1520 0469 Jas3385.1
1520 0469 Jas3385.1
947
*Laboratory for Atmospheres, NASA Goddard Space Flight Center, Greenbelt, Maryland
⫹Laboratoire d’Optique Atmospherique, Université des Sciences et Technologies de Lille, Villenueve d’Ascq, France
#Laboratory for Atmospheres, NASA Goddard Space Flight Center, Greenbelt, Maryland
@Laboratory for Terrestrial Physics, NASA Goddard Space Flight Center, Greenbelt, Maryland
ABSTRACT
The Moderate Resolution Imaging Spectroradiometer (MODIS) aboard both NASA’s Terra and Aqua
satellites is making near-global daily observations of the earth in a wide spectral range (0.41–15 m). These
measurements are used to derive spectral aerosol optical thickness and aerosol size parameters over both
land and ocean. The aerosol products available over land include aerosol optical thickness at three visible
wavelengths, a measure of the fraction of aerosol optical thickness attributed to the fine mode, and several
derived parameters including reflected spectral solar flux at the top of the atmosphere. Over the ocean, the
aerosol optical thickness is provided in seven wavelengths from 0.47 to 2.13 m. In addition, quantitative
aerosol size information includes effective radius of the aerosol and quantitative fraction of optical thickness
attributed to the fine mode. Spectral irradiance contributed by the aerosol, mass concentration, and number
of cloud condensation nuclei round out the list of available aerosol products over the ocean. The spectral
optical thickness and effective radius of the aerosol over the ocean are validated by comparison with two
years of Aerosol Robotic Network (AERONET) data gleaned from 132 AERONET stations. Eight thou-
sand MODIS aerosol retrievals collocated with AERONET measurements confirm that one standard
deviation of MODIS optical thickness retrievals fall within the predicted uncertainty of ⌬ ⫽ ⫾0.03 ⫾0.05
over ocean and ⌬ ⫽ ⫾0.05 ⫾ 0.15 over land. Two hundred and seventy-one MODIS aerosol retrievals
collocated with AERONET inversions at island and coastal sites suggest that one standard deviation of
MODIS effective radius retrievals falls within ⌬reff ⫽ ⫾0.11 m. The accuracy of the MODIS retrievals
suggests that the product can be used to help narrow the uncertainties associated with aerosol radiative
forcing of global climate.
& Additional affiliation: Science Systems and Applications, Inc., Lanham, Maryland.
** Current affiliation: Joint Center for Earth Systems Technology, University of Maryland, Baltimore County, Baltimore,
Maryland.
⫹⫹ Additional affiliation: Joint Center for Earth Systems Technology, University of Maryland, Baltimore County, Baltimore,
Maryland.
## Current affiliation: Goddard Earth Sciences and Technology Center, University of Maryland, Baltimore County, Baltimore,
Maryland.
@@ Additional affiliation: Goddard Earth Sciences and Technology Center, University of Maryland, Baltimore County, Baltimore,
Maryland.
&& Additional affiliation: Department of Geography, University of Maryland, College Park, College Park, Maryland.
Corresponding author address: Dr. Lorraine Remer, Code 913, NASA Goddard Space Flight Center, Greenbelt, MD 20771.
E-mail: [email protected]
as described in Kaufman et al. (1997a,b). A retrieval pendently. The retrieved optical thicknesses along with
following path A is given a quality control value of 3, the continental model’s single scattering albedoes
“very good.” An alternative path used for brighter sur- (o0.47, o0.66) and phase functions (P0.47, P0.66) at the
faces is described below. appropriate scattering angle are used to calculate the
path radiance in each wavelength using the single scat-
3) CHOOSING THE AEROSOL MODELS tering approximation:
The estimated surface reflectances (s0.47, s0.66) and o0.47 ⫽ o0.470.47P0.47;
the measured mean top-of-atmosphere reflectances 共2兲
o0.66 ⫽ o0.660.66P0.66,
(0.47, 0.66) are used as input into the continental model
lookup table (LUT) to retrieve values for the aerosol where o0.47 and o0.66 are the path radiances at 0.47 and
optical thickness at 0.47 and 0.66 m (0.47, 0.66). In the 0.66 m, respectively. The spectral dependence of the
land algorithm, the two wavelengths are derived inde- path radiance distinguishes between dust (dominated
by coarse mode) and nondust (dominated by fine more, Eq. (4) magnifies relative errors in the retrievals
mode). Kaufman et al. (1997a) describes how the algo- of individual path radiances (o 0.47, o 0.66) by up to 4 to
rithm uses the ratio of path radiances, o0.66/o0.47, to 8 times the original percentages. This is especially pro-
make a three-branched decision whether the aerosol is nounced for situations of low aerosol loading. As a rule,
pure dust, nondust, or mixed. Using the continental satellite retrievals of aerosol optical thickness are more
model does not impact the resulting ratio. It is simply robust than corresponding retrievals of aerosol size,
used to remove the molecular and surface contributions and retrievals of size parameter require sufficient aero-
and to isolate the aerosol reflectance in both channels. sol loading in order to be valid (Ignatov et al. 1998;
The thresholds for the decision tree are Remer et al. 2002).
In practice the “mixed” aerosol criterion [Eq. (3c)] is
o0.66Ⲑo 0.47 ⬍ 0.72 THEN pure nondust 共3a兲
seldom found, and is usually either 0 or 1 over land.
o0.66Ⲑo0.47 ⬎ 0.9 ⫺ 0.01共⌰ ⫺ 150⬚兲 THEN pure dust However, in a monthly mean analysis, at least qualita-
tively, the pattern of fine-mode fraction corresponds to
共3b兲
the global distributions of dust and nondust sources and
0.72 ⱕ o0.66Ⲑo0.47 ⱕ 0.9 transport. Figure 2 shows four monthly mean values of
the fraction of total aerosol optical thickness attributed
⫺ 0.01共⌰ ⫺ 150⬚兲 THEN mixed 共3c兲 to the “nondust” aerosol model. Red shades indicate
for scattering angles ⌰ ⫽ 150° to 168°. For scattering that “nondust” dominates over the monthly mean.
angles ⬍150°, ⌰ is simply set to 150° for the boundary Purple shades indicate that “pure dust” dominates.
to collapse to 0.9 in that angle range. If the aerosol is Blank areas in black are where no retrievals were made
mixed, then the fraction that the fine mode contributes due to overly bright surfaces, monthly domination by
to the total optical thickness, ⫽ f/tot, is given by clouds, or snow. Note that these plots do not differen-
tiate between high and low aerosol loading. Sensitivity
⫽1⫺
冋
o0.66
o0.47
⫺ 0.72 册
, 共4兲
to aerosol size decreases in very clean regions.
Note that nondust is a misnomer because even
0.90-0.01共⌰-150⬚兲 ⫺ 0.72 though the nondust aerosol models are dominated by
their fine modes, each also contains a coarse mode as
again, where ⌰ is set to 150° when ⌰ ⬍ 150°. Equation well. Three nondust models are available and are de-
(4) gives an approximation to the fine-mode fraction scribed in Table 1, along with a description of the con-
over land. Its accuracy is dependent on the assumptions tinental model and the dust model. The urban/
of aerosol models and surface reflectances. Further- industrial model remains unchanged from Kaufman et
FIG. 2. Monthly mean plots of fraction of total aerosol optical thickness attributed to nondust or fine-mode aerosol over land.
Fraction 1.0 indicates all fine mode. Fraction of 0.0 indicates all coarse mode.
al. (1997a) and Remer and Kaufman (1998). The new 4) DETERMINING AEROSOL OPTICAL THICKNESS
developing world—moderate absorption model is
In both the nondust and pure dust cases, the esti-
based on the biomass burning model of Kaufman et al.
mated surface reflectances (s0.47, s0.66) and the mea-
(1997a) and Remer et al. (1998), slightly modified to
sured mean top-of-atmosphere reflectances (0.47, 0.66)
reflect the more recent study of Dubovik et al. (2002a).
are used as input into the chosen model’s LUT to re-
The developing world—heavy absorption model uses
trieve values for the aerosol optical thicknesses, fluxes
the same size parameters as the other developing world
and other parameters. A full second retrieval is made
model, but allows for the greater aerosol light absorp-
from the appropriate model’s LUT. This second full
tion noted in Africa (Ichoku et al. 2003; Eck et al. 2003;
retrieval differs from Kaufman et al. (1997a) that de-
Dubovik et al. 2002a). Similar to the original concep-
scribes, instead, a correction based on the single scat-
tion of the algorithm, the current version uses season
tering approximation to the preliminary continental
and geography to choose between the three nondust
model retrieval described above. The final step in the
models (Dubovik et al. 2002a). However, the bound-
process is to interpolate the values at 0.47 and 0.66 m
aries have changed. Figure 3 gives the new distribution
using an Ångström law in order to report optical thick-
of the three nondust models.
ness and flux values at 0.55 m. Note that the algorithm
If the aerosol is identified as dust by Eq. (3), then the
does not retrieve aerosol directly at 0.55 m over land
dust model of Table 1 is used. However, pure dust
because there is no established relationship between
poses a problem with the dark target method. The as-
that channel and the surface reflectance at 2.13 m
sumption that 2.13 is transparent to aerosols and pro-
analogous to Eq. (1) and, therefore, no method for es-
vides direct information from the surface does not hold
timating surface reflectance at 0.55 m. However, 0.55
when the aerosol is composed of large particles. There-
m is an important wavelength often used in global
fore, Eq. (1) is not expected to hold in the pure dust
climate modeling and analysis, and therefore MODIS
case for very dark surfaces. However, over moderately
reports a value for that wavelength even though there is
bright surfaces, near the point of critical reflectance
no direct retrieval.
