Atmos. Meas. Tech., 4, 379–408, 2011
www.atmos-meas-tech.net/4/379/2011/
doi:10.5194/amt-4-379-2011
© Author(s) 2011. CC Attribution 3.0 License.
Atmospheric
Measurement
Techniques
An over-land aerosol optical depth data set for data assimilation
by filtering, correction, and aggregation of MODIS Collection 5
optical depth retrievals
E. J. Hyer1, J. S. Reid1, and J. Zhang2
1Naval Research Laboratory, 7 Grace Hopper Avenue, Monterey, CA 93943, USA
2University of North Dakota, 4149 University Avenue Stop 9006, Grand Forks, ND 58202, USA
Received: 12 August 2010 – Published in Atmos. Meas. Tech. Discuss.: 14 September 2010
Revised: 4 February 2011 – Accepted: 4 February 2011 – Published: 1 March 2011
Abstract. MODIS Collection 5 retrieved aerosol optical
depth (AOD) over land (MOD04/MYD04) was evaluated
using 4 years of matching AERONET observations, to as-
sess its suitability for aerosol data assimilation in numeri-
cal weather prediction models. Examination of errors re-
vealed important sources of variation in random errors (e.g.,
atmospheric path length, scattering angle “hot spot”), and
systematic biases (e.g., snow and cloud contamination, sur-
face albedo bias). A set of quality assurance (QA) filters
was developed to avoid conditions with potential for signifi-
cant AOD error. An empirical correction for surface bound-
ary condition using the MODIS 16-day albedo product cap-
tured 25% of the variability in the site mean bias at low
AOD. A correction for regional microphysical bias using the
AERONET fine/coarse partitioning information increased
the global correlation between MODIS and AERONET from
r2 = 0.62–0.65 to r2 = 0.71–0.73. Application of these fil-
ters and corrections improved the global fraction of MODIS
AOD within (0.05± 20%) of AERONET to 77%, up from
67% using only built-in MODIS QA. The compliant frac-
tion in individual regions was improved by as much as 20%
(South America). An aggregated Level 3 product for use in a
data assimilation system is described, along with a prognos-
tic error model to estimate uncertainties on a per-observation
basis. The new filtered and corrected Level 3 product has im-
proved performance over built-in MODIS QA with less than
a 15% reduction in overall data available for data assimila-
tion.
Correspondence to: E. J. Hyer
(edward.hyer@nrlmry.navy.mil)
1 Introduction
The Moderate Resolution Imaging Spectroradiome-
ter (MODIS) instruments on board the Terra and Aqua
platforms provide nearly daily global coverage of key atmo-
spheric and land surface parameters. Retrievals of aerosol
optical depth (AOD) by MODIS are the most commonly
used of any satellite AOD product. Although MODIS is
technically a research instrument, its use in operational
applications is increasingly widespread. Remote sensing
and modeling technology has progressed to a point where
operational aerosol data assimilation methods are being
implemented for forecasting applications (e.g., Zhang et al.,
2008; Hollingsworth et al., 2008).
The use, aggregation, and statistical reduction of remotely
sensed aerosol data are application specific. Crucial to the
application of AOD data in a data assimilation system is a
thorough understanding of the data’s error characteristics.
Not only is a global estimate of measurement uncertainty
needed, but also a point by point uncertainty determination
gained through a thorough understanding of retrieval pro-
cesses that can cause both random and systematic errors.
AOD data with consistent biases or poorly characterized un-
certainties can degrade analyses and forecasts if used in a
data assimilation system (Reid et al., 2009).
For over ocean data, Zhang and Reid (2006) identified
MODIS MOD04/MYD04 retrieval bias from sources includ-
ing lower boundary condition (e.g., white-capping and glint),
cloud contamination, and microphysics. From these anal-
yses, a series of quality control/assurance procedures and
empirical corrections were applied to devise a data assimi-
lation grade Level 3 product. These data are now assimi-
lated operationally at the Fleet Numerical Meteorology and
Published by Copernicus Publications on behalf of the European Geosciences Union.

380 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
Oceanography Center (FNMOC) through the Navy Varia-
tional Data Assimilation System for AOD (NAVDAS-AOD)
into the Navy Aerosol Analysis and Prediction System
(NAAPS) (Zhang et al., 2008).
Assimilation of over-land AOD data, however, is a very
different problem, both in terms of the retrieval of AOD from
satellite observations and assimilation into an atmospheric
model. Even without systematic bias, the precision of re-
trieved AOD over land is lower than over the ocean. The land
surface has strong aerosol sources, and observations in prox-
imity to these sources will have stronger spatial and temporal
gradients. These gradients reduce the accuracy and precision
of AOD retrievals, and interfere with pair-wise comparison
to point validation datasets. But the principal challenge of
aerosol retrieval over the land surface is the surface itself.
Whereas the ocean surface is dark and well characterized, the
land surface is brighter and more heterogeneous with strong
temporal variability.
The strongest aerosol signal in visible wavelengths comes
from scattering of incoming solar radiation. Inverting to ob-
tain the signal of aerosol effects from reflected light requires
simultaneous retrieval of the surface reflectance. As surface
brightness increases, the relative contribution of the atmo-
spheric radiance decreases, leading to a differential signal-
to-noise limitation of the atmospheric component. Thus, re-
trieval over a brighter surface requires a tighter constraint on
surface reflectance to achieve a comparably precise retrieval
of AOD. The complexity and variability of the lower bound-
ary condition make over-land retrieval of AOD with passive
optical observations very challenging.
The MODIS Collection 5 over-land aerosol retrieval (Levy
et al., 2007b; Remer et al., 2006) corrects many of the short-
comings of earlier versions, and now represents the state
of the art for global retrievals of AOD. It is made publi-
cally available in near real time through the NASA Land-
Atmosphere Near real-time Capability for EOS (LANCE)
(http://lance.nasa.gov). Based on a preliminary analysis,
Collection 5 reached the minimum efficacy requirements to
be considered for inclusion into the US Navy’s operational
aerosol modeling. Preliminary studies without major correc-
tion have shown that the inclusion of MODIS over-land data
can improve model scores in some regions (Reid et al., 2009).
This study expands the Navy’s over ocean data assimila-
tion quality dataset (Zhang and Reid, 2006) to include a data
assimilation quality Level 3 over land data product based on
data collection 5 for ultimate use in NAVDAS-AOD (Zhang
et al., 2008). This study is concerned with identifying the
various sources of random and systematic error in retrieval
of AOD, developing filters and corrections to improve data
quality when possible, and characterizing residual uncer-
tainty for use in aerosol data assimilation systems.
The center of the current study is an evaluation of the
MODIS Collection 5 data product using 4 years of Aerosol
Robotic Network (AERONET, Holben et al., 1998) Sun pho-
tometer data and available ancillary data sources. Random
and systematic errors in the MODIS Collection 5 AOD re-
trieval are examined in detail, with a detailed consideration
of both global and regionally specific sources of error. Cor-
rections are developed and evaluated for errors caused by the
parameterization of the surface boundary condition and by
regional microphysical bias. Based on this analysis, a data
assimilation ready level 3 product and error model for oper-
ational use is presented as the outcome of this study.
2 Data and methods
The methods and rationales for this analysis are similar to
the over-ocean study of Zhang and Reid (2006), investigat-
ing lower boundary condition, microphysical and cloud bias.
However, due to differences between the ocean and land
problems two important differences dictate modifications to
the protocol. First, the MODIS over-land aerosol retrieval
yields very little information about aerosol size, even com-
pared with the limited coarse-mode/fine-mode fractionation
derived by the over-ocean retrieval (Kahn et al., 2009). Sec-
ond, surface conditions affecting AOD retrieval will co-vary
geographically with aerosol properties. For instance, high
aerosol loads over brighter surfaces (arid or barren ecosys-
tems) will tend to be made up of coarser particles (dust).
These differences require some simplifications to be made
to the analysis, and the resulting QA/QC system.
2.1 MODIS aerosol retrievals
The MODIS data analyzed in this study was the Collection 5
MODIS Level 2 over-land aerosol product MOD04/MYD04
(Levy et al., 2007b) from 2005–2008. The over-land al-
gorithm uses 0.47 µm, 0.66 µm, 1.24 µm, and 2.12 µm radi-
ances. Unlike the ocean retrieval, aerosol properties in the
land retrieval are specified regionally and seasonally (Levy
et al., 2007a). Radiance inputs are mapped to AOD using a
look-up table based on radiative transfer simulations (Levy et
al., 2007a). AOD is retrieved at 0.47 µm, 0.55 µm, 0.66 µm,
and 2.12 µm. A diagnostic fine mode fraction at 0.55 µm is
also generated, although it lacks skill over land (Kahn et al.,
2009).
Similar to older versions, the Collection 5 algorithm
uses the MODIS 2.12 µm reflectance to estimate surface re-
flectance at 0.47 and 0.66 µm, and retrieves aerosol prop-
erties by inverting the observed reflectance in the visi-
ble wavelengths. The Collection 5 retrieval uses a lin-
ear model of the relationship between infrared and visi-
ble surface reflectances, informed by two empirical rela-
tionships estimated from a database of atmospherically cor-
rected MODIS surface reflectance data from cloud-free ar-
eas with τM < 0.2 (Levy et al., 2007b). The first correction
relates the 0.66 µm/2.12 µm regression line to a reflectance-
based Normalized Difference Vegetation Index of vegetation
condition, NDVISWIR, calculated as NDVISWIR = (ρ1.24 µm−
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 381
ρ2.12 µm)/(ρ1.240 µm + ρ2.12 µm), chosen for its relatively
greater sensitivity to vegetation condition and lower sensi-
tivity to atmospheric conditions (including aerosol particles).
The second correction modifies the slope and intercept esti-
mated from the first relationship by as much as ±50% ac-
cording to scattering angle, based on the observations of Re-
mer et al. (2001). Note that because the information used in
the MODIS AOD retrieval all comes from the same viewing
geometry, the phenomenon corrected by the scattering angle
dependence is the relative anisotropy between the visible and
near-infrared reflectances.
MODIS Collection 5 products have a ground footprint spa-
tial resolution of 10× 10 km at nadir, increasing to more than
20× 40 km at the edge of the swath. Datasets are divided into
5-min granules covering approximately 2330 km across the
satellite ground track and 2030 km along the ground track.
Data used in this study are the 0.55 µm “Corrected Opti-
cal Depth – Land” from the Level 2 product (hereafter τM).
One year of data from one sensor is approximately 105 000
Level 2 granules.
The MODIS Level 2 aerosol product includes numerous
pieces of auxiliary information about the retrieval condi-
tions. In this study, the following were used: (1) the MODIS
Mandatory quality assurance (QA) flag, which assigns each
retrieval an estimated quality of “Bad”, ”Marginal”, ”Good”
or “Very Good,” (2) the cloud fraction information, indicat-
ing the fraction of pixels within the retrieval footprint with
MODIS-detected cloud, (3) viewing angle and (4) the scat-
tering angle for each retrieval.
For the purposes of data assimilation MODIS products
from Terra and Aqua behave similarly, and in this paper joint
statistics are sometimes presented. When potentially sig-
nificant differences between Terra and Aqua are diagnosed,
separate statics are presented and discussed. Appendix D
presents discussion and analysis specifically targeting differ-
ences between Terra and Aqua MODIS AOD.
Note that the MODIS retrieval algorithm permits negative
values of retrieved AOD, and in certain locations this condi-
tion can be quite common (e.g. forested area in South Amer-
ica, see Sect. 4.3). Our statistical analysis of the MODIS
Level 2 data uses these negative retrievals. The aggregated
product produced after correction of the Level 2 data does
not include negative AODs; if the AOD is negative after ag-
gregation, it is truncated to 0.
2.2 AERONET sun photometer data
The basis for our evaluation of MODIS retrieved AOD is di-
rect measurements of AOD from the Aerosol Robotic Net-
work (AERONET) (Holben et al., 1998) of Sun photome-
ter instruments. This study uses AERONET Level 2 quality
controlled data collected throughout the entire network for
2005–2008. Quality control procedures are as described in
Smirnov et al. (2000). Depending on the exact instrument
used, there are a variety of wavelengths collected ranging
from 0.34–1.6 µm. AOD at the MODIS retrieval wavelength
of 0.55 µm (hereafter τA) was derived from the quadratic
interpolation method of O’Neill et al. (2003). Typically
AERONET level 2 AOD data have uncertainties of <±0.015
(Eck et al., 1999; Schmid et al., 2003) spectral deconvolution
algorithm and are used in this analysis.
MOD04 retrievals were matched to AERONET measure-
ments of AOD with tolerances of ±30 min and ±30 km.
Errors resulting from transport or localization conditions
were found to be small at the broad statistical level (see
Appendix A). Every possible match between instantaneous
AERONET measurements and MODIS retrievals was in-
cluded in the matched dataset. Because many AERONET
stations have a high sampling rate (15 min or less), individual
MODIS retrievals are often paired with multiple AERONET
retrievals. For the final step of estimating instrument er-
ror variance for data assimilation, as described in Sect. 7,
AERONET data from each site were aggregated into six-hour
bins.
The AERONET Level 2.0 algorithm also includes steps to
remove cloud contamination, creating a potential sampling
bias in our analysis of cloud effects in the MODIS retrieval.
This is discussed further in Sect. 3.2.
2.3 MODIS albedo and snow products
The lower boundary condition is an important component
of the aerosol retrieval, and the complexity and variability
of the land surface pose significant challenges for accurate
retrievals. This study is concerned with two potential is-
sues: (1) Surface albedo characterization and in particular
how it relates to the IR-visible regression, and (2) poten-
tial for snow contamination. Surface properties were char-
acterized for this study using the MODIS MCD43 albedo
product (Schaaf et al., 2002). This product is produced as
a Level 3 16-day composite product based on the MOD09
atmospherically corrected surface reflectance product (Ver-
mote et al., 1997). The 16-day compositing algorithm is de-
signed to systematically eliminate the influence of clouds and
aerosols on the observations. The algorithm uses observa-
tions from multiple viewing geometries to estimate the bi-
directional reflectance function (BRDF), by fitting observa-
tions to a RossThick-LiSparse model of surface albedo and
BRDF (Lucht et al., 2000). The model parameters can be
used to calculate black- and white-sky albedoes for the seven
primary MODIS bands, as well as integrated visible, near-
infrared, and shortwave albedoes. This study used the black-
sky hemispheric albedoes at 0.47 µm, 0.66 µm, and 2.12 µm
from the MODIS albedo product (product MCD43C3), cal-
culated using the mean solar zenith angle for each location
and time. The QA information included with the MODIS
albedo product was used to exclude all albedo data with qual-
ity other than “very good.”
