The Advanced Baseline Imager (ABI) on board the
Geostationary Operational Environmental Satellite-R (GOES-R) series enables
retrieval of aerosol optical depth (AOD) from geostationary satellites using
a multiband algorithm similar to those of polar-orbiting satellites'
sensors, such as the Moderate Resolution Imaging Spectroradiometer (MODIS)
and Visible Infrared Imaging Radiometer Suite (VIIRS). However, this work
demonstrates that the current version of GOES-16 (GOES-East) ABI AOD has
diurnally varying biases due to limitations in the land surface reflectance
relationships between the 0.47
Aerosols in the atmosphere such as dust, smoke, pollutants, volcanic ash,
and sea spray can affect climate through scattering and absorption of
radiation directly and through interaction with clouds indirectly
(Albrecht, 1989; Rosenfeld and Lensky, 1998; Mahowald, 2011). In addition,
aerosols impact air quality and thus affect human health (e.g., Pope and
Dockery 2006). Satellite-retrieved aerosol optical depth (AOD), a
quantitative measure of the amount of aerosols present in the atmosphere, is
useful for evaluating aerosols' effect on climate change (e.g., Yu et al.,
2006) and for estimating and forecasting ambient PM
AOD from polar-orbiting satellite sensors, such as the Moderate Resolution
Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer
Suite (VIIRS), is retrieved using multi-channel algorithms (Levy et al.,
2007, 2010; Sayer et al., 2014; Jackson et al., 2013; Liu et
al., 2014; Laszlo and Liu, 2016). As a result, AOD from MODIS and VIIRS has
high accuracy, e.g., MODIS dark target AOD has an expected error of
The NOAA ABI AOD product has a spatial resolution of 2 km at nadir, compared
to 3 and 10 km from MODIS Collection 6 and 750 m (NOAA product) and 6 km
(NASA product) from VIIRS. The GOES-16 ABI AOD product was released on 25 July
2018, while the GOES-17 ABI AOD product reached provisional maturity on
January 1, 2019; Definition for provisional maturity can be found in EOSDIS
Glossary (
The accuracy and precision of VIIRS and MODIS AOD is well documented for use in various decision support systems (Laszlo and Liu, 2016; Sawyer et al., 2020; Levy et al., 2013; Sayer et al., 2014). The geometries of observations from a geostationary satellite are quite different from a polar-orbiting satellite; this can lead to differences in the quality of retrieved AOD despite the similarity of the AOD retrieval algorithms. It is therefore very important to evaluate the new ABI AOD product and demonstrate its accuracy and precision at daily and sub-daily timescales. This should allow users to interpret the ABI AOD product correctly and apply it appropriately in research and operational applications.
In this study, we compare GOES-16 ABI AODs to AERONET AODs for a 5-month period in 2018 and document a diurnal bias in the ABI AOD due to deficiencies in the land surface reflectance relationship currently applied in the retrieval algorithm. The presence of the bias is consistent across the CONUS, but its magnitude varies by location. We describe a novel method that corrects the bias for each AOD pixel and time step. The resultant corrected ABI AOD shows little to no diurnal bias over a variety of surface types (e.g., urban, rural).
The GOES-16 ABI AOD data used in this work is from the period of 6 August to
31 December 2018 over the CONUS region. The ABI AOD data have 2 km spatial
resolution at nadir and 5 min temporal resolution. Similar to MODIS and
VIIRS AOD, ABI AOD are retrieved using separate algorithms over ocean and
over land, due to the different surface characteristics of ocean and land
(Kondragunta et al., 2020; GOES-R AOD ATBD, 2018). Over land, three ABI
channels are used in the retrieval, i.e., 0.47, 0.64, and
2.2
Other atmospheric and geographic parameters needed for the retrieval are
also inputted, such as surface pressure, surface height, total column ozone,
etc. The algorithm only retrieves AOD over dark surfaces, when the TOA
reflectance in the 2.2
The retrieval algorithm assigns the pixel level AOD to one of three
qualities: high, medium, and low. AOD quality is determined on conditions of
the pixels, such as solar–satellite zenith angle, cloud–shadow adjacency, and
standard deviation of measured reflectance at a specific band. The full set
of criteria used for assigning a quality level is listed in Table 1. High-quality AOD is the most accurate and the one recommended for scientific
applications. However, the ABI AOD retrieval algorithm uses such strict
criteria to remove potential erroneous pixels that the number of pixels with
high-quality AOD is usually very small. For example, the ratio between the
number of the top 2 qualities and the high-quality matchup with AERONET is
about 2 (see the following section), while the ratio is 1.2 for NOAA VIIRS
AOD (Laszlo and Liu, 2016). The following criteria are used to degrade a
pixel from high quality to medium quality: (1) adjacent to a cloudy pixel;
(2) adjacent to a snow pixel within 3 pixels distance; (3) 3
Conditions for quality levels of ABI AOD pixels.
