid to semiarid regions surface covers, including water, natural vegetation, agricultural
like the Sahel) and low value of a (e.g., dust, which is common crops, and soils (data are taken from [4]) is plotted. The
in these regions [26]), and moves on to a = 1.3-1.9 for average difference between the reflectances in the two channels
vegetation fraction of 0.3-1.0, which is typical for vegetated is 0.08 for all the data, 0.04 for vegetation only, and 0.13
regions [28], [35], [39], [42]. for soils only (see Table 11). The dependence, of an error in
There are some exclusions to this remarkably suitable the ARVI - SARVI resulting from a variation in the aerosol
variation of the zero isoline for y = 1.0. For larger values optical thickness of 0.2, on the reflectance in the blue and
of vegetation cover (more than 60%), and for very humid the red channels is shown by isolines for constant SARVI
conditions, which increase the particle size [35], [43] and values. Computations of ARVI were performed for continental
decrease the value of a, the error in ARVI is large though model (a = 1.3) and for y = 1.0 as a function of the surface
still half as large as that for the NDVI (y = 0). For very reflectance in the blue and the red channels. The reflectance
small vegetation cover, and for unknown aerosol particle size, in the near IR is 0.4, the average value in Table 11. The error
a value of y = 0.5 leads to better results, since for this value
in ARVI as a function of the surface reflectance in the two
the isoline for zero error is very steep, thus resulting in a very bands, for y = 1.0 covers most of the surface types in Fig. 5,
good correction for any value of a. for a range of SARVI between -0.03 and tO.01. Therefore,
As a conclusion, unless the aerosol model is know apriori, a ARVI should be suitable to be applied to other surface covers
single value of y (y = 1.0) should be adopted for minimizing than those shown in Fig. 2.
the atmospheric effects. In specific conditions (e.g., the Sahel),
a better value of y (y = 1.0) can be used (e.g., y = 0.5). The
single value of y gives better results than the NDVI in any v. APPLICATION AND OPTIMIZATION
possible remote sensing condition. The sensitivity, presented in the previous section, demon-
KAUFMAN AND TA"? ATMOSPHERICALLY RESISTANT VEGETATION INDEX 267

TABLE I1
RELATIONBETWEEN (0.66k0.025 n m )
IN THE RED CHANNEL
THE REFLECTANCE . I
AND I N THE BLUECHANNEL
(0.47k0.01 jtm)

surface/ reflectances [4] Ratio

property blue red NIR blue/red difference NDVI ARVI
0.47 pm 0.66 pm 0.86 pm
all surfaces 0.11=0.11 0.19-cO.17 0.41-cO.19 0.64k0.24 0.08k0.08 0.38-cO.33 0.26k0.40
vegetation 0.06-cO.04 0.10=0.07 0.45k0.18 0.7120.25 0.04k0.05 0.63k0.25 0.55*0.32
soils 0.18-cO.14 0.31eo. 18 0.35k0.18 0.56k0.19 0.13k0.09 0.09k0.06 -0.08~0.08

- 0.05
E, 0.40 L
8
e 0.04 - t
\
+ CONTINENTAL
X MARITIME
9
a
0.30 -
> -
w a 0.03
2
0.20 0.02 -
-
w
0
= 0.10
0.01
2
0
w -- %
-I
ya o.ooa.!35n ,O.k
0.k 0.;0'
e ,O.k
a ,o.;o ' , O . L

REFLECTANCE IN THE BLUE (0.47pm)