For snow, the MODIS albedo product incorporates tests
for snow contamination in each 500-m MODIS pixel that
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

382 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
goes into the MCD43 albedo product. The MCD43 product
calculates albedo using only snow-free pixels, and includes
a diagnostic variable indicating the fraction of 500 m pixels
in each 0.05◦ cell where snow was indicated. This informa-
tion was used to diagnose residual snow-related bias in the
MODIS AOD.
Because the footprint of MODIS aerosol retrievals has
a highly variable size, shape, and orientation, no attempt
was made to match MODIS AOD retrievals precisely to the
Level 3 snow/albedo data. Instead, the center of each AOD
retrieval was matched to the nearest 0.05◦ grid cell from the
MCD43C3 climate modeling grid (CMG) albedo product.
For the snow analysis, a spatial and temporal window around
the matched observation was used (see details in Sect. 3.3).
Since both the AOD and snow/albedo products are derived
from the same MODIS observations, there is the possibility
that the interaction between AOD and albedo is reciprocal.
Appendix B to this paper includes an analysis to determine
whether MODIS albedo data were contaminated by aerosol.
Results in Appendix B show that persistent high AOD often
resulted in failed albedo retrievals, but that AOD contami-
nation in albedo retrievals with “very good” QA flags was
negligible.
2.4 Statistical evaluation of MODIS AOD
To evaluate the utility of AOD information for use in a data
assimilation grade product, standards of accuracy must be
established. In terms of interaction with an aerosol transport
model, both random and systematic errors will affect analysis
and forecast outcomes. Producing a dataset for data assim-
ilation applications involves three goals: (a) elimination of
outliers, (b) elimination of systematic bias, (c) quantitative
characterization of residual errors. Because of the physics
of the retrieval problem, the relative strengths of the land
surface, microphysics and cloud biases are AOD dependent.
Nominally aerosol regimes can be broken down into: <0.2,
low AOD near signal to noise thresholds; 0.2–0.6, sufficient
aerosol signal in the single scattering regime; 0.6–1.4 multi-
ple scattering regime; and >1.4 extreme AOD events. These
categories are used as necessary in the statistical analysis.
A simple intuitive metric for analysis of the retrieved AOD
values is the tolerance which bounds AOD errors. For this
study, a target accuracy is defined by:
τM = τA ±
(
0.05 + τA
5
)
, (1)
and sets of retrievals are evaluated according to the fraction
that fall within, above, and below this target accuracy. This
is a weaker constraint than the proposed 0.05± 0.15×AOD
used in some studies (e.g., Levy et al., 2007b, 2010); how-
ever, those studies apply spatial and temporal averaging to
the MODIS and AERONET AOD data before calculation of
errors. This metric provides a simple accounting for the fre-
quency of positive and negative outliers, but provides little
information about the magnitude of errors. Where appro-
priate, cumulative distribution functions (CDF) of errors are
used for a more quantitative description.
Positive and negative errors in retrieved AOD are of-
ten asymmetrical, which violates a key assumption of or-
dinary least-squares regression. Linear fits of MODIS and
AERONET AOD consistently have intercepts significantly
different from zero. These non-zero intercepts are driven
by errors in retrieved AOD in relatively clean conditions,
due to problems with the surface boundary condition, cloud
contamination, or other problems. Because the error bud-
get of retrieved AOD varies as a function of AOD, the non-
zero intercepts often artificially skew the regression slope
away from the relationship indicated by the high-AOD data.
Therefore, to derive a linear representation of the relation-
ship between AOD values, the slope forced through zero is
calculated using the following equation:
SLOPE
(
τM
τA
)
=
∑
τM · τA
∑
τ 2A
(2)
Slope calculations in this analysis are made using only mod-
erate to high AOD (0.2<τA < 1.4). Where data volume is
sufficient, a separate slope calculation is made for extreme
AOD (τA > 1.4).
Finally, this analysis makes extensive use of the Root
Mean Squared Error (RMSE), calculated as:
RMSE =
√
1
n
∑
n
(τA − τM)2 (3)
RMSE is sensitive to both systematic and random errors and
provides an estimate of the expected error in AOD in the ab-
sence of information about specific sources of errors. In non-
background aerosol conditions, RMSE has a strong relation-
ship with AOD (see Sect. 3.1). For this analysis, the back-
ground observations (τAERONET < 0.2) are used to calculate
a “noise floor” for RMSE, and then a linear relationship of
the form
RMSE = a τ + b (4)
is estimated for higher AOD. Prognostically, “estimated
RMSE” (ε) is calculated for a MODIS observation of a given
optical depth as the higher of the regressed value or the
RMSE “noise floor”:
ε (τM) = max [RMSE (τA < 0.2), a τM + b] (5)
When data volumes are sufficient, this approach is extended
to use a separate linear relationship for very high AOD val-
ues.
2.5 Aggregation and textural filtering
For assimilation into an aerosol model, AOD observations
are aggregated into a gridded Level 3 product for assimila-
tion. This is done for two reasons: first, to avoid artifacts
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 385
Table 1a. Distribution of AOD values and linear regression results for matched MODIS-AERONET dataset, stratified by MODIS QA value.
AOD data in matched dataset
Sensor QA τM < 0.2 0.2<τM < 0.6 0.6<τM < 1.4 τM > 1.4 Slope r2
Terra All 1 437 755 485 807 83 224 11 173 0.97 0.54
bad 145 482 78 837 11 852 983 0.98 0.35
marginal 146 174 64 311 9251 1027 1.12 0.45
good 235 306 93 972 13,123 1404 1.08 0.50
very good 910 793 248 687 48 998 7759 0.93 0.64
Aqua All 1 159 361 427 808 72 875 10 154 1.00 0.51
bad 137 130 78 170 12 664 1137 1.01 0.39
marginal 128 620 66 576 10 041 1141 1.14 0.46
good 198 042 88 644 13 663 1947 1.11 0.50
very good 695 569 194 418 36 507 5929 0.94 0.60
Table 1b. Compliance of AOD values to the error limits of Eq. (1) (|τM−τA|<= (0.05 + 0.20τA) for matched MODIS-AERONET dataset,
stratified by MODIS QA value.
Compliance (Below/Within/Above)
Sensor QA τM < 0.2 0.2<τM < 0.6 0.6<τM < 1.4 τM > 1.4 All
Terra ALL 20/66/12 08/52/39 08/53/37 03/50/46 17/62/20
bad 16/57/25 09/34/56 07/37/55 02/44/52 13/49/37
marginal 11/65/23 03/39/56 03/36/59 02/42/55 08/56/34
good 13/67/18 04/45/50 05/43/51 01/45/53 10/60/28
very good 24/68/07 10/63/25 11/62/26 03/53/42 21/66/12
Aqua ALL 17/68/13 06/49/43 08/44/47 01/44/53 14/62/22
bad 16/59/23 08/33/57 06/36/56 02/39/57 13/49/37
marginal 10/66/23 03/38/57 04/34/61 00/33/65 07/55/36
good 13/67/18 03/44/51 04/38/57 01/37/61 10/59/30
very good 20/71/07 08/61/29 11/52/35 01/50/47 17/68/13
3.1 Factors affecting global signal-to-noise ratio of
retrieved AOD
For an atmospheric modeling application, systematic errors
are often more damaging than random errors. In the case
of a global model, the model resolution is typically coarser
than the AOD observations, permitting extensive averaging
to reduce random noise. High levels of random noise can
also obscure systematic bias.
The dominant signal of aerosols for optical retrieval is the
scattering of direct sunlight back to the sensor. The MODIS
aerosol retrieval is based on inversion of a full radiative trans-
fer simulation that includes aerosol effects on all pathways
except multiple scattering (Levy et al., 2007a), but the pri-
mary determinant of the signal strength for retrieval of AOD
is the ratio of scattered direct sunlight to light reflected from
the surface. This is a reason why retrieval of AOD over
ocean is more precise than over land, and also a reason why
this MODIS algorithm is unable to retrieve aerosol properties
over bright surfaces.
Figure 3 shows how the compliance of the MODIS re-
trieved AOD varies as a function of surface albedo at 0.47,
0.66, and 2.12 µm. The gray bars at the top and bottom of the
graph indicate the fraction of retrievals above and below the
target accuracy. Thus, the area between the shaded regions
is indicative of the compliant fraction of the data. The solid
lines (each vertex is an average of 20 000 retrievals) indicate
the mean AOD bias, and the dotted lines indicate the 25th and
75th percentiles of each bin. The graphs in Fig. 3 illustrate
two effects. First, the increase in MODIS-AERONET bias
with increasing albedo at 0.47 µm and 0.66 µm clearly points
to mis-parameterization of the lower boundary condition in
the retrieval. This is examined in detail in Sect. 4. Second,
the increased spread of the errors at higher albedo illustrates
the decrease in differential signal to noise ratio.
Figure 4a illustrates the differential signal-to-noise con-
sideration for AOD retrievals through the optical path length
of the atmosphere. Compliance statistics and mean AOD
bias are shown as a function of the sensor scan angle (angle
from nadir), which determines the atmospheric path length
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

386 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
Table 1c. Prognostic and diagnostic regression of RMS error in MODIS AOD as a function of AOD, stratified by MODIS QA value. A
single RMS error estimate for low-AOD conditions (τA < 0.2) is also shown.
RMSE RMSE vs. τA RMSE vs. τM
Sensor QA τA < 0.2 0.2<τA < 1.4 τA > 1.4 0.2<τA < 1.4 τA > 1.4
Terra ALL 0.10 0.03 + 0.22τA −0.18 + 0.36τA 0.05 + 0.22τM −0.57 + 0.53τM
bad 0.15 0.05 + 0.24τA −0.17 + 0.41τA 0.06 + 0.33τM −0.80 + 0.70τM
marginal 0.13 0.04 + 0.23τA −0.21 + 0.39τA 0.04 + 0.33τM −0.52 + 0.54τM
good 0.11 0.04 + 0.22τA −0.24 + 0.38τA 0.04 + 0.27τM −0.43 + 0.47τM
very good 0.08 0.02 + 0.22τA −0.17 + 0.34τA 0.04 + 0.18τM −0.61 + 0.54τM
Aqua ALL 0.10 0.03 + 0.26τA −0.19 + 0.38τA 0.04 + 0.27τM −0.58 + 0.56τM
bad 0.15 0.06 + 0.25τA 0.18 + 0.19τA 0.06 + 0.32τM −0.36 + 0.47τM
marginal 0.13 0.05 + 0.24τA −0.19 + 0.43τA 0.03 + 0.34τM −0.63 + 0.65τM
good 0.11 0.03 + 0.27τA −0.05 + 0.30τA 0.03 + 0.30τM −0.62 + 0.62τM
very good 0.07 0.01 + 0.26τA −0.29 + 0.42τA 0.03 + 0.22*τM −0.62 + 0.56τM
from the surface to the sensor. Near nadir (scan angle< 5◦),
56% of MODIS AOD retrievals are compliant, compared
with 83% at the scan edge (scan angle> 60◦). This means
that comparisons of MODIS aerosol retrievals with narrow-
swath instruments such as the Multi-Angle Scanning Ra-
diometer (MISR) or the Cloud-Aerosol Lidar with Orthog-
onal Polarization (CALIOP) will overestimate random error
against the whole MODIS product. This discrepancy be-
tween nadir and scan edge is caused by the relative contri-
bution of surface reflected light to the total radiance at the
sensor. Thus, it diminishes at increasing optical depth: for
τA > 0.6, compliance fractions are 60% and 61% at nadir
and scan edge, respectively. At even higher τA, spatial mis-
match between MODIS and AERONET is a larger factor; for
τA > 1.0, compliance is better at nadir.
A related phenomenon is shown in Fig. 4b, which depicts
the bias and compliance statistics as a function of the scat-
tering angle. In the global dataset the retrieval has almost
no systematic bias associated with scattering angle, indicat-
ing that the model used to account for anisotropy in the sur-
face reflectance (Levy et al., 2007b) appears to be sufficient.
But unavoidably, as scattering angle increases, shadows over
vegetated surfaces diminish, the surface brightness increases,
and the precision of the retrieval declines. The interaction be-
tween solar geometry and scattering angle means that scatter-
ing angle distributions are not stationary with latitude, which
may also influence this result. At scattering angles smaller
than 100◦, 84% of retrievals are compliant. At very high
scattering angles, where the sun is almost directly behind the
sensor, there is a sharp spike in retrieved AOD: compliance
is only 41% for scattering angles above 170◦. This is caused
by the “hot spot” of vegetation reflectance (Vermote and Roy,
2002), and retrievals with scattering angles over 170◦ (0.5%
of our matched retrieval data set) should be avoided because
of this problem.
3.2 Influence of MODIS-detected clouds on retrieved
AOD
Undetected clouds, subpixel or otherwise, can cause a pos-
itive bias in the retrieval. Conversely, cloud shadows can
result in a negative bias. Overall, however, we expect a pre-
dominantly positive bias from cloud effects. The MODIS
retrieval includes auxiliary information on the fraction of
pixels within the retrieval footprint with MODIS-detected
cloud. Only 16% of successful AOD retrievals have MODIS-
detected cloud cover within the retrieval footprint. This
fraction decreases with increasingly strict QA, from 26% of
“Bad” retrievals to 10% of “Very Good” retrievals.
Retrievals with MODIS-detected cloud have a slight posi-
tive bias relative to the complete dataset (Fig. 4c). The differ-
ence in mean τM for retrievals with indicated clouds versus
no detected clouds is +0.04, while the corresponding differ-
ence in τA is less than 0.01. While this elevated AOD may in-
dicate undiagnosed subpixel clouds affecting the reflectances
used in the retrieval (Zhang and Reid, 2006), or may be an
artifact of three-dimensional scattering not included in the re-
trieval model (Varnai and Marshak, 2009), other studies (Ko-
ren et al., 2007) contend that this may be the result of elevated
aerosol particle concentrations in the vicinity of clouds. But,
given the scale of such aerosol features, we might expect a
larger response in τA. For over water cases, τA did not show
an increase in AOD with increasing cloud cover nearly as
large as τM (Zhang et al., 2005), suggesting artifact may be
the more dominant factor.
One important caveat of this analysis is the cloud-clearing
undertaken in the processing of the AERONET data. While
not 100%, the cloud-clearing algorithms used in AERONET
Level 2 processing are highly effective. Because no
AERONET retrieval is made in conditions flagged as cloudy,
the matched MODIS-AERONET dataset systematically ex-
cludes many observations in partially cloudy or thin cirrus
cloud conditions. Thus, the true impact of clouds on MODIS
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

388 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
observed radiance in near-infrared and thermal wavelengths.
This technique was shown to be a great improvement over
the snow masking used in the Collection 4 MODIS AOD re-
trieval (Li et al., 2005).
