The MODerate resolution Imaging Spectroradiometer (MODIS) instruments, aboard the two Earth Observing System (EOS) satellites Terra and Aqua, provide aerosol information with nearly daily global coverage at moderate spatial resolution (10 and 3 km). Almost 15 years of aerosol data records are now available from MODIS that can be used for various climate and air-quality applications. However, the application of MODIS aerosol products for air-quality concerns is limited by a reduction in retrieval accuracy over urban surfaces. This is largely because the urban surface reflectance behaves differently than that assumed for natural surfaces. In this study, we address the inaccuracies produced by the MODIS Dark Target (MDT) algorithm aerosol optical depth (AOD) retrievals over urban areas and suggest improvements by modifying the surface reflectance scheme in the algorithm. By integrating MODIS Land Surface Reflectance and Land Cover Type information into the aerosol surface parameterization scheme for urban areas, much of the issues associated with the standard algorithm have been mitigated for our test region, the continental United States (CONUS). The new surface scheme takes into account the change in underlying surface type and is only applied for MODIS pixels with urban percentage (UP) larger than 20 %. Over the urban areas where the new scheme has been applied (UP > 20 %), the number of AOD retrievals falling within expected error (EE %) has increased by 20 %, and the strong positive bias against ground-based sun photometry has been eliminated. However, we note that the new retrieval introduces a small negative bias for AOD values less than 0.1 due to the ultra-sensitivity of the AOD retrieval to the surface parameterization under low atmospheric aerosol loadings. Global application of the new urban surface parameterization appears promising, but further research and analysis are required before global implementation.
In large concentrations, aerosols near the surface (also called particulate matter or PM) are air pollutants. As urbanization and industrialization have amplified many folds during the last few decades (United Nations, 2014), air quality has become a global public health concern, especially in densely populated urban areas. In some cities, urban PM concentrations are at dangerous levels, 5 to 10 times higher than World Health Organization (WHO) guidelines. Although urban areas only represent about half a percentage of the total Earth's surface and about 3 % of Earth's land surface, half of the human population lives in these areas. According to the new projections, two-thirds of the human population will live in urban areas by 2025; therefore, it is critical to monitor air quality (aerosol or PM), especially as relating to human exposure in populated regions around the world.
In economically developed countries such as the United States, and some European nations, surface PM concentrations and air quality are monitored by thousands of ground-based monitoring stations (Cooper et al., 2012). While the density of measurements in the USA may be sufficient for metropolitan-scale mapping, they are not dense enough for local or neighborhood scales. Also, about 30 % of counties in the USA are without any PM monitoring. In contrast, most other countries, especially in developing nations, have few or no surface PM monitors and cannot measure the urban population's exposure to PM.
In the last few decades, satellites are increasingly being used to offer a
global perspective on many atmospheric variables. One of these is aerosol
optical depth (AOD), which is a measure of aerosol loading, integrated
through the atmospheric column. As retrievals of AOD are increasing in their
spatial resolution, coverage and accuracy, environmental monitoring agencies
are increasingly looking to satellites to cover the gaps in aerosol
monitoring. Although it is not straightforward, many studies (Wang and
Christopher 2003; Gupta et al., 2006; Gupta and Christopher, 2009; Hoff and
Christopher, 2009; van Donkelaar et al., 2010, 2015; Liu et al., 2010) are
trying to link satellites retrievals of AOD to surface concentrations of
PM
In addition to its role in air quality and public health, aerosols are considered an essential climate variable (e.g., IPCC, 2007), and remote sensing of AOD has evolved, primarily, to address climate-related questions. Decisions about aerosol products, product resolution and tolerance of poorer quality retrievals have been made to maximize the effectiveness of these products for climate, not air-quality, applications. For example, poor quality retrievals over urban surfaces that represent so little of the Earth's surface have a negligible effect on the climate question, and so retrieval algorithms originally were tuned to ignore the special nature of the urban environment. There are several satellite-based AOD products and each combination of satellite/retrieval algorithm has its own advantages and limitations. One satellite/algorithm pair with a climate-oriented aerosol product is the MODerate resolution Imaging Spectroradiometer (MODIS) Dark Target (MDT) algorithm. The MDT algorithm is mature, having been developed 20 years ago (Kaufman et al., 1997b; Tanré et al., 1997) for retrieval of AOD over primarily vegetated (e.g., dark) land surfaces and remote oceans. It is now running as Collection 6 (Levy et al., 2007a, b, 2013), and the standard AOD product (nominal 10 km spatial resolution) is generally unbiased over global land regions, a requirement for climate applications.