0.06
Fig. 5. The relation between the reflectance in the red (0.66k0.025 pm)
and the blue (0.47-cO.01 pm) bands for soils and water ( 0 )and for vegetation
- 0.05
( A ) taken from Bowker et al. [4]. The sensitivity of the new vegetation index, 0.04
ARVI, to the values of the reflectance in these two channels is also shown by
isolines for constant hARVl values (lines) computed for a difference in the 0.03
aerosol optical thickness of 0.2. Computations were performed for continental 0.02
model ( 0 = 1.3) and for -/ = 1.0. The reflectance in the near IR was kept
constant at 0.4. 0.01
0
-0.5 0 0.5 1 1.5 2
strated the dependence of the optimum value of the parameter Y
y, introduced in the new vegetation index ((5)-(7)), on the
Fig. 6. The average difference, CIARVI, between the value of ARVI for the
surface and atmospheric properties. In this section, the simu- continental and maritime atmospheric models and that for no atmospheric
lation model (5s radiative code), used for the demonstration effect as a function of -,.
The simulation is performed using the S S code. (a)
of the new vegetation index in Fig. 2, and described in Section OARVI averaged over six values of the fraction of vegetation cover between
0 and 1, for the three vegetation types (forest, grass, and alfalfa). For the
111, is also used to compute the optimum value of y for continental model the data were averaged for the two visibilities (10 and 25
which the error (SARVI) in the derived index is minimum. km). The average between the three atmospheric models is also shown (0).(b)
The optimization is for several surface covers and atmospheric OARVI averaged over the three atmospheric models and the three vegetation
covers, for several fractions of the vegetation cover.
models. In Fig. 6 the average difference between the value of
ARVI for the continental and maritime atmospheric models
and that for the actual surface reflectance is plotted as a MODIS images. In specific regions and seasons with defined
function of the value of y. In Fig. 6(a) the difference is vegetation and atmospheric properties, a special value may be
averaged over six values of the fraction of vegetation between applied. These possibilities are further explored in Fig. 6(b).
0 and 1, for the three vegetation types (forest, grass, and As was already indicated in the sensitivity study of Section
alfalfa), and for two soil types. For the continental model IV, a lower fraction of vegetation cover, corresponds to a
the results are averaged for the two visibilities (10 and lower optimum value yopt. Since the value of SARVI in Fig.
25 km). For Maritime aerosol only visibility of 25 km is 6(b) is asymmetrical around Topt = 1 the resulting range of
used. The continental model represents moderate particle size variation in ARVI around yopt = 1 is small, between SARVI
with Angstrom exponent cy = 1.3, and the maritime model = 0.005 and 0.015. As a result, application of a constant value
represents large particles with cy = 0.2. The average between yopt = 1, even in cases where y = 0.5 is more appropriate, will
the value of SARVI for the two continental models and the result in a small error SARVI for all surface covers (SARVI
maritime model is also plotted. While for the continental model = 0.005 to 0.015) while a smaller value, y = 0.5, which is
the optimum value of y is around Topt = 0.9, for maritime better for soils, would result in large errors for pixels covered
model, with larger particle size (smaller value of cy) the value by vegetation (SARVI up to 0.06) that are located in the
is larger, Topt = 1.7. The plot that corresponds to the average same image. Therefore, based on the present simulation it is
of all these models shows minimum SARVI for yopt = 1. recommended, that as a rule a single value of y should be
In future application of the ARVI to MODIS data, an used. Unless the surface and atmospheric conditions are well
average value of yopt will be probably used to analyze most of known and require a substantially different value of y (e.g., for
268 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 30, NO. 2, MARCH 1992

0.1
b) 80-10096 VEGETATION COVER
MARmMEAEROSOL
0.08

2 0.06
5
h 0.04
I2
0.02

d) @20%VEGETATION COVER
MRmMEAEROSOL

0 0.4 0.8 1.2 1.6 2 0 0.4 0.8 1.2 1.6 2

Y Y

Fig. 7. Analysis of the dependence of the ratio 6ARVI/AARVI (ratio of noise to signal) on the value of -( for the two atmospheric models and for the
humid vs. arid regions. The effect of the continental model ( a = 1.3) is shown in Figs. 7(a) and (c), and the effect of maritime model ( a = 0.2) in Figs.
7@) and (d). Humid regions, are simulated by 80% and 100% vegetation cover (Figs. 7(a) and (b)) and arid conditions by 0 and 20% vegetation cover
(Figs. 7(c) and (d)). The simulation is performed using the 5s code. For 7(a) and (b) the curves are: (-) - alfalfa, (- - -) - forest, (- - - - - - - ) -
grass [4],(- - - -) - 80% alfalfa, (- - - -) - 80% forest, (- ... - ...-) - 80% grass. For 7(c) and (d) the curves are: (-) - soil (41, (- - -) -
soil [37], (- - - - - - - ) - 20% alfalfa, (- - - -) - 20% forest, (- ._.- ... -) - 20% grass.