Snow contamination is extremely difficult to completely
eliminate because snow can cover any fraction of the retrieval
footprint. For this study, the snow contamination checks in
the MODIS Collection 5 algorithm (Li et al., 2005) were ex-
tended using the snow flag in the MODIS 0.05◦ albedo prod-
uct, which indicates the fraction of 500-m pixels affected by
snow during the 16-day compositing period of this product.
Using this product makes it possible to extend the test for
snow contamination both in time, by considering data from
a period before the AOD retrieval, and in space, by looking
for snow over a wider area. These spatially and temporally
extended checks do not give proof of snow contamination of
a given retrieval. However, as will be shown, they permit
identification and removal of retrievals where the probability
of snow contamination is greater.
For AERONET retrievals in Boreal North America (above
49◦ N) with τA < 0.2, compliance is 70%, and the mean bias
in the matched dataset is +0.013, or 14%. Of the non-
compliant retrievals, 70% are biased high. Note that the
matched dataset includes no retrievals from this region dur-
ing December through February, and only 4% of retrievals
are from November through April. This relates to limita-
tions in both datasets: AERONET sites in northern regions
are customarily shut down during winter months to prevent
damage by snow, and MODIS retrieves AOD only when the
sun is less than 72◦ from nadir.
Only 0.2% of matched retrievals from this region are
flagged as having snow in the corresponding MODIS 16-
day albedo product, but these flagged points are 59% compli-
ant and have a mean bias of +0.052, and the non-compliant
retrievals are 95% biased high. Thus, the snow flag in the
MODIS albedo product is consistent with snow contamina-
tion in the MODIS AOD, but the number of points flagged
is very small, and their effect on the bulk error statistics is
minimal. This indicates that this test for snow is effective but
likely incomplete.
If we extend the snow test in time by checking the snow
flag in the MODIS albedo product for the 32 days prior to
the AOD observation, and in space by looking for snow in
a 0.35× 0.35◦ box around the AOD retrieval, we capture
nearly 12% of the retrievals in the matched dataset, includ-
ing nearly all the retrievals in March and April. The retrievals
captured in this wider net have 59% compliance, a positive
bias of +0.031, and 85% of non-compliant retrievals are bi-
ased high. Thus, even this larger set of retrievals is con-
sistent with snow contamination. When these retrievals are
removed, the compliance of the remaining retrievals in Bo-
real NA is 73%, 3% higher than before filtering. The mean
bias is reduced to +0.009 (9%). Non-compliant retrievals are
still 66% biased high. These results indicate that some snow
contamination has been removed, but positive biases persist
Table 2. Statistical analysis of filters to remove snow contami-
nation in MODIS AOD. Data shown were calculated using only
MODIS AOD with MODIS QA of “Very Good”.
All Data
Region Bias Compliance RMSE
N. American Boreal 0.013 08/70/20 0.10
E. CONUS −0.025 22/71/05 0.07
W. CONUS 0.017 13/64/22 0.11
Europe–Mediterranean −0.010 14/75/09 0.08
Eurasian Boreal −0.016 15/77/06 0.09
East Asia Mid-Latitudes 0.028 13/64/21 0.18
Snow-Matched
Bias Compliance RMSE %data
N. American Boreal 0.052 02/58/38 0.16 0.09
E. CONUS 0.012 01/83/14 0.25 0.05
W. CONUS 0.061 01/53/44 0.17 0.08
Europe–Mediterranean −0.011 03/96/00 0.01 0.05
Eurasian Boreal 0.006 01/84/13 0.03 0.07
East Asia Mid-Latitudes 0.092 00/57/42 0.24 0.13
Snow-Extended Match
Bias Compliance RMSE %data
N. American Boreal 0.031 06/59/34 0.09 17.87
E. CONUS 0.026 07/70/21 0.08 7.58
W. CONUS 0.023 08/67/24 0.09 8.32
Europe–Mediterranean −0.009 13/75/10 0.07 4.05
Eurasian Boreal −0.010 14/79/05 0.07 10.35
East Asia Mid-Latitudes 0.085 01/57/40 0.14 5.20
in the data, whether related to snow contamination or other
causes.
Table 2 shows statistics for retrievals excluded using the
matched and extended filters based on the MODIS albedo
product, for all regions that extend to mid-latitudes or above.
3.4 Basic QA filtering for subsequent testing
Based on the results in this section, the remainder of our anal-
ysis at the retrieval level (Sects. 4–6) will exclude the follow-
ing MODIS AOD data:
1. Data with MODIS mandatory QA other than “Very
Good”;
2. Data with MODIS-indicated cloud;
3. Data with scattering angle above 170◦;
This represents the extent of filtering that can be achieved
using only the auxiliary data included in the MODIS aerosol
product. This level of filtering will serve as the baseline for
our evaluation of the additional filtering and correction steps
described in the next sections.
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 389
Table 3. Statistical evaluation of MODIS AOD by region, after application of basic QA filtering.
Compliance RMSE vs. τA vs. τM
Region NAOD Nsites Slope r2 τM < 0.2 0.2<τM < 0.6 0.6<τM < 1.4 τM > 1.4 All τA < 0.2 0.2<τA < 1.4 0.2<τA < 1.4
Global 1 797 997 257 0.92 0.62 24/69/06 10/64/24 12/58/29 02/52/44 21/67/10 0.07 0.02 + 0.25τA 0.04 + 0.20τM
N. American Boreal 168 961 14 1.17 0.62 11/78/09 00/50/49 02/43/54 00/22/77 10/74/14 0.06 −0.06 + 0.51τA 0.01 + 0.31τM
E. CONUS 258 728 33 1.01 0.74 28/69/02 03/80/15 02/74/22 00/14/85 24/71/04 0.06 0.04 + 0.09τA 0.03 + 0.12τM
W. CONUS 176 530 29 1.31 0.25 16/71/11 01/30/67 01/23/74 00/00/100 13/64/21 0.10 0.02 + 0.31τA −0.01 + 0.53τM
Central America 14 146 10 0.96 0.54 58/40/01 15/66/18 06/56/37 00/19/80 44/46/08 0.09 0.07 + 0.20τA 0.05 + 0.16τM
South America 59 765 11 1.00 0.81 71/27/00 12/80/07 01/78/19 00/42/56 57/37/05 0.11 0.05 + 0.17τA 0.04 + 0.11τM
S. South America 90 946 7 1.07 0.68 30/51/18 05/28/65 00/73/25 00/61/38 27/49/22 0.09 0.09 + 0.10τA 0.15 + 0.02τM
Africa below equator 79 314 3 0.88 0.55 22/72/04 10/69/20 27/66/06 00/00/00 19/72/07 0.06 0.00 + 0.27τA 0.01 + 0.23τM
Equatorial Africa 7411 3 0.99 0.45 36/62/00 03/84/12 00/35/64 00/00/00 32/65/02 0.08 0.07 + 0.04τA 0.02 + 0.15τM
Africa above equator 102 001 19 0.69 0.68 47/47/05 48/44/07 47/49/02 10/79/09 47/46/06 0.09 0.03 + 0.31τA 0.13 + 0.18τM
Europe–Mediterranean 389 898 44 1.04 0.46 19/77/03 03/73/22 00/51/47 19/50/30 14/76/08 0.06 0.03 + 0.17τA −0.00 + 0.25τM
Eurasian Boreal 170 230 26 1.05 0.65 20/77/01 03/79/16 01/45/53 00/26/73 16/77/06 0.06 0.00 + 0.24τA 0.00 + 0.21τM
East Asia Mid-Latitudes 89 265 31 1.02 0.65 25/68/06 09/67/22 04/55/39 02/59/38 15/65/19 0.09 0.05 + 0.18τA 0.02 + 0.21τM
Peninsular SE Asia 35 094 7 0.90 0.62 65/33/00 23/68/08 09/66/24 00/63/36 38/53/07 0.11 0.07 + 0.13τA 0.06 + 0.13τM
Indian Subcontinent 18 209 7 1.04 0.71 07/91/00 04/79/15 02/74/23 02/42/55 03/77/18 0.10 0.05 + 0.11τA 0.01 + 0.16τM
Australian Continent 128 848 8 0.99 0.19 24/69/05 06/55/38 00/50/50 00/00/00 23/68/07 0.06 −0.04 + 0.52τA 0.02 + 0.25τM
4 Description of regional biases
Global statistics average out considerable regional variabil-
ity in the retrieval quality, and are not geographically repre-
sentative because they are weighted to locations with high
AERONET coverage. Even at the regional level, the covari-
ance of particle properties with AOD results in mixed in-
dividual site efficacy as a function of AOD. Table 3 shows
statistics, calculated using QA filtering as described in the
previous section, for 14 different land regions. Statistics for
each AERONET site for the 2005–2008 study period are pro-
vided in the Supplement associated with this paper, or can be
sent by request.
Higher compliance regions include northern Eurasia,
Southern Europe-Mediterranean, and the Eastern Contiguous
United States (E CONUS). Low compliance regions include
South America, Africa above the equator, Central America,
and Peninsular Southeast Asia. Just as important, the bal-
ance of positive and negative non-compliance shows large re-
gional variability. In Northern Africa, 89% of non-compliant
retrievals are biased low, while in the Indian subcontinent,
86% of non-compliant retrievals are biased high. Regional
slope is also variable, ranging from a low of 0.69 for Africa
above the equator to 1.17 for North American Boreal (the
slope for the Western continental US is 1.31, but the regres-
sion is weak, with r2 = 0.25 for that region).
Just as the global mean statistics are composed of regions
with different behavior, so the slopes calculated for each re-
gion are the aggregate of the sites within that region, whose
behavior is not always homogeneous. Our analysis is neces-
sarily coarse-grained; in addition to variations between sites,
aerosol properties in our selected regions will also have sea-
sonal and interannual variation that we do not explicitly ad-
dress (seasonal breakdowns for each site are included in the
Supplement). This section provides a few key pieces of con-
text from our analysis and from the literature that can aid in
interpreting the statistics for each region. Regional biases
will be analyzed and quantified in the next section.
4.1 North America
North America is split between a Northern Boreal Region,
Eastern CONUS and Western CONUS. Owing to its back-
ground nature, the northern boreal AOD is typically low,
lowest of all of the regions compared (mean τA = 0.11). In-
dividual site performance varies widely. For very clean
background sites – e.g., Yellowknife (62◦ N, 114◦ W), Opal
(80◦ N, 86◦W) – AOD rarely exceeds 0.15, yielding no infor-
mation from regressions. Sporadic high AOD events occur in
the spring and summer months at some sites – e.g., Bonanza
Creek (65◦ N, 148◦ W), Bratts Lake (50◦ N, 105◦W), Pickle
Lake (51◦ N, 90◦ W) – due to boreal biomass burning activity
(Eck et al., 2009), which can drive r2 values to nearly 0.9. In
very rare occasions, AOD can be extremely high (>8), over-
whelming the AERONET sensor (O’Neill et al., 2002) and
driving aerosol layer reflectivity to the AOD “Semi-infinite”
regime . In addition to these real events, the boreal is also
susceptible to potential snow bias which some sites – e.g.,
Kelowna (50◦ N, 119◦W) – clearly show as strong early sea-
son high biases. Regressions are dominated by burning ac-
tivity, and a fairly consistent high slope bias is present among
most sites. These slopes are event driven, and are dependent
on the microphysical properties of smoke particles constantly
evolving in size and absorption (Reid et al., 2005a,b). As a
consequence, some sites show slope biases higher than 1.5
(e.g., Bratts Lake), and others close to 1.0 (e.g., Bonanza
Creek). Terra and Aqua appear to perform similarly, with
slope deviations between them at ±0.1. For the sites affected
by extreme AOD events, slope biases ranged from 1.4 to 1.9,
with Aqua slopes higher than Terra by 0.2 or more in all
cases.
MODIS performs best in the Eastern CONUS. With the
more uniform mix of sulfate and organic pollution, slopes
for most sites are within within ±0.2 of unity. Forty per-
cent of sites have over 80% compliance. RMS error is also
very good, one of the lowest regionally. Conversely, Western
CONUS is one of MODIS’s greater challenges and shows
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

390 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
one of the highest RMSEs. Previous studies have reported a
significant bias (+0.2 or greater) over the arid regions due to
shortcomings in the lower boundary condition (Drury et al.,
2008). This is manifested in our statistics as extremely poor
compliance and strong positive bias for above-background
values of τM, which are often retrieved in conditions where
the true AOD is a low background value. Regression fits
are generally weak because the AOD in the pristine desert
has little range. Heavily urbanized sites such as around
Southern California or the San Francisco Bay Area perform
nearly as poorly. Like the boreal zone, large biomass burn-
ing events can dominate individual regressions, with slopes
ranging from 0.6 to more than 1.6. Remote sites with less
arid landscapes – e.g., HJ Andrews (44◦ N, 122◦ W), Mis-
soula (47◦ N, 114◦ W), and Fresno (37◦ N, 120◦ W) – have
more reasonable performance, with slopes generally within
±0.1 of unity and >70% compliance.
4.2 Central America
Central America has only four sites with a significant num-
ber of data points of AOD> 0.2 (e.g., >500 data points):
Tuxtla Gutierrez (17◦ N, 93◦ W), Tenosique (17◦ N, 91◦ W),
and Mexico City (19◦ N, 99◦W) in Mexico and La Par-
guera (18◦ N, 67◦ W) in Puerto Rico. With its urban na-
ture, Mexico City shows very poor scores (r2 < 0.30). The
other two sites in Mexico show higher r2 values (0.55–0.85),
but variable slopes: slopes for moderate to high AOD range
from 0.8 to 1.1, and slopes for extreme AOD range from 1.06
(associated correlation is very low) to 1.67 (MODIS-Aqua
at Tenosique, r2 = 0.66). As with other regions, estimated
slopes for MODIS-Aqua are generally somewhat higher for
moderate to high AOD values, and consistently much higher
for extreme AOD values.
La Parguera is an excellent site for monitoring the trans-
port of African dust over the Caribbean. In the matched data
set, only MODIS-Aqua has sufficient range of AOD to es-
tablish a robust correlation (slope = 0.94, r2 = 0.63). Compli-
ance for MODIS-Aqua is also much better (72% for all AOD
vs 57% for MODIS-Terra).
4.3 South America
South America has four key aerosol regimes: The August–
October burning in Rhondonia and Mato Grosso Brazil;
February–May northern biomass burning in Columbia and
Venezuela; an Argentinean dust regime; and urban super
plumes from such major cities as San Paulo and Buenos
Aires.
At low AOD, compliance is highly variable within the
region: for AOD< 0.2, compliance varies from 14% at
Cuiaba-Miranda (15◦ S, 56◦ W) to 52% at Petrolina-SONDA
(9◦ S, 41◦ W) and Santa Cruz – UTEPSA (18◦ S, 63◦ W).