The standard MDT product, while well suited to answer climate questions, has many shortcomings when used for air-quality monitoring. The first, and most important, is that PM is defined as the concentration of particles in the surface layer of the atmosphere where people can be affected by the pollution, while the MDT product measures the aerosol loading (AOD), integrated from this surface layer all the way to the top of the atmosphere. Correlation between column AOD and surface PM depends on the vertical profile of aerosol concentration, which is not measured by MODIS. However, the other problem is that the MDT retrieval, while nearly unbiased compared with the full set of sun photometer (SP) measurements in the Aerosol Robotic Network (AERONET) database, has strong biases for particular surface types.
MDT over land was designed for retrieval over vegetated and other “dark” surface regions. MDT does not provide aerosol retrieval over very bright surface (i.e., desert) and over snow and ice regions. In addition, several validation studies have shown that MDT AOD retrievals over urban area are positively biased with respect to AERONET AODs (Levy et al., 2010; Jethva et al., 2007; Hyer et al., 2011; Gupta et al., 2013; Munchak et al., 2013). These studies have shown that the major source of bias in the MDT over cities is that the city surface does not behave as a vegetated “dark” target.
Several other research attempts have been made to change the surface scheme in the MDT for particular urban regions and produce better AOD retrievals for those specific cities (de Almeida Castanho et al.,2007, 2008; Li et al., 2005; Wong et al., 2011; Zha et al., 2011; Li et al., 2012; Escribano et al., 2014; Jäkel et al., 2015). Instead of a retrieval focused on only one city, we seek a surface parameterization that is valid for cities across the globe. This will lead to more accurate AOD retrievals, which can be utilized for air-quality applications and research, including estimating urban population exposure to aerosols. In this paper, we have developed a surface characterization for cities in the continental US (CONUS) region, applied it to the MDT algorithm and evaluated the results. In Sect. 2 we introduce MODIS, the MDT algorithm and its current limitations over cities. Section 3 discusses study region and various data sets utilized in this study. Section 4 describes the new surface parameterization, whereas results and impact of new surface scheme on MDT AOD retrievals over CONUS regions are discussed in Sect. 5. Section 6 covers implications of the new surface scheme over global regions and its limitations and challenges. We summarize the results and future directions in Sect. 7.
MODIS sensors have been observing the Earth system (atmosphere, land and
ocean) on board two satellites, since 1999 on Terra and since 2002 on Aqua.
MODIS observes top-of-the-atmosphere (TOA) radiance in 36 spectral bands,
which are used to derive geophysical information about atmosphere, land and
ocean. The spatial resolution varies from 250 m to 1 km depending on
spectral band. The large swath width of MODIS (
Since MODIS measures the reflectance of the Earth's system, the measured radiation contains information about atmospheric properties as well as Earth's surface. In some spectral bands, the surface dominates the signal, whereas the atmosphere dominates in other bands. The strategy then is to use the right combinations of spectral bands to retrieve a particular aspect of the combined Earth's system. Specifically, the MDT is used to derive global aerosol properties under cloud/snow/ice-free conditions.
The theoretical basis of the MDT algorithm has remained constant from its original at-launch version, although individual sub-modules have been continuously evolving. While a summary of the algorithm is provided here, the reader is encouraged to read the references for details of the algorithm assumptions and structure and its evolution through different versions over time (Remer et al., 2005; Levy et al., 2007a, b, 2013).
The five main components of the MDT aerosol retrieval algorithm are (1) pixel
selection and aggregation, including cloud masking and other filtering;
(2) separating the surface from the atmosphere; (3) assumptions about the
aerosol (e.g., aerosol models); (4) matching observed TOA spectral
reflectance to lookup tables; and (5) inferring the ambient aerosol
conditions (model weightings and total AOD). There are two separate
retrievals, one over ocean and one over land, and the results of these
retrievals and associated diagnostics are available in a single data file.
Each retrieval result and each diagnostic are independent science data sets
(SDSs) within the produced file. In production, collectively these SDSs
within the single file are known as the MxD04
product, where x is “O” for Terra and “Y” for Aqua, and the “04” denotes the level 2 aerosol products. Products (SDSs) include the total AOD (at
0.55
The MODIS retrieval codes are run in an operational environment to create C6 products. This includes infrastructure for managing file formats and also for processing entire granules of MODIS data. In this work, we also make use of the so-called “stand-alone” version of the MODIS Dark Target (S-MDT) retrieval code (e.g., Levy and Pinker, 2007). The S-MDT is stripped of all routines for cloud masking, pixel selection and pixel aggregation, instead operating on a pixel-by-pixel basis. Inputs are a single set of TOA spectral reflectance values, plus sun–satellite geometry and geo-location. Outputs are the retrieved AOD and most of the diagnostics contained in the standard output. Since the standard C6 data (e.g., the 10 km retrievals) include the TOA spectral reflectance used for each retrieved AOD value, the data provided in the C6 output file (e.g., MxD04_L2) can be recycled through the S-MDT algorithm to retrieve the same AOD value as provided within the C6 product. Thus, the S-MDT can be easily modified to test different assumptions within the retrieval, including surface reflectance assumptions. The insights gleaned from the S-MDT exercises can then be transferred back to modify the full (operational) retrieval code, and tests (and statistics) can be performed for global data.