vegetated area with large aerosol particles or for low vegetation and (d)). The conclusions are very similar to the one drawn
fraction with small particle size), this value of yopt = 1 may from the analysis of Figs. 4 and 6. The main advantage in the
be used. presentation of Fig. 7 is that we can see the individual errors
The data in Fig. 6 can also be used to summarize the for dry and humid regions for specific values of y.
comparison between the residual dependence of ARVI and the The consequences of using a single value of y,y = 1.0, for
NDVI on the atmosphere. For the selected value y = 1, ARVI varying vegetation cover and atmospheric conditions, is shown
is in average four times less dependent on the atmospheric in Fig. 8. The data are averaged for the three vegetation types
effects (SARVI four times smaller) than the NDVI (e.g., ARVI and for the three atmospheric models, therefore, this figure
for y = 0.0). For maritime atmosphere (or other larger aerosol represents the conditions for which the aerosol model and
particles), the improvement is only by a factor of two. For the type of vegetation that covers the surface are unknown.
continental aerosol, the improvement is by more than factor The value of y for which the error SARVI is within 0.005
of 4. Again, the improvement of the resistance of ARVI to from the minimal value, is plotted in this figure. Due to the
atmospheric effects is much greater for vegetated regions than relatively small atmospheric effect over bare soil, the range
over arid regions (for y = 1). of y for which SARVI is within 0.005, is larger for low
The ratio of SARVIIAARVI, shown in Fig. 7, is the ratio vegetation cover. Therefore, for the whole range of vegetation
between the residual noise in the index (SARVI) due to fraction, the additional error SARVI is always less than 0.005
the atmospheric effect and the signal that is sensitive to for y = 1.0.
surface characteristics (AARVI). Here AARVI is defined as
the difference between the value of ARVI for 100% vegetation VI. RELATIONTO OTHER VEGETATION INDEXES
cover and the value of ARVI for no vegetation. A detailed A change in the definition of the vegetation index in
analysis of the dependence of this ratio on the value of y for remote sensing applications may cause a discontinuity in
the two atmospheric models and for the humid vs. arid regions global vegetation patterns. Therefore, it is important to check
is shown in the figure. The effect of the continental aerosol if it is possible to relate previous studies of the NDVI to
is shown in Figs. 7(a) and (c), and the effect of maritime vegetation properties derived from the new index - ARVI.
aerosol in Figs. 7 (b) and (d). Humid regions, are simulated The comparison in Fig. 2 shows that both indexes have a
by 80% and 100% vegetation cover (Figs. 7(a) and (b)) and similar range for the five surface covers used in that figure.
arid conditions by 0 and 20% vegetation cover (Figs. 7(c) To expand the comparison between the two indexes for the
KAUFMAN AND TANRE: ATMOSPHERICALLY RESISTANT VEGETATION INDEX 269