At low AODs, negative retrieved AOD is common. Over-
all RMSE is good in South America for moderate AOD. The
massive Brazilian biomass burning signal is the primary de-
terminant of validation statistics for this region. Each of the
10 sites in the biomass burning region have among the high-
est r2 values over the globe, yet each of the sites has a sig-
nificant slope bias, and these biases vary from site to site;
this leads to the overall poor r2 and RMSE values for the
continent. Also, differences between Terra and Aqua de-
rived AOD are highly localized, with Aqua being generally
higher. The heavily impacted Ji-Parana site (11◦ S, 62◦ W)
in central Brazil (average AOD> 1.0 for all data where
AOD> 0.2), shows the strongest biases, with a moderate-
AOD and extreme-AOD slopes of 1.13 and 1.54 for Terra,
and 1.28 and 1.93 for Aqua. For other smoke receptor sites,
the slopes for Terra vs. Aqua show differences on the order
of 0.1, with Aqua consistently higher. The positive bias in τM
is considerably stronger at high AOD; the high-AOD slope is
always higher than the moderate-AOD slope (e.g., Fig. 7p
where slopes are presented for data where τM > 1.4). The
regional mix of forest, cerrado, and pasture burning cou-
pled with the rapid evolution of the smoke plume makes
smoke aerosol properties highly variable (Ferek et al., 1998).
Since the MODIS retrieval cannot resolve these variations in
aerosol properties, the MODIS microphysical bias is highly
variable.
No AERONET sites are available in Northern South
America. Southern South America is prone to occasional
dust emissions in an otherwise pristine environment (Gasso
and Stein, 2007). Background AODs are typically low
(<0.1). In such conditions over a semi-arid region, we are
near the noise threshold for MODIS, which shows poor skill
in general, often overestimating AODs by a factor of three.
Compliance at most sites in this region is low and is driven
by biases at low optical depth, whether high biases (48%
of retrievals at Trelew (43◦ S, 65◦ W)with τM < 0.2 are bi-
ased high) or low biases (55% of retrievals at CEILAP-UTN
(35◦ S, 58◦ W) with τM < 0.2 are biased low).
Lastly, we examined the urban regions of San Paulo
(Brazil) (24◦ S, 47◦ W) and La Paz (Bolivia) (17◦ S, 68◦ W).
The La Paz site is extremely clean: in the matched MODIS-
AERONET dataset, less than 3% of τA were above 0.2. The
MODIS AOD has no skill in regression against AERONET
at this site, and compliance for τM > 0.2 is poor (<20%), but
overall compliance is good because of the dominance of very
low AOD. Sao Paulo performs moderately well in both re-
gression and compliance, but results for low AOD are worse
than most other stations (34% compliant for τM < 0.2).
4.4 Europe, the Mediterranean, and northern Eurasia
Like Eastern CONUS, there are many AERONET sites in
Europe. As expected, urban sites often show poorer perfor-
mance relative to more remote areas, but in general the frac-
tion of complaint data points is very high. RMSEs are quite
good, both in the noise floor (0.06 for Terra and Aqua) and
the diagnostic slope (0.03 + 0.17 τA for Terra, 0.03 + 0.16 τA
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 391
for Aqua). Regression slopes for moderate AOD are typi-
cally within +/0.1 of unity. Only a few sites (e.g., Forth Crete
– 35◦ N, 25◦ W), have slope biases larger than 1.2. Site spe-
cific Terra and Aqua slopes also correlate very well, within
±0.1 of each other for 34/44 sites. Scores are also excellent
in the Eurasian Boreal, with greater than 75% compliance for
17/26 sites.
4.5 Africa
African aerosol regimes observed by MODIS can be broken
down into the northern dust/smoke impacted Sahelian region,
biomass burning impacted equatorial Africa, and the burn-
ing and pollution region of southern Africa (No retrievals are
made in the pure dust regimes of the Sahara). The Sahe-
lian Africa environment strains the MODIS algorithms, with
both high background albedo and variable fine/coarse parti-
tion particle size. RMSEs are more than 50% for all τA < 0.8.
Slopes for moderate AOD vary between 0.52 and 1.20, al-
though in general they are in the 0.6–0.8 range. AOD is neg-
atively biased in all ranges at all sites except for Izana (28◦ N,
16◦ W).
The three sites in equatorial Africa – ICIPE-Mbita (0◦ S,
34◦ E), Nairobi (1◦ S, 37◦ E), and Kibale (1◦ N, 30◦ E), typi-
cally exhibited AOD< 0.2. They are typically compliant for
50–80% of retrievals with estimated slopes ranging from 0.8
to 1.1. Given the small dynamic range of data for these
sites, however, r2 values are <0.5. Central Africa hosts the
world’s largest biomass burning features in the months of
July–October. Despite the size of these features, there are
very few AERONET sites in the region and it is difficult to
determine if there is the same performance heterogeneity as
South America. However, there is some advection into south-
ern Africa.
In addition to episodic smoke events from the north, high
sulfate pollution is present year around in the South African
Highveld and the Johannesburg regions. For biomass burn-
ing, Mongu (15◦ S, 23◦ W) is the only representative site of
the central biomass burning region, with moderate r2 value
(∼0.6 for both Terra and Aqua), nearly 50% RMS errors,
and a slope of 0.8. Bias between MODIS and AERONET
at southern African sites appears to be correlated to particle
absorption (Tom Eck, NASA GSFC, personal communica-
tion, 2010). In the polluted Skukuza site (25◦ S, 32◦ E) in
the Highveld r2 values are good (>0.75), but a low bias is
also persistent (slope = 0.88). The urbanized Witswatersrand
University site in Johannesburg (26◦ S, 28◦ E) shows positive
slope bias 1.32 (r2 = 0.30) but still exhibits 70% compliance
overall.
4.6 Indian sub-continent
Over our the study period, the number of AERONET sites
around the Indian Sub-continent was not sufficient to de-
termine spatial performance. The major aerosol features of
the Indo-Gangenic Plain do appear to be well represented
in MODIS AOD data. At the Kanpur (27◦ N, 80◦ E) and
Gandhi College (26◦ N, 84◦ E) sites correlations are good
(r2 > 0.7), and slopes are within ±0.1 of unity. Compliance
is above 77% for both sites and satellites. The limited data
(N ∼ 500) at Nainital (29◦ N, 79◦ E) and Pantnagar (29◦ N,
80◦ E), right at the base of the Himalayan mounts also per-
forms adequately, with some slope bias indicated (ranging
from 0.95 to 1.39), and r2 values of ∼0.25–0.75. All other
sites, including all sites outside of the Indo-Gangenic plane
(rest of India, Pakistan) have very limited data points (<200)
and very poor scores. It is noteworthy, however, that stud-
ies with handheld Sun photometers have been conducted in
western India for Aqua (Misra et al., 2008). They showed
interseasonal variability and moderate r2 values (∼0.5) in
MODIS scores, with in general a slope of 0.8 (but see dis-
cussion of linear regression in Sect. 2.4).
4.7 East and Southeast Asia
East and Southeast Asia encompasses several of the most
complex aerosol and land surface environments. Land sur-
face varies from extreme arid to dense vegetation to urban,
with occasional strong pollution, smoke, or dust events aloft.
Values for individual site r2 range from 0.38 – Taipei, Taiwan
(25◦ N, 122◦ E) to 0.84 – Xianghe, China (40◦ N, 117◦ E),
and slopes range from 0.75 (Taipei, Taiwan) to 1.2 – Bei-
jing (40◦ N, 116◦ E). Regions susceptible to dust, smoke, and
sulfate aerosol particles can have considerable scatter and ex-
traction of exact biases is well outside the scope of this paper.
Because of the environmental heterogeneity of Asia, regres-
sions can be reasonable, but compliance fraction low, and
visa versa. Readers are encouraged to review the individual
site statistics given in the Supplement.
SE Asia also shows mixed results. For most sites r2 values
are above 0.7, yet compliance is typically less than 60%. Fine
mode dominates here, yet slopes ranged from 0.76 – Mukda-
han (17◦ N, 105◦ E) to 1.52 – Bac Lieu (9◦ N, 106◦ E). Terra
and Aqua are reasonably consistent (slope difference <0.2
for most sites).
4.8 Australia
The bulk of Australian AERONET sites are in arid regions
and typically have very low AODs (<0.2). Consequently,
regression statistics are poor, and calculated slopes are not
meaningful. Jabiru (13◦ S, 133◦ E), the only location in the
matched data set impacted by biomass burning smoke, is
the only site with enough moderate AOD retrievals in the
matched dataset for a meaningful comparison (slope near 1.0,
r2 = 0.24, compliance of 56% for τM > 0.2). Australia ex-
hibits the same pattern seen in other arid landscapes where
many sites have predominantly low bias at τM < 0.2, but
almost 100% positive bias for 0.2<τM < 0.6, indicating a
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

392 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
boundary condition bias as well as significant positive errors
in retrieved AOD under clean conditions.
5 Diagnosis and correction of errors related to surface
boundary condition and microphysics
The site by site biases described in Sect. 4 imply that bi-
ases in the MODIS aerosol products are spatially and tem-
porally correlated, thus violating a common assumption of
many data assimilation schemes (Dee and Da Silva, 1998;
Daley and Barker, 2001). Because the RMSE includes both
the variance and the bias, systematic biases in the product
greatly increase the magnitude of RMSE and hence reduce
product impact in the model. Further, systematic bias can re-
sult in “hot” and “cold” spots in model analyses which in turn
propagate in model forecasts. To the fullest extent possible,
data must be debiased before assimilation. Regions which
cannot be adequately debiased must be masked.
Two effects that warrant a diagnostic debiasing involve the
lower boundary condition and aerosol microphysical bias.
Because the lower boundary condition and aerosol micro-
physical bias often co-vary, we must be careful to examine
them independently. In this section we isolate lower bound-
ary condition biases which are subsequently removed from
the signal for evaluation of residual microphysical biases.
5.1 Lower boundary condition
Figure 3 illustrates how shortcomings in the parameteriza-
tion of the lower boundary condition result in systematic
biases in the MODIS aerosol products. As discussed in
Sect. 2.1, the MODIS AOD retrieval uses a simple model to
describe the relationship between surface reflectance in the
near-infrared (2.12 µm) and visible (0.66 µm) wavelengths of
MODIS. This relationship is affected by numerous environ-
mental factors, not all of which are captured in the param-
eterization. Viewing conditions such as the scattering an-
gle are important, but reflectance is primarily determined by
the properties of the surface. Further, Fig. 4b suggests that
with the exception of the “hotspot” the treatment of surface
anisotropy is sufficient.
Using MODIS albedo data, an empirical correction for the
biases shown in Fig. 3 was estimated. Numerous correction
schemes were examined, and the best results were obtained
with the following form:
τA − τM = m1 A0.66 µm + m2 A2.12 µm + b, (6)
where A0.66 µm and A2.12 µm are the MODIS black-sky
albedo values at those wavelengths, and m1, m2 and b are
parameters to be fit empirically.
To estimate these parameters and test the correc-
tion, matched AOD retrievals where τA < 0.2 were subdi-
vided into estimation and validation subsets (selected by
AERONET site name: “A–K” for estimation, “L–Z” for val-
idation). Subsets were made geographically independent to
avoid overfitting to confounding regional variation not re-
lated to surface properties. Regressing the estimation subset,
we obtained a correction of the following form:
τM,corrected = τM−2.66A0.66 µm+1.25A2.12 µm+0.056 (7)
The regression coefficient for this relationship (r2 = 0.24) is
of the same order as the r2 for τM vs. τA for the low-AOD
estimation subset. Regression of albedo data at other wave-
lengths (as well as using the 0.66 µm/2.1 µm albedo ratio)
was also attempted, but the correlation was not improved (re-
sults not shown). Regression results using only data from
MODIS-Terra or MODIS-Aqua were very similar (results
not shown). The contribution of the surface reflectance to
observed radiance and thus to AOD error decreases with in-
creasing AOD. Therefore, we applied the calculated albedo
correction only to retrievals with τM < 0.6.
Evaluation of this correction with the geographically in-
dependent validation subset showed robust reduction of dif-
ferences between τM and τA. For 124 AERONET sites in
the validation subset where the mean correction to τM was
at least 0.005, the albedo correction reduced the mean AOD
error in 85 sites. For those 124 sites, the mean site AOD er-
ror and the estimated AOD correction were correlated with
r =−0.65, indicating the importance of the surface proper-
ties in the error budget at low AOD.
Figure 5 illustrates the nature and effects of the correction
shown in Eq. (7). As an example, Fig. 5b shows the magni-
tude of the correction estimated from one global data day of
MCD43 albedo data from May 2008. Well known high bi-
ases in desert regions are visible, especially in such regions
as the desert western United States, the Andes Mountains,
arid Australia and the desert belt across Africa, and the Tak-
lamakan and Gobi deserts. Low biases are visible over the
more forested regions over the globe. To demonstrate impact
over all AERONET sites, Fig. 5a shows the overall effect
of the correction on matched MODIS-AERONET site statis-
tics from the complete 2005–2008 data set. Contour lines
in Fig. 5a show the magnitude of the correction as a func-
tion of the 0.66 µm and 2.12 µm albedo values. Symbols in-
dicate specific AERONET stations, color coded to indicate
the mean AOD bias for τA < 0.2 (opposite in sign to the cor-
rection. That is, we want cold colors to counteract warm).
The dashed line indicates a 2:1 ratio of near-infrared to vis-
ible albedo, a standard assumption in older over-land AOD
retrievals (Kaufman et al., 1997).
Of 224 sites where the mean absolute AOD correction was
greater than 0.005, 70% had reduced absolute bias, 77% had
improved compliance, and 79% had reduced RMS error. For
most albedo regimes, this correction effectively debiases the
albedo error term. Figure 6 presents the same series of plots
as Fig. 3 after correction. These graphs show a much weaker
relationship between AOD errors and surface albedo proper-
ties.
As applied, the correction increases the range of surface
conditions over which errors are small; however, problems
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 393
Strong ThresholdWeak ThresholdAll Valid AlbedoNo Valid Albedo
No Valid Albedo
-0.20 -0.15 -0.10 -0.05 -0.02 0.02 0.05 0.10 0.15 0.20Magnitude of Albedo Correction to AOT
0.0 0.1 0.2 0.32100nm MODIS Albedo
0.00
0.05
0.10
0.15
0.20
650n
m MO
DIS A
lbedo
12BONDVILLE(-0.05)
34CARTEL(-0.06)
54Cart_Site(-0.05)
77Egbert(-0.04)
93HJAndrews(-0.02) 94Halifax(-0.02)
115KONZA_EDC(-0.02)
172OK_St_Univ(-0.09)
245Trinidad_Head(-0.04)
Boulder(+0.07)
39COVE(+0.05)
84Frenchman_Flat(+0.29)
85Fresno(+0.01) 99Hermosillo(+0.03)
129La_Jolla(+0.01)
163NASA_LaRC(+0.03)
193Railroad_Valley(+0.21)
197Richland(+0.04)
200Rogers_Dry_Lake(+0.16)
214Sevilleta(+0.12)
249UCLA(+0.04)
A B
C
Fig. 5. Albedo correction and filtering and effects on MODIS AOD. (a) Surface albedo and AOD bias for individual sites. Symbols
indicate the mean surface albedo for each AERONET site, with colors indicating the mean bias in AOD for τA < 0.2 at that site. Contours
behind indicate the estimated AOD correction based on the surface albedo (Eq. 7). The dashed line indicates the relationship Albedo
(0.65 µm) = Albedo(2.1 µm)/2. The symbols are placed to indicate the mean albedo for each AERONET site for the matched MODIS-
AERONET dataset. The colors of the symbols indicate the mean bias of τM at each site, on a scale that is the reverse of the contours.