The land surface is too variable to apply an explicit model to describe its
spectral optical properties. Thus MDT uses an empirical parameterization
based on only three MODIS bands. Kaufman et al. (1997a, 2002) noted that for
most vegetated and dark-soiled land surfaces, observations showed that
surface reflectance in a blue wavelength (0.47
The current operational version (C6) of the MDT algorithm is still based on
the SWIR–VIS surface relationships, but it also adjusts the relationship for
vegetation amount and geometry as determined by a variant of the normalized
difference vegetation index based on SWIR bands
(NDVI
Although the SWIR–VIS assumption characterizes “dark” surfaces on a global
scale, it fails to account for all surface types, especially for
anthropogenic modifications to natural land surfaces. As a land surface
transitions from natural vegetation to manmade structures such as buildings
and roads, the global SWIR–VIS relationship is violated. This was noted in
Levy et al. (2007b, 2009). In addition, de Almeida Castanho et al. (2007) analyzed the
SWIR–VIS relationship over Mexico City and found that the ratios of SWIR–red
are much higher (0.73–0.76) than what is assumed in the MDT global algorithm
(0.50–0.55). They also found that the SWIR–VIS relationship strongly depends
on differences in urbanization fraction. Using the global values for surface
ratio when the actual ratio is much higher will underestimate the visible
surface reflectance and the resulting aerosol contribution will be
overestimated. Munchak et al. (2013) used the dense AERONET Distributed
Regional Aerosol Gridded Observation Networks (DRAGONs) (Holben et al., 2011)
that operated in Washington DC–Maryland during 2011 to identify the urban
high bias and link that bias to the urban percentage (UP) around each AERONET
site. While there have been attempts to improve the MDT aerosol retrieval on
the local or even regional scale, no attempt has been made to date to develop
a general improved surface reflectance parameterization that can be applied
to the global retrieval. In this study, we develop a new SWIR–VIS scheme that
accounts for NDVI
We explore whether MDT surface parameterization can be modified to account
for urban surface properties. We start with the research S-MDT version (Levy
and Pinker, 2007; and
To test the C6U retrieval, we compare the retrieved AOD with the SP measurements taken at standard AERONET sites, as well as AERONET's DRAGONs during the Maryland deployment of the “Deriving Information on Surface conditions from Column and Vertically Resolved Observations Relevant to Air Quality” (i.e., DISCOVER-AQ) (Holben et al., 2010) experiment.
AERONET is NASA's global ground network of SPs that make measurement of direct transmitted solar light during daylight hours (Holben et al., 1998), and from that measurement we derive spectral AOD. There are currently about 300 sites around the globe with more than 1 year of regular observations; in addition, dense DRAGONs, while more limited in time, make dense measurements over urban/suburban areas. DRAGONs have tended to be deployed in support of large field experiments, including DISCOVER-AQ.
The spectral measurements from the SPs are used to derive AODs in respective spectral bands. The typical frequency of measurements is every 15 min and the spectral bands are generally centered at 340, 380, 440, 500, 675, 870 and 1020 nm. Here we use the Ångström coefficient to interpolate AERONET AOD to 550 nm. AERONET data products are available as unscreened (level 1.0), cloud screened (level 1.5) and cloud screened and quality assured (level 2.0). In this study, level 2.0 AERONET AOD data are considered as ground truth to validate the satellite-retrieved AOD data. The reported uncertainty in AERONET AOD is of the order of 0.01–0.02 (Eck et al., 1999). There are about 135 CONUS AERONET stations collocated with MODIS Aqua for the 2003 to 2012 time period. This includes 39 DRAGON sites, permanent AERONET sites and temporary AERONET sites operated for different periods between 2003 and 2012. Only the sites located over land are considered in this study.
The DRAGON sites in Maryland DISCOVER-AQ deployment were operated for about 6 weeks from 1 July to 15 August 2011. The sites were located in the Washington DC–Baltimore metropolitan area. The network provided useful AOD measurements over urban, agricultural, coastal and mountain landscapes over the Washington DC metropolitan area.