(Kaufman 1989) the surface contribution is negligible
and the procedure can continue with minimal uncer-
5) ALTERNATIVE FOR BRIGHT SURFACES
tainty introduced from the surface. Therefore, in the
pure dust case retrievals are made only when 2.13 falls The standard dark target retrieval path, described
between 0.15 and 0.25. above, that uses specific aerosol models requires a
TABLE 1. Size distribution parameters and single scattering albedo used in the MODIS lookup table for the land algorithm.
Mode rg(m) rv(m) Vo(m) o(470) o(660)
Continental aerosol model
Water soluble 0.005 0.176 1.09 3.05 0.96 0.96
Dustlike 0.50 17.6 1.09 7.364 0.69 0.69
Soot 0.0118 0.050 0.693 0.105 0.16 0.16
Urban/industrial
Accumulation 1 0.036 0.106 0.6 F1 0.96 0.96
Accumulation 2 0.114 0.21 0.45 F2 0.97 0.97
Coarse 1 0.99 1.3 0.3 F3 0.92 0.92
Coarse 2 0.67 9.5 0.94 0.045 0.88 0.88
Developing world—moderate absorption
Accumulation 0.061 0.13 0.50 F4 0.91 0.89
Coarse F5 F6 F7 F8 0.84 0.84
Developing world—strong absorption
Accumulation 0.061 0.13 0.50 F4 0.86 0.85
Coarse F5 F6 F7 F8 0.84 0.84
Desert dust
Mode 1 0.0010 0.0055 0.755 6.0 ⫻ 10⫺8 0.015 0.015
Mode 2 0.0218 1.230 1.160 0.01 0.95 0.95
Mode 3 6.24 21.50 0.638 0.006 0.62 0.62
FIG. 3. Distribution of the nondust models used in the derivation of aerosol over land. Single scattering
albedo values given in parentheses. Uncolored regions use the urban/industrial aerosol model. Solid
black regions use the moderate absorption aerosol model all year. Large checkerboard pattern in East
Asia denotes a region of strong absorption aerosol model all year. Africa is divided by region and season.
North of the equator, during the burning season (Nov–May) the strong absorption aerosol model is used,
while the moderate absorption model is used the remainder of the year. South of the equator, the
burning season shifts to Jun–Oct when the strong absorption model is used, while in the remainder of
the year the algorithm uses the urban/industrial model.
minimum of 12 dark pixels in every 20 ⫻ 20 pixel nomi- tinental model is used in the retrieval. The aerosol op-
nal 10-km box. Path B, of Fig. 1, represents an alterna- tical thickness and flux are derived from the LUT for
tive if the surface is too bright to support 12 dark pixels 0.47 m. These parameters are extrapolated to 0.55 and
in the standard manner. The upper limit of the 2.13 0.66 m using the spectral dependence of the continen-
value is permitted to increase as a function of the slant tal model.
path until a final upper bound of 2.13 ⫽ 0.40 is reached. Figure 4 shows a scene from the eastern part of
When the sun is at zenith and the satellite view nadir, southern Africa, where the surface reflectance is mod-
path B collapses back to path A requirements. How- erately bright. The top panel shows the MODIS re-
ever, as the photon path increases, more and more sig- trieval of aerosol optical thickness at 0.55 m following
nal originates from the atmosphere, and the contribu- path A of the algorithm, which insists on 12 dark target
tion from the surface reflectance becomes less and less pixels. The bottom panel shows the modified version of
important. This is especially true at the 0.47-m chan- the algorithm, which allows retrievals over brighter sur-
nel where atmospheric signal is highest and the surface faces by following both path A and path B. From the
usually darkest. For this reason, the alternative path B image we see how the extension to brighter surfaces
retrieves aerosol only in the 0.47-m channel. At least fills in holes without introducing suspicious artifacts.
12 pixels must again meet the path B criteria, otherwise In this example, permitting path B increases the num-
the procedure ends with no retrieval made, and fill ber of retrievals over land from 7060 to 17 849. As we
values are placed in the output fields. Path B is consid- extend to brighter surfaces, we move away from the
ered to be less accurate than path A, and the quality biomass burning regions into cleaner regions. Thus
control (QC) is set to 0, representing “poor quality.” the mean optical thickness of the granule decreases
Because of the greater uncertainty over these brighter from 0.20 to 0.15, but the standard deviation of the
surfaces and because we retrieve in only one wave- optical thickness remains constant at 0.15. For the 285
length and cannot use the path radiance ratio to distin- granules collected over southern Africa during the
guish between dust and nondust aerosol, only the con- Southern African Regional Science Initiative (SAFARI
FIG. 4. MODIS-derived aerosol optical thickness at 0.55 m for an image of the east coast of southern
Africa. (top) The traditional dark target method described by path A in Fig. 1 is used. (bottom) The
results after extending the retrieval to brighter surfaces as described by path B of Fig. 1. By extending
to brighter surfaces the number of retrievals over land in this image increases from 7060 to 17 849.
2000) campaign (Swap et al. 2003), extending to to the process described in Tanré et al. (1997), the
brighter surfaces increases the number of land retriev- masking of clouds and sediments, the special hand-
als by 130%. ling of heavy dust including dust retrievals over glint,
and revisions of the lookup table are new. As in the
b. The ocean algorithm
land algorithm, after the water vapor, ozone, and
The mechanics of the ocean algorithm are illustrated carbon dioxide corrections are applied, the first step
in Fig. 5. Although the core inversion remains similar in the ocean algorithm is to organize the reflectance
from the six wavelengths used in the procedure (0.55, 1) CLOUD AND SEDIMENT MASKING
0.66, 0.86, 1.24, 1.6, 2.13) into nominal 10-km boxes of
20 ⫻ 20 pixels at 500-m resolution. This requires If all 400 pixels in the box are identified as water
degrading the resolution of the 250-m channels (0.66 pixels, the algorithm then begins the arduous task of
and 0.86). The ocean algorithm requires all 400 pixels separating “good” pixels from “cloudy” pixels. The
in the box to be identified as ocean pixels by the MOD/ standard MOD35 cloud mask includes using the bright-
MYD35 mask. This helps to minimize problems in- ness in the visible channels to identify clouds. This pro-
troduced by shallow water near the coasts. If any land cedure will mistake heavy aerosol as “cloudy” and miss
is encountered, the entire box is left for the land al- retrieving important aerosol events over ocean. On the
gorithm, but quality is decreased for coastal land re- other hand, relying on IR tests alone permits low-
trievals. altitude, warm clouds to escape and be misidentified as
“clear,” introducing cloud contamination in the aerosol THEN “not cloudy”, but the quality of the retrieval
products. Thus, our primary cloud mask is based on the
is “poor” 共QC ⫽ 0兲.
difference in spatial variability between aerosols and
clouds (Martins et al. 2002). The algorithm marches Rayleigh
IF 共0.66 ⬍ 1.5 0.66 兲 OR 共1.38 ⬍ 0.01兲
through the 10 km ⫻ 10 km box, examining the stan-
dard deviation of 0.55 in every group of 3 ⫻ 3 pixels. AND NOT 共1.38Ⲑ1.24 ⬎ 0.3兲
Any group of nine pixels with standard deviation THEN “not cloudy”, and the quality of the retrieval
greater than 0.0025 is labeled as “cloudy,” and all nine
pixels in the group are discarded (Martins et al. 2002). is “good” 共QC ⫽ 3兲.
The only exception to this rule is for heavy dust, which A quality flag of QC ⫽ 0 permits a retrieval at the
may at times be as spatially inhomogeneous as clouds. orbital level (level 2), but prohibits the retrieval from
Heavy dust is identified by its absorption at 0.47 m contributing to the long-term global aerosol statistics
using the ratio (0.47/0.66). This quantifies the differ- (level 3). Only retrievals with QC ⬎ 0 contribute to the
ence that our eyes witness naturally. Dust absorbs at level 3 quality weighted products.
blue wavelengths and appears brown. Clouds are spec- The final mask applied to the data is the sediment
trally neutral and appear white to our eyes. If 0.47/0.66 mask, which identifies which ocean scenes are contami-
⬍ 0.75, then the central pixel of the group of nine is nated by river sediments (Li et al. 2003), discarding
identified as “dust” and will be included in the retrieval those pixels. The sediment mask takes advantage of the
even if it is inhomogeneous. This is a conservative strong absorption by water at wavelengths longer than
threshold that requires very heavy dust in order to 1 m. Spectral reflectances over water with suspended
avoid clouds. Less restrictive thresholds would permit sediments show elevated values in the visible, but not in
more dust retrievals, but might accidentally permit the longer wavelengths. This is a spectral signature that
cloud contamination. is quite different from clear ocean water and also dif-
The spatial variability test separates aerosol from ferent from airborne dust.
most cloud types, but sometimes fails at the centers of All pixels that have evaded the cloud mask tests and
large, thick clouds and also with cirrus, both of which the sediment mask are sorted according to their 0.86
can be spatially smooth. The centers of large, thick value. The darkest and brightest 25% are discarded,
clouds are very bright in the visible, so we identify these thereby leaving the middle 50% of the data. The filter
clouds when 0.47 ⬎ 0.40. This is an extremely high is used to eliminate residual cloud contamination, cloud
threshold that could be a nonabsorbing aerosol optical shadows, or other unusual extreme conditions in the
thickness greater than 5.0, but only for nonabsorbing box. Because the ocean cloud mask and the ocean sur-
aerosol. Absorbing aerosol never produces that high face are expected to be more accurate than their coun-
value of reflectance and will pass this cloud test un- terparts over land, the filter is less restrictive than the
scathed. Some high values of nonabsorbing aerosol may one used in the land retrieval. Of the 400 pixels in the
be discarded along with bright clouds, but this confu- original box, at least 10 must remain for the 0.86-m
sion is rare. Heavy aerosol loading, with ⬎ 5.0, ab- channel after the masking and filtering. Otherwise, no
sorbs somewhat at 0.47 m and fails to reach the 0.40 retrieval is attempted and all aerosol products in the
threshold value exhibited by very bright white clouds. 10-km box are given fill values. If there are at least 10
Cirrus clouds are identified with a combination of good pixels, the mean reflectance and standard devia-
infrared and near-infrared tests. Three infrared tests tion are calculated for the remaining good pixels at the
provided by the standard MODIS cloud mask, MOD35, six pertinent wavelengths.
are examined. These tests are IR cirrus test (byte 2, bit
4), 6.7-m test (byte 2, bit 8), and Delta IR test (byte 3, 2) OCEAN GLINT AND INTERNAL CONSISTENCY
bit 3) (Ackerman et al. 1998). If any one of the three The glint angle is defined as
tests indicates clouds, we label the pixel as “cloudy.”