The surface reflectance relationship used in the operational ABI AOD
algorithm was derived from AERONET matchup dataset using strict criteria,
with cloud screening using that for high-quality, low AERONET AOD (
The NASA dark target (DT) aerosol algorithm team applied their DT algorithm on geostationary satellite data such as ABI and AHI (Advanced Himawari Imager, Gupta et al., 2019). In order to test the bias correction algorithm on other AOD retrieval algorithms, the GOES-16 ABI DT AOD was obtained from NASA for the month of July 2019. The product covers the full disk with 10 min temporal resolution and 10 km pixel resolution (at nadir).
The AErosol RObotic NETwork (AERONET) is a global ground-based aerosol
remote sensing network (Holben et al., 1998). It uses CIMEL sun photometers
to measure spectral sun irradiance and sky radiances. The measurements are
then used to calculate and retrieve aerosol properties. Among them, AOD is
one of the main products; it is measured at a subset of 22 different
wavelengths from ultraviolet to infrared, i.e., 340, 380, 400, 412, 440, 443,
490, 500, 510, 531, 532, 551, 555, 560, 620, 667, 675, 779, 865, 870, 1020,
and 1640 nm, depending on the specific instrument. Ångström Exponent
(AE) can be calculated from the multispectral AOD. Besides AOD, AERONET also
retrieves other aerosol properties, such as volume size distribution,
refractive index, phase function, and single-scattering albedo (SSA).
AERONET AOD is considered ground truth for satellite AOD (Holben et al.,
1998) and is used to evaluate the ABI AOD retrievals. AERONET AOD at 550 nm
is obtained through interpolation from other spectral bands so that it can
be compared against ABI AOD, which is reported at 550 nm. In this work,
AERONET AOD version 3 level 1.5 is used. Although level 2.0 data have higher
quality, they have time delays such that the latest data were not available
during the analysis period. Level 1.5 AERONET AOD data are cloud screened and
quality controlled, with an up to
The diurnal bias of ABI AOD is evident when it is compared to coincident
measurements of AERONET AOD. The diurnal bias is most apparent on “clear”
days, when AERONET AOD is
Details about the representative AERONET sites used as examples to illustrate the range of the observed diurnal bias in GOES-16 ABI AOD.
Time series of GOES-16 ABI AOD and AERONET AOD at six representative
AERONET sites:
To illustrate the problem of the diurnal bias of ABI AOD, the time series of ABI AOD and AERONET AOD for clear days are plotted at the representative AERONET sites in Fig. 1. As demonstrated in the figure, the number of the ABI top 2 qualities (high and medium quality) data points are much larger than that of the high-quality AOD. For example, on 18 October 2018 at the CCNY site (Fig. 1a), which is located in New York City, New York, no high-quality ABI AOD data matchup data are available, but top 2 qualities AOD matchup points exist at nearly all time steps.
The diurnal variation of the ABI AOD bias is observed at all six sites, but
the magnitude of the bias varies, with higher bias observed at the
urban and suburban sites (Fig. 1a, c, d, and e) compared to the rural
sites (Fig. 1b and f). For all sites, the bias peaks around 17:00 UTC,
when the Sun moves from the east of the satellite to the west of the
satellite, as determined by the location of the satellite, i.e.,
75.2
There are several potential causes of the diurnal bias observed in ABI AOD,
including known sources of uncertainty associated with calibration,
cloud and snow contaminations, aerosol models, and errors in the surface
reflectance model (Li, et al., 2009). In the cases shown in Fig. 1, all
days have low AOD values and continuous AOD measurements from AERONET,
indicating that the influences of the aerosol model selection and cloud
contamination are small. Snow contamination is not an issue because the
analysis days are mostly in September and October, before it was cold enough
for widespread snowfall. The one case in December (University of Houston)
was not contaminated by snow through visual inspection of the true color
(RGB) images of VIIRS or GOES, which are available on the AerosolWatch
website (
The diurnal pattern of biases in GOES-16 ABI AOD at GSFC on 2
additional clear days:
The diurnal pattern of biases is also found to be different on different days. As an example, Fig. 2 shows the diurnal bias at GSFC on two additional days in 2018, 18 October and 30 October. Although the peak of the bias occurs at approximately the same time on both days, around 17:00 UTC, the magnitudes of the peaks are different. On 12 October (Fig. 1a) the maximum ABI AOD is about 0.25, while it is 0.2 on 18 October (Fig. 2a) and only 0.1 on 30 October (Fig. 2b).