The MODIS Land Cover Type product (MCD12Q1) for the year 2011, at 500 m
resolution, has been used to identify urban surfaces in the new surface
scheme. This is a yearly product, based on a trained classification algorithm
that uses five different classification schemes (Friedl et al., 2010), and is
derived using observations from both MODIS sensors. There are 17 land cover
classes as defined by the International Geosphere-Biosphere Program (IGBP).
The land cover class defined as “urban and built-up area” has been
extracted and UP at 0.1
The urban percentage map for the continental United States derived
from MODIS Land Cover Type product (MCD12Q1) at 0.1
The surface parameterization in the MDT algorithm was based on performing atmospheric correction of MODIS-measured reflectance near AERONET sites, which led to formulation of the empirical SWIR–VIS relationships (e.g., Levy et al., 2007b, 2013). Rather than repeating this exercise, we rely on the MODIS-derived land spectral surface reflectance product because it is available and has been validated with similar atmospheric correction exercises. Specifically, we use the MODIS 8-day, clear-sky surface reflectance product (MOD09A1/MYD09A1; Vermote and Kotchenova, 2008) for the first seven bands of MODIS, which are gridded at 500 m resolution. These are same bands used for the MDT aerosol retrieval.
The MxD09A1 product is created by identifying the “clearest” observations
of a scene during an 8-day period and performing generic atmospheric
corrections (estimating aerosol type and AOD) to obtain surface reflectance
values. These pixels are selected for low view angle, the absence of clouds
or cloud shadow, and low aerosol loading. Pixels identified as snow/ice,
adjacent to cloud, fire, cirrus, inland water or high aerosols loading are
excluded. This surface reflectance product is officially validated as a stage
3 product. Validation of the MxD09A1 product has been performed over
150 AERONET sites (
Uncertainty in surface reflectance product (MYD09) as obtained by
validation exercise and reported at
Validation results show that the red and SWIR bands are more accurate than
the blue band, and uncertainty increases over urban areas in all bands, but
especially in the blue. Early validation results suggested that for urban
surfaces, typified by Hamburg, Germany, the percentage of surface reflectance
retrievals falling within the above stated error bars were 70 % and
100 % for the red and SWIR channels, respectively, but less than 10 %
for the blue channel (Vermote and Kotchenova, 2008). Also, this site exhibited
a consistent high bias of 0.01 reflectance in band 4 (0.54
Despite these uncertainties, we will make use of all three wavelengths bands, because the MxD09A1 provides the most complete characterization of surface reflectance at the spatial scales necessary for our study. We are fully aware that the high uncertainty in the blue band can lead to errors in surface reflectance ratios and thus to errors in retrieved AODs, specifically under low aerosol conditions when the algorithm is most sensitive to accurate surface reflectance values. Further discussion of the MxD09A1 product from the perspective of this study is provided in Sect. 7.
For the MDT algorithm to include an urban surface reflectance scheme, it
must
reduce the bias between MDT retrievals and SP measurements
in urban areas; improve the retrievals in urban areas without degrading retrieval
quality in non-urban areas; have an operational path that identifies whether or not to apply the
new scheme; slip into the structure of the existing operational algorithm without
requiring extensive additional modifications.
The first step to meeting these requirements is to characterize the unique surface reflectance behavior of urban surfaces. Although there have been previous studies which have characterized surface reflectance over urban areas, they are limited to being near AERONET sites, where atmospheric correction could be applied. These local relationships have not been shown to represent the SWIR–VIS relationship over larger areas. Relationships derived in this limited database are skewed towards surface characteristics of AERONET locations.
By using the MODIS Land Surface Reflectance product (MOD09) as described in Sect. 3, we have a data set with global coverage, as opposed to atmospheric correction exercises for individual sites (which use observed ground-based aerosol properties to derive surface reflectance).
To create our urban surface parameterization, we start with the MOD09 product
in its native resolution (e.g., 500 m pixels). The MODIS 500 m resolution
surface reflectance values are quality controlled and averaged into a
0.1
We expect that different cities will exhibit different surface reflectance
relationships for the same UP, because the natural vegetation background is
different. Therefore, we employ both UP and NDVI
Regression coefficients for the new surface scheme introduced in C6U MDT algorithm.
Figure 2 provides the surface reflectance spectral relationships between SWIR
and VIS, defined for the four different categories based on the combinations
of NDVI
The 0.65
The regression statistics provide the information needed to relate the
surface reflectance (
As shown in Table 2, the new slopes of SWIR–VIS over urban areas are
significantly higher than those assumed in the C6 algorithm for global
application, which is consistent with other studies (i.e., Levy et al., 2007b;
de Almeida Castanho et al., 2007). The red–SWIR ratios are also higher
for less-vegetated areas (NDVI
The new surface scheme is expected to provide better surface reflectance estimates over urban areas than the existing operational scheme (i.e., C6). Figure 2 shows very tight relationships between SWIR–VIS reflectances, but even the presence of the small amount of scatter in the correlation can cause errors in certain conditions and seasons. Bidirectional reflectance function (BRDF) effects over urban surfaces and other factors as discussed in Levy et al. (2007b) may also introduce error. Because surface reflectance dominates the TOA signal for low aerosol conditions (AOD < 0.1) as compared to high aerosol loading (AOD > 0.4), relative uncertainties in retrieving AOD are larger under clean conditions.
