Validation of MODIS aerosol optical depth retrieval over land
D. A. Chu,
1,2
Y. J. Kaufman,
2
C. Ichoku,
1,2
L. A. Remer,
2
D. Tanre´,
3
and B. N Holben
4
Received 20 March 2001; revised 11 October 2002; accepted 23 November 2002; published 29 June 2002.
[1] Aerosol optical depths (t
a
) are derived operationally for the
first time from the MODIS (Moderate Resolution Imaging
Spectroradiometer) measurements over vegetated and partially
vegetated land at 0.47 and 0.66 mm wavelengths. The extensive
validation made during July – September 2000 encompasses 315
co-located t
a
in space and time derived by MODIS and AERONET
(Aerosol Robotic Network) from more than 30 AERONET sites.
The lack of AERONET measurements in East Asia, India and
Australia makes this validation unavailable for those regions. The
MODIS aerosol retrievals, except in coastal zones, are found within
the retrieval errors of
t
a
= ±0.05 ±0.2 t
a
. The root mean square
(RMS) errors are
0.1 in the continental inland regions and up to
0.3 in the coastal regions (attributed mainly to water contaminated
signals). With this validation we believe that MODIS aerosol
products can be used quantitatively in many applications with
caution for possible residual clouds, snow/ice, and water
contamination. INDEX TERMS: 1610 Global Change:
Atmosphere (0315, 0325); 1640 Global Change: Remote sensing;
3360 Meteorology and Atmospheric Dynamics: Remote sensing
1. Introduction
[2] The application of satellite data to derive global aerosol
properties has advanced dramatically in the last few years [King
et al., 1999]. One of the main advancements is the systematic
derivation of aerosol over land from MODIS onboard the EOS-
Terra satellite launched on December 18, 1999. Since February 24,
2000, MODIS has continuously acquired daily global measure-
ments with thirty-six spectral bands (0.41 – 14 mm) at three
different spatial resolutions (250 m, 500 m and 1 km) [Salomonson
et al., 1989]. We retrieve aerosol properties over both land [Kauf-
man et al., 1997a] and ocean [Tanre´ etal., 1997] using seven well
calibrated spectral channels in the solar spectrum (0.47 – 2.1 mm).
In this paper we present the first comprehensive validation of the
MODIS-derived t
a
over land.
[3] The MODIS retrieval of t
a
over land employs primarily
three spectral channels centered at 0.47, 0.66, and 2.1 mm wave-
length at 500m resolution. A short description of the MODIS
aerosol algorithm presented here follows Kaufman et al. [1997a].
In a 10 km  10 km grid box, cloud-free pixels are first selected
using the multi-spectral MODIS cloud mask [Ackerman et al.,
1998]. The cloud mask uses more than twenty tests, including two
cirrus detection tests, to indicate a cloudy or clear pixel at 1 1km
resolution. Fine-mode aerosols are transparent at 2.1 mm wave-
length, allowing the direct observation of land surface. The
empirical relationships developed over vegetated surfaces are used
to estimate the surface reflectance (r
s
) at 0.47 mm and at 0.66 mm
from the measurements at 2.1 mm(r
s
0.47mm
/r
s
2.1mm
=0.25and
r
s
0.66mm
/r
s
2.1mm
=0.5)[Kaufman et al., 1997b]. To minimize the
error, the MODIS aerosol retrievals over land are limited to pixels
of r
s
2.1mm
< 0.15. Snow/ice and water covered surfaces are excluded
because the empirical relationships given above are invalid over
those regions. The selected cloud-free dark pixels in the grid box
may still be partially contaminated by sub-pixel clouds, snow/ice,
or soil types that do not fit the empirical relationship (e.g., red soil
[Gatebe et al., 2001]). Thus only the 10–40 percentile of MODIS
measured radiance is used. The overall retrieval errors were
estimated to be
t
a
= ±0.05 ± 0.2t
a
(100% error for t
a
=
0.05). Larger error of ±0.3 is found for dust particles when using
the 2.1 mm channel [Kaufman et al., 2000] as opposed to ±0.05 for
urban/industrial and biomass burning aerosols. To distinguish
between dust and non-dust aerosols, the ratio of aerosol path
radiance at 0.66 and 0.47 mm is used. Sulfate and smoke aerosols
that cannot be distinguished by the path radiance ratio are
separated a priori according to the geographic locations and
seasons of their emission sources. The details of the determination
of aerosol types (including the mixture of different aerosols) and
the aerosol models used by the retrieval algorithm can be found in
Kaufman et al. [1997a].
[4] The release of MODIS level-2 (10 km  10 km) granule-
based (granule: a 5-minute segment of one MODIS orbital data)
aerosol products in August 2000 and later level-3 (1  1) gridded
products was to provide a preview before quality assurance. The
validation data presented here are the basis for the release of the
first version of validated products.