The near-infrared cirrus test is based on the reflectance ⌰glint ⫽ cos⫺1关共coss cosv兲 ⫹ 共sinssinvcos兲兴, 共5兲
in the 1.38-m channel and the ratio 1.38 /1.24 (Gao et where s, v, and are the solar zenith, the satellite
al. 2002). It is applied in the algorithm as a three-step zenith, and the relative azimuth angles (between the
process: sun and satellite), respectively (Levy et al. 2003). Note
IF 共1.38Ⲑ1.24 ⬎ 0.3兲 THEN “cloudy” that Fresnel reflection corresponds to ⌰glint ⫽ 0. The
ocean algorithm was designed to retrieve only over
IF 共0.10 ⱕ 1.38Ⲑ1.24 ⱕ 0.30兲 AND 共1.38 ⬎ 0.03兲 dark ocean, away from glint (except in one special case
Rayleigh
AND 共0.66 ⬎ 1.50.66 兲 described below). If ⌰glint ⬎ 40°, we can avoid glint
contamination and proceed with the retrieval. The al-
THEN “cloudy” gorithm performs several consistency checks of the
IF 共0.10 ⱕ 1.38Ⲑ1.24 ⱕ 0.30兲 AND 共0.01 ⱕ 1.38 ⱕ 0.03兲 spectral reflectances. Depending on the outcome of
Rayleigh
these consistency checks, the algorithm may either de-
AND 共0.66 ⬎ 1.50.66 兲 clare the reflectances to be beyond the range necessary
for a successful inversion and exit the procedure or Equation (6) means that the spectral reflectance
continue onto the inversion after assigning quality flags measured from the satellite that corresponds to the
(QC values) to each wavelength. LUT value, LUT (tot
0.55) for the determined values of
tot
and 0.55, is a weighted average of the reflectance values
3) INVERSION PROCEDURE for an atmosphere with a pure fine mode “f” and opti-
cal thickness tot
0.55 and the reflectance of an atmosphere
The inversion procedure is aptly described in Tanré
with a pure coarse mode “c” also with the same tot 0.55. In
et al. (1997) and Levy et al. (2003). Following Tanré et
appendix A, we show that ⫽ f0.55/tot 0.55, the fraction of
al. (1996), we know that the six reflectances measured
total optical thickness at 0.55 m contributed by the
from MODIS and used in the ocean retrieval (0.55–2.13
fine mode.
m) contain three pieces of information about the
For each of the 20 combinations of one fine mode
aerosol. From this information we derive three param-
and one coarse mode, the inversion finds the pair of
eters: the optical thickness at one wavelength (tot
0.55), the
tot
0.55 and 0.55 that minimizes the error () defined as
reflectance weighting parameter at one wavelength
(0.55), and the effective radius, which is the ratio of the
third and second moments of the aerosol size distribu-
冉 冊
冪
6
m ⫺ LUT 2
tion. The inversion is based on a LUT that now consists
of four fine modes and five coarse modes [Table 2,
兺
⫽1
N
m ⫹ 0.01
⫽ , 共7兲
following Levy et al. (2003), which differs from the 11 6
TABLE 2. Refractive indices, median, standard deviation, and effective radius for the aerosol models used in the MODIS lookup
table for the ocean algorithm. Models 1–4 are fine modes and models 5–9 are coarse modes. From Levy et al. (2003).
⫽ 0.47→0.86 m ⫽ 1.24 m ⫽ 1.64 m ⫽ 2.13 m rg reff Comments
1 1.45–0.0035i 1.45–0.0035i 1.43–0.01i 1.40–0.005i 0.07 0.40 0.10 Wet water soluble type
2 1.45–0.0035i 1.45–0.0035i 1.43–0.01i 1.40–0.005i 0.06 0.60 0.15 Wet water soluble type
3 1.40–0.0020i 1.40–0.0020i 1.39–0.005i 1.36–0.003i 0.08 0.60 0.20 Water soluble with humidity
4 1.40–0.0020i 1.40–0.0020i 1.39–0.005i 1.36–0.003i 0.10 0.60 0.25 Water soluble with humidity
5 1.45–0.0035i 1.45–0.0035i 1.43–0.0035i 1.43–0.0035i 0.40 0.60 0.98 Wet sea salt type
6 1.45–0.0035i 1.45–0.0035i 1.43–0.0035i 1.43–0.0035i 0.60 0.60 1.48 Wet sea salt type
7 1.45–0.0035i 1.45–0.0035i 1.43–0.0035i 1.43–0.0035i 0.80 0.60 1.98 Wet sea salt type
8 1.53–0.003i (0.47) 1.46–0.000i 1.46–0.001i 1.46–0.000i 0.60 0.60 1.48 Dustlike type
1.53–0.001i (0.55)
1.53–0.000i (0.66)
1.53–0.000i (0.86)
9 1.53–0.003i (0.47) 1.46–0.000i 1.46–0.001i 1.46–0.000i 0.50 0.80 2.50 Dustlike type
1.53–0.001i (0.55)
1.53–0.000i (0.66)
1.53–0.000i (0.86)
TABLE 3. Values of the normalized extinction coefficients, asymmetry parameter, single scattering albedo for the nine ocean models
of Table 2.
Model 0.47 m 0.55 m 0.66 m 0.87 m 1.24 m 1.61 m 2.13 m
Normalized extinction coefficients
1 1.538 1.0 0.661 0.286 0.085 0.046 0.016
2 1.300 1.0 0.764 0.427 0.169 0.081 0.030
3 1.244 1.0 0.796 0.483 0.211 0.104 0.042
4 1.188 1.0 0.836 0.549 0.269 0.140 0.060
5 0.963 1.0 1.037 1.081 1.055 0.919 0.745
6 0.980 1.0 1.034 1.100 1.177 1.166 1.081
7 0.986 1.0 1.025 1.079 1.162 1.225 1.215
8 0.977 1.0 1.023 1.086 1.185 1.192 1.124
9 0.964 1.0 1.000 1.039 1.098 1.117 1.105
Asymmetry parameter
1 0.5755 0.5117 0.4478 0.3221 0.1773 0.1048 0.0622
2 0.6832 0.6606 0.6357 0.5756 0.4677 0.3685 0.2635
3 0.7354 0.7183 0.6991 0.6510 0.5590 0.4715 0.3711
4 0.7513 0.7398 0.7260 0.6903 0.6179 0.5451 0.4566
5 0.7450 0.7369 0.7328 0.7316 0.7330 0.7411 0.7282
6 0.7770 0.7651 0.7503 0.7358 0.7314 0.7461 0.7446
7 0.8035 0.7912 0.7738 0.7506 0.7335 0.7443 0.7461
8 0.7534 0.7200 0.6979 0.6795 0.7129 0.7173 0.7190
9 0.7801 0.7462 0.7234 0.7065 0.7220 0.7176 0.7151
Single scattering albedo
1 0.9735 0.9683 0.9616 0.9406 0.8786 0.5390 0.4968
2 0.9782 0.9772 0.9757 0.9704 0.9554 0.8158 0.8209
3 0.9865 0.9864 0.9859 0.9838 0.9775 0.9211 0.9156
4 0.9861 0.9865 0.9865 0.9855 0.9819 0.9401 0.9404
5 0.9239 0.9358 0.9451 0.9589 0.9707 0.9753 0.9774
6 0.8911 0.9026 0.9178 0.9377 0.9576 0.9676 0.9733
7 0.8640 0.8770 0.8942 0.9175 0.9430 0.9577 0.9669
8 0.9013 0.9674 1.0000 1.0000 1.0000 1.0000 1.0000
9 0.8669 0.9530 1.0000 1.0000 1.0000 1.0000 1.0000
the accompanying tot0.55 and 0.55 that minimizes . The choices, the inversion is less sensitive to calibration er-
solution may not be unique. The average solution is the rors. We know now that the true uncertainty in the
average of all solutions with ⬍ 3%, or if no solution input reflectances is 1.8%–1.9% (Guenther et al. 2002),
has ⬍ 3%, then the average is of the three best solu- not the 1% assumed in the study. This will have little
tions. Once the solutions are found, then the chosen additional effect on the accuracy of the optical thick-
combination of modes is the de facto derived aerosol ness retrievals, but may further reduce the accuracy of
model. A variety of parameters can be inferred from the size retrievals.
the chosen size distribution including spectral optical
thickness, effective radius, spectral flux, mass concen- 5) FINAL CHECKING
tration, etc.