UP is the only new parameter added to the surface scheme. This is a globally available parameter calculated from a standard MODIS Land Product (MCD12Q1) and is updated annually. The C6U algorithm inputs UP and uses this parameter to decide whether to apply the old or the new surface parameterization. Dependence on UP assures that there will be no changes to the retrieved aerosol products for non-urban surfaces (UP < 20 %).
We have developed this surface scheme using data from the continental USA, expecting the scheme to be optimized for this region. Global implementation may be challenging and we discuss these challenges in Sect. 6. Also, we note that this application of UP surface correction is applied only for the MODIS 10 km aerosol product. We discuss future application to 3 km retrieval in Sect. 6.
The data from MODIS Aqua from 2003 to 2012 over the CONUS region are processed
using the C6U algorithm and compared against C6 retrievals and AERONET
measurements. MODIS C6 and C6U AODs were collocated over AERONET sites in
the CONUS region, using the same spatiotemporal technique (Ichoku et al.,
2002) used for MODIS validation exercises (Levy et al., 2010). In this
method, AOD values from all the pixels within a 0.5
Root mean square error (RMSE) and mean bias is estimated using Eqs. (4)
and 5.
In addition, we have computed linear regression statistics, including
correlation coefficient (
Note that MODIS also reports AOD values at 0.47 and 0.67
Figure 3 shows 2-D density scatter plots, representing all
collocations of MODIS-retrieved AOD and AERONET SP measured
AOD over CONUS urban sites. Here, only retrievals where UP > 20 % and the retrieval quality assurance flag indicates “very good
quality” (QAF
The frequency scatter plot for AOD at 0.55
As discussed in Sect. 4, aerosol retrieval from passive satellite measurements is sensitive to underlying surface reflectance; the relative uncertainty becomes greater for lower AOD conditions, when the surface reflectance dominates the signal. By improving the overall urban surface reflectance parameterization, the bias is reduced. However, the small negative bias indicates that there is still uncertainty in estimating visible reflectance. We note that most of the negative AODs from the C6U retrieval are correctly identifying low values of AODs (< 0.1), such that a retrieval of “clean” conditions is correct. The other source of uncertainty in AOD retrievals comes from selection of proper aerosol model, but this effect should be minimal at such low optical depths. The number of retrievals within pre-defined expected error envelope (EE %) has also increased from 63 % for C6 to 85 % for the C6U retrievals. Most of the statistical parameters in comparing MODIS and SP AODs demonstrate improvement in C6U AODs as compared to C6 retrievals.
Previous validation and inter-comparison studies (Munchak et al., 2013) have pointed out the positive correlation of C6 AOD biases with respect to urban land cover amount. The C6U surface characterization accounts for change in urban area within each MODIS AOD pixel, such that the positive correlation should be reduced in our data set. We would expect that for UP > 20 % the bias should be flat with respect to UP if the new parameterization is doing its job. Figure 4 shows bias in MODIS AOD (MODIS–AERONET) as a function of change in the UP. Again side-by-side comparisons for C6 in Fig. 4a and C6U in the Fig. 4b are presented. Here, we have considered all AOD pixels irrespective of urban percentage, which is different from the source of Fig. 3, which limited us to cases with UP larger than 20 %. Here the objective is to evaluate the impact of new surface scheme on the quality of the MDT-retrieved AODs as a whole, beyond urban areas. We recall that the new surface scheme is applied only on selective pixels (UP > 20 %), which consist of only about 3 % of the total retrieved AODs in the CONUS region. Red dots show the equal number of points bin averaged value with 1 standard deviation in AOD error as vertical blue line. The biases in C6 AODs linearly increase as land cover becomes more urbanized (larger UP), whereas biases in C6U AODs do not show significant dependence on UP. This analysis indicates that the C6U surface parameterization successfully removes AOD biases over cities and should be applied to MDT aerosol retrievals over the CONUS region.