Sites marked with “+”, are labeled with the name of the site and the mean bias of τM for the matched data at that site. (b) Example of
the estimated albedo correction calculated using Eq. (7), using the 16-day MODIS albedo product for days 177–193 of 2008. (c) Effect on
geographic coverage of albedo filtering of MODIS AOD product. The map above illustrates 3 zones based on the MODIS albedo product
from days 177–193 of 2008. The dark blue area highlights regions where the surface albedo falls within the “strong” limits shown in Fig. 11.
The light green area shows regions that fall outside the “strong” limits (50% or more of cases) but within the “weak” limits of Fig. 11. The
gray area shows area where valid albedo data were available, but albedo fell outside the “weak” constraints of Fig. 11 (50% or more of cases).
Unshaded areas had no valid albedo for the dates shown.
persist at high albedo. Exclusion of certain retrievals based
on surface properties is necessary even after correction. The
graphs in Figs. 3 and 6 show two cutoffs for albedo val-
ues, shown as vertical lines. Based on the error regimes,
it is clear that the retrieval works best in the darkest back-
grounds to the left of the dashed line (0.06 µm, 0.11 µm,
0.25 µm, and 0.50 for 0.47, 0.66, 2.1 and A0.66 µm/A2.12 µm,
respectively). This “Strict” threshold (dashed lines, to the
left) reflects the surface conditions corresponding to the best
retrieval performance in the uncorrected product; the more
lax “Weak” threshold (dotted lines, to the right) provides sig-
nificantly higher coverage while still maintaining acceptable
error statistics in most regions. The geographic ramifications
of these thresholds are shown in Fig. 5c, for the same data
as shown in Fig. 5b: blue areas have at least 50% of albe-
dos within the “Strong” thresholds, green areas have at least
50% within the “Weak” thresholds, and gray areas indicate
places where albedo data are available, but fewer than 50%
fall within the thresholds of Figs. 4 and 6. Note that gray
areas do not indicate no usable retrievals; areas in green and
gray lose more than 50% of data volume when albedo thresh-
olds are applied.
The albedo filtering and correction described here, as well
as the snow filtering described in Sect. 3, depend on datasets
not available in a timely fashion for operational use. Ap-
pendix C to this paper discusses the creation and evaluation
of an alternative approach to filtering and correction using an
8-year data record of MODIS snow and albedo data.
5.2 Residual microphysical bias
To support bias attribution, we examined each site’s
AERONET derived fine/coarse partition as well as regional
surface albedo properties. The noise floor statistics, included
for each site in the Supplement, are often indicative of issues
with the lower boundary condition. Figure 7 shows the r2
versus slope for the regression of all τM (with albedo correc-
tion applied) to τA pairs for the range 0.2<τM < 1.4. Terra
and Aqua MODIS are in red and blue, respectively. Slopes
are largely determined by the higher AOD values, which
should be more sensitive to microphysical bias. Within some
regions, the variability in estimated slope can be large, even
for sites with strong correlations to AERONET observa-
tions. This suggests spatially and temporally correlated bias
in AOD at scales finer than our regional analysis can resolve.
However, for many regions, the slope statistics for individ-
ual sites are clustered around values significantly different
from 1.0. It is these broad slope biases that we hope to quan-
tify and, to the extent possible, correct.
Once we make our albedo correction as described in
Sect. 5.1, we must remove the residual slope bias. How-
ever, as shown in Fig. 7, there are significant site by site
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 397
0.00 0.10 0.20 0.30 0.40 0.50 0.70 0.80 0.90 0.60 1.00Mean MODIS AOD (MODIS-Terra)
JJA: BASE
-0.10 -0.06 -0.02 +0.02 +0.06 +0.08 -0.04 -0.08 +0.04 +0.10Difference of AOD, NEW-BASE
JJA: NEW - BASE
Fig. 9. Mean AOD for 1-◦ grid cells. The upper map shows the ab-
solute mean AOD for the BASE scenario of basic QA filtering and
no corrections applied. The lower map indicates the relative effect
on seasonal mean AOD from the filtering and correction applied for
the NEW scenario. Data used were from MODIS-Terra for the pe-
riod June–August 2008 (results for other seasons are included in the
Supplement).
6.2 Effects of specific filters and corrections on AOD
error and data volume
Any exclusion of data from the assimilation system must
consider the balance of data volume and coverage and data
errors. This section addresses the effects of specific filter-
ing steps and presents the evidence for the efficacy of these
steps. This analysis uses the error statistics from compari-
son between gridded MODIS data and AERONET data. The
compliance statistics before and after filtering are used to cal-
culate the “effective compliance” of the excluded retrievals,
which combines the effect of filtering on errors and data vol-
ume in a single metric. Table 5 shows these statistics for
a cumulative series of filters and corrections. Note that for
generation of the gridded MODIS AOD product, the textural
filters are applied last, after all other filtering steps.
Textural filters
The textural filters impose a cost in gridded data volume
between 10% (0.2<τM < 0.6) and 6% (τM < 0.2). After
Ratio of DATADAYS NEW/BASE 0.00 0.10 0.20 0.30 0.40 0.50 0.70 0.80 0.90 0.60 1.00
0 20 30 40 60 80 90 10 50 70 100Days with Data
JJA: BASE
JJA: NEW / BASE
Fig. 10. Number of days with data for 1-◦ grid cells. The upper
map shows the absolute number of days with data in each season
for the BASE scenario of basic QA filtering and no corrections
applied. The lower map indicates the relative effect on data fre-
quency from the filtering and correction applied for the NEW sce-
nario. Data used were from both MODIS instruments for the period
June–August 2008.
filtering, compliance in all AOD ranges is improved. Tex-
tural filtering has little effect on the distribution of positive
and negative errors and is nearly neutral for τM < 0.2, but
removes substantial positive biases at higher AOD ranges.
Quality assurance filtering
Quality assurance filtering includes both the exclusion of
MODIS retrievals with mandatory QA values other than
“Very Good” as well as the exclusion of retrievals with scat-
tering angles above 170◦ (see Sect. 3.1). These filters, ap-
plied after textural filtering, result in exclusion of gridded
data ranging from 28% of τM > 1.4 to 50% of τM 0.2–0.6.
For τM < 0.2, the excluded data have overall compliance only
slightly worse than the unfiltered dataset, but exhibit a strong
surplus of positive errors. For higher AOD ranges, compli-
ance is poor and positive errors dominate in the excluded
data.
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

398 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
Table 4. Effect of filtering and correction steps on overall compliance and correlation in the matched MODIS-AERONET data set. Slope
correction in parentheses for South America was applied only to data with τM > 1.4 (see Fig. 7p).
Fraction Compliant (r2)
Region Satellite Slope Correction All Data Basic QA Snow/Albedo Filtering Albedo Correction Slope Correction
Global Terra 1 62% (0.54) 67% (0.63) 67% (0.63) 75% (0.65) 77% (0.73)
Aqua 1 62% (0.51) 69% (0.59) 69% (0.60) 76% (0.62) 78% (0.71)
N. American Boreal Terra 1.15 53% (0.58) 72% (0.64) 74% (0.66) 82% (0.68) 84% (0.73)
Aqua 1.25 54% (0.59) 77% (0.59) 80% (0.70) 84% (0.73) 88% (0.72)
E. CONUS Terra 1.05 70% (0.63) 70% (0.76) 70% (0.73) 83% (0.77) 84% (0.76)
Aqua 1.05 70% (0.58) 71% (0.73) 72% (0.71) 84% (0.76) 85% (0.75)
W. CONUS Terra 1.25 50% (0.10) 63% (0.26) 65% (0.28) 72% (0.33) 74% (0.32)
Aqua 1.25 53% (0.11) 66% (0.25) 68% (0.26) 75% (0.31) 77% (0.30)
Central America Terra 0.9 57% (0.36) 44% (0.58) 45% (0.57) 64% (0.62) 63% (0.56)
Aqua 1 54% (0.37) 50% (0.51) 50% (0.46) 59% (0.50) 59% (0.50)
South America Terra 1 (1.35) 43% (0.81) 35% (0.80) 37% (0.81) 56% (0.82) 57% (0.83)
Aqua 1 (1.35) 49% (0.79) 39% (0.82) 40% (0.80) 58% (0.81) 59% (0.82)
S. South America Terra 1.05 50% (0.52) 48% (0.70) 49% (0.69) 66% (0.70) 66% (0.70)
Aqua 1.1 50% (0.40) 50% (0.62) 50% (0.62) 64% (0.63) 65% (0.70)
Africa below equator Terra 0.9 68% (0.50) 71% (0.58) 72% (0.56) 82% (0.63) 80% (0.63)
Aqua 0.95 69% (0.49) 73% (0.49) 73% (0.51) 83% (0.56) 83% (0.56)
Equatorial Africa Terra 1 68% (0.53) 62% (0.53) 61% (0.56) 76% (0.62) 76% (0.62)
Aqua 1.1 69% (0.42) 68% (0.35) 68% (0.43) 76% (0.46) 78% (0.45)
Africa above equator Terra 0.7 47% (0.59) 45% (0.69) 45% (0.67) 42% (0.68) 54% (0.61)
Aqua 0.7 48% (0.57) 48% (0.67) 48% (0.65) 44% (0.66) 53% (0.57)
Europe–Mediterranean Terra 1 73% (0.38) 76% (0.48) 76% (0.48) 82% (0.51) 82% (0.51)
Aqua 1 70% (0.37) 76% (0.45) 76% (0.45) 80% (0.49) 80% (0.49)
Eurasian Boreal Terra 1.05 77% (0.63) 76% (0.65) 77% (0.65) 85% (0.67) 86% (0.67)
Aqua 1.15 76% (0.65) 77% (0.68) 77% (0.68) 84% (0.70) 87% (0.72)
East Asia Mid-Latitudes Terra 1 61% (0.62) 66% (0.65) 66% (0.65) 73% (0.67) 73% (0.67)
Aqua 1.05 58% (0.58) 64% (0.66) 65% (0.66) 71% (0.68) 73% (0.69)
Peninsular SE Asia Terra 0.9 52% (0.60) 52% (0.63) 52% (0.63) 58% (0.65) 58% (0.64)
Aqua 0.9 54% (0.57) 56% (0.58) 57% (0.61) 62% (0.64) 60% (0.62)
Indian Subcontinent Terra 1 74% (0.64) 81% (0.74) 81% (0.74) 82% (0.73) 82% (0.73)
Aqua 1 66% (0.59) 72% (0.68) 73% (0.67) 73% (0.68) 73% (0.68)
Australian Continent Terra 0.95 66% (0.08) 68% (0.20) 68% (0.14) 75% (0.16) 75% (0.16)
Aqua 1.05 65% (0.10) 69% (0.18) 70% (0.14) 76% (0.10) 76% (0.09)
Exclusion of partially cloudy retrievals
Because this filter is applied at the retrieval level, it will both
reduce the data volume of the gridded dataset, and also mod-
ify the AOD values in grids with both cloudy and cloud-free
retrievals. Exclusion of partially cloudy retrievals imposes
an overall data cost of 7–9% of gridded data in the matched
dataset, but the spatial distribution of these data are highly
non-uniform. Figure 11 shows the impact of cloud exclusion
from the BASE data set in terms of change in seasonal mean
gridded AOD for Terra and Aqua.
Retrievals excluded by application of this filter have vari-
able compliance, and a large surplus of positive AOD er-
rors. Application of this filter does not change the overall
compliant fraction, but shifts the balance of errors from pos-
itive to negative.
Albedo filters
QA filtering, albedo filtering has the most dramatic effects
on data volume among the suite of filters tested in this study.
The “Weak” albedo filter (matched) reduces the data volume
by 7% for AOD< 0.2, and as much as 13% for AOD> 1.4.
“Weak” albedo filtering results in large improvements to
compliance of the gridded product. Excluded retrievals have
relatively high compliance, but errors are overwhelmingly
biased positive. Application of the “weak” albedo filter im-
proves compliance at all AOD levels.
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 399
-0.05 -0.03 -0.01 +0.01 +0.03 +0.04 -0.02 -0.04 +0.02 +0.05Difference of AOD, BASE - (BASE+partially cloudy)
JJA: MODIS-Aqua
JJA: MODIS-Terra
Fig. 11. Effect of excluding retrievals with MODIS-detected cloud
cover on mean MODIS AOD. Maps show the difference between
the BASE scenario and a scenario which is identical but includes
retrievals with MODIS-detected cloud. Upper map shows data from
MODIS-Terra; lower map shows MODIS-Aqua. All maps use data
from June–August 2008.
The “Strong” albedo filters are extremely aggressive, and
result in a reduction of data volume by 27% for AOD< 0.2
and by up to 53% for AOD> 1.4 (compared with the “weak”
filters). This filter also has a definite spatial pattern of exclu-
sion, as illustrated by Fig. 5c, which leaves large areas with
little or no available AOD data. This extremely selective fil-
tering delivers some additional improvement in compliance,
but the quality of the excluded data is generally high. For
some highly demanding applications, this strict filtering may
be appropriate.
Effect of albedo and regional microphysical corrections
Corrections applied to the data have no effect on data volume,
although they can change the distribution of AOD values.
The albedo correction increases compliance for τM < 0.2
from 78% (80%) to 83% (84%) for MODIS-Terra (Aqua).
Compliance at higher optical depths is slightly improved as
well.
Regional slope corrections improve AOD for all but the
highest AOD values (τM > 1.4). Correlation between grid-
ded MODIS and AERONET AOD is also improved by the
regional slope correction, from r2 = 0.61 (0.60) to r2 = 0.65
(0.65) for MODIS-Terra (Aqua).
0.60 0.70 0.90 1.05 1.20 1.30 0.80 0.95 1.10 1.40AOD error ratio NEW/BASE
JJA: NEW / BASE 0.00 0.10 0.15 0.05 0.20 0.25Mean Seasonal AOD Uncertainty (MODIS-Terra)
JJA: BASE
Fig. 12. Mean estimated uncertainty in AOD for 1-◦ grid cells.