Scatterplots of surface reflectance on 0.47
To further illustrate the reasons that cause the diurnal variation of the
ABI AOD biases, atmospheric corrections were performed to obtain the surface
reflectance at different times and days for the pixels near GSFC site, i.e.,
at 17:02 and 20:02 UTC on 12, 18, and
30 October. The atmospheric correction uses the LUT from the ABI AOD
retrieval and the input of the TOA reflectance from ABI, geometries, and
AERONET AOD, along with the assumptions of standard column ozone, water
vapor, and surface pressure. Because there are four aerosol models in the
LUT, the four surface reflectance values were averaged. In the ABI AOD
retrieval algorithm, 0.47 and 2.2
In the scatterplots, the average of the 3 d of solar zenith angle is
used to calculate the coefficients of the linear relationships for each time
step for illustration purposes because the solar zenith angles are close in
value for the 3 d at each time step with about
At 17:02 UTC on 30 October 2018, nearly all the pixels fall into the
category of 0.3
Comparing the two time steps, pixels have lower NDVI at 20:02 UTC than those
at 17:02 UTC on the same days. The surface reflectance is significantly
lower at 20:02 UTC, i.e., with mean surface reflectance reduced from 0.06 to
0.04 in the 0.47
The change in NDVI between 12, 18, and 30 October is most likely due to the change in the colors of the vegetation during fall, when the leaves of trees turn reddish. Within the same day, due to the change in geometry, NDVI changed. It should be pointed out that even though at 20:02 UTC the surface relationships used are close to reality, there is still a lot of scatter in the individual pixels. This can introduce pixel level uncertainty that cannot be observed when averaged over the area around AERONET site.
Now that the source of the diurnal bias in ABI AOD has been identified, the
next step is to develop an algorithm to correct it by taking advantage of
the special characteristics of geostationary satellites. Because the GOES-16
satellite is stationary, the locations of the image pixels are fixed and the
satellite zenith and azimuthal angles remain unchanged. In addition, the
solar zenith and azimuthal angles at a given time of day change little
during a relatively short time period (
We designed a GOES-16 ABI AOD bias correction algorithm similar to the GASP AOD retrieval algorithm. However, instead of reflectance space, the composite bias correction algorithm works in AOD space. The basic idea to derive the ABI AOD bias is that the minimum of a month's ABI AOD at each time step should be close to the background AOD. Therefore, deviation of the minimum of ABI AOD retrievals during the 1-month period from the AERONET-derived background AOD is assumed to represent a systematic bias. The AOD bias at higher AOD load is estimated to be the same as the one obtained at the background AOD, which will be proved in Sect. 6 through radiative transfer simulation. The AOD bias can then be removed from the original ABI AOD by subtraction.
The flowchart of the algorithm is shown in Fig. 4. GOES-16 ABI AOD top 2 qualities, i.e., high quality and medium quality, are used to generate the bias curves in the algorithm, because the criteria for high-quality AOD are very conservative and the standard deviation test that moves data from high quality to medium quality is very stringent and throws away a lot of good retrievals. The top 2 qualities data have much larger area coverage than the high-quality data alone. For example, it is not possible to build a bias curve for pixels near CCNY using high-quality AOD data as there are too few data points, as seen in Fig. 1.
Flow chart of the ABI AOD bias correction algorithm.
In the bias correction algorithm, ABI AOD (top 2 qualities) over the CONUS with 5 min temporal resolution is first aggregated into 15 min temporal resolution. This is because GOES can operate in different modes and the observation times are different for different modes, even though the time interval between the time steps stays the same for the CONUS region. Averaging AOD into 15 min intervals reorganizes the AOD data into regular time steps. In addition, averaging AOD also increases data coverage at each time step. At each time step, the algorithm loops through a 30 d period to look for the lowest AOD for each pixel. In this work, the 30 d time period was selected based on Prados et al. (2007). For real-time bias correction, the most recent past 30 d are used because future AOD observations after the date of interest are not yet available. If the bias correction is being done as part of reprocessing, such that all the AOD data after the date of interest are available, a 30 d period is used with the date of interest placed at the center; this period may estimate the AOD bias more accurately. As shown in Knapp et al. (2005), the optimal time period to obtain a clear day background is not fixed and is dependent on seasons.
