An Evaluation Of MODIS-Retrieved Aerosol Optical Depth Over A Mountainous AERONET Site In The Southeastern US

The literature shows that aerosol optical depth (AOD) derived from the MODIS Collection 5 (C5) dark target algorithm has been extensively validated by spatiotemporal collocation with AERONET sites on both global and regional scales. Although generally comparing well over the eastern US region, poor performance over mountains in other regions indicate the need to evaluate the MODIS product over a mountain site. This study compares MODIS C5 AOD at 550nm to AOD measured at the Appalachian State University AERONET site in Boone, NC over 30 months between August 2010 and September 2013. For the combined Aqua and Terra datasets, although more than 70% of the 500 MODIS AOD measurements agree with collocated AERONET AOD to within error envelope of ± (0.05 + 15%), MODIS tends to have a low bias (0.02–0.03). The agreement between MODIS and AERONET AOD does not depend on MODIS quality assurance confidence (QAC) value. However, when stratified by satellite, MODIS-Terra data does not perform as well as Aqua, with especially poor correlation (r = 0.39) for low aerosol loading conditions (AERONET AOD less than 0.15). Linear regressions between Terra and AERONET possess statistically-different slopes for AOD < 0.15 and AOD ≥ 0.15. AERONET AOD measured only during MODIS overpass hours is highly correlated with daily-averaged AERONET AOD. MODIS monthly-averaged AOD also tracks that of AERONET over the study period. These results indicate that MODIS is sensitive to the day-to-day variability, as well as the annual cycle of AOD over the Appalachian State AERONET site. The complex topography and high seasonality in AOD and vegetation indices allow us to specifically evaluate MODIS dark target algorithm surface albedo and aerosol model assumptions at a regionally-representative SE US mountain site. Sherman, J., Gupta, P., Levy, R., & Sherman, P. (2016). An Evaluation of MODIS-Retrieved Aerosol Optical Depth over a Mountainous AERONET Site in the Southeastern US. Aerosol and Air Quality Research, 16: 3243– 3255. doi: 10.4209/aaqr.2015.09.0568. Publisher version of record available at: https://app.dimensions.ai/ details/publication/pub.1072361174 Aerosol and Air Quality Research, 16: 3243–3255 ISSN: 1680-8584 print / 2071-1409 online doi: 10.4209/aaqr.2015.09.0568 An Evaluation of MODIS-Retrieved Aerosol Optical Depth over a Mountainous AERONET Site in the Southeastern US James P. Sherman, Pawan Gupta, Robert C. Levy, Peter J. Sherman 1 Department of Physics and Astronomy, Appalachian State University, Boone, NC 28608, USA 2 NASA Goddard Space Flight Center, Greenbelt, MD, USA 3 Department of Aerospace Engineering, Iowa State University, Boone, IA, USA


INTRODUCTION
Due to sparse sampling by ground monitors, satellite remote sensing of aerosol optical depth (AOD) is used for both air quality and climate applications. To be useful for these applications on a local scale, one needs to characterize how well a satellite product represents the daily average AOD, as well as the seasonal and annual AOD cycles. To quantify this information, one begins by comparing the satellite product to ground-truth observation at a site that is sufficiently representative of the region. For validation of such satellite-retrieved AOD, it is common to rely on collocated measurements by ground-based sunphotometers, such as those provided by NASA's Aerosol Robotic Network (AERONET).
A well-known satellite dataset is obtained by the "darktarget" (DT) algorithm that retrieves AOD from spectral reflectance observed by MODerate resolution Imaging Spectro-radiometers (MODIS) aboard the polar-orbiting Terra and Aqua satellites (Levy et al., 2007a). The "Collection 5" (C5) DT product covers the entire lifetime of the two MODIS sensors (since 2000 for Terra, and 2002 for Aqua), and covers both global oceans and dark surfaces (primarily vegetated) over land. Levy et al. (2010) compared the MODIS C5 product with hundreds of AERONET sites around the globe, to derive error estimates (EE) for AOD (at 550 nm). Over land, this meant that 66% (or approximately one standard deviation) of all high-quality retrievals of AOD, matched with AERONET-observed AOD within EE of ±(0.05 + 15%).