The upper map shows the mean uncertainty estimated for the BASE
scenario using the prognostic error model. The lower map shows
the ratio of the mean uncertainty for the NEW scenario to the BASE
scenario. Data used were from MODIS-Terra for the period June–
August 2008.
6.3 Prognostic error model for gridded product
Uncertainty in the gridded MODIS AOD data is estimated
using Eq. (5) (Sect. 2.4). For each region, a linear estimate of
RMSE as a function of τM and a “noise floor” RMSE used as
a minimum value were calculated. Table 6 shows the param-
eters of the uncertainty estimation calculation for each region
with sufficient data volume, for the RAW, BASE, and NEW
scenarios. Note that while Fig. 3 shows a bi-linear estimation
to account for the different error characteristics of very high
AOD values, the uncertainty estimates for the gridded prod-
uct use only a single linear estimate, because data volumes
are insufficient for a robust calculation at high AOD. Re-
gional estimates of the “noise floor” RMSE are made for all
cases having at least 100 points with τM < 0.2 in the matched
aggregated data set. Estimates of the linear relationship are
made only for cases with at least 100 points with τM > 0.2.
The parameters in Table 6 can be used to estimate the un-
certainty (ε) of any gridded MODIS AOD. Figure 12 shows
maps, based on the same data as Figs. 9–10, representing the
mean estimated uncertainty for the BASE scenario, and the
fractional difference in uncertainty for the NEW scenario.
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

400 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
Table 5. Effects of filtering and correction steps on compliance and data volume of gridded MODIS AOD product. For each step, compliance
statistics are given before and after. For filtering steps, the before and after compliance is used to calculated an effective compliance of
excluded retrievals.
All Data Context QA Filters Cloud = 0 “Weak” Albedo Slope “Strong”
(RAW) Filtering (BASE) Albedo Correction Correction Albedo
Filter (NEW) Filter
(STRONG)
TERRA
<0.2 Compliance 12/75/12 11/76/12 13/79/07 15/78/06 15/78/05 10/83/06 09/83/06 08/86/04
Excluded 26/49/24 03/69/26 −06/86/19 04/75/19 13/76/09
Fraction 100% 94.40% 70.10% 66.20% 62.50% 46.10%
0.2–0.6 Compliance 08/52/39 08/54/36 10/64/25 10/66/22 11/68/19 10/70/18 07/74/17 08/77/13
Excluded 03/30/66 06/43/49 06/52/41 01/46/51 12/58/28
Fraction 100% 91.50% 47.90% 42.00% 37.90% 26.40%
0.6–1.4 Compliance 09/51/39 09/53/36 09/59/31 10/59/29 11/61/26 11/61/26 07/64/28 09/70/19
Excluded 05/29/64 09/46/44 04/50/45 03/52/44 11/53/34
Fraction 100% 90.40% 51.80% 46.90% 39.40% 24.00%
>1.4 Compliance 06/46/47 06/47/46 09/46/43 09/46/44 08/47/43 08/47/43 09/43/46 14/44/40
Excluded 05/35/58 00/49/50 19/46/34 12/36/51 12/48/39
Fraction 100% 93.60% 56.40% 51.50% 46.60% 22.80%
All Compliance 10/66/23 10/68/21 12/74/13 13/74/11 14/75/10 10/79/10 09/80/10 08/83/07
Excluded 15/40/44 05/56/38 00/70/30 03/63/32 13/69/17
Fraction 100% 93.2% 61.7% 57.1% 53.0% 38.2%
AQUA
<0.2 Compliance 10/76/13 09/77/12 12/80/07 13/80/06 14/80/05 09/84/06 08/84/06 07/87/04
Excluded 21/53/24 02/71/25 −02/81/21 00/76/23 13/76/09
Fraction 100% 94.00% 66.30% 61.60% 58.20% 43.40%
0.2–0.6 Compliance 06/48/44 07/50/42 08/62/29 09/64/26 10/66/23 09/68/21 08/72/19 08/76/15
Excluded 03/29/67 05/39/54 01/49/48 02/41/55 07/62/30
Fraction 100% 90.30% 42.80% 36.80% 33.40% 22.80%
0.6–1.4 Compliance 08/47/44 08/49/42 09/55/35 09/56/34 10/58/30 10/59/30 07/63/28 09/69/20
Excluded 04/34/61 07/43/48 10/47/42 01/44/53 05/57/36
Fraction 100% 89.10% 44.00% 38.80% 31.60% 17.20%
>1.4 Compliance 05/41/53 04/43/51 05/49/45 05/50/44 05/53/41 05/53/41 07/52/39 09/58/31
Excluded 05/16/77 04/37/58 09/40/50 03/36/60 05/48/45
Fraction 100% 92.50% 50.80% 44.20% 37.30% 17.10%
All Compliance 08/64/26 08/66/25 11/74/14 12/74/12 13/76/10 09/79/11 08/80/10 07/84/07
Excluded 12/40/47 04/53/41 00/65/33 01/60/38 10/71/18
Fraction 100% 92.4% 56.4% 51.2% 47.6% 34.2%
The filtering and correction described in this study re-
duces the estimated uncertainty in the MODIS AOD over
nearly the entire globe. For τM = 0.2, estimated uncertainty
is lower in all regions except for Australia (εBASE = 0.064,
εNEW = 0.072). For τM = 1.0, ε is lower in all regions except
South America, Southern Europe/Mediterranean, Eurasian
Boreal, and Peninsular SE Asia. In each of those cases, the
difference between εBASE and εNEW is less than 0.05.
7 Conclusions
This study presents an examination of the MODIS Collec-
tion 5 AOD retrieval over-land based on comparison with
AERONET Sun Photometer AOD for 4 years of data (2005–
2008). The results presented here give a more detailed pic-
ture of the global and regional structure of errors in the re-
trieval than any previous study. Key findings are:
1. Global RMSE: The findings of our global data anal-
ysis are mostly consistent with previous assessments.
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 401
Table 6. Prognostic RMS error model based on comparison of L3 gridded MODIS AOD to AERONET. Model has the form εEST = MAX(A,
B + C τM) (Eq. 4). “N/A” indicates regions where data volume was insufficient to calculate model parameters. Error for these regions is
estimated using the Global parameters.
Global NA Boreal E. CONUS W. CONUS Cent. Am. Central SA South SA Australia
Terra RAW 0.11|0.04 + 0.24 t 0.11|0.06 + 0.33 t 0.08|0.01 + 0.27 t 0.15| − 0.02 + 0.73 t 0.08|0.09 + 0.07 t 0.07|0.03 + 0.17 t 0.09|0.15 + 0.05 t 0.07| − 0.03 + 0.54 t
BASE 0.07|0.03 + 0.21 t 0.09|0.09 + 0.09 t 0.05|0.01 + 0.20 t 0.10|0.03 + 0.46 t 0.08|0.07 + 0.16 t 0.07|0.03 + 0.14 t 0.07|0.16± 0.03 t 0.04|0.03 + 0.17 t
NEW 0.06|0.02 + 0.20 t 0.08|0.07 + 0.06 t 0.04|0.01 + 0.17 t 0.07|0.07 + 0.21 t 0.06|0.06 + 0.11 t 0.05|0.02 + 0.16 t 0.05|0.11 + 0.02 t 0.04|0.05 + 0.11 t
Aqua RAW 0.11|0.04 + 0.25 t 0.10|0.07 + 0.29 t 0.08|0.02 + 0.25 t 0.15| − 0.03 + 0.74 t 0.08|0.08 + 0.10 t 0.08|0.04 + 0.16 t 0.10|0.09 + 0.23 t 0.08| − 0.07 + 0.71 t
BASE 0.07|0.03 + 0.22 t 0.06|0.03 + 0.28 t 0.06|0.02 + 0.17 t 0.10|0.03 + 0.47 t 0.07|0.03 + 0.18 t 0.07|0.03 + 0.13 t 0.08|0.10 + 0.16 t 0.05|0.00 + 0.39 t
NEW 0.06|0.03 + 0.19 t 0.05|0.03 + 0.15 t 0.05|0.02 + 0.18 t 0.06| − 0.01 + 0.51 t 0.06|0.03 + 0.18 t 0.06|0.01 + 0.16 t 0.06|0.05 + 0.27 t 0.04|0.00 + 0.34 t
S. Africa Eq. Africa N. Africa S. Europe Eurasian Boreal East Asia Peninsular Indian
Mid-Latitudes SE Asia Subcontinent
Terra RAW 0.06|0.00 + 0.28 t 0.07|0.00 + 0.26 t 0.12|0.08 + 0.20 t 0.08| − 0.02 + 0.37 t 0.08|0.01 + 0.26 t 0.16|0.04 + 0.24 t 0.17|0.07 + 0.19 t 0.25|0.03 + 0.22 t
BASE 0.05|0.02+0.20t N/A 0.08|0.11 + 0.14 t 0.06|0.00 + 0.26 t 0.05|0.02 + 0.13 t 0.13|0.02 + 0.25 t 0.11|0.06 + 0.14 t 0.22|0.04 + 0.17 t
NEW 0.05|0.03 + 0.11 t N/A 0.09|0.05 + 0.15 t 0.05| − 0.01 + 0.28 t 0.04|0.02 + 0.15 t 0.09|0.02 + 0.24 t 0.08|0.02 + 0.21 t 0.21|0.04 + 0.15 t
Aqua RAW 0.07| − 0.02 + 0.39 t 0.06| − 0.02 + 0.30 t 0.13|0.08 + 0.20 t 0.09| − 0.02 + 0.39 t 0.08|0.01 + 0.27 t 0.17|0.04 + 0.25 t 0.20|0.07 + 0.18 t 0.27|0.03 + 0.25 t
BASE 0.06|0.04 + 0.16 t 0.05| − 0.03 + 0.27 t 0.07|0.08 + 0.21 t 0.06| − 0.01 + 0.29 t 0.05|0.02 + 0.13 t 0.17|0.03 + 0.24 t 0.11|0.08 + 0.10 t 0.22|0.03 + 0.19 t
NEW 0.05|0.01 + 0.21 t 0.05| − 0.02 + 0.26 t 0.09|0.07 + 0.14 t 0.06| − 0.01 + 0.28 t 0.04|0.03 + 0.11 t 0.10|0.04 + 0.19 t 0.09|0.05 + 0.13 t 0.20|0.04 + 0.17 t
Globally, one standard deviation of “Very Good” data
falls within the (0.05 + 0.2× τA) error thresholds, and
we concur with the MODIS science team recommen-
dation that only “Very Good” data be used (Table 1a).
However, for most applications a prognostic RMS er-
ror model with a noise floor is more appropriate. For
global applications using MODIS Level 2 data over
land, we recommend the use of the greater of 0.08 or
0.02 + 0.22× τM for Terra and 0.07 or 0.01 + 0.26× τM
for Aqua (Table 1c).
2. Global noise issues: The amount of scatter in the
MODIS-AERONET comparison was sensitive to ob-
serving conditions, especially viewing geometry. Ow-
ing to an increase in optical path length and pixel size at
larger scan angles, MODIS products have higher com-
pliance at higher scan angles (Fig. 3). Conversely, er-
rors are largest at nadir owing to shorter optical path
length and perhaps increases in BRDF gap probabili-
ties. One consequence of this is that comparisons be-
tween MODIS and MISR are not necessarily indicative
of behavior of the MODIS product at all scan angles.
Scattering angle does not appear to affect retrieval error
except for scattering angles >170◦ owing to the hotspot
effect (Fig. 3).
3. Global cloud bias: The MODIS retrieval is highly
clear sky conservative with very few retrievals with
“Very Good” QA and MODIS-detected cloud fraction
>15%. Even so, retrievals with non-zero MODIS-
detected cloud fractions still have perceptible high bi-
ases (Fig. 3). Elimination of partially cloudy retrievals
reduces mean seasonal AOD by as much as 0.05 over
large regions of the tropics (Fig. 11).
4. Global snow bias: While Collection 5 has a much im-
proved snow filter, we still find periodic positive per-
turbations at high latitudes. By using a spatially and
temporally extended snow filter based on the snow flag
in the MODIS MCD43 albedo product, we can reduce
the incidence of positive biased AOD in northern lati-
tudes (Table 2).
5. Regional variability: Despite good overall global statis-
tics, MODIS products have widely varying regional
efficacy. Widely diverging statistics for adjacent
AERONET sites suggest spatially and temporally cor-
related bias. Regions such as the Eastern CONUS and
Europe perform best. Sahelian Africa shows the poor-
est performance. Sites in East Asia have highly mixed
efficacy. Urban located sites also tend to have poor ef-
ficacy. Regions which experience intermittent smoke
events show significant slope bias on a per event basis,
which is manifested in highly variable (but uniformly
positive) slope bias for sites in South America and bo-
real North America. This regional variation in retrieved
AOD has been seen in other AOD products; compari-
son of these products has been used to identify regions
and seasons where more detailed observations, in situ or
otherwise, might be necessary to understand the behav-
ior of the satellite retrievals.
6. Albedo correction: Some bias at low AOD can be
corrected based on the use of surface albedo maps.
An empirical relationship between surface albedo at
0.66 µm and 2.1 µm and MODIS-AERONET differ-
ences in AOD explained more than 20% of the variance
in those differences (Eq. 7). This correction was shown
to be robust using geographically independent data
(Sect. 5.1), and improves compliance of MODIS AOD
by around 8% (Table 4).
7. Regional slope correction: A clear, if noisy, relationship
between AERONET estimated fine/coarse partitioning
and MODIS slope bias was identified (Fig. 8), and used
as the basis of a regional correction. This correction
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

402 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
improved the compliance of the MODIS AOD in many
regions, and increase the global correlation between
MODIS-Terra(-Aqua) and AERONET AOD from 0.65
(0.62) to 0.73 (0.71), by eliminating systematic biases
between regions.
8. Level 3 products: Based on various aggregation and
empirical correction schemes, operational and research
grade Level 3 products were generated. Basic QA
screening reduces data volume by 45–50%, while re-
ducing the fraction of non-compliant data by 50%.
Screening and correction for snow, albedo, and regional
slope bias, as described in this study, reduces data vol-
ume by a further ∼10% while improving the compliant
fraction to 80% for both Terra and Aqua (Table 5).
9. Bias between Terra and Aqua: Lastly, using these prod-
ucts, small but perceivable time dependent differences
between Terra and Aqua (globally on the order of 0.02)
are visible in statistical analysis of the Level 2 products,
but are within the uncertainty of the products. These
differences can be seen more clearly using the Level 3
products with extensive averaging (see Appendix D).
The regional uniformity of this difference suggests a
shortcoming in radiometric calibration of the two instru-
ments, which must be resolved or at least corrected to
perform a long-baseline study of global aerosol (Zhang
and Reid, 2010).