Once the optimum 30 d period has been selected, the bias at each pixel and
at each time step is estimated using the lowest AOD during the 30 d period
minus the background AOD. The background AOD over the CONUS area is obtained
through an analysis of multi-year AERONET AOD data using the method
described in Zhang et al. (2016). The main steps are summarized here for
reference. At each AERONET site
An example of a 2 km pixel close to GSFC is shown in Fig. 5, where AOD is plotted as a function of time for the 30 d period from 12 September to 11 October 2018. The AOD lower bound is derived from the time period and is shown as red curve. The bias is estimated using the lower bound minus the background AOD of 0.025. It is then subsequently used to correct the bias for that pixel for 12 October 2018.
AOD at a pixel close to GSFC over time period of 12 September–11 October 2018 (blue dots) vs. UTC time and the lower bound of the AOD (red curve).
GOES-16 ABI AOD data and AERONET AOD data for the time period from 6 August to 31 December 2018 are used to validate the bias correction
algorithm. The diurnal bias of ABI AOD data across the CONUS domain was
corrected using the algorithm described in Sect. 4 and compared to
coincident AERONET AOD. The original ABI AOD and the bias-corrected ABI AOD
were matched with AERONET AOD using the following criteria: (1) ABI AOD are
averaged within the circle of 27.5 km radius around an AERONET site,
requiring at least 120 valid AOD pixels within the circle; (2) AERONET AOD
are averaged within
For the first 30 d of the validation period (6 August to 4 September), the bias correction curves are derived from the same 30 d period. For the remainder of the validation period, the bias correction curves are derived from the 30 d period immediately prior to the day of interest.
Figure 6 shows scatterplots of GOES-16 ABI AOD vs. AERONET AOD for high-quality ABI AOD and top 2 qualities of ABI AOD, before and after bias correction,
averaged over the entire validation period and across the CONUS domain.
Scatterplots for both high quality and top 2 qualities are shown, although
the bias curves were derived using the top 2 qualities data. In order for a
valid comparison, the AOD pixels in the plots have one-to-one correspondence
before and after bias corrections, i.e., the quality flag does not change and
all the pixels are kept even though some of them may be below the lower
bound of the operational GOES-16 ABI AOD product (
Scatterplots of GOES-16 ABI AOD vs. AERONET AOD for 6 August 2018
to 31 December 2018 across the CONUS domain:
Validation statistics for comparisons between GOES-16 ABI AOD (top 2 qualities) and AERONET AOD at the six representative AERONET sites listed in
Table 2 for 6 August 2018 to 31 December 2018 across the CONUS domain,
both before and after bias correction.
Table 3 shows validation statistics for GOES-16 ABI AOD vs. AERONET AOD at the six AERONET sites listed in Table 2. After applying the bias correction, most of the statistics for ABI AOD improve at the six sites, demonstrating the success of the bias correction algorithm. For example, five out of six sites have RMSE improved to 0.05 or below. The exception is the University of Houston site, where the RMSE is still as high as 0.08 after correction, although it is improved from 0.19. This result may indicate there is still some bias left uncorrected at this site due to its complicated surface with respect to geometries. The sites in the eastern US have a geometry symmetric to the local noon and therefore the AOD biases are symmetric to the local noon. The sites in the western US do not have such symmetry and therefore the splitting of parameterization at noon and using second-order polynomials may introduce some errors. The complexity of surfaces over the University of Houston can be seen in Fig. 1e, where two AOD bias peaks are observed, one in the morning and the other at noon, indicating that the diurnal variation of surface reflectance relationship is different from the other sites, such as GSFC and CCNY, where AOD biases only peak at noon.
Figure 7 demonstrates the scattering angle dependence of the ABI AOD errors
for high quality and top 2 qualities. It can be seen that the errors before
bias correction have strong scattering angle dependency: AODs have positive
bias when the scattering angle is greater than 110
Comparisons of ABI AOD error vs. scattering angle before
and after bias correction for
To evaluate the performance of the algorithm for a range of AODs, Fig. 8 shows the ABI AOD error and standard deviation in different AERONET AOD bins, with an equal amount of matchup data in each bin. For high-quality AOD, bias correction reduces bias in the highest two AOD bins, with a center around 0.3 and 0.57. In the range [0.1, 0.3], bias correction overcorrects and introduces negative mean bias with slightly larger magnitude than the original mean bias, around 0.01 in magnitude difference. In the range [0,0.1], AOD mean biases are close to zero both before and after correction, but the bias correction AOD error has a smaller standard deviation. For the top 2 qualities ABI AOD, bias correction reduces the bias for all ranges of AODs with slight over corrections of magnitude of about 0.02 when AOD is greater than 0.1.