(5), replaces the red (0.66-pm) channel in the present NDVI

with a combination of the red and the blue channels. This
combination has self-correction properties for the atmospheric
effect.
Simulations, for various atmospheric conditions, of the veg-
etation indices for Lambertian surfaces using the 5s radiative
transfer code, show that ARVI is, on average, four times
less sensitive to atmospheric effects than the NDVI. The
J improvement is much better for vegetated surfaces for which
the atmospheric effect is larger than for soils. It is much
~ ~ ~ I ~ ~ ~ I . ~ ~ I . ~ ’ ’ . ~ ~ I ’ ” I ’ ’ ’
0 better for moderate or small size of the aerosol particles (e.g.,
-0.2 0 0.2 0.4 0.6 0.8 1 1.2 continental, urban, or smoke aerosol) than for large particle
FRACTION OF VEGETATION size (e.g., maritime aerosol or dust). Due to a nice coincidence,
Fig. 8. The value of -/ for which the error ClARVI is within 0.005 from the the same optimal value of the parameter y (y = 1) that defines
minimal value, plotted as a function of the vegetation cover. The data are the weighing of the blue band radiance in the ARVI definition,
averaged for the three vegetation types and for the three atmospheric models.
Therefore, this figure represents selection of -( in conditions for which the is found for vegetated areas with small to moderate aerosol
aerosol model is unknown, and the type of vegetation that covers the surface particle size and for arid regions with large particle size.
- -,
-,
is also unknown. Solid line - values for minimum 6ARV1, dashed lines
values for 6ARVI 0.005 above the minimum.
Therefore, a single value of y may be used in all or most
remote sensing applications.
In this study only results of simulation for a few AFGL
pure surface reflectance, we plotted a scatter diagram of the atmospheric models [41] and a power law size distribution
values of ARVI as a function of the corresponding values of the and Lambertian surfaces were reported, for one direction of
NDVI. The vegetation indexes were computed for 44 surface illumination and observation. It is important to apply the
covers from Bowker et al. [4] without atmospheric effects. new index to simulations using different characteristics of
The correlation coefficient between the two indexes is 0.98 the surface, atmospheric aerosol and observation geometry.
and the relation is: The index has to be tested against field data where the
radiances are recorded over the same surface for varying
ARVI = -0.18 + 1.17 * NDVI (13) atmospheric conditions. Experience with the analysis of ARVI
The similarity and high correlation between the two indexes in these conditions will verify how good ARVI is for global
should make the transfer from index to index relatively easy. applications and whether a single value of the weighing y
The concept of resistance to atmospheric variations in should be used or y should vary with surface and atmospheric
ARVI, which is based on the atmospheric information con- conditions.
tained in the blue channel, can be applied also to other
vegetation indexes than the NDVI. For example Huete [44]
ACKNOWLEDGMENT
suggested to modify the definition of NDVI in order to make
it less dependent on variation of soil properties, for partial We would like to thank Shana Mattoo and Vincent Rose
vegetation cover. The resultant Soil Adjusted Vegetation Index from ARC for part of the computations used in this work.
(SAVI) is (eq. (4) in [44] ) in present notations: The name of the new index, ARVI, was proposed by Brent
Holben from NASNGSFC. The concern of Jim Tucker, Brent
Holben, Chris Justice, and others with atmospheric effects on
where C is a constant that is chosen as to minimize the soil the NDVI provided the inspiration for this research.
effect (C = 0.5). Following the same analysis used in the
derivation of (14) by Huete [44], the combined Soil Adjusted REFERENCES
and Atmospherically Resistant Vegetation Index, SARVI, is
C. 0. Justice, J. G. R. Townshend, B. N. Holben, and C. J. Tucker,
obtained from (5) and (14): “Analysis of the phenology of global vegetation using meteorological
satellite data,” Int. J. Remote Sensing,vol. 6, pp. 1271-1318, 1985.
SARVI = (1 + C)(pI;JIR - C. J. Tucker, J. R. G . Townshend, and T. E. Goff, “African land
cover classification using satellite data,” Science N. Y., vol. 227, pp.
The value of C in (15) is expected to be similar to that in (14). 369.3587-3594, 1985.
C. J. Tucker, C. L. Vapraet, M. J. Sharman, and T. Van Ittersum,
“Satellite remote sensing of total herbaceous biomass production in the
VII. CONCLUSIONS Senegalese Sahel: 198G1984,” Remote Sensing Environment, vol. 17,
pp. 233-249, 1985.
In order to reduce the dependence of the vegetation index D. E. Bowker, R. E. Davis, D. L. Myrick, K. Stacy, and W. T.
NDVI on the atmospheric properties, a new atmospherically Jones, “Spectral reflectances of natural targets for use in remote sensing
resistant vegetation index, ARVI, is proposed and developed studies,” NASA Reference Publication 1139.
P. J. Sellers, “Canopy reflectance photosynthesis and transpiration,” Int.
to be used for remote sensing of vegetation from the Earth J . Remote Sensing, vol. 6, p. 1335, 1985.
Observing System MODIS sensor. Same index can be used C. J. Tucker and P. J. Sellers, “Satellite remote sensing of primary
productivity,” In?.J . Remote Sensing, vol. 7, pp. 1395-1416, 1986.
for remote sensing from the Landsat TM sensor, and the S. E. Nicholson, M. L. Davenport, and A. D. Malo, “A comparison of
Earth Observing System HIRIS sensor. The index, defined by the vegetation response to rainfall in the Sahel and east Africa, using
270 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 30, NO. 2, MARCH 1992