Binned bias in MODIS AODs compared to AERONET AODs as a function of
UP, using all collocated data sets with QAF
The MODIS AOD data presented in Figs. 3 and 4 represent collocations for
where the quality assurance flag identifies the retrieval as
“science quality” (QAF
Statistics of MODIS and AERONET inter-comparisons using collocated
data sets. Comparisons are performed for different quality flags. Three
different MODIS pixel selection schemes based on surface type are used. All: all
MODIS pixels were considered irrespective of underlying surface type; urban % > 0.0: MODIS pixels with urban surface were
selected;
urban % > 20: MODIS pixels with urban percentage larger than
20 % were selected in collocation. Statistical parameters, number of data
points (
In order to analyze and validate MODIS AODs with lower quality flags, we have
grouped the data in several different ways to represent different retrieval
conditions. Table 3 presents the statistical analysis of the two retrievals
(C6 and C6U) for three categories of underlying surface type (i.e., UP).
These are
ALL: all retrievals irrespective of UP; UP > 0.0 %: retrievals that have some urban fraction; UP > 20 %: retrievals with UP larger than 20 %.
For the third category (UP > 20 %), this includes only retrievals
where C6U retrieval would be applied and different from C6, while the
second category includes retrievals that may be suburbs or small towns,
and the first category (ALL) includes everything.
Since QAF value is most strongly connected to the number of pixels used in the retrieval, the difference between C6U and C6 would not be reflected in reported QAF value.
Each surface category in Table 3 is further broken down by QAF level. Note
that the case of ALL and QAF
For the C6 data, there is a clear reduction in regression quality (decreased
correlation, increased bias, %EE reduction) as QAF criteria are relaxed
from 3 to 2 to 1. This is true for the set of ALL retrievals, but especially
when the retrieval is performed over even a small fraction of urban surface
type (UP > 0). For C6U, the immediate effect is to cut bias to a
negligible value for QAF
The improved statistics for AOD retrievals with lower assigned quality flags are encouraging and suggest opportunity for overall increase in high-quality sampling with the MDT algorithm. This will definitely help characterize aerosol for air-quality applications in densely populated areas. However, further research and dedicated evaluations of quality flag assignment criteria in the algorithm are required before we suggest making use of lower-quality data even in the C6U retrievals.
Statistics of MODIS and AERONET inter-comparisons over selected
urban sites (Fig. 5). Statistical parameters are the number of coincident
points (
The main reason for the development of the C6U retrieval algorithm is to
reduce the biases in AOD retrievals over cities where a large portion of the
human populations lives. In this section, we evaluate both C6 and C6U
retrievals over selected cities where an AERONET station is available.
Figure 5 shows how C6 (red dots) and C6U (blue dots) compare with AERONET
AODs over eight selected AERONET stations covering various parts of CONUS
and Canada. Only MODIS AOD retrievals with QAF
Inter-comparison of MODIS AODs at 0.55
The lowest correlation is observed over Caltech with values of 0.42 and 0.58
for C6 and C6U, respectively, likely because of the complex terrain within the
0.5
In Fig. 6 we have evaluated the spatial distribution of AOD over the region covering two large urban cities, Baltimore and Washington DC. The map shows averaged AODs for the period of June–August 2011. The main purpose of the figure is to demonstrate, spatially, that the C6U algorithm reduces the high AOD bias over urban surfaces, such as the Washington DC area in the figure. There we see that the seasonal mean values in the C6U algorithm are more spatially consistent with the surrounding suburban and rural area than are the values from C6. Figure 6c shows the difference between C6 and C6U AODs, which is correlated with UP (Fig. 6d) and could be as high as 0.12.
Seasonal (June–July–August 2011) maps of MODIS AOD at
0.55
The secondary purpose of the figure is to demonstrate that the C6U algorithm has not solved all problems associated with the retrieval over cities. There are still artificially high seasonal mean values for Baltimore and the Chesapeake Bay shoreline. The reason these seasonal mean values remain artificially inflated is because of sampling. Figure 6e presents the number of averaging days (or number of retrievals) for each grid box, and it is apparent that some grids near city centers and along the shoreline have very limited sampling (1–5) days. Coincidently, these available days correspond to high aerosol loading days, creating an illusion of high seasonal mean aerosol loading in the city centers and along the Chesapeake Bay. The low number of retrievals in these squares is caused by a combination of clouds and the additional issue of the algorithm choosing not to retrieve over very bright urban surfaces under low aerosol loading and at certain sun–satellite geometry. While the new C6U algorithm will be able to produce a better urban retrieval when an urban pixel is selected for processing, it will continue to be affected by the algorithm's pixel selection process that makes it difficult for urban pixels to be chosen. This work focuses on the parameterization of the surface reflectance relationships and not on the upstream pixel selection and masking processes.