Before the final results are output, additional consis-
4) NOTE ON ERRORS tency checks are employed. In general, if the retrieved
optical thickness at 0.55 m is greater than ⫺0.01 and
Tanré et al. (1997) explore the sensitivity of the in- less than 5, then the results are output. Negative optical
version procedure to various sources of error. For ex- depths are possible, occurring only in situations with
ample, they tested the retrieval sensitivity by introduc- low optical depth. This situation arises from errors in
ing a random calibration error of 1%, finding no sys- assumptions of surface conditions, aerosol properties,
tematic bias and negligible impact upon the optical or calibration expectations. We choose to report small
thickness retrievals. The effects on the retrieved size negative values in order not to introduce a positive bias
parameters were much greater. Errors due to the 1% in long-term statistics for clean marine conditions, but
calibration error translated to uncertainties of ⫾0.25 negative optical depths are given lower quality flags.
for 0.55⫽1 and ⫾0.50 for 0.55⫽0. For small particles Quality flags may be adjusted during this final checking
less than 0.40 m, the effective radius was retrieved to phase.
within ⫾0.10 m, but for large particles greater than 1.0
m the effective radius was severely underestimated.
6) SPECIAL CASE: HEAVY DUST OVER GLINT
Since that sensitivity study, the lookup table has been
changed and the number of possible aerosol models If ⌰glint ⱕ 40°, then we check for heavy dust in the glint.
available to the inversion has been reduced. With fewer Heavy dust has a distinctive spectral signature because
FIG. 6. Examples of MODIS aerosol products at the level 2 stage (MOD04). The data represent a 5-min granule collected on 7 Jul
2002 from 1835 to 1840 UTC when smoke from Canadian fires had been transported south over the mid-Atlantic states and then out
to sea. (Upper left) A true color image created from level 1b reflectances, (upper right) aerosol optical thickness at 550 nm, (lower left)
reflected flux at 550 nm, and (lower right) fraction of fine mode. The sun glint over the Gulf of Mexico can be seen in the true color
red–green–blue (RGB) image. The other panels exhibit a sharp dividing boundary in this region due to the application of the glint mask.
fine mode (0.55), and the reflected flux at the top of the TABLE 4. Contents of MODIS aerosol level 2 hdf file (MOD04):
atmosphere at 0.55 m. The flux is defined as the hemi- Time and geometric information. Two-dimensional arrays of 204
⫻ 135 are indicated.
spherical irradiance at a particular wavelength. Over
land, the flux is defined for zero surface reflectance and Name of product Dimension Status
computed consistently with the same aerosol param- Longitude 2D Diagnostic
eters used in the optical thickness retrieval. Over ocean, Latitude 2D Diagnostic
the flux is defined for the same rough ocean surface Scan_Start_Time 2D Diagnostic
model and the same aerosol parameters derived in the Solar_Zenith 2D Diagnostic
Solar_Azimuth 2D Diagnostic
optical thickness retrieval. Errors in the retrieved aero- Sensor_Zenith 2D Diagnostic
sol optical thickness that may be introduced by inaccu- Sensor_Azimuth 2D Diagnostic
rate assumptions of the aerosol model are subsequently Scattering_Angle 2D Diagnostic
canceled when those same aerosol model assumptions Cloud_Mask_QA 2D Diagnostic
are used with the optical thickness to calculate top-of-
atmosphere fluxes; errors introduced by assumptions of
surface reflectance will remain. For a monthly average, dated,” “not yet validated,” “derived,” “experimental,”
the consistent calculations produce a highly accurate or “diagnostic.” A validated product indicates that sub-
measure of the flux more accurate than the optical stantial comparison was made to ground-based data
thickness itself. and that the retrieval is well characterized so that error
Figure 6 shows orbital-based level 2 products cut bars can be defined and comfortably applied to the
from 5-min segments of a satellite orbit, called granules. retrieval product (Ichoku et al. 2002; Chu et al. 2002;
The level 2 products, designated as MOD04 files, con- Remer et al. 2002). Examples of validation are shown in
tain 64 separate products, all connected to the aerosol section 4. “Not yet validated” indicates that the re-
retrieval. The level 2 files are produced every day and trieved parameter has not yet been well characterized,
represent the first level of MODIS aerosol retrieval. In but that data are being collected and analysis is under-
addition, statistics based on the level 2 aerosol retriev- way. “Derived” is a parameter that follows from the
als can also be found in level 3 files, designated as retrieval’s choice of aerosol model and the magnitude
MOD08 files. These level 3 files contain parameters of the retrieved optical thickness. Definitions of some
produced from the entire MODIS atmospheres team of the derived parameters are given in appendix B. A
and include such parameters as water vapor and cloud derived parameter is not directly retrieved and there
characteristics along with the aerosol information. The are no expectations of ever validating a derived param-
level 3 data are averaged to a 1° latitude/longitude grid eter with independent data. “Experimental” is a scien-
and are produced every day (MOD08_D3), averaged tific product that may have future applications but,
every 8 days (MOD08_E3) or averaged on a monthly as of now, is too innovative to be well characterized.
basis (MOD08_M3). They include both statistics calcu- “Diagnostic” refers to output that is either an auxillary
lated equally from all the data, and also statistics or intermediate parameter. Diagnostic parameters are
weighted by the quality of each individual retrieval. meant to aid in understanding the final product, but
Quality weights of 0 will prevent poor retrievals from will never themselves become validated Recommenda-
affecting the calculated statistics of the quality weighted tions for choosing particular products are given in ap-
quantities. Further information about the level 3 prod- pendix C.
ucts can be found in King et al. (2003) and online at
http://modis-atmos.gsfc.nasa.gov.
All together there are 64 aerosol products at level 2: 4. Validation of aerosol products
9 products describe geometry and location, 3 products
are joint land and ocean products, 23 are land-only Our primary means of validation is comparison with
products, and 29 are ocean-only products. Tables 4–7 equivalent measurements from Aerosol Robotic Net-
lists all 64 products. The three joint land and ocean work (AERONET) ground-based sun/sky radiometers
products are simple two-dimensional arrays of one (Holben et al. 1998). The AERONET instruments
wavelength (Fig. 6). The land-only and ocean-only
products contain an additional dimension. In many
cases this additional dimension is wavelength. Tables TABLE 5. Contents of MODIS aerosol level 2 hdf file (MOD04):
4–7 list the wavelengths for each product where appli- Global land and ocean products, at 550 nm.
cable. The additional dimension in the ocean-only Name of product Dimension Status
products can designate either the “best” solution or the
Optical_Depth_
“average” solution from the ocean retrieval as de- Land_And_Ocean 2D Validated
scribed above in section 2b(3). Both solutions are re- Optical_Depth_Ratio_
ported for some parameters, although they are often Small_Land_And_Ocean 2D Not yet validated
identical. Reflected_Flux_
Land_And_Ocean 2D Derived
Tables 4–7 also list whether the product is “vali-
TABLE 6. Contents of MODIS aerosol level 2 hdf file (MOD04): Land products.
Name of product Dimension Status
Corrected_Optical_Depth_Land 0.47, 0.55, 0.66 m Validated
Optical_Depth_Ratio_Small_Land 0.55 m Not yet valid
Mass_Concentration_Land 2D Derived
Ångström_Exponent_Land 0.66/0.47 Not yet valid
Reflected_Flux_Land 0.47, 0.55, 0.66 m. Derived
Transmitted_Flux_Land 0.47, 0.66 m Derived
Aerosol_Type_Land 2D Diagnostic
Continental_Optical_Depth_Land 0.47, 0.66 m Diagnostic
Estimated_Uncertainty_Land 0.47, 0.66 m Diagnostic
Mean_Reflectance_Land_All 0.47, 0.66, 2.13 m Diagnostic
Standard_Deviation_Reflectance_Land_All 0.47, 0.66, 2.13 m Diagnostic
Cloud_Fraction_Land 2D Diagnostic
Number_Pixels_Percentile_Land 0.47, 0.66 m Diagnostic
Mean_Reflectance_Land 0.47, 0.66, 0.87, 2.13, 3.75 m Diagnostic
STD_Reflectance_Land 0.47, 0.66, 0.87, 2.13, 3.75m Diagnostic
Quality_Assurance_Land See QA plan Diagnostic
Path_Radiance_Land 0.47, 0.66 m Experimental
Error_Path_Radiance_Land 0.47, 0.66 m Diagnostic
Critical_Reflectance_Land 0.47, 0.66 m Experimental
Error_Critical_Reflectance_Land 0.47, 0.66 m Diagnostic
Quality_Weight_Path_Radiance 0.47, 0.66 m Experimental
Quality_Weight_Critical_Reflectance_Land 0.47, 0.66 m Experimental
Quality_Assurance_Crit_Ref_Land 0.47, 0.66, 0.87, 2.13, 3.75 m Diagnostic
measure spectral aerosol optical thickness, to within reff, whenever conditions are favorable (Dubovik et al.
⬃0.01 for the channels 0.38, 0.44, 0.50, 0.67, 0.87, and 2000). The methodology of comparing temporally vary-
1.02 m (Eck et al. 1999). They also can derive ambi- ing AERONET data with spatially varying MODIS
ent, total atmospheric column aerosol effective radius, data is described in Ichoku et al. (2002). In the follow-
TABLE 7. Contents of MODIS aerosol level 2 hdf file (MOD04): Ocean products.