Taken together, the MODIS Collection 5 aerosol product,
with the QA procedures laid out in this study, can be used
to produce a product with desirable qualities for data assim-
ilation: low systematic bias and random error and relatively
well-characterized residual uncertainties. The next step is
testing of this product in the NAVDAS-AOD data assimi-
lation system, which is currently underway. The MODIS
aerosol science team is also currently producing an updated
version of the MODIS AOD retrieval (Collection 6) (L. Re-
mer, personal communication, 2010), which will include sig-
nificant changes that hopefully will address some of the bi-
ases identified in this study. The lessons from this data eval-
uation exercise can also be applied to data from other sensors
both current and future.
Appendix A
Sampling considerations for matching MODIS 10 km
AOD retrievals to AERONET
Representativeness error is an unavoidable consideration for
comparison of datasets with different sampling properties. In
the case of AERONET and MODIS AOD, the former is a
direct-sun measurement representing atmospheric conditions
along a line between the sensor and the sun, while the latter
is a product made using data covering an area 10 km square
on the ground (although much larger away from nadir), but
with limited sampling within that footprint due to pixels re-
jected in the retrieval process. Nominally, one could specify
a distance between the nominal center of the MODIS AOD
retrieval and the AERONET station that would ensure that
the AERONET station fell within the ground footprint of the
MODIS retrieval, but this would not ensure a spatial match,
and depending on the sun angle, the atmospheric column
sampled could still be completely different for the two sen-
sors.
Rather than attempt to capture all of these factors in build-
ing a comparison dataset, they are instead rolled into ran-
dom error which indicates the precision with which indi-
vidual retrievals can be analyzed in the comparison dataset.
Of greater concern are the systematic biases in a compar-
ison dataset caused by the interaction between the scales
of the observations and the scales of the phenomena ob-
served. Over land, these biases can potentially be large as
fine plumes near sources will often have characteristic di-
mensions smaller than 10 km. The overall impact of these
biases is examined in this section.
Figure A1 shows 4 compliance plots, indicating the
change in mean bias and spread between MODIS and
AERONET as a function of spatial and temporal separation
of the observations. Effects of spatial separation are evalu-
ated for 3 different ranges of τA. In all cases, a slight decrease
in τM is seen with distance from the AERONET station. This
reflects the location of AERONET stations; while system-
atic attempts are made to place these stations in locations
at some remove from local sources, the requirements of ac-
cessibility compromise those attempts to a small degree. At
moderate AOD values (Fig. A1b), the bias crosses over from
positive to negative at a distance of approximately 15 km; at
high τA (Fig. A1c)), the bias is negative at all separations.
AERONET retrievals with high AOD (Fig. A1c) are often
indicative of near-source features; the strong gradients asso-
ciated with these features are expected to result in higher rep-
resentativeness error and negative bias of the large-footprint
MODIS AOD compared to the AERONET AOD.
For this study, a maximum distance threshold of 30 km
between the nominal center of the MODIS retrieval and the
AERONET station is used. This threshold balances increas-
ing random variability with distance, and the need for as
many comparison data points as possible.
Figure A1d shows bias and compliance as a function of
temporal separation, calculated using only matched retrievals
within 10 km of AERONET stations. For a time period of
one hour before or after the AERONET observation, only
very slight changes in bias and compliance can be observed.
For this study, MODIS and AERONET retrievals were paired
with a maximum delay of ±30 min.
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

404 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
“high-AOD” case was 0.006 lower than for the “low-AOD”
case. For 89% of cases, the absolute difference in the calcu-
lated corrections was less than 0.02. The more positive cor-
rection corresponded to the low-AOD case in 67% of cases.
These results suggest that elevated AOD can result in a small
bias to retrieved albedo, but this effect is generally within the
range of uncertainty of the MODIS albedo. The interaction
with the albedo correction to MODIS AOD calculated in this
study is slight, and very rarely larger than the noise level of
the MODIS AOD.
Appendix C
Use of climatological data for snow and albedo
filtering and albedo correction
The primary purpose of our work with the MODIS AOD
product is the improvement of the NAVDAS-AOD opera-
tional aerosol data assimilation system. The demands of
this system emphasize removal of outliers and quantitative
characterization of observation errors. The error correction
demonstrated in Sect. 5 expands the range of surface condi-
tions over which MODIS can retrieve albedo within accept-
able error limits (compare Figs. 3 and 6). The statistics in
Sect. 4.2 indicate that with the albedo correction in place, we
can use the less stringent (“weak”) albedo thresholds. How-
ever, this correction is dependent on the MODIS MCD43
albedo product, which is not available in a sufficiently timely
fashion for operational use. Thus, the initial operational ver-
sion of the MODIS AOD data for data assimilation will need
an alternative means of surface albedo correction. In addi-
tion, the snow detection algorithm described in Sect. 3 above
also relies on the MODIS albedo product, and so will also re-
quire a climatological substitute method for operational im-
plementation.
In order to develop a filtering and correction method
usable for real-time applications, we used the MODIS
MCD43C3 data record from 2000–2007. Because of the
8-day production schedule for the 16-day product, this pe-
riod yields as many as 16 observations for each 0.05◦ grid
cell. This albedo climatology included data points matching
98% of the matched MODIS-AERONET dataset for 2008.
Note that the occurrence of positive non-compliant errors in
the retrievals with no climatological albedo data is very high
(89%). This is because nearly all of the locations without any
successful albedo retrievals during 2000–2007 are in perma-
nently snow- and ice-covered regions (see Appendix B).
Our potential snow climatology was constructed by flag-
ging all locations and dates where 2000–2007 data indicated
snow cover at any time. The matched dataset indicated snow
contamination in less than 0.1% of 2007 MODIS AOD re-
trievals (Table 2). The climatological snow filter captures
more than 99% of these retrievals, while reducing overall
2007 data volume by 4%. The fraction of retrievals flagged
Table C1. Analysis of snow filter based on MODIS MCD43 clima-
tology. Compare to Table 2.
Snow-Climatology
Bias Compliance RMSE %data
N. American Boreal 0.040 05/56/38 0.10 7.34
E. CONUS 0.010 04/83/12 0.05 9.40
W. CONUS 0.046 01/63/34 0.08 3.97
Europe–Mediterranean −0.013 12/80/06 0.06 2.39
Eurasian Boreal −0.001 06/88/05 0.06 5.85
East Asia Mid-Latitudes 0.040 05/64/30 0.09 5.46
by the climatological filters with real snow contamination is
likely to be low, but the excluded data points still have a pos-
itive bias, and a poor correlation with AERONET, relative to
the entire dataset (see Table C1, and compare Table 2). Note
that some of the error statistics for retrievals flagged by the
climatological snow filter are better than the overall dataset.
This is because many areas susceptible to snow contamina-
tion are in darker, denser vegetation, where the AOD retrieval
is generally more accurate (in the absence of snow or other
disrupting factors). The climatological snow filter captures
fewer retrievals than the spatially and temporally extended
filter. This could indicate a more effective filter, but further
evaluation with the data assimilation system is needed to de-
termine how best to eliminate snow contamination without
excessive loss of data.
We used the albedo data to calculate a correction based
on all valid albedo measurements for each date from 2000–
2007. Because missing data are less common with the cli-
matological approach, the multi-year dataset can be used to
implement an additional consistency check without exces-
sive data loss. A filter was created to exclude retrievals
where fewer than three data points were used to calculate
the mean albedo correction, as well as retrievals where the
range of corrections exceeded 0.1. These two exclusions
together accounted for 7% of data volume in the matched
dataset. With this filter applied, the climatological correction
was within 0.03 of the matched correction for 98% of cases.
With the quality filter, the climatology also delivers a
slightly better correction. When the estimated correction is
larger than 0.05, the correction moves τM closer to τA in 73%
of cases, compared with 69% of cases for the matched albedo
correction.
Any locations/dates where the climatology indicated
albedo above the threshold in 20% of more of cases were
excluded. The combination of the consistency checks ap-
plied for application of the albedo correction and the “weak”
albedo thresholds reduced data volume by 17%, compared
with 16% using the matched data. Data volume reduction
using the “strong” albedo thresholds was much greater.
The Level 3 product generated with this climatological
approach (CLIM scenario) has similar characteristics to the
NEW scenario made using matched snow/albedo data (see
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 405
Table C2. Prognostic error model for Level 3 product using climatological data for snow and albedo filters and albedo correction. Compare
to Table 6.
Global NA Boreal E. CONUS W. CONUS Cent. Am. N. SA S. SA Australia
Terra CLIM 0.06|0.02 + 0.19τ 0.05|0.06 + 0.10τ 0.04|0.00 + +0.22τ 0.07|0.07 + 0.20τ 0.06|0.07 + 0.11τ 0.05|0.02 + 0.16τ 0.066|0.11 + 0.01τ 0.04|0.05 + 0.11τ
Aqua CLIM 0.06|0.02 + 0.19τ 0.05|0.06 + 0.11τ 0.05|0.02 + 0.17τ 0.06|0.04 + 0.44τ 0.06|0.03 + 0.17τ 0.05|0.02 + 0.14τ 0.06|0.06 + 0.22τ 0.04| − 0.03 + 0.42τ
S. Africa Eq. Africa N. Africa S. Europe Eurasian Boreal East Asia Mid-Lat. Penin. SE Asia Indian
Subcontinent
Terra CLIM 0.05|0.02 + 0.15τ N/A 0.09|0.6 + 0.14τ 0.05| − 0.01 + 0.26τ 0.04|0.01 + 0.15τ 0.10|0.02 + 0.22τ 0.08|0.02 + 0.19τ 0.22|0.05 + 0.13τ
Aqua CLIM 0.05|0.02 + 0.15τ 0.05| − 0.01 + 0.20τ 0.09| − 0.01 + 0.29τ 0.06| − 0.01 + 0.29τ 0.04|0.03 + 0.20τ 0.10|0.03 + 0.20τ 0.09|0.06 + 0.11τ 0.22|0.04 + 0.17τ
DJF
MAM
JJA
SON
CLIM
0.60 0.70 0.90 1.05 1.20 1.30 0.80 0.95 1.10 1.40AOD error ratio vs. BASE
Fig. C1. Estimated uncertainty in CLIM scenario relative to BASE
scenario. Compare to Fig. 12. Rows indicate different seasons.
Data used were from MODIS-Terra for the period December 2007–
November 2008.
Sect. 7 for descriptions of Level 3 product scenarios). Ta-
ble C2 shows the prognostic error model derived for this
product, and Fig. C1 shows the seasonal mean estimated un-
certainty relative to the BASE scenario (compare to Fig. 12).
Remaining uncertainty for the CLIM scenario is generally
higher than the NEW scenario, but degradation from the
BASE scenario, as seen in Australia in Fig. 12, is absent from
the CLIM scenario.
Appendix D
Bias between terra and aqua MODIS retrieved AOD
The existence of two MODIS sensors, sampling late morn-
ing and early afternoon conditions, has led to the applica-
tion of MODIS data to study diurnal variation in important
Earth system processes. Important processes with significant
diurnal variability include fires (e.g., Giglio, 2007; Roberts
et al., 2009), cloud cover (e.g., Jin et al., 2009), and cloud
properties (e.g., Meskhidze et al., 2009). The interaction of
variation between aerosol sources and cloud properties is an
object of intense scientific interest, as it may shed light on
the interactions between aerosol particles and clouds that can
strongly affect radiation budgets in polluted regions.
Comparison of Terra and Aqua MODIS to examine diur-
nal cycles of Earth system processes is dependent on detec-
tion and correction of any artifacts of calibration between the
two sensors. Because the two sensors never overlap except
near the poles, and even there only with large differences in
viewing geometry, indirect methods must be used. There is a
large literature on these methods, and studies generally con-
clude that Terra and Aqua MODIS calibration in visible and
infrared bands is basically within the precision of the indirect
methods used, generally cited as on the order of 2% in top-
of-atmosphere reflectance (e.g., Wu et al., 2004). However,
the AOD retrieval is extremely sensitive to small differences
in reflectance under some conditions.
The analyses performed in this study, for virtually all cases
where Terra and Aqua results differ, have shown results con-
sistent with τM from MODIS-Aqua being slightly higher than
τM from MODIS-Terra. For instance, in every region, the
fraction of low-τA retrievals with negative errors below the
compliance threshold is higher or equal for MODIS-Terra
in every region except Mid-Latitude East Asia (Table 3).
The same is true for every QA level (Table 1). In general,
these differences could not pass tests of statistical signifi-
cance, and they are often beyond the least significant digit of
our results. However, the accumulation of multiple statisti-
cal results gives qualitative support to the idea that these two
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

406 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
Daily Mean of matched 1-D cells
Jan2005 Jan2006 Jan2007 Jan2008 Jan2009 Jan2010Date
-0.04-0.02
0.000.02
0.04
Aqua
-Terr
a AO
D
-0.05 -0.04 -0.03 -0.02 -0.01 0.01 0.02 0.03 0.04 0.05 0.00Difference of AOD, Aqua - Terra
JJA2
006
JJA2
007
JJA2006
JJA2007
Fig. D1. Comparison of Terra and Aqua using gridded data. All
data are from the BASE scenario, with QA filters applied and par-
tially cloudy retrievals excluded. Time series plot is based on 1-◦
grid cells with both Terra and Aqua data in a single day. The time
series shown usings a 32-day moving window to better show the
trend. Maps are comparisons using all seasonal averages for each
location using all data (not pairwise). Time ranges for the two maps
are highlighted in gray on the time series plot. All analyses are
based on the BASE scenario.
sensors have some systematic bias relative to each other. Be-
cause this bias is very small relative to the random variation
in AOD for individual retrievals, averaging of large numbers
of retrievals is the only way to see it clearly. The time se-
ries plot in Fig. D1a represents one such averaging: daily
AOD values from grid cells with both Terra and Aqua data
for a given date (BASE scenario) are averaged globally (pair-
wise), and then a 32-day moving average is used to smooth
the data for plotting. Thus, each point on this plot represents
a comparison made with more than 106 × 10 km retrievals.
With this averaging applied, it can be seen that bias between
Terra and Aqua τM is variable over the time series examined
and persistent on a scale of months.
The spatial variation averaged over in this time series plot
is shown in the maps in Fig. D1b–c, which show the differ-
ence in seasonal mean AOD (not pairwise) for two 3-month
periods one year apart, highlighted in gray on the time series
plot. Persistent regional variability is expected for this com-
parison, because many factors affecting both aerosol loading
and aerosol detection have diurnal cycles that vary among
regions. However, because of the year-over-year change in
overall differences, the regional signal for JJA2006 looks
very different from JJA2007.