Comparisons of ABI AOD error vs. AERONET AOD before and
after bias correction for
Monthly mean AOD (top 2 qualities) for September 2018 at three time steps, i.e., 15:00, 17:00, and 20:00 UTC, for the original ABI AOD (left column) and bias-corrected AOD (right column).
Figure 9 shows the monthly mean AOD for September 2018 at three time steps,
i.e., 15:00, 17:00 and 20:00 UTC. At each time step AOD is first
composited within
Maps of correlation coefficients, mean biases, and RMSEs at AERONET sites with more than 400 matchups for the time period 6 August–31 December 2018 for the original ABI AOD (top 2 qualities) vs. AERONET AOD (left column) and for the bias-corrected ABI AOD (top 2 qualities) vs. AERONET AOD (right column).
Figure 10 shows the maps of the statistical metrics over AERONET sites with
more than 400 matchups for the correlation coefficients, mean biases, and
RMSEs for the original ABI AOD (top 2 qualities) vs. AERONET AOD and for the
bias-corrected ABI AOD (top 2 qualities) vs. AERONET AOD. As can be seen,
over most of the sites, the performances of bias-corrected AOD improve
compared to the original AODs. In the original ABI AOD, no geographical
pattern of the performances is observed. Especially noteworthy is that AOD
retrievals for some sites that are very close to each other have very
different performance metrics. There are no AERONET matchups in the western
US because the ABI AOD restricts the satellite-view zenith angle to those
below 60
Most of the sites with high bias (around 0.1 or above) and RMSE (around 0.15
or above) before bias correction are urban sites. For example, Tucson,
University of Houston, and CCNY, which have already been shown in previous
analysis. There are two sites in Florida that have high RMSE, one is Key Biscayne
(25.732
Overcorrection, under-correction, and/or reduction of correlation are
observed at several sites. For example, at NEON_TALL
(32.950
The same as Fig. 1 but after correcting the GOES-16 ABI AOD for the diurnal bias.
Figure 11, analogous to Fig. 1, shows the time series comparisons between
bias-corrected ABI AOD and AERONET AOD for clear days at the same
representative AERONET sites used in Fig. 1. Almost all of the large
biases in Fig. 1 are reduced to a magnitude
Figure 12 shows maps of the top 2 qualities of ABI AOD over the Northeast US at 17:42 UTC on 18 October 2018 before (Fig. 12a) and after (Fig. 12b) bias correction, illustrating the effects of the bias correction on observed ABI AOD. The black areas in the figures are locations where no AOD was retrieved, primarily caused by cloud coverage. This is a clear day, with no major sources of ambient atmospheric aerosols. However, before the bias correction, Fig. 12a shows that the ABI AOD field is noisy, due to the effects of the surface reflectance on the AOD retrievals. For example, over New York City, uncorrected ABI AOD values are as high as 0.5, while the coincident AERONET AOD measurement at the CCNY site is only 0.02. After the bias correction, Fig. 12b shows that the ABI AOD field is mostly cleared from the surface effects. Some isolated pixels of slightly higher AODs are still observed in the bias-corrected ABI AOD map, which likely originate from cloud contamination, with a few due to incomplete bias correction caused by outliers in fitting the bias correction with a second-order polynomial. For comparison, Fig. 12c and d show MODIS AOD dark target and deep blue retrievals from Aqua for this day with an overpassing time 17:55 UTC. The bias-corrected high- and medium-quality ABI AOD compares well with MODIS deep blue AOD in both magnitude and data coverage. MODIS dark target AOD has much less data coverage, but ABI AOD also compares well in magnitude in the areas with MODIS dark target AOD data.