Name of product Dimension Status
Effective_Optical_Depth_Average_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Validated
Optical_Depth_Small_Average 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Not yet validated
Optical_Depth_Large_Average 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Not yet validated
Effective_Radius_Ocean Best, average Validated
Optical_Depth_Ratio_Small_Ocean_0.86micron Best, average Not yet validated
Mass_Concentration_Ocean Best, average Derived
Cloud_Condensation_Nuclei_Ocean Best, average Derived
Ångström_Exponent_1_Ocean 0.55/0.87 Not yet validated
Ångström_Exponent_2_Ocean 0.87/2.13 Not yet validated
Reflected_Flux_Average_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Derived
Transmitted_Flux_Average_Ocean 0.47, 0.55, 0.66, 0.87 Derived
1.24, 1.63, 2.13 m
Asymmetry_Factor_Average_Ocean 0.47, 0.55, 0.66, 0.87 Derived
1.24, 1.63, 2.13 m
Backscattering_Ratio_Average_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Derived
Solution_Index_Ocean_Small Best, average Diagnostic
Solution_Index_Ocean_Large Best, average Diagnostic
Least_Squares_Error_Ocean Best, average Diagnostic
Optical_Depth_by_models_Ocean Nine models Diagnostic
Effective_Optical_Depth_Best_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Diagnostic
Optical_Depth_Small_Best 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Diagnostic
Optical_Depth_Large_Best 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Diagnostic
Reflected_Flux_Best_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Diagnostic
Transmitted_Flux_Best_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Diagnostic
Asymmetry_Factor_Best_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Diagnostic
Backscattering_Ratio_Best_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Diagnostic
Cloud_Fraction_Ocean 2D Diagnostic
Number_Pixels_Used_Ocean 2D Diagnostic
Mean_Reflectance_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Diagnostic
STD_Reflectance_Ocean 0.47, 0.55, 0.66, 0.87 1.24, 1.63, 2.13 m Diagnostic
Quality_Assurance_Ocean See QA plan Diagnostic
ing validation, we use AERONET level 1.5 data, which algorithms’ development and history, starting from the
are cloud screened but not quality assured, primarily most recent version and going backward in time can be
because final calibration is not applied (Smirnov et al. found at the MODIS atmospheres Web site (http://
2000). The procedure that collocates MODIS and modis-atmos.gsfc.nasa.gov/MOD04_L2/history.html).
AERONET data is applied during a very short window Much of the algorithm modifications concern experi-
of opportunity (⬃24 h) after the operational MODIS mental products, improved cloud/snow/water masking,
aerosol product has been processed and before that internal bookkeeping, or minor adjustments meant to
data are transferred to the archive. The data are purged improve long-term statistics. However, the last adjust-
from the operational facility after transfer. Level 2.0 ment over land to version 4.2.1 modifies the land
AERONET data become available only several months lookup tables to increase aerosol absorption as needed
behind real time depending on site. During a system- in certain regions (Ichoku et al. 2003). This is a signifi-
wide reprocessing of MODIS data, which may take cant change, which will not be reflected in the valida-
place months to years after real time, if AERONET tion plots described below.
level 2.0 data are available then, they are extracted and A preliminary validation of the aerosol products was
added to the MODIS validation dataset. The two years made of the data collected in the first months of opera-
of collocated data in this analysis do not benefit from tion. The results are reported in Ichoku et al. (2002),
reprocessing and are therefore taken from the real-time Chu et al. (2002), and Remer et al. (2002). The prelimi-
processing that use AERONET level 1.5. Figure 7 nary validation compared 2–3 months of MODIS aero-
shows the distribution of the 132 AERONET stations sol optical thickness and effective radius retrievals to
used in the comparisons to be described below. Al- the same parameters observed (optical thickness) or
though North American and European stations domi- derived (effective radius) from AERONET radiom-
nate the database, all continents (except Antarctica), all eters. The preliminary validation from the limited
oceans, and all aerosol types are represented. dataset showed the MODIS-derived parameters agreed
Validation is an ongoing effort. Not only do aerosol with the AERONET parameters to within the expected
conditions vary in location and time, requiring a con- prelaunch uncertainties: ⫾0.05 ⫾ 0.15 for optical
tinued effort to validate the algorithms under various thickness over land, and ⫾0.03 ⫾ 0.05 for optical
conditions, but the algorithms themselves evolve. The thickness and ⫾25% for effective radius over ocean.
FIG. 7. The distribution of the 132 AERONET stations used to validate MODIS land and ocean aerosol
retrieval algorithms.
a. Validation of aerosol optical thickness all wavelengths are greater than the expected offset of
0.05 at low optical thickness. Furthermore, these plots
1) VALIDATION OVER LAND show a positive bias at low optical thickness, suggesting
At the time of this analysis there were two years of a possible instrument calibration issue or, more likely,
MODIS aerosol products collocated with AERONET that surface reflectance may be improperly represented
retrievals (1 August 2000 to 1 August 2002). Figure 8 in a systematic way at certain locations and seasons.
shows plots of 5906 collocated points over land at The regression equations also show that in all wave-
wavelengths 0.47, 0.55, and 0.66 m. There are no lengths the slopes are less than one. Ichoku et al. (2003)
AERONET measurements at the MODIS wavelengths demonstrate that underprediction of aerosol optical
of 0.47 and 0.55 m; therefore, the AERONET values thickness at higher aerosol loadings can be attributed to
in the plots of Fig. 8 have been interpolated from the insufficient light absorption in the aerosol models in
values at 0.44 and 0.87 m. The 0.50-m AERONET certain regions of the world, specifically Africa. This is
channel is not used for interpolation because not all the reason why the strong absorption model (Table 1)
AERONET stations have that channel and the proce- was introduced to the algorithm in the version 4 deliv-
dure that matches MODIS and AERONET data must ery. The data shown in Fig. 8 and other figures in this
paper are combined from Terra Collections 003 and
be uniform and automatic. Although AERONET does
004. We expect the underprediction of optical thickness
make measurements at 0.675 m, the values at this
at high aerosol loading to be less of an issue after re-
wavelength have also been interpolated from 0.44 and
processing with the updated algorithm and after per-
0.87 m, due to occasional calibration drift at this chan-
forming analyses on only Collection 004 data, but the
nel in the AERONET instruments. These calibration
issue of the offset at low aerosol loading will remain.
issues due to gradual filter degradation are identified
Even with the deviations described above, the results
and corrected in the quality assured level 2 AERONET of Fig. 8 indicate that the algorithm is retrieving aerosol
data, but unfortunately the collocation with MODIS is optical thickness over land to roughly within the ex-
done in real-processing time and cannot wait for the pected accuracy. On a global basis, 61%, 68%, and 71%
postdeployment AERONET calibration corrections. of the retrievals at the 0.47, 0.55, and 0.66 m, respec-
Therefore, the 0.675 m is not used and, instead, the tively, fall within expected error. From a global per-
information is transferred from the more reliable 0.44- spective, only the blue channel is falling outside of the
and 0.87-m AERONET channels. Also, the MODIS error bars slightly more often than the prelaunch ex-
value at 0.55 m is not a direct retrieval but an inter- pectations of 66%. The average at 0.55 m is 0.18 for
polation from the 0.47- and 0.66-m retrievals; thus, the the land global database at defined AERONET
plot at 0.55 m in Fig. 8 is a comparison of two inter- stations. The percent error (relative error) between
polated values. The interpolation of AERONET data is MODIS retrievals and AERONET observations at
done on a log–log plot assuming linearity between 0.44 0.55 m is 41%, showing a positive bias in which
and 0.87 m. The error in the interpolation varies be- MODIS overestimates . The overestimate corrobo-
tween 0% and ⬃10% depending on the aerosol type rates the positive offsets seen at low to moderate values
(due to nonlinear spectral dependence), with fine- of optical thickness in Fig. 8. Table 8 shows the percent
mode-dominated aerosol at high optical thickness in- of retrievals falling within the expected error lines for
troducing the most error, and a mixed- or coarse-domi- the entire dataset, as well as grouped by specific region.
nated aerosol introducing the least (Eck et al. 1999). In some regions the retrievals are poorer than in other
Figure 8 represents the scatterplot between MODIS regions. Specifically the North American continent, es-
retrievals and AERONET observations, collocated in pecially Alaska/Canada, is proving to be difficult. This
space and time. The data were sorted according to region also exhibits the lowest , which contributes to
AERONET aerosol optical thickness. An average was the high relative error. The absolute error in Alaska/
then calculated for every 300 points and plotted. At Canada is comparable to other regions.
higher optical thickness where the data become sparser,
2) VALIDATION OVER OCEAN
fewer points are used in the average, as indicated. The
standard deviation in each bin is shown by error bars. Figure 9 represents the scatterplot of 2052 MODIS
The regression equation and correlation given at the retrievals over ocean collocated with an AERONET
top of each plot were calculated from the full scatter- station either on the coast or on an island. Note that the
plots, before binning. The solid black line is the 1:1 line, dashed lines, denoting expected uncertainty, are nar-
and the dashed lines denote the expected uncertainty rower than those over land. The MODIS over-ocean
calculated from prelaunch analysis. These dashed lines algorithm is expected to be more accurate than the
should encompass one standard deviation (66%) of the over-land algorithm (⌬ ⫽ ⫾0.03 ⫾ 0.05) (Tanré et al.
aerosol retrievals. The prelaunch expected uncertainty 1999; King et al. 1999). AERONET values at 0.55 and
over land is ⌬ ⫽ ⫾ 0.05 ⫾ 0.15 (Chu et al. 1998; King 0.66 m, are interpolated as in Fig. 8. The MODIS
et al. 1999). The regression equations in Fig. 8 indicate values are not interpolated for these plots. The 0.87-m
that MODIS aerosol optical thickness (AOT) offsets at plot is the only one showing a directly retrieved
3) DISCUSSION OF VALIDATION
←
FIG. 8. MODIS aerosol optical thickness retrievals over land at
470, 550, and 660 nm as a function of AERONET observations
collocated in space and time. The data were sorted according to
AERONET aerosol optical thickness and averaged for every 300
points. At higher optical thickness where the data become sparser,
fewer points are used in the average, as indicated. The standard
deviation in each bin is shown by error bars. The regression equa-
tions given at the top of each plot were calculated from the full
scatterplots before binning. The dashed lines denote the expected
uncertainty calculated from prelaunch analysis.