NDVI from NOAA AVHRR,” Climate Change, vol. 17, pp. 209-214, M. C. Pereira, and R. Rasmussen, “Biomass Burning Airborne and
1990. Spaceborne Experiment in the Amazonas (BASE-A),’’ J. Geophys. Res.,
[8] C. J. Tucker, I. Y. Fung, C. D. Keeling, and R. H. Gammon, “Re- in press, 1992.
lationship between atmospheric CO2 variations and a satellite derived [29] J. T. Peterson, E. C. Flowers, G. J. Berri, C. L. Reynolds, and J. H.
vegetation index,” Nature, vol. 319, pp. 195-199, 1986. Rudisil, “Atmospheric turbidity over central north Carolina,” J. Appl.
[9] S. N. Goward, D. G. Dye, and C. J. Tucker, “North American vegetation Meteorol., vol. 20, pp. 22’+241, 1981.
patterns observed by NOAA-7 AVHRR,” Vegetation, vol. 64, p. 3, 1985. [30] S. Chandrasekhar, Radiative Transfer. New York: Dover, 1960.
[lo] S. D. Prince and C. J. Tucker, “Satellite remote sensing of rangelands in [31] R. S. Fraser and Y. J. Kaufman, “The relative importance of scattering
Botswana 11: NOAA AVHRR and herbaceous vegetation,” Int. J. Remote and absorption in remote sensing,” IEEE Trans. Geosci. Remote Sensing,
Sensing, vol. 7, pp. 1555-1570, 1986. vol. 23, pp. 625-633, 1985.
[ll] J. R. G. Townsend and C. 0. Justice, “Analysis of the dynamics of [32] Deschamps, M. Herman and D. Tanrt, “Modeling of the atmospheric
African vegetation using the NDVI,” Int. J. Remote Sensing, vol. 7, pp. effect and its application to the remote sensing of ocean color,” Appl.
1435-1446, 1986. Opt., vol. 23, pp. 3751-3758, 1983.
[12] B. N. Holben, Y. J. Kaufman, and J. D. Kendall, “NOAA-11 AVHRR [33] Y. J. Kaufman, “Measurements of the aerosol optical thickness and the
visible and near-IR inflight calibration bands,” Int. J . Remote Sensing, path radiance -implications on aerosol remote sensing and atmospheric
vol. 11, pp. 1511-1519, 1990. corrections,’’ submitted for publication, J. Geophys. Res., 1992.
[13] Y. J. Kaufman and B. N. Holben, “Calibration of the AVHRR visible [34] C. E. Junge, Air Chemistry and Radiochemistry. New York: Academic
and near-IR bands by atmospheric scattering, ocean glint and desert Press, 1963.
reflection,” Inf. J. Remote Sensing, in press, 1992. [35] Y. J. Kaufman and R. S. Fraser, “Light extinction by aerosols during
[14] Y. J. Kaufman, “Atmospheric effect on remote sensing of surface summer air pollution,” J . Appl. Mefeorol., vol. 22, pp. 1694-1706, 1983.
reflectance,” SPIE, vol. 475, pp. 20-33, 1984. [36] D. Williams and C. L. Walthcell, “Helicopter based multispectral data
[15] R. S. Fraser, Y. J. Kaufman, and R. L. Mahoney, “Satellite measure- collection over the northern experimental forest: Preliminary results
ments of aerosol mass and transport,” J. Atmos. Environ., vol. 18, pp. from the 1989 field season,” in Proc. IGARSS 90. New York: IEEE,
2577-2584, 1984. 1990, pp. 875-878.
[16] B. N. Holben and R. S. Fraser, “Red and near IR sensor response to [37] D. W. Deering, “Field measurements of bidirectional reflection,” Chap-
off-nadir viewing,” Int. J. Remote Sensing,vol. 5, pp. 145-160, 1984. ter 2 in Optical Remote Sensing, Technology and Application, G. Asrar,
[17] B. N. Holben, “Characteristics of maximum value composite images for Ed. New York: Wiley, 1989.
temporal AVHRR data,” Int. J. Remote Sensing, vol. 7, pp. 1417-1437, [38] D. Tanrt, C. Deroo, P. Duhaut, M. Herman, J. J. Morcrette, J. Perbos,