These results do not establish the “true” bias between
Terra and Aqua MODIS AOD values. They do, how-
ever, identify a systematic variation between the sensors that
(a) does not follow a seasonal pattern and (b) disrupts the ob-
served spatial patterns for individual seasons. It is difficult
to prove conclusively that this difference results from cali-
bration differences between the instruments, but it may be
indicative of the magnitude of signal that can be reasonably
attributed to variations in instrument calibration.
Supplementary material related to this
article is available online at:
http://www.atmos-meas-tech.net/4/379/2011/
amt-4-379-2011-supplement.zip.
Acknowledgements. This work was supported by the Office of
Naval Research, Code 322. We are grateful to the all of the
investigators who participate in the AErosol RObotic NETwork for
the use of network wide data. E. J. Hyer would like to thank all the
members of the MODIS Aerosol team for helpful comments during
this research, as well as access to pre-publication results. E. J. Hyer
would also like to thank James Campbell and Elizabeth Reid for
helpful comments on the manuscript.
Edited by: S. Kinne
References
Daley, R. and Barker, E.: Navdas: Formulation and diagnostics,
Mon. Weather Rev., 129, 869–883, 2001.
Dee, D. P. and Da Silva, A. M.: Data assimilation in the presence
of forecast bias, Q. J. Roy. Meteor. Soc., 124, 269–295, 1998.
Drury, E., Jacob, D. J., Wang, J., Spurr, R. J. D., and Chance, K.:
Improved algorithm for modis satellite retrievals of aerosol opti-
cal depths over western north america, J. Geophys. Res.-Atmos.,
113, D16204 doi:10.1029/2007jd009573, 2008.
Eck, T. F., Holben, B. N., Reid, J. S., Dubovik, O., Smirnov,
A., O’Neill, N. T., Slutsker, I., and Kinne, S.: Wavelength de-
pendence of the optical depth of biomass burning, urban, and
desert dust aerosols, J. Geophys. Res.-Atmos., 104, 31333–
31349, 1999.
Eck, T. F., Holben, B. N., Reid, J. S., Sinyuk, A., Hyer, E. J.,
O’Neill, N. T., Shaw, G. E., Vande Castle, J. R., Chapin, F. S.,
Dubovik, O., Smirnov, A., Vermote, E., Schafer, J. S., Giles, D.
M., Slutsker, I., Sorokine, M., and Newcomb, W. W.: Optical
properties of boreal region biomass burning aerosols in alaska
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/

E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation 407
and transport of smoke to arctic regions, J. Geophys. Res., 114,
D11201, doi:10.1029/2008JD010870, 2009.
Ferek, R. J., Reid, J. S., Hobbs, P. V., Blake, D. R., and Liousse,
C.: Emission factors of hydrocarbons, halocarbons, trace gases
and particles from biomass burning in brazil, J. Geophys. Res.-
Atmos., 103, 32107–32118, 1998.
Gasso, S. and Stein, A. F.: Does dust from patagonia reach the
sub-antarctic atlantic ocean?, Geophys. Res. Lett., 34, L01801,
doi:10.1029/2006GL027693, 2007.
Giglio, L.: Characterization of the tropical diurnal fire cycle using
virs and modis observations, Remote Sens. Environ., 108, 407–
421, 2007.
Gupta, P. and Christopher, S. A.: Particulate matter air quality
assessment using integrated surface, satellite, and meteorologi-
cal products: 2. A neural network approach, J. Geophys. Res.-
Atmos., 114, D20205, doi:10.1029/2008jd011497, 2009.
Holben, B. N., Eck, T. F., Slutsker, I., Tanre, D., Buis, J. P., Setzer,
A., Vermote, E., Reagan, J. A., Kaufman, Y. J., Nakajima, T.,
Lavenu, F., Jankowiak, I., and Smirnov, A.: Aeronet – a federated
instrument network and data archive for aerosol characterization,
Remote Sens. Environ., 66, 1–16, 1998.
Hollingsworth, A., Engelen, R. J., Textor, C., Benedetti, A.,
Boucher, O., Chevallier, F., Dethof, A., Elbern, H., Eskes, H.,
Flemming, J., Granier, C., Kaiser, J. W., Morcrette, J. J., Rayner,
P., Peuch, V. H., Rouil, L., Schultz, M. G., Simmons, A. J., and
Consortium, G.: Toward a monitoring and forecasting system for
atmospheric composition: The gems project, B. Am. Meteorol.
Soc., 89, 1147–1164, 2008.
Jin, X., Wu, T. W., Li, L., and Shi, C. X.: Cloudiness character-
istics over southeast asia from satellite fy-2c and their compari-
son to three other cloud data sets, J. Geophys. Res.-Atmos., 114,
D17207, doi:10.1029/2008JD011422, 2009.
Kahn, R. A., Nelson, D. L., Garay, M. J., Levy, R. C., Bull, M. A.,
Diner, D. J., Martonchik, J. V., Paradise, S. R., Hansen, E. G.,
and Remer, L. A.: Misr aerosol product attributes and statistical
comparisons with modis, IEEE T. Geosci. Remote, 47, 4095–
4114, 2009.
Kaufman, Y. J., Tanre, D., Remer, L. A., Vermote, E. F., Chu, A.,
and Holben, B. N.: Operational remote sensing of tropospheric
aerosol over land from eos moderate resolution imaging spectro-
radiometer, J. Geophys. Res.-Atmos., 102, 17051–17067, 1997.
Koren, I., Remer, L. A., Kaufman, Y. J., Rudich, Y., and Martins, J.
V.: On the twilight zone between clouds and aerosols, Geophys.
Res. Lett., 34, L08805, doi:10.1029/2007GL029253, 2007.
Lary, D. J., Remer, L. A., MacNeill, D., Roscoe, B., and Par-
adise, S.: Machine learning and bias correction of modis
aerosol optical depth, IEEE Geosci. Remote S., 694–698,
doi:10.1109/lgrs.2009.2023605, 2009.
Levy, R. C., Remer, L. A., and Dubovik, O.: Global aerosol op-
tical properties and application to moderate resolution imaging
spectroradiometer aerosol retrieval over land, J. Geophys. Res.-
Atmos., 112, D13210, doi:10.1029/2006jd007815, 2007a.
Levy, R. C., Remer, L. A., Mattoo, S., Vermote, E. F., and Kauf-
man, Y. J.: Second-generation operational algorithm: Retrieval
of aerosol properties over land from inversion of moderate res-
olution imaging spectroradiometer spectral reflectance, J. Geo-
phys. Res.-Atmos., 112, D13211, doi:10.1029/2006JD007811,
2007b.
Levy, R. C., Remer, L. A., Kleidman, R. G., Mattoo, S., Ichoku,
C., Kahn, R., and Eck, T. F.: Global evaluation of the Collection
5 MODIS dark-target aerosol products over land, Atmos. Chem.
Phys., 10, 10399–10420, doi:10.5194/acp-10-10399-2010, 2010.
Li, R. R., Remer, L., Kaufman, Y. J., Mattoo, S., Gao, B. C., and
Vermote, E.: Snow and ice mask for the modis aerosol products,
IEEE Geosci. Remote S., 2, 306–310, 2005.
Lucht, W., Schaaf, C. B., and Strahler, A. H.: An algorithm for the
retrieval of albedo from space using semiempirical brdf models,
IEEE T. Geosci. Remote, 38, 977–998, 2000.
Meskhidze, N., Remer, L. A., Platnick, S., Negro´n Jua´rez, R., Licht-
enberger, A. M., and Aiyyer, A. R.: Exploring the differences in
cloud properties observed by the Terra and Aqua MODIS Sen-
sors, Atmos. Chem. Phys., 9, 3461–3475, doi:10.5194/acp-9-
3461-2009, 2009.
Misra, A., Jayaraman, A., and Ganguly, D.: Validation of modis de-
rived aerosol optical depth over western india, J. Geophys. Res.-
Atmos., 113, D04203, doi:10.1029/2007jd009075, 2008.
O’Neill, N. T., Eck, T. F., Holben, B. N., Smirnov, A., Royer, A.,
and Li, Z.: Optical properties of boreal forest fire smoke derived
from sun photometry, J. Geophys. Res.-Atmos., 107(D11), 4125,
doi:10.1029/2001JD000877, 2002.
O’Neill, N. T., Eck, T. F., Smirnov, A., Holben, B. N., and
Thulasiraman, S.: Spectral discrimination of coarse and fine
mode optical depth, J. Geophys. Res.-Atmos., 108(D11), 4559,
doi:10.1029/2002JD002975, 2003.
Radosavljevic, V., Vucetic, S., and Obradovic, Z.: A data-
mining technique for aerosol retrieval across multiple ac-
curacy measures, IEEE Geosci. Remote S., 7, 411–415,
doi:10.1109/lgrs.2009.2037720, 2010.
Reid, J. S., Eck, T. F., Christopher, S. A., Koppmann, R., Dubovik,
O., Eleuterio, D. P., Holben, B. N., Reid, E. A., and Zhang, J.:
A review of biomass burning emissions part III: intensive optical
properties of biomass burning particles, Atmos. Chem. Phys., 5,
827–849, doi:10.5194/acp-5-827-2005, 2005a.
Reid, J. S., Koppmann, R., Eck, T. F., and Eleuterio, D. P.: A review
of biomass burning emissions part II: intensive physical proper-
ties of biomass burning particles, Atmos. Chem. Phys., 5, 799–
825, doi:10.5194/acp-5-799-2005, 2005b.
Reid, J. S., Hyer, E. J., Prins, E. M., Westphal, D. M., Zhang, J.,
Wang, J., Christopher, S. A., Curtis, C. A., Schmidt, C. C., Eleu-
terio, D. P., Richardson, K. A., and Hoffman, J. P.: Global moni-
toring and forecasting of biomass-burning smoke: Description of
and lessons from the fire locating and modeling of burning emis-
sions (flambe) program, IEEE J. Sel. Top. Appl., 2, 144–162,
doi:10.1109/JSTARS.2009.2027443, 2009.
Remer, L., Tanre, D., and Kaufman, Y.: Algorithm for remote sens-
ing of tropospheric aerosol from modis: Collection 5, NASA
Goddard Space Flight Center, Greenbelt, MD, 87, 2006.
Remer, L. A., Wald, A. E., and Kaufman, Y. J.: Angular and sea-
sonal variation of spectral surface reflectance ratios: Implications
for the remote sensing of aerosol over land, IEEE T. Geosci. Re-
mote, 39, 275–283, 2001.
Remer, L. A., Kaufman, Y. J., Tanre, D., Mattoo, S., Chu, D. A.,
Martins, J. V., Li, R. R., Ichoku, C., Levy, R. C., Kleidman, R.
G., Eck, T. F., Vermote, E., and Holben, B. N.: The modis aerosol
algorithm, products, and validation, J. Atmos. Sci., 62, 947–973,
2005.
www.atmos-meas-tech.net/4/379/2011/ Atmos. Meas. Tech., 4, 379–408, 2011

408 E. J. Hyer et al.: An over-land aerosol optical depth data set for data assimilation
Roberts, G., Wooster, M. J., and Lagoudakis, E.: Annual and diur-
nal african biomass burning temporal dynamics, Biogeosciences,
6, 849–866, doi:10.5194/bg-6-849-2009, 2009.
Salomon, J. G., Schaaf, C. B., Strahler, A. H., Gao, F., and Jin, Y.
F.: Validation of the modis bidirectional reflectance distribution
function and albedo retrievals using combined observations from
the aqua and terra platforms, IEEE T. Geosci. Remote, 44, 1555–
1565, 2006.
Schaaf, C. B., Gao, F., Strahler, A. H., Lucht, W., Li, X. W., Tsang,
T., Strugnell, N. C., Zhang, X. Y., Jin, Y. F., Muller, J. P., Lewis,
P., Barnsley, M., Hobson, P., Disney, M., Roberts, G., Dun-
derdale, M., Doll, C., d’Entremont, R. P., Hu, B. X., Liang, S.
L., Privette, J. L., and Roy, D.: First operational brdf, albedo
nadir reflectance products from modis, Remote Sens. Environ.,
83, 135–148, 2002.
Schmid, B., Michalsky, J., Halthore, R., Beauharnois, M., Har-
rison, L., Livingston, J., Russell, P., Holben, B., Eck, T., and
Smirnov, A.: Comparison of aerosol optical depth from four so-
lar radiometers during the fall 1997 arm intensive observation
period, Geophys. Res. Lett., 26, 2725–2728, 1999.
Smirnov, A., Holben, B. N., Eck, T. F., Dubovik, O., and Slutsker,
I.: Cloud-screening and quality control algorithms for the aeronet
database, Remote Sens. Environ., 73, 337–349, 2000.
Varnai, T. and Marshak, A.: Modis observations of enhanced clear
sky reflectance near clouds, Geophys. Res. Lett., 36, L06807,
doi:10.1029/2008GL037089, 2009.
Vermote, E. F. and Roy, D. P.: Land surface hot-spot observed by
modis over central africa, Int. J. Remote Sens., 23, 2141–2143,
2002.
Vermote, E. F., ElSaleous, N., Justice, C. O., Kaufman, Y. J., Priv-
ette, J. L., Remer, L., Roger, J. C., and Tanre, D.: Atmospheric
correction of visible to middle-infrared eos-modis data over land
surfaces: Background, operational algorithm and validation, J.
Geophys. Res.-Atmos., 102, 17131–17141, 1997.
Wu, A. S., Xiong, J., Cao, C. Y., Wu, F., and Barnes, B.: Inter-
comparison of radiometric calibration of terra and aqua modis
11 µm and 12 µm bands, in: Earth observing systems ix, Pro-
ceedings of the society of photo-optical instrumentation engi-
neers (SPIE), Spie-Int Soc Optical Engineering, Bellingham,
248–258, 2004.
Zhang, J. and Reid, J. S.: A decadal regional and global trend
analysis of the aerosol optical depth using a data-assimilation
grade over-water MODIS and Level 2 MISR aerosol products,
Atmos. Chem. Phys., 10, 10949–10963, doi:10.5194/acp-10-
10949-2010, 2010.
Zhang, J. L. and Reid, J. S.: Modis aerosol product analysis for
data assimilation: Assessment of over-ocean level 2 aerosol opti-
cal thickness retrievals, J. Geophys. Res.-Atmos., 111, D22207,
doi:10.1029/2005JD006898, 2006.
Zhang, J. L., Reid, J. S., and Holben, B. N.: An analysis
of potential cloud artifacts in modis over ocean aerosol op-
tical thickness products, Geophys. Res. Lett., 32, L15803,
doi:10.1029/2005GL023254, 2005.
Zhang, J. L., Reid, J. S., Westphal, D. L., Baker, N., and Hyer,
E. J.: A system for operational aerosol optical depth data as-
similation over global oceans, J. Geophys. Res., 113, D10208,
doi:10.1029/2007JD009065, 2008.
Atmos. Meas. Tech., 4, 379–408, 2011 www.atmos-meas-tech.net/4/379/2011/