Figure 7a–h show the evaluation statistics of the two algorithms over a dense network of AERONET instruments in the Baltimore–Washington DC metropolitan area during summer 2011. The DRAGON is a mesoscale network of sun/sky radiometers that encompasses, urban, suburban, agricultural and hilly landscapes over the Washington DC metropolitan area. There were about 39 AERONET DRAGON stations operating during this deployment. This dense SP network provides an excellent aerosol measurements data set covering different types of landscapes. This AERONET DRAGON deployment also provides an excellent opportunity to evaluate the high AOD values near the cities as observed in Fig. 6. The data have been utilized to validate satellite aerosol retrievals and spatial variability in the aerosol fields (Munchak et al., 2013). Munchak et al. (2013) reported that the MODIS C6 AOD retrievals were positively biased against AERONET values in (and near) urban areas with a high degree of correlation with UP. We now revisit the Munchak et al. (2013) data sets to verify whether the new C6U retrievals alleviate the issues noted by the previous study.
Inter-comparison statistics of MODIS–AERONET AODs over DRAGON
network during DISCOVER-AQ field campaign (June–July 2011) in the Washington
DC–Baltimore area. This analysis used data from AERONET stations operated
as part of DRAGON network. Scatter plot between AERONET and MODIS for C6
Figure 7a, c, e, g (left panels) and b, d, f, h (right panels)
represent comparisons between MODIS C6 and C6U AOD validation statistics,
respectively. The two scatter plots show MODIS-retrieved AOD plotted against
AERONET measurements. Each point is color-coded with the UP estimated for
each DRAGON site using the MODIS Land Cover Type information. The C6 AODs
show (Fig. 7a) positive biases specifically for AOD values larger than 0.15,
which results in an overall positive bias of 0.04 with very high correlation
(
The new surface scheme presented here is designed to work only over CONUS (and perhaps other regions with similar surface
properties) and with the 10 km aerosol retrieval. Implementing the new
scheme into the global algorithm, as well as at a different spatial
resolution (e.g., 3 km), may be challenging. In CONUS we had a wealth of
data to work with: 135 AERONET stations with several in highly urban
locations and a well-analyzed DRAGON network to evaluate the small-scale
variability in the aerosol fields. We were able to parameterize the surface
reflectance relationships by dividing the surfaces into only four categories
depending on UP and NDVI
Binned bias in MODIS AODs compared to AERONET AODs as a function of
UP using all collocated data sets with QAF
For years the MDT algorithm refrained from using the MxD09
because of the question of “circularity”. The MODIS land atmospheric
correction and aerosol retrieval algorithms evolved from the same basic root
(Vermote et al., 1997; Kaufman et al., 1997). Using the land reflectance
derived from the aerosol algorithm to derive the AOD that is used to produce
the land reflectance would create an incestuous circular relationship, tuned
to agree at AERONET stations and nowhere else. However, over time the land
atmospheric correction and the Dark Target aerosol retrieval evolved
significantly into very different second-generation algorithms, using a
different set of wavelengths, a different set of assumptions of aerosol
properties and minimizing a different cost function in the inversion (Vermote
and Kotchenova, 2008; Levy et al., 2007a, b). To reduce the possible
circularity even further, we do not attempt to directly use daily surface
reflectance product in our MDT algorithm but rather look for spectral
relationships. Mirroring the logic within the C6 MDT algorithm, we use
1 year of MOD09 over CONUS to form the SWIR–VIS relationships as a
function of NDVI
The next question concerns the uncertainty in the MxD09 described above in Sect. 3.3. The MxD09 algorithm attempts to match the atmospherically corrected surface reflectance ratios with pre-determined values by adjusting simultaneously the AOD and surface reflectance. The pre-determined surface reflectance ratios are derived from a global database of atmospherically corrected surface reflectance at AERONET sites. There is one global value per pair of wavelengths. Because, as we have shown, the surface reflectance ratios over urban surfaces differ from the vast majority of land surface types, we would expect that MxD09 values over urban surfaces to show higher errors and greater uncertainty than more rural and typical surfaces. Indeed that is the case as shown in Sect. 3.3. If this error is not spectrally uniform and involves a bias as well as random error, as is suggested by the urban example in Vermote and Kotchenova (2008), then the ratios that we derive from MxD09 in Fig. 2 will not represent that actual surface reflectances in the MDT retrieval. Errors and biases will be introduced into the C6U results. Comparisons against AERONET will be partially contaminated because the original MxD09 pre-determined surface reflectance ratios were based on the corrected surface reflectances surrounding AERONET sites.