Although Levy et al. (2010) and others (e.g., Hyer et al., 2011) demonstrated the global "validation" of MODISretrieved AOD, these studies also picked out regions and conditions for which the AOD product did not meet the requirements of accuracy and correlation with AERONET data. In general, the MODIS C5 product compared well for AERONET sites over the eastern United States, western Europe, and other regions with vegetation and relatively flat surfaces (Levy et al., 2010). On the other hand, there was poorer correlation for brighter surfaces, including semi-arid, urban, and mountain sites. There were no mountainous eastern US AERONET sites during the time period studied by Levy et al. (2010).
The DT algorithm assumes that surface albedo (and surface reflectance) is characteristic of vegetation (which appears dark for visible wavelengths) and that it can be easily constrained. However, for the less vegetated (semi-arid or urban), and/or complex terrain (e.g., mountains), the DT assumption may be violated. We hypothesize that the complex terrain of the Appalachian Mountains may introduce challenges for the MODIS spectral surface albedo model (Levy et al., 2007b). Surface assumption errors represent the largest source of error for low AOD less than ~0.15 (Levy et al., 2010), which occurs for most non-summer months over a majority of the southeastern US.
After the Levy et al. (2010) publication and other global validation studies, an AERONET site was deployed on the campus of Appalachian State University in Boone, NC. For the first time we can assess the performance of the MODIS DT algorithm over a site that represents the mountainous Southern Appalachian region of the southeastern US. At the same time, we determine whether the MODIS data are representative of daily mean AOD, and assess whether sampling of the MODIS retrievals (less than daily due to clouds and other obstructions) are able to capture the annual variation of AOD.
The current study compares the MODIS C5 AOD product to spatiotemporally collocated AOD measurements from the Appalachian_State AERONET site to evaluate the performance of the MODIS AOD product above the mountainous site. Although a new collection (Collection 6) of the MODIS product is now available with some upgrades to the DT algorithm (Levy et al., 2013), there has not yet been comprehensive global evaluation analogous to Levy et al. (2010). Therefore, we report only C5 products in this paper and briefly discuss how the results may be different in C6, based on initial C6 studies (Levy et al., 2013). MODIS temporal resolution is then evaluated by two means: (1) the correlation and level of agreement between AERONET AOD measured during MODIS overpass hours with daily-averaged AERONET AOD; and (2) the ability of monthly-averaged MODIS AOD to track AERONET over the 3+ year study period.

AppalAIR Site
Established in 2009, the Appalachian Atmospheric Interdisciplinary Research facility (AppalAIR) at Appalachian State University in Boone, NC (36.21°N, 81.69°W, 1080 m asl) is home to the only AERONET site in the Appalachian Mountain region and the only collocated NASA AERONET and NOAA Earth System Research Laboratory (NOAA-ESRL) aerosol monitoring sites in the entire eastern US. The region is heavily forested and possesses a diversity of elevations (< 300 m to > 2000 m) and a variety of weather regimes (e.g., winter storms, convective cells, dying tropical cyclones, and stagnant summertime episodes). The region also includes a diversity of anthropogenic and biogenic aerosol sources. Lower tropospheric aerosol light scattering and absorption measured at AppalAIR is dominated by particles with diameter less than 1 µm  and sub-1 µm aerosol mass consists primarily of organics, with lower levels of sulfates (Link et al., 2015). Summer aerosol optical depth in the southeastern US is influenced by isoprene-derived secondary organic aerosol (Goldstein et al., 2009). A biomass-burning influence is also present in winter aerosol mass concentrations measured at AppalAIR (supplement to Link et al., 2015), likely due to residential wood-burning in the region.