Maps of GOES-16 ABI AOD, top 2 qualities (high and medium), over
the northeastern US at 17:42 UTC on 18 October 2018
Histograms of original (uncorrected) and bias-corrected GOES-16
ABI AOD at the
Figure 13 shows histograms of original (uncorrected) and bias-corrected ABI
AOD pixels over the areas within a 27.5 km radius circle around the CCNY
AERONET site (Fig. 13a) and the Wallops AERONET site (Fig. 13b) at 17:42 UTC on 18 October 2018 (the same observation time as the AOD data shown in
Fig. 12). At the urban CCNY site, ABI AOD before bias correction ranges
from 0 to 0.5, with an average of 0.25, which is much higher than the
AERONET AOD value of 0.02. After correction, the ABI AOD distribution
narrows down to a very small range with a peak and average at 0.02 – the
same value as AERONET. Wallops is a site with mixed pixels of rural, small
town, and water, and therefore its surface is darker and more favorable for
AOD retrievals. Figure 13b shows that uncorrected ABI AOD at the Wallops
site ranges from
Time series of original (uncorrected) GOES-16 ABI AOD, bias-corrected ABI AOD, and AERONET AOD at the CCNY AERONET site on
Statistics of original (uncorrected) ABI AOD and AERONET AOD retrievals at the CCNY AERONET site on 15 and 16 August 2018 for the four aerosol models used in the ABI AOD algorithm.
As hypothesized, the results discussed thus far suggest that the surface reflectance parameterization in the ABI AOD algorithm is the main source of the diurnal bias when ABI AOD is close to zero. However, when AOD is higher, such as during periods of high aerosol concentration, the aerosol model in the ABI AOD algorithm becomes a larger source of bias. As an example, a case with a moderate aerosol loading is examined. On 15–16 August 2018, smoke aerosols were transported to the New York City metropolitan area from wildfires burning in the western US and Canada, resulting in AERONET AODs in the range of 0.4–0.7 at the CCNY site. As shown in Fig. 14, the bias-corrected ABI AOD is very close to the AERONET AOD on 15 August (Fig. 14a), but much lower than the AERONET AOD on 16 August (Fig. 14b). To investigate the reason for this discrepancy in the bias-corrected ABI AOD, the statistics of the ABI AOD retrievals were examined for the 18:12 UTC time step. These statistics are listed in Table 4 for the original ABI AOD pixels within a 27.5 km radius circle of the CCNY AERONET site, which are involved in the average of the matchup with AERONET AOD. AERONET AOD increases from 0.35 on 15 August to 0.55 on 16 August, but the uncorrected ABI AOD remains the same on 16 August as on 15 August. The reason for this discrepancy is that the aerosol models retrieved within the 27.5 km circle are not the same between the 2 d. Table 4 indicates that on 15 August the dust model was retrieved primarily (46 %), but on 16 August the urban aerosol was predominant. This aerosol event in August 2018 was dominated by smoke, so it is surprising that the ABI AOD algorithm did not select the smoke model a majority of the time on these days. The results for ABI AOD in this case are not unprecedented. The selection of the aerosol model in AOD retrievals over land sometimes does not perform very well in the NOAA VIIRS AOD retrieval either, e.g., over China (Huang et al., 2016; Wang et al., 2020). The ABI retrieval uses only four aerosol models for retrieval over land, and the real model may be different from every one of them. Wagner et al. (2018) showed that smoke often carries dust and therefore the aerosol may be a mixture of smoke and dust, which makes the aerosol selection in the AOD retrieval algorithm more challenging, especially because we do not have LUTs for mixtures of aerosols.
Uncertainties in the bias correction algorithm can also be caused by the geometry change within the 30 d period. During the 30 d period, the position of the Sun and therefore the solar geometry does change for a given time step. Hence, the surface reflectance relationship and AOD bias are not constant in the time period. The magnitude of AOD bias variation during the time period determines the magnitude of the uncertainty of the algorithm. Besides the change in solar geometry, the surface vegetation color change during seasonal variation may also be a source of uncertainty through its influence on surface reflectance relationships. The choice of 30 d time period to search for lowest AOD in a given pixel was made with extensive research done to minimize the solar zenith angle changes and maximize the chance of finding the lowest AOD (Prados et al., 2007).