Despite these causes of concern, we proceeded with the use of MxD09 in this study because this is the only data set having robust statistics with any hope of providing surface reflectance ratios in urban settings at the spatial scale needed. Any alternative method of performing our own atmospheric correction at specific AERONET locations with variable UP to determine our own urban surface reflectance ratios would suffer from limited statistics. Our decision to proceed with MxD09 has been justified with the results shown in Figs. 3, 4, 5 and especially 6, 7 and 8. The C6U results not only bring the urban AOD closer to the AERONET values in a general sense but also reduce the urban bias beyond the original AERONET stations used to derive the pre-determined surface reflectance ratios. We see this across the DRAGON network of Fig. 7, where C6U brings down the bias and captures the mesoscale variation of the aerosol. None of these DRAGON stations existed at the time that the universal surface reflectance ratios were determined for the MxD09 algorithm, and therefore this is a clean validation. We also note that AERONET has grown significantly since the MxD09 pre-determined ratios were calculated, and while our validation data set may overlap with the formulation one, it is not identical. Figure 8 also bolsters our confidence in this data set. Note that the urban surface reflectance ratios we derive in Fig. 2 were derived using only values over CONUS. There was no guarantee that these values would improve retrievals over non-CONUS cities, but they do.
Thus, while the MxD09 product has significant uncertainty in the blue channel, especially over urban surfaces, that uncertainty appears to be manageable. There is no doubt that C6U is an improvement over C6 in urban settings.
The MDT aerosol retrieval algorithm, shaped by continuing
research and influenced by application needs, has been operating successfully
for 15 years. The MODIS C6U algorithm presented here reflects the MODIS
science team's commitment to keeping the algorithm updated and relevant. In
this spirit, we address the AOD biases in the current operational product
related to improper surface parameterization over urban areas. We develop a
revised surface parameterization scheme over urban regions using the MODIS
Land Surface Reflectance and Land Cover Type products. The new
parameterization parallels the current Collection 6 surface scheme, where
visible surface reflectances are estimated for each pixel using the value of
the SWIR surface reflectances at that pixel, modified by NDVI
MODIS Aqua data sets from 2003 to 2012 over all AERONET stations in the USA including a dense DRAGON network deployed during the DISCOVER-AQ field campaign in the Baltimore–Washington DC metropolitan area, have been utilized to evaluate the revised AOD retrievals. The side-by-side comparison of C6 and C6U retrieval against AERONET measurements provided quantitative estimates of improvements in the MODIS AOD retrievals. Over urban areas where the C6U retrieval has been applied (UP > 20 %), we find an increase of more than 20 % in the number of retrievals falling within EE %. The strong positive correlation between bias in AOD and amount of urban surface near the AERONET site that was observed in C6 is gone in C6U. The C6U retrieval does introduce a small negative bias in the retrieved AOD for AOD values less than 0.1 due to ultra-sensitivity of the AOD retrieval to the surface parameterization under low atmospheric aerosol loadings.
While the new C6U algorithm successfully reduces the high biases in AOD seen over urban pixels when one of these pixels is selected for retrieval, C6U does not affect the pixel selection process. Pixel selection itself is affected by the properties of urban surfaces, and under low aerosol loading conditions and certain geometries, urban pixels are disproportionally rejected for retrieval. Thus, C6U will improve retrievals when retrievals are made but will not increase the number of retrievals attempted.
In general, the MODIS science team recommends using AOD data with the best-quality flag (QAF
While the formulation of the C6U algorithm is based on surface characterization of stations in the continental USA, we tested the new algorithm on the global data set and compared with AERONET AOD. Even when excluding the CONUS AERONET stations to avoid the mistake of validating the formulation data set, the results show the elimination of AOD bias as a function of urban percentage. These are unexpected, but encouraging, results that suggest that the parameterization developed from the CONUS data may be implemented soon into the global operational algorithm for a significant improvement over urban centers worldwide. Additional testing will be necessary first.
As populations flock to urban centers, causing the urban landscape around the world to grow continuously, it becomes obvious that these regions can no longer be treated with second-class status by the MDT aerosol algorithm. It is crucial to have accurate retrievals of AOD over the urban landscape, which translate into more accurate estimates of particulate matter concentrations for air-quality purposes over the regions where most people live. The revised C6U algorithm improves the quality of MODIS AOD retrievals over urban regions, which will be extremely useful for air-quality applications. We expect that this improvement will open up new opportunities for the research community to apply the MDT AOD data to address other pressing issues such as urban-scale spatial variability, gradients between rural and urban areas, more accurate long-term trends and air-quality–health links.
AERONET data were obtained from the NASA AERONET data server; we would like to thank the AERONET team for maintaining the network and data archive. We could not do this study without the AERONET and DRAGON teams' continuing support of quality-controlled, easy-access data. This project is supported through NASA ROSES grants under Terra–Aqua: NNH13ZDA001N-TERAQEA MODIS maintenance project. Edited by: M. Wendisch Reviewed by: two anonymous referees