2. Validation Approach
[5] In order to take into account both spatial and temporal
variabilities of aerosol distribution, the MODIS retrievals at 10 km
 10 km resolution and the AERONET direct Sun measurements
at 15-minute intervals [Holben et al., 1998] need to be co-located
Figure 1. Frequency map of MODIS aerosol retrievals over land
derived from MODIS 1  1 level-3 daily products between July
and September 2000. The white circles depict the locations of
AERONET Sun photometer sites.
GEOPHYSICAL RESEARCH LETTER, VOL. 29, NO. 12, 10.1029/2001GL013205, 2002
1
Science Systems and Applications Inc., Lanham, Maryland, USA.
2
Laboratory for Atmospheres, NASA Goddard Space Flight Center,
Greenbelt, Maryland, USA.
3
Laboratorire d’Optique Atmospherique, Universite des Sciences et
Technologies de Lille, Villeneuve d’Ascq, France.
4
Laboratory for Terrestrial Physics, NASA Goddard Space Flight
Center, Greenbelt, Maryland, USA.
Copyright 2002 by the American Geophysical Union.
0094-8276/02/2001GL013205$05.00
MOD2 - 1

in space and time. We require at least 2 out of possible 5 AERONET
measurements within ±30 min of MODIS overpasses and at least 5
out of possible 25 MODIS retrievals in a square box of 50 km 
50 km centered over AERONET sites. The mean values of the co-
located spatial and temporal ensemble are then used in linear
regression analysis and in calculating RMS errors. The AERONET
level 1.5 data are cloud screened. Though the level 2.0 data provide
final calibration, they are not available in real time. Therefore, the
level 1.5 data (instead of level 2.0) are used in the operational
MODIS aerosol validation scheme [Ichoku et al., 2002].
3. Validation of Aerosol Optical Depth
[6] Figure 1 displays the frequency map of MODIS aerosol
retrievals over land as derived using the level-3 daily aerosol
products from July to September 2000. Superimposed are the
locations of the AERONET sites included in this validation. The
MODIS aerosol retrievals cover approximately 70% of the land
surface. The frequency lower than 100% of a given 1  1 grid
box is due to cloud cover, non-vegetated surfaces, or missing data.
Most of the regions with dust occurrence are excluded due to high
brightness of desert surface (e.g., the Sahara Desert). Also
excluded are snow/ice-covered regions (e.g., Antarctica and Green-
land)—too bright in the visible wavelength to derive aerosol
optical depth. At high latitudes, more retrievals are seen because
of the overlapped satellite orbits.
[7] A total of 315 points representing more than 30 AERONET
sites meet our match-up criteria for the MODIS- and AERONET-
derived t
a
in the period of July – September 2000. Small islands,
such as Barbados, Bermuda, Cape Verde, Hawaii, etc., are too
small for aerosol land validation. The slopes (S
l
) of linear regres-
sion represent systematic biases if differing from 1 and the
intercepts (I
c
) represent the errors in the r
s
estimates. Large errors
in r
s
lead to large I
c
. The scatter plots in Figures 2a and 2b depict
overall a very good agreement between MODIS and AERONET
with S
l
0.86, I
c
0.02–0.06, and high correlation coefficients (R)
0.85–0.91. Nearly all the points fall within the retrieval errors of

t
a
= ±0.05 ± 0.2t
a
with RMS errors ranging from 0.07 to 0.11.
Venice and El Arenosillo coastal sites excluded in Figure 2 with
larger RMS errors (0.2–0.3) and I
c
(0.2) will be discussed later.
The systematic biases in MODIS retrievals are mainly due to
aerosol model assumptions (deviation of 0–20%), instrument
calibration (2–5%), or the choice of the lowest 10–40 percentile
Figure 2. Global comparisons of MODIS- and AERONET-
derived t
a
at 0.47 and 0.66 mm wavelengths, encompassing 315
points from more than 30 AERONET sites excluding Venice and El
Arenosillo coastal sites. The solid lines represent the slopes of
linear regression and the dot lines the retrieval errors of
t
a
=
±0.05 ± 0.2t
a
. Temporal and spatial standard deviations are shown
as the error bars in x(AERONET)- and y(MODIS)-direction
respectively.
Figure 3. Regional comparisons of MODIS- and AERONET-derived t
a
at 0.47 and 0.66 mm wavelengths in the continental inland
regions: (a) eastern US, (b) Brazil, (c) western Europe, and (d) southern Africa. The solid lines represent the slopes of linear regression
and the dot lines the retrieval errors of
t
a
= ±0.05 ± 0.2t
a
. Temporal and spatial standard deviations are shown as the error bars in
x(AERONET)- and y(MODIS)-direction respectively.
MOD2 - 2 CHU ET AL.: MODIS AEROSOL VALIDATION OVER LAND