TABLE 8. Number of retrievals (N ), percentage of retrievals (%) falling within expected uncertainty for each wavelength, average
optical thickness from AERONET (0.55), and percent difference between MODIS and AERONET at 0.55 m (Diff) for the global
dataset and for each region, land, and ocean separately.
Region N % 0.47 % 0.55 % 0.66 % 0.87 0.55 Diff
All land 5906 61 68 71 0.18 41
China: land 205 71 75 76 0.28 17
India: land 70 70 90 79. 0.46 ⫺2
Mediterranean: land 781 65 67 66 0.20 32
Alaska/Canada: land 178 49 57 67 0.10 122
South Africa: land 347 79 86 84 0.19 ⫺10
South America: land 762 63 72 75 0.18 21
West United States: land 901 56 61 63 0.14 53
East United States: land 1385 55 64 68 0.17 54
East Europe: land 192 79 83 72 0.24 ⫺10
West Europe: land 793 66 72 77 0.17 48
All ocean 2052 62 66 70 0.18 1
Indian Ocean 47 64 77 77 0.16 7
Asian Pacific Ocean 57 56 53 60 0.21 13
Pacific Island ocean 163 70 74 79 0.08 ⫺6
West Mediterranean ocean 334 52 62 68 0.21 ⫺6
East Mediterranean ocean 205 57 63 71 0.23 ⫺7
Saharan ocean 184 58 56 51 0.31 1
Atlantic Isles ocean 146 64 71 71 0.13 8
Australia ocean 70 83 81 83 0.05 2
North Europe ocean 150 65 72 81 0.16 ⫺8
Caribbean ocean 242 62 67 68 0.14 20
East Pacific Ocean 160 52 61 69 0.18 ⫺6
U.S. Atlantic Ocean 288 72 68 70 0.15 7
that play a significant role in the quality of the retrieved time do not exist, and the monthly AERONET values
aerosol products. Figures 8 and 9 represent the collo- are calculated from AERONET level 2.0 data. The
cated points only for the events in which both the data from MODIS and AERONET need not be simul-
MODIS and the AERONET cloud-masking algorithms taneous. The observations from both instruments were
indicated that no clouds were present. It is possible that designed to represent monthly mean aerosol optical
MODIS might retrieve, while AERONET identified thickness for each region independently. In this com-
clouds and did not. Those points would not show up on parison, MODIS does not benefit from AERONET’s
the scatterplots. If those cases were numerous, long- cloud-clearing algorithm. If MODIS retrievals were
term MODIS aerosol statistics could be cloud contami- systematically cloud contaminated, we would expect
nated despite the good agreement in Figs. 8 and 9. the MODIS monthly mean values to be systematically
Figures 10a and 10b present an alternative method of higher than AERONET’s. In most cases, MODIS and
validation that checks the long-term statistics for pos- AERONET exhibit very similar annual cycles, often
sible cloud contamination. Here we plot monthly mean with very similar magnitudes of optical thickness. Two-
values in a 3° latitude ⫻ 3° longitude box centered on thirds of the differences in optical thickness over land
the AERONET station. The data are for the year 2001 are less than 0.10. There is some indication that
at eight selected stations, four with land components MODIS retrievals over land may be systematically bi-
and four with only ocean retrievals. Within the four ased high, but in most cases the difference is still well
land sites, three of them are near enough to the coast to within the estimated uncertainty of ⫾0.05 ⫾ 0.15.
contain sufficient ocean retrievals within the 3° ⫻ 3° When optical thickness is high and magnitudes signifi-
box and can be used for testing both land and ocean cantly differ, as in Cuiaba-Miranda of Fig. 10a, it is
retrievals. For each location, the monthly mean values AERONET that systematically exceeds MODIS. Over
of the aerosol optical thickness are plotted in the upper ocean, two-thirds of the differences between MODIS
panel, while the difference between the MODIS values and AERONET monthly means are less than 0.065.
and the AERONET values along with prelaunch un- The spatial variability across the 3° ⫻ 3° box can ex-
certainty estimates are plotted in the lower panel. plain some of these differences, especially at sites like
Red and blue indicate MODIS retrievals over land Goddard Space Flight Center (GSFC) and Rome_
and ocean, respectively, while black represents the Tor_Vergata where the ocean retrievals are at least 100
AERONET observations. The MODIS monthly mean km from the land-based AERONET station. Over the
values were calculated from archived MODIS level 3 four ocean-only sites the agreement between MODIS
data (MOD08) on a 1° resolution. Thus, the difficulties and AERONET optical thickness is even better with
with matching MODIS with AERONET in near–real two-thirds of the monthly means having less than 0.035
←
FIG. 9. MODIS aerosol optical thickness retrievals over ocean
at 550, 660, and 870 nm as a function of AERONET observations
collocated in space and time. The data were sorted according to
AERONET aerosol optical thickness and averaged for every 100
points. At higher optical thickness where the data become sparser,
fewer points are used in the average, as indicated. The standard
deviation in each bin is shown by error bars. The regression equa-
tions given at the top of each plot were calculated from the full
scatterplots before binning. The dashed lines denote the expected
uncertainty calculated from prelaunch analysis.
FIG. 10. Monthly mean aerosol optical thickness at 0.55 m for the year 2001 at (a) four sites with land retrievals and (b) four sites with only ocean retrievals. The top portion of
each plot shows the monthly means. The bottom portion shows the difference between MODIS and AERONET values. Also shown by thin dashed lines in the bottom portions are
the prelaunch estimated uncertainties of optical thickness retrievals, ⫾0.03 ⫾ 0.05 over ocean and ⫾0.05 ⫾ 0.15 over land. Blue denotes MODIS ocean retrievals, red denotes MODIS
land retrievals, and black denotes AERONET. The MODIS values are calculated from level 3 daily statistics and represent a 3° latitude by 3° longitude box centered on the AERONET
station.
967
FIG. 11. Monthly mean aerosol fine-mode fraction, , at 0.55 m for the year 2001 at (a) four sites with land retrievals and (b) four sites with only ocean retrievals, where is defined
as f/tot. The top portion of each plot shows the monthly means. The bottom portion shows the difference between MODIS and AERONET values. Also shown by thin dashed lines
JOURNAL OF THE ATMOSPHERIC SCIENCES—SPECIAL SECTION
in the bottom portions are the prelaunch estimated uncertainties of retrievals over the ocean, ⫾30%. Blue denotes MODIS ocean retrievals., red denotes MODIS land retrievals,
and black denotes AERONET. The MODIS values are calculated from level 3 daily statistics and represent a 3° latitude ⫻ 3° longitude box centered on the AERONET station.
VOLUME 62
tical thickness. The derivation of size parameters at low parameters. Figure 12 shows that 62% of the points fall
optical thickness, when aerosol signal is small, will be within the ⫾0.10 m errors, which is a smaller percent-
very sensitive to instrument calibration. In June 2001, age than what Remer et al. (2002) reported for a more
the MODIS instrument suffered an anomaly and the limited dataset. Note that the MODIS and AERONET
data processing was switched from the B-side to the retrievals both assume spherical particles in deriving
A-side electronics (http://modis.gsfc.nasa.gov/news/). size distribution. This assumption causes both MODIS
This switch created a small aberration in the calibration and AERONET to underpredict particle effective ra-
that did not affect the more robust optical thickness dius when nonspherical dust is present (Dubovik et al.
retrievals, but did affect the size parameter derivations 2002b). Thus, the agreement in Fig. 12 may be better at
when optical thickness was low. some sites if AERONET retrievals had assumed spher-
Comparison of MODIS-derived particle effective ra- oids instead of spheres (Dubovik et al. 2002b). The
dius and AERONET-derived particle effective radius is MODIS algorithm is being modified to include the op-
shown in Fig. 12. This plot matches MODIS retrievals tion of empirical phase functions that do not require
to daily averages of level 2 quality assured AERONET any assumption of particle shape.
sky radiance inversions. The MODIS effective radius
parameter is for the total bimodal size distribution in
which each of the two modes is weighted by (appen- 5. Results
dix B). There were 492 MODIS effective radius retriev-
als collocated with a daily average AERONET re- Figure 13 illustrates the MODIS aerosol retrievals at
trieval. However, only the 271 matchups with optical the global scale. The images are constructed from the
thickness greater than 0.15 are plotted in Fig. 12. At low aerosol optical thickness and size parameter products
optical thickness, because of less signal, there is greater both derived from observed MODIS radiances. Red
susceptibility to all algorithmic and sensor uncertain- indicates aerosol dominated by small particles (less
ties. These uncertainties include small calibration er- than 0.5 m) and greenish tints indicate aerosol with a
rors (discussed above) and retrieval errors for both in- higher proportion of large particles (greater than 0.5
struments (Ignatov et al. 1998; Remer et al. 2002). m). We can see that aerosol from natural sources,
These errors make little difference to retrievals of op- such as sea salt and desert dust, contain larger particles
tical thickness but may create large errors in the size than aerosols emanating from human-produced com-
bustion sources such as agricultural and deforestation
burning or urban/industrial pollution. Therefore, aero-
sol size easily separates aerosols into natural and man-
made components [with the exceptions of lightning-
initiated forest fire smoke and ocean dimethyl sulfide
(DMS) production]. Thus we see that MODIS’s ability
to separate aerosols by size can be used as a proxy for
separating anthropogenic aerosol from natural sources
and increases the accuracy of estimating human-
induced aerosol forcing (Kaufman et al. 2002).
6. Conclusions
FIG. 13. MODIS aerosol optical thickness at 0.55 m, representing global aerosol distribution on 12
Aug 2001. The two-dimensional color bar describes both magnitude of optical thickness (along bottom
axis) and fraction of optical thickness contributed by smaller fine-mode particles (along vertical axis).