To demonstrate the effectiveness of the bias correction algorithm and its
general applicability, we tested it on an independent geostationary
satellite AOD product. The bias correction algorithm is applied to the DT
ABI AOD provided by NASA (Gupta et al., 2019). The data used in this study
is for the time period of July 2019. Figure 15 shows the scatterplots of DT
ABI AOD vs. AERONET AOD before and after the bias correction for AERONET
sites over CONUS. The original ABI AOD has a correlation of 0.91, mean bias
of 0.07 and RMSE 0.09. After the bias correction, the correlation improves
to 0.93, the mean bias reduces to
Scatterplot of NASA's dark target ABI AOD vs. AERONET AOD for
July 2019 over CONUS:
Diurnal time series over three selected AERONET sites and days
with low AERONET AOD for NASA ABI AOD:
Figure 16 shows the diurnal variation of the ABI AOD before and after bias correction for three AERONET sites on the days with low AERONET AOD, i.e., GSFC on 13 July 2019, Tucson on 4 July 2019, and CCNY on 1 July 2019. All three sites show a diurnal variation of the AOD biases. Although Tucson and CCNY only have retrievals at certain times of the day, the upward trend in the morning at CCNY and downward trend in the afternoon at Tucson of dark target ABI AOD are similar to what has been observed in NOAA's ABI AOD product in Fig. 1. GSFC has a smaller magnitude of peak at noon than the other two sites, but there is an overall positive bias. The diurnal variation at GSFC is also similar to NOAA ABI data shown in Figs. 1 and 2. After the bias correction, biases at all the three sites are reduced. The examples here demonstrate that the biases observed in NOAA's ABI AOD product also exist in other geostationary satellite AOD products because the underlying fundamental question is how well the algorithms can account for surface reflectance contributions to the observed top of the atmosphere (TOA) reflectances. The procedures developed for polar-orbiting satellites that worked so well are not adequate for geostationary satellite geometries. Either the spectral surface reflectance relationships need to be frequently updated in the retrieval algorithms or external empirical bias correction to AOD need to be applied.
A diurnal AOD bias variation pattern was also observed over Asian land surface and over ocean when the DT algorithm was applied to Himawari 8 AHI geostationary satellite data (Gupta et al., 2019). The AHI AOD retrieved from the DT algorithm is found to be higher in the morning and lower in the afternoon compared against the daily mean. The biases are observed to be as high as 0.2 and are more serious over ocean for high solar zenith angles. They speculate that the problem may be caused by the errors in radiative transfer code that do not fully account for the curvature of the Earth. Although they claim that they did not find any systematic artifact over land, such an artifact is expected because it exists in DT ABI AOD over CONUS, as shown in Fig. 16 at Tucson and CCNY. Because the bias found in DT AHI AOD is a systematic error, the bias correction algorithm can also potentially be applied on that product, even if it is caused by radiative transfer model.
NOAA generates AOD products from its polar-orbiting and geostationary
satellites for operational use by the National Weather Service and
the Environmental Protection Agency (EPA) air quality monitoring and
forecasting applications. For air quality applications, AOD is often used as
a proxy for surface PM
Scatterplots of hourly PM
In this section, a further analysis of the behavior of surface reflectance bias and its effect on AOD is performed to demonstrate that it is the source of the AOD bias and the validity of the bias correction algorithm. A radiative transfer simulation is performed using 6SV (Kotchenova et al., 2006; Kotchenova and Vermote, 2007) to demonstrate the equivalence of bias correction in AOD and surface reflectance bias reduction.
The surface reflectance relationships used in the operational ABI AOD
retrieval algorithm are described in the following equations (ABI AOD ATBD,
2018):
Surface reflectance relationship coefficients used in Eqs. (4) and (5) (Tables 3–12 of ABI AOD ATBD, 2018).
The coefficients are obtained using a training dataset of a full disk
ABI–AERONET matchup in the time period of 29 April 2017–15 January
2018. The reflectances used as training data to generate Eqs. (4) and
(5) were cleared for clouds, screened for low AODs (
Because the 0.47
Surface reflectance error at 0.47
A radiative transfer simulation study is performed to investigate the AOD
retrieval biases due to the surface reflectance errors. A forward
calculation is first performed to obtain TOA reflectance with a set of
parameters: surface reflectance at 0.47
Parameters used in AOD uncertainty simulation due to the surface reflectance uncertainty.
The AOD biases obtained in each configuration are grouped by input AOD, surface reflectance, and surface reflectance bias. The mean and standard deviation are calculated and the results are shown in Fig. 19. As expected, a negative surface reflectance error introduces a positive AOD error. The corresponding mean AOD bias does not change much with respect to AOD load when AOD is small (less than or equal to 0.5). However, there is a positive increase in the mean bias and a larger standard deviation when AOD is 1.0. This is due to the uncertainty in aerosol model selection. This can be seen in Fig. 19, where surface reflectance bias is 0 and the AOD bias is exclusively coming from aerosol model selection error, which tends to give a positive mean AOD bias (about 0.06) and a larger standard deviation (about 0.4).