1986. and P. Y. Deschamps, “Description of a computer code to simulate the
[18] D. T a d , B. N. Holben, and Y. J. Kaufman, “Atmospheric correction satellite signal in the solar spectrum: 5s code,” Inr. J. Remote Sensing,
algorithm for NOAA-AVHRR products, theory and application,” IEEE vol. 11, pp. 659-668, 1990.
Trans. Geosci. Remote Sensing, vol. 30, pp. OO(MO0, 1992. [39] B. N. Holben, Y. J. Kaufman, A. Setzer, D. Tanrt, and D. E. Ward,
[19] Y. J. Kaufman, “The effect of subpixel clouds on remote sensing,” fnt. “Optical properties of aerosol from biomass burning in the tropics,
J. Remofe Sensing, vol. 8, pp. 839-857, 1987. BASE-A,” in Global Biomass Burning. Cambridge, MA: MIT Press,
[20] V. V. Salomonson, W. L. Barnes, P. W. Maymon, H. E. Montgomery, 1991, pp. 403-411.
and H. Ostrow, “MODIS: Advanced facility instrument for studies of [40] R. A. McClathchey, R. W. Fenn, J. E. A. Selby, F. E. Volz, and J. S.
the earth as a system,” IEEE Trans. Geosci. Remote Sensing, vol. 27, Garing, “Optical properties of the atmosphere,” AFCRL 72-0497, AD
pp. 145-153, 1989. 75075, 1972.
[21] M. D. King, Y. J. Kaufman, P. Menzel, and D. TanrC, “Determination of [41] E. P. Shettle and R. W. Fenn, “Models for the aerosol of the lower
cloud, aerosol, and water vapor properties from the Moderate Resolution atmosphere and the effect of humidity variations on their optical
Imaging Spectrometer (MODIS).” IEEE Trans. Geosci. Remote SensinP, I,
properties,” AFGL-TR 790214, Opt. Phys. Div., Air Force Geophysics
vol. 30, pp. 2-27, 1992. Laboratory, Hanscom AFB bfA, 1979.
1221
-~ Y. J. Kaufman and C. Sendra. ‘‘Algorithm for automatic atmosoheric
L [42] K. T. Whitby, “The physical characteristics of sulfur aerosols,” Atmos.
corrections to visible and near-IR satellite imagery,” Int. J. Remote Environ., vol. 12, pp. 135-159, 1978.
Sensing, vol. 9, pp. 1357-1381, 1988. [43] G. Hand, “An attempt to intercept the humidity dependencies of the
[23] D. Tanre, P. Y. Deschamps, C. Devaux, and M. Herman, “Estimation
of Saharan aerosol optical thickness from blurring effects in Thematic
-
aerosol extinction and scattering coefficient,” Atmos. Environ., vol. 15,
pp. 403-406, 1981.
Mapper data,” J. Geophys. Res., vol. 93, pp. 15955-15964, 1988. [44] A. R. Huete, “A soil-adjusted vegetation index (SAVI),” Remote Sensing
[24] Y. J. Kaufman, R. S. Fraser, and R. A. Ferrare, “Satellite remote sensing Environment, vol. 25, pp. 295-309, 1988.
of large scale air pollution - method,” J. Geophys. Res., vol. 95, pp.
9895-9909, 1990.
[25] Y. J. Kaufman, “The atmospheric effect on remote sensing and its
correction,” Chapter 9 in Optical Remote Sensing, Technology and
Application G. Asrar, Ed. New York: Wiley, 1989. Yo” J. Kalmnsn, for a photog@ and biography, please sec page 222 of
[26] B. N. Holben, T. F. Eck, and R. S. Fraser, “Temporal and spatial TRANSACTIONS.
this issue of the
variability of aerosol optical depth in the Sahel region in relation
to vegetation remote sensing,” Inr. J. Remote Sensing, vol. 12, pp.
1147-1163, 1991.
[27] Y. J. Kaufman, C. J. Tucker, and I. Fung, “Remote sensing of biomass
burning in the tropics,” J. Geophys. Res., vol. 95, pp. 9927-9939, 1990. Didier Tmd, for a photograph and biography. please see page 222 of this
[28] Y. J. Kaufman, A. Setzer, D. Ward, D. TanrC, B. N. Holben, P. Menzel, issue of the TRANSACTIONS.