AERONET Aerosol optical Depth and Ångström Exponent
The CIMEL sunphotometer, deployed at the AppalAIR site (known as 'Appalachian_State' within the AERONET network), collected data over 30 months during the period August 2010-September 2013. There are no data available between Nov. 2011-May 2012 and for Oct. 2012, due to calibrations and instrument maintenance. The CIMEL measures direct solar radiance at nine wavelengths (λ = 340, 380, 440, 500, 670, 875, 940, 1020, and 1640 nm) and sky radiance at four of these wavelengths (λ = 440, 670, 870, and 1020 nm), using standard AERONET protocols (Holben et al., 1998). The direct solar radiance measurements are used to calculate AOD at each of the nine wavelengths except 940 nm using the Beer-Lambert-Bouguer equation (Holben et al., 1998). Direct solar radiance measurements are made at optical air mass intervals of 0.25, corresponding to every ~15 minutes near noon and more often near dawn and dusk. Only Level 2 AERONET AOD (cloud-screened, calibrated) is used in this study. The uncertainty for Level 2 AOD is small enough (0.01-0.02; Eck et al., 1999) so that AERONET serves as ground-truth for comparisons with satellite-derived AOD (Levy et al., 2010;Hyer et al., 2011).
Sky radiance measurements are used to derive columnaveraged aerosol properties including size distributions and single-scattering albedo (SSA). Single-scattering albedo can only be reliably retrieved to within ~0.03 for AOD (λ = 440 nm) ≥ 0.40 (Dubovik et al., 2000). This high-loading condition is only satisfied on 2-4 days per year at the Appalachian_State site and therefore AERONET SSA is not available to use in this study. Ångström exponent in the visible spectral range is typically computed as the slope of a linear fit of log (AOD) versus log (λ) using available wavelengths between 440-870 nm and is used in Sect. 4.4 as a semi-quantitative indicator of aerosol size.

MODIS Aerosol Optical Depth
The "Level 2" (derived-geophysical) MODIS aerosol product is derived at 10 km spatial resolution (at nadir), and known as MOD04_L2 (for Terra) and MYD04_L2 (for Aqua). Collectively, referred to as MxD04, the data used here (Aug 2010-Sept 2013) are products from consistent application of the DT retrieval algorithm (Levy et al., 2007a ,b), instrument calibration, and computer processing environment. Although the data are from Collection 5.1 (C51), the DT portion is identical to C5, so we refer to the set as Ç5. A short description of the algorithm, products, and validation follows here.
MODIS, aboard the Terra and Aqua polar-orbiting satellites, measures top-of-the-atmosphere (TOA) spectral reflectance or radiance in 36 channels ranging from visible to infrared wavelengths, with spatial resolution ranging from 250 m to 1 km. Terra (Aqua) crosses the equator going north to south (south to north) near 10:30am (1:30 pm) local solar time (LST). With wide swath (~2330 km), there is normally twice-daily overpass over Appalachian_State.
The DT algorithm attempts to interpret the contrast of aerosol (relatively bright) against the (dark) surface background. The retrieval algorithm (Levy et al., 2007b) works by comparing the observed spectral reflectance to a lookup table (LUT) that simulates possible surface/molecular/ aerosol scenarios. More specifically, the algorithm uses a subset of the spectral reflectance information to filter out cloudy pixels, and then aggregates remaining pixels into boxes that represent 10 km spatial resolution (at nadir). The (aggregated, 10 km) spectral reflectances in seven MODIS bands are used as the observations to drive the aerosol retrieval. These seven bands (Bands #1-#7 or B1-B7) are centered near 645, 855, 466, 553, 1243, 1628, and 2113 nm, respectively.
The LUT is represented by a prescribed aerosol model "type" (aerosol optical properties), along with a model of spectral surface reflectance appropriate for the regional vegetation indices and season. The prescribed aerosol "type" is one of three global aerosol models (Levy et al., 2007b), which has been assigned to each 1° latitude × 1° longitude grid point, as a function of season. These three aerosol types differ primarily in SSA, with a weakly-absorbing aerosol type (SSA ~0.95 at 553 nm) used to represent the eastern US (Levy et al., 2007b). As a "dark-target" retrieval, the algorithm attempts to retrieve when the observed reflectance at 2113 nm is between 0.01 and 0.25. For the surface properties, the algorithm makes a major assumption: specifically, that for primarily vegetated surfaces, the surface reflectance (that would be measured) in a shortwave-infrared (SWIR) MODIS channel (e.g., 2113 nm) is linearly correlated with surface reflectance in blue (466 nm) and red (645 nm) MODIS channels (e.g., Kaufman et al., 1997). Levy et al. (2007b) noted also that this VIS/SWIR relationship also depends on scattering angle and on surface greenness, and that surface greenness could be parameterized by the Normalized Differential Vegetation Index (NDVI- Karnieli et al., 2001) calculated using MODIS SWIR channels centered at 1243 nm and 2113 nm (Levy et al., 2007a).