Blue indicates low aerosol loading. Red indicates heavy loading of small particles such as pollution and
smoke. The greener tones indicate a greater percentage of large particles such as desert dust and sea salt.
The image was created from the MODIS daily 10-km resolution data after smoothing the raw data with
Gaussian filters applied both in the temporal and spatial domains.
extensive validation effort that collocated over 8000 over-ocean MODIS values agree to within 20%, which
MODIS retrievals with AERONET measurements of exceeds the prelaunch estimate of ⫾30% for individual
optical thickness show that globally, the MODIS prod- retrievals. However, at low aerosol optical thickness (
ucts are accurate to within prelaunch expectations, ⬍ 0.15) the MODIS size retrievals are susceptible to
namely, ⫾0.05 ⫾ 0.15 over land and ⫾0.03 ⫾0.05 small aberrations in the calibration and other factors,
over ocean. In particular, the retrieval of aerosol over which introduce greater uncertainty. In addition, dust,
oceans consistently shows remarkably good agreement with its nonspherical shapes, introduce uncertainty in
with virtually no offset or bias through the range of both the optical thickness and size parameter retrievals.
optical thickness where most observations occur. Re- This latter issue will be addressed with the incorpora-
gional analysis, however, shows specific issues for cer- tion of nonspherical phase functions into the next ver-
tain locations. Comparison of MODIS and AERONET sion of the algorithms. In the meantime, the MODIS
monthly means at eight specific locations scattered glo- aerosol products are sufficiently accurate for a variety
bally demonstrates that the MODIS retrievals are not of applications, including improved estimates of obser-
affected by cloud contamination at those sites, and that vationally based aerosol radiative effects.
MODIS long-term statistics agree with AERONET to
within 0.10 over land and to within 0.035 at oceanic Acknowledgments. We thank the many AERONET
island sites. MODIS-derived aerosol size parameters principal investigators and site managers who provided
are in general agreement with the same quantities de- the data used in the validation analysis. In particular we
rived by AERONET instruments on the ground. For acknowledge the contribution of Charles McClain and
moderate optical thickness, one standard deviation of the SIMBIOS project, which maintains many of the
MODIS effective radius retrievals falls within ⫾0.11 AERONET sites used in the over-ocean validation.
m of AERONET measurements. Comparison of The MODIS aerosol product is the culmination of ef-
MODIS and AERONET monthly mean values of , fort by many diverse people and groups including Vin-
the ratio of fine-mode aerosol optical thickness to total cent Salomonson, MODIS team leader, Michael King,
optical thickness at eight specific sites suggests that EOS senior project scientist, Ed Masuoka and Richard
冉 冊
where fand care the fine- and coarse-mode atmo-
3VO公2 ⫺3 9
spheric reflectances for the same optical thickness as NO ⫽ rg exp ⫺ 2 .
the total spectral reflectance, LUT
, respectively, and 4 2
is the reflectance weighting parameter. Note that The scattering coefficient s is specific to each model
LUT ⫽ f ⫹ c, and that the total and component mode, is the density of the particle assumed to be 1 g
reflectances all are functions of the total optical thick- cm⫺3, erf( ) is the error function, is ln g where g is
ness (tot f
0.55), not the component optical thicknesses (0.55 the geometric mean standard deviation of the lognor-
and c0.55). All optical thicknesses are defined at 0.55 mal distribution, is cos , where is the scattering
m. This is by definition in constructing the lookup angle, P( ) is the phase function, is the optical thick-
tables. ness and, unless designated specifically for wavelength
Define the total optical thickness (tot
0.55) equal to the or large or small mode, represents the total optical
sum of the fine (f0.55) and coarse (c0.55) components. thickness at 0.55 m. Extcoeff is the extinction coeffi-
Using the single scattering approximation, cient, and unless designated specifically for wavelength
f ⫽ C0.55
tot
Pf , c ⫽ C0.55
tot
Pc 共A2兲 is understood to be 0.55 m. Cloud condensation nuclei
in units of cm⫺2 is:
and
LUT ⫽ C共0.55
f
Pf ⫹ 0.55
c
Pc 兲,
where C is a constant depending on geometry, and
共A3兲
Pf
CCN ⫽ 冕 ⬁
ro⫽0.03m
n共r兲dr
lookup table are calculated for the collection of indi- Asymmetry factor:
vidual fine and coarse modes, not for any “total” aero-
sol size distribution. Solving for in Eq. (A1) gives
⫽ 共LUT ⫺ c 兲Ⲑ共f ⫺ c 兲. 共A4兲
g ⫽ 0.5 冕⫺1
1
P共兲d ⫽ 0.5 冕 0
cos共兲P共兲 sin d.
冕 冕
f
⫽ 关C共0.55 Pf ⫹ 0.55
c
Pc 兲 ⫺ C0.55
tot
Pc 兴Ⲑ 1 11
1
 ⫽ P共兲d ⫽ P共兲 sin d.
tot
共C0.55 Pf ⫺ C0.55
tot
Pc 兲. 共A5兲 2 ⫺1 2 0
Dropping the constant C and using the definition of Number of particles in each mode (1 cm⫺3), and
tot⫽ f0.55 ⫹ c0.55 gives
0.55 and extcoeff defined at 0.55 m:
Mk ⫽ 冕 ⬁
o
rkn共r兲dr ⫽ 共rg兲k exp共0.5k22兲.
and 2.13 m.
Aerosol_Type under the land products is a function
mostly of geography and season and should not be con-
Effective radius (m): sidered a retrieved quantity. The Cloud_Fraction listed
in the tables is not a true cloud fraction, but instead an
3 3 2
reff ⫽ 共NsmallMsmall ⫹ NlargeMlarge 兲Ⲑ共NsmallMsmall indication of the fraction of pixels not used in the re-
2 trieval due to a combination of clouds, surface issues, or
⫹ NlargeMlarge 兲. internal inconsistencies. Likewise, Mean_Reflectance is
Mass concentration (g per cm2), the mean reflectance only of those pixels that survive
3 3
the masking and elimination procedures and are actu-
Mass_conc ⫽ NsmallMsmall ⫹ NlargeMlarge , over ocean ally used in the retrievals. Solution_Index tells which
Mass_conc ⫽
4
3
再冉
Nor3g0.66
S
冊 small
⫹ 共1 ⫺ 兲
fine and coarse aerosol models were chosen in the re-
trieval. Least_Squares_Error reports the fitting error of
the inversion, , from Eq. (7).
冉 Nor3g0.66
S
冊 冎
large
, over land
The Quality_Assurance parameters are five-byte
codes that hold information concerning the retrievals
and the overall quality. Details of the Quality_Assur-
Ångström exponent 1 (0.55/0.87) and Ångström ex- ance code are given by the MODIS Atmosphere’s
ponent 2 (0.87/2.13): Quality Assurance Plan, which can be found online at
冉 冊
http://modis-atmos.gsfc.nasa.gov.
0.55
ln
0.87 REFERENCES
冉 冊
AngExp_1 ⫽ Ackerman, S. A., K. I. Strabala, W. P. Menzel, R. A. Frey, C. C.
0.55
ln Moeller, and L. E. Gumley, 1998: Discriminating clear sky
0.87 from clouds with MODIS. J. Geophys. Res., 103, 32 139–
冉 冊
32 140.
0.87 Ahmad, Z., and R. S. Fraser, 1982: An iterative radiative transfer
ln
2.13 code for the ocean–atmosphere system. J. Atmos. Sci., 39,
冉 冊
AngExp_2 ⫽ . 656–665.
0.87 Chu, D. A., Y. J. Kaufman, L. A. Remer, and B. N. Holben, 1998:
ln Remote sensing of smoke from MODIS airborne simulator
2.13
during the SCAR-B experiment. J. Geophys. Res., 103,
31 979–31 988.
APPENDIX C ——, ——, C. Ichoku, L. A. Remer, D. Tanre, and B. N. Holben,
2002: Validation of MODIS aerosol optical depth retrieval
Recommendations for Using Products over land. Geophys. Res. Lett., 29, 8007, doi:10.1029/
2001GL013205.
There are many choices for aerosol optical thickness. Dubovik, O., A. Smirnov, B. N. Holben, M. D. King, Y. J. Kauf-
The products in Tables 4–7 labeled as validated, not yet man, T. F. Eck, and I. Slutsker, 2000: Accuracy assessments
validated, or derived are recommended. Correct- of aerosol optical properties retrieved from AERONET sun
and sky-radiance measurements. J. Geophys. Res., 105, 9791–
ed_Optical_Depth_Land is the recommended spectral
9806.
product over land, and Effective_Optical_Depth_Av- ——, B. N. Holben, T. F. Eck, A. Smirnov, Y. J. Kaufman, M. D.
erage_Ocean is the recommended spectral product King, D. Tanré, and I. Slutsker, 2002a: Variability of absorp-
over ocean. Products such as Continental_Opti- tion and optical properties of key aerosol types observed in
cal_Depth_Land are intermediate parameters and worldwide locations. J. Atmos. Sci., 59, 590–608.