AOD retrieval uncertainty due to the uncertainty of surface
reflectance at 0.47
One can show that the bias correction procedure proposed in this work is valid through this simulation study. In the bias correction algorithm, the AOD bias for a pixel at 0.025 background AOD load is obtained from a 30 d composite procedure, which corresponds to the simulated AOD bias when AOD is 0.025. As shown in Fig. 19, the AOD biases at higher AOD are of the similar magnitude as that at 0.025 background AOD if the surface reflectance bias is the same, especially for the negative bias of surface reflectance. For a given pixel, the surface reflectance bias originating from the surface reflectance model is assumed to remain constant during the 30 d period and does not change with AOD load. Therefore, AOD biases at higher AOD load can be estimated by the AOD bias obtained at background AOD of 0.025.
In our validation work of GOES-16 ABI AOD, we noticed a substantial diurnal bias in AOD that needed to be fixed for our operational users. Analysis shows that the bias is caused by errors in the land surface reflectance relationship between the spectral bands used in the ABI AOD retrieval algorithm. To remove the biases, an empirical algorithm is developed that utilizes the lowest AOD in a recent 30 d period in conjunction with the background AOD to derive a smooth bias curve at each ABI AOD pixel. The ABI AODs are then corrected by subtracting the derived bias curves at each time step.
The bias correction algorithm is validated for 5 months of GOES-16 ABI AOD data through comparisons against coincident AERONET AODs. The results demonstrate that the bias correction algorithm works successfully for the top 2 qualities of ABI AODs and that the correlation with AERONET AOD, average bias, and RMSE all improve. As a result of the bias correction, top 2 qualities ABI AOD performs as well as uncorrected high-quality ABI AOD. Therefore, bias-corrected top 2 qualities ABI AOD data are recommended for use in research and operations. The bias-corrected AODs cover twice the area of high-quality ABI AOD data alone with the same accuracy.
The ABI AOD bias correction process is most effective when AOD is low because under those conditions, the surface reflectance relationship is the main source of uncertainty in the ABI AOD retrieval. When AOD is higher, the uncertainty from the aerosol model selection in the ABI AOD retrieval algorithm becomes as large as or larger than that from the surface reflectance relationship, and therefore the bias correction for high AOD conditions is not as effective as that for low AOD conditions.
The surface reflectance relationships in the ABI AOD retrieval algorithm will be improved when more GOES-16 data are accumulated and analyzed. However, these relationships are based on AERONET sites and they are statistical models. Therefore, individual AOD pixels will always suffer to some degree from deviation in the statistical relationship and some bias will always exist, although it may be reduced by a more accurate surface reflectance relationship. Hence, future versions of the GOES ABI AOD product may still benefit from applying the bias correction algorithm, unless the AOD retrieval algorithm uses pixel-level surface reflectance relationships that are routinely updated. Such an exercise in an operational setting is prohibitive.
The bias correction algorithm has a general applicability. It can also be applied to other geostationary AOD products, which may also suffer the bias described in this research, especially if the AOD algorithms are similarly relying on deriving surface reflectance from relationships between the blue band and short-wave infrared (SWIR) band. We tested and demonstrated that the performance of NASA's dark target ABI AOD is improved by applying the bias correction algorithm. The existence of bias in NASA's dark target algorithm indicates that the bias issue is a more general problem rather than only existing in NOAA's ABI AOD product. Therefore, other geostationary AOD products can benefit by applying the bias correction technique introduced in this research.
GOES-16 ABI AOD can be obtained at NOAA CLASS (
HZ worked on the developing and analyzing activities described and led the manuscript writing. MZ worked on surface reflectance relationship analysis. SK and IL are co-leads of the aerosol algorithm development and guided the work. SK, IL, and MZ reviewed the algorithm science and analysis and contributed to the paper revisions. MZ and IL provided the AOD retrieval code that is used in the atmospheric correction for the surface reflectance analysis.
The authors declare that they have no conflict of interest.
The contents of this paper are solely the opinions of the authors and do not constitute a statement of policy, decision, or position on behalf of NOAA or the U.S. Government.
The authors thank the AERONET principal investigators and site managers for providing the data used in this work and Amy Huff (IM Systems Group) for providing internal review. The authors thank the NASA DT team (Rob Levy and Pawan Gupta) for providing their DT GOES-16 ABI AOD retrievals.
This research has been supported by the NOAA GOES-R program.
This paper was edited by Andrew Sayer and reviewed by three anonymous referees.