In theory, the assumed surface reflectance relationships, coupled with a prescribed model of aerosol properties (aerosol type), provides enough constraint to retrieve the total aerosol loading, in addition to some estimate of the aerosol size. Therefore, the products of the retrieval include AOD (at 550 nm) and some qualitative measures of the aerosol size distribution. Note that due to uncertainties in estimating surface reflectance, it is possible to retrieve a negative (non-physical) AOD value (allowed down to -0.05). As long as there are enough non-cloudy pixels within the 10 km box and the retrieval inversion finds an acceptable solution (see Levy et al., 2007b for details), we have "confidence" in the retrieved AOD values. For conditions with fewer acceptable pixels, poor fitting to observed reflectance, or other retrieval issues, there is lower confidence in the retrieved product. Therefore, according to Levy et al. (2007b), each MODIS (10 km) AOD retrieval is assigned a quality assurance confidence (QAC) value ranging from 0 (lowest) to 3 (highest) (Levy et al., 2007b).
Since QAC value does not indicate accuracy of the retrieved AOD product, the MODIS team turned to collocation with ground-based AERONET data to validate the MODIS product. Following Ichoku et al. (2002), MODIS AOD uncertainty over land is estimated based on global comparisons with AERONET observations. For the C5 version of the MODIS dataset, Levy et al. (2010) determined that the error envelope (EE) was EE = ± (0.05 + 0.15 × AOD AERONET ) (1) Note that while AOD is determined at 10 km resolution, EEs are determined using averages of MODIS AOD retrievals over a 5 pixel by 5 pixel box, corresponding to 50 km × 50 km), centered at the AERONET site (Sect. 3). This reduces noise in the MODIS retrievals, as well as allowing for non-ideal representation of the area by the AERONET site. The primary sources of MODIS AOD retrieval errors over a region result from uncertainties in (1) surface reflectance; and (2) aerosol model (e.g., optical properties) used to construct the LUTs (Kaufman et al., 1997). Although sensor calibration drift and inadequate cloud screening also contribute to errors in MODIS AOD, we will concentrate on the first two sources. Following Levy et al. (2010) and recommendations for the use of MODIS Level 2 AOD (http://modis-atmos.gsfc.nasa.gov/ MOD04_L2/format.html), we retain negative MODIS AOD values down to -0.05 so as not to introduce an artificial positive bias under clean air conditions.

MODIS Surface Albedo
In addition to aerosol products, there are many algorithms to derive other geophysical parameters from the MODIS observations. One of these is the spectral surface reflectance product (Vermote and Kotchenova, 2008), known as MYD09A1 (derived from MODIS-Aqua data). MYD09 products are created by analyzing MODIS spectral observations over 8-day periods and identifying the invariant contributions (e.g., the surface). These products are gridded, reported at 500 m spatial resolution, and have their own quality assurance and error characteristics. Here, we concentrate on the same MODIS Bands 1-7 used for the aerosol retrieval. Each MYD09A1 pixel contains the best possible observation (with atmospheric correction applied) during an 8-day period as selected by high observation coverage, low view angle, absence of clouds and cloud shadow, and low aerosol loading.
To compare with the aerosol products (50 km × 50 km), we utilize a similar-sized box of MYD09A1 data centered at the Appalachian_State AERONET site. Only 8-day surface reflectance products with at least 50% of pixels in the 50.5 km × 50.5 km box passing MODIS quality assurance tests are used in this study and the mean surface reflectance of these pixels is calculated for each wavelength. In addition to surface reflectance in the seven bands that are compared with values used in the MxD04_L2 aerosol retrieval, we can estimate scene brightness (based on 2113 nm reflectance) and surface "greenness" ( based on 1243 nm and 2113 nm reflectance), defined by the NDVI_swir Values of NDVI swir greater than ~0.6 correspond to active "green" vegetation and values less than ~0.2 correspond to dormant or sparse vegetation (Levy et al., 2007a).