——, ——, T. Lapyonok, A. Sinyk, M. I. Mishchenko, P. Yang,
should be used only as a diagnostic. and I. Slutsker, 2002b: Non-spherical aerosol retrieval
The word “small” in the product name indicates fine method employing light scattering by spheroids. Geophys.
mode so that Optical_Depth_Small is the fine-mode Res. Lett., 29, 1415, doi:10.1029/2001GL014506.
optical thickness, f, from appendix A. Optical_Depth_ Eck, T. F., B. N. Holben, J. S. Reid, O. Dubovik, A. Smirnov,
N. T. O’Neill, I. Slutsker, and S. Kinne, 1999: Wavelength
Ratio_Small is the fine-mode ratio, . Likewise, the
dependence of the optical depth of biomass burning, urban
word “large” indicates coarse mode. The word “aver- and desert dust aerosols. J. Geophys. Res., 104, 31 333–
age” indicates the solution averaged from all retrieval 31 349.
solutions with fitting error less than 3% or the average ——, and Coauthors, 2003: Variability of biomass burning aerosol
of the three best solutions if all ⬎ 3%. The word optical characteristics in southern Africa during the SAFARI
2000 dry season campaign and a comparison of single scat-
“best” indicates the single solution with the least error, tering albedo estimates from radiometric measurements. J.
, no matter how large. See Eq. (7). The recommenda- Geophys. Res., 108, 8477, doi:10.1029/2002JD002321.
tion is to use those products labeled as average. Gao, B.-C., Y. J. Kaufman, D. Tanré, and R.-R. Li, 2002: Distin-
guishing tropospheric aerosols from thin cirrus clouds for im- of the MODIS aerosol retrievals over ocean and land during
proved aerosol retrievals using the ratio of 1.38-m and 1.24- CLAMS. J. Atmos. Sci., 62, 974–992.
m channels. Geophys. Res. Lett., 29, 1890, doi:10.1029/ Li, R.-R., Y. J. Kaufman, B.-C. Gao, and C. O. Davis, 2003: Re-
2002GL015475. mote sensing of suspended sediments and shallow coastal
Guenther, B., X. Xiong, V. V. Salomonson, W. L. Barnes, and J. waters. IEEE Trans. Geosci. Remote Sens., 41, 559–566.
Young, 2002: On-orbit performance of the Earth Observing Martins, J. V., D. Tanré, L. A. Remer, Y. J. Kaufman, S. Mattoo,
System Moderate Resolution Imaging Spectroradiometer: and R. Levy, 2002: MODIS cloud screening for remote sens-
First year of data. Remote Sens. Environ., 83, 16–30. ing of aerosol over oceans using spatial variability. Geophys.
Herman, J. R., P. K. Bhartia, O. Torres, C. Hsu, C. Seftor, and E. Res. Lett., 29, 8009, doi:10.1029/2001GL013252.
Celarier, 1997: Global distribution of UV-absorbing aerosols MCST, 2000: MODIS level 1B product user’s guide, for level 1B
from Nimbus 7/TOMS data. J. Geophys. Res., 102, 16 911– version 2.3.x, release 2. MCST Document PUB-01-U-DNCN,
16 922. 47 pp. [Available online at http://www.mcst.ssai.biz/mcstweb/
Holben, B. N., and Coauthors, 1998: AERONET—A federated L1B/product.html.]
instrument network and data archive for aerosol character- ——, 2002: MODIS level 1B product user’s guide, for level 1B
ization. Remote Sens. Environ., 66, 1–16. version 4.0.9 (Terra) and version 4.1.1 (Aqua). MCST Docu-
Houghton, J. T., Y. Ding, D. J. Griggs, M. Noguer, P. J. van der ment PUB-01-U-0202-REV B, MCST Internal Memo,
Linden, and D. Xiaosu, Eds., 2001: Climate Change 2001: The M1039, 61 pp. [Available online at http://www.mcst.ssai.
Scientific Basis. Cambridge University Press, 944 pp. biz/mcstweb/L1B/product.html.]
Husar, R. B., L. L. Stowe, and J. M. Prospero, 1997: Character- Prins, E. M., J. M. Feltz, W. P. Menzel, and D. E. Ward, 1998: An
ization of tropospheric aerosols over the oceans with the overview of GOES-8 diurnal fire and smoke results for
NOAA Advanced Very High Resolution Radiometer optical SCAR-B and 1995 fire season in South America. J. Geophys.
thickness operational product. J. Geophys. Res., 102, 16 889– Res., 103, 31 821–31 836.
16 910. Remer, L. A., and Y. J. Kaufman, 1998: Dynamical aerosol model:
Ichoku, C., D. A. Chu, S. Mattoo, Y. J. Kaufman, L. A. Remer, D. Urban/industrial aerosol. J. Geophys. Res., 103, 13 859–
Tanré, I. Slutsker, and B. N. Holben, 2002: A spatio-temporal 13 871.
approach for global validation and analysis of MODIS aero- ——, ——, B. N. Holben, A. M. Thompson, and D. McNamara,
sol products. Geophys. Res. Lett., 29, doi:10.1029/ 1998: A model of tropical biomass burning smoke aerosol
2001GL013206. size distribution. J. Geophys. Res., 103, 31 879–31 892.
——, L. A. Remer, Y. J. Kaufman, R. Levy, D. A. Chu, D. Tanré, ——, and Coauthors, 2002: Validation of MODIS aerosol re-
and B. N. Holben, 2003: MODIS observation of aerosols and trieval over ocean. Geophys. Res. Lett., 29, 8008, doi:10.1029/
estimation of aerosol radiative forcing over southern Africa 2001GL013204.
during SAFARI 2000. J. Geophys. Res., 108, 8006,
Rosenfield, D., and I. M. Lensky, 1998: Satellite-based insights
doi:10.1029/2002JD002366.
into precipitation formation processes in continental and
——, ——, and T. F. Eck, 2005: Quantitative evaluation and in-
maritime convective clouds. Bull. Amer. Meteor. Soc., 79,
tercomparison of morning and afternoon MODIS aerosol
2457–2476.
measurements from the Terra and Aqua satellites. J. Geo-
Samet, J. M., S. L. Zeger, F. Dominici, F. Curriero, I. Coursac, D.
phys. Res., in press.
W. Dockery, J. Schwartz, and A. Zanobetti, 2000: The na-
Ignatov, A., L. Stowe, and R. Singh, 1998: Sensitivity study of the
tional morbidity, mortality, and air pollution study, Part II:
Angstrom exponent derived from AVHRR over oceans.
Morbidity, mortality, and air pollution in the United States.
Adv. Space Res., 21, 439–442.
Health Effects Institute Research Rep. 94, Cambridge, MA,
——, and Coauthors, 2005: Two MODIS aerosol products over
87 pp.
ocean on the Terra and Aqua CERES SSF datasets. J. Atmos.
Sci., 62, 1008–1031. Smirnov, A., B. N. Holben, T. F. Eck, O. Dubovik, and I. Slutsker,
Kaufman, Y. J., 1989: The atmospheric effect on remote sensing 2000: Cloud-screening and quality control algorithms for the
and its correction. Theory and Applications of Optical Re- AERONET database. Remote Sens. Environ., 73, 337–349.
mote Sensing, G. Asrar, Ed., John Wiley & Sons, 336–428. Swap, R. J., H. J. Annegarn, J. T. Suttles, M. D. King, S. Platnick,
——, D. Tanré, L. A. Remer, E. Vermote, A. Chu, and B. N. J. L. Privette, and R. J. Scholes, 2003: Africa burning: A
Holben, 1997a: Operational remote sensing of tropospheric thematic analysis of the Southern African Regional Science
aerosol over land from EOS Moderate Resolution Imaging Initiative (SAFARI 2000). J. Geophys. Res., 108, 8465,
Spectroradiometer. J. Geophys. Res., 102, 17 051–17 067. doi:10.1029/2003JD003747.
——, A. E. Wald, L. A. Remer, B.-C. Gao, R.-R. Li, and L. Flynn, Tanré, D., M. Herman, and Y. J. Kaufman, 1996: Information on
1997b: The MODIS 2.1 m Channel—Correlation with vis- aerosol size distribution contained in solar reflected spectral
ible reflectance for use in remote sensing of aerosol. IEEE radiances. J. Geophys. Res., 101, 19 043–19 060.
Trans. Geosci. Remote Sens., 35, 1286–1298. ——, Y. J. Kaufman, M. Herman, and S. Mattoo, 1997: Remote
——, D. Tanré, and O. Boucher, 2002: A satellite view of aerosols sensing of aerosol properties over oceans using the MODIS/
in the climate system. Nature, 419, 215–223. EOS spectral radiances. J. Geophys. Res., 102, 16 971–16 988.
King, M. D., Y. J. Kaufman, D. Tanré, and T. Nakajima, 1999: ——, L. A. Remer, Y. J. Kaufman, S. Mattoo, P. V. Hobbs, J. M.
Remote sensing of tropospheric aerosols from space: Past, Livingston, P. B. Russell, and A. Smirnov, 1999: Retrieval of
present, and future. Bull. Amer. Meteor. Soc., 80, 2229–2259. aerosol optical thickness and size distribution over ocean
——, and Coauthors, 2003: Cloud and aerosol properties, precipi- from the MODIS Airborne Simulator during TARFOX. J.
table water, and profiles of temperature and humidity from Geophys. Res., 104, 2261–2278.
MODIS. IEEE Trans. Geosci. Remote Sens., 41, 442–458. Torres, O., P. K. Bhartia, J. R. Herman, A. Sinyuk, P. Ginoux, and
Levy, R. C., and Coauthors, 2003: Evaluation of the MODIS re- B. Holben, 2002: A long-term record of aerosol optical depth
trievals of dust aerosol over the ocean during PRIDE. J. from TOMS observations and comparison to AERONET
Geophys. Res., 108, 8594, doi:10.1029/2002JD002460. measurements. J. Atmos. Sci., 59, 398–413.
——, L. A. Remer, J. V. Martins, Y. J. Kaufman, A. Plana-Fattori, Twomey, S. A., 1977: The influence of pollution on the shortwave
J. Redemann, P. B. Russell, and B. Wenny, 2005: Evaluation albedo of clouds. J. Atmos. Sci., 34, 1149–1152.