NOAA-ESRL Single-Scattering Albedo
Ground-based measurements of aerosol optical properties at the collocated NOAA-ESRL site can be used to evaluate the aerosol model assumptions in the DT retrieval. Specifically, monthly-averaged single scattering albedo (SSA) at 550 nm is derived from continuous sampling of aerosol light scattering and absorption coefficients from a 34 m tower at AppalAIR . In order to decouple the relative humidity (RH) dependence on light scattering (aerosol swelling), the aerosols are heated as needed to attain RH ≤ 40%. In-depth discussions of NOAA-ESRL aerosol sampling protocols, scattering and absorption coefficient measurements, and data analysis techniques are provided in Sheridan et al. (2001). A scanning humidograph (Sheridan et al., 2001) is employed to measure the RH dependence of light scattering coefficient but SSA is not corrected to ambient RH in this study, as column-averaged RH measurements were not available for most of the study period. This likely leads to an SSA under-estimation, which may be up to of ~0.02-0.03 during humid summer months (based on RH values from radiosonde launches at AppalAIR during summer 2013). Vertical profiles of aerosol attenuated backscatter measured by collocated micro-pulsed lidar indicate that most of the aerosols above this site are contained in the boundary layer (not shown) and monthlyaveraged SSA from the NOAA-ESRL site likely serves as a reasonable approximation of column-averaged SSA.

COLLOCATION STRATEGY
We apply a similar collocation strategy to that used by Levy et al. (2010) in their global validation of MODIS C5. We first interpolate AERONET AOD to match the MODISreported wavelength (550 nm) by applying a quadratic fit (on a log-log scale) to spectral AERONET AOD versus wavelength (Eck et al., 1999). We then use the method (Fig. 1) of spatiotemporal collocation similar to that described by Ichoku, et al. (2002). For each Terra and Aqua overpass, we calculate the mean MODIS AOD over a 50 km × 50 km box (5 × 5 MODIS Level 2 pixels) centered at the Appalachian_State AERONET site, to compare with the average of AERONET AOD measured within ± 30 minutes of MODIS overpass (typically 2-4 measurements). We also keep track of the MODIS QAC (confidence) value. The difference between our collocation method and that of Ichoku et al. (2002) is that we removed the restrictions that at least five MODIS pixels are used to calculate MODIS box-averaged AOD and that at least two AERONET AOD measurements are used to calculate temporally-averaged AOD. The agreement between MODIS and AERONET AOD did not degrade, yet there were nearly double the number of collocations (from 285 to 500 for QAC = 3 cases). For all QAC cases there are 581 "valid" MODIS/AERONET collocations that span the 30 data months (e.g., Table 1).
Following the logic of previous studies (e.g., Levy et al., 2010;Hyer et al., 2011), we stratify the collocations by QAC, by satellite sensor (Terra versus Aqua), and by a threshold for "moderate" aerosol loading (AOD = 0.15; Levy et al., 2010). Our 581 collocations are reduced to 566 if we require QAC ≥ 1 (moderate confidence), and to 500 if we require QAC = 3 (high confidence). The number of collocations is similar between Terra and Aqua (~290 for QAC ≥ 0). For the cases receiving QAC = 3, more than 80% are for AERONET reporting AOD < 0.15.
For each of the categories of stratification (rows in Table 1), we evaluate the performance of MODIS in capturing the AERONET AOD. We create a scatterplot and compute linear regression parameters (slope, intercept, and correlation coefficient), along with ninety-five percent confidence intervals for these parameters (Wonnacott and Wonnacott, 1981;Miller and Miller, 2012). Similar to other MODIS validation studies (Levy et al., 2010;More et al., 2013), we also calculate the percentage of MODIS AOD values lying within EE of AERONET AOD (Eq. (1)), along with the root-mean-squared error (RMSE), defined as the RMS difference in MODIS and AERONET AOD. MODIS AOD is 'validated' in this study if at least 2/3 of the spatially-averaged MODIS AOD retrievals lie within EE of the temporally-averaged AERONET retrievals, in addition to high correlation between the two (Levy et al., 2010). In addition to the overall regressions, we use the collocated data (500 points with QAC = 3) to calculate monthly-averaged AOD for each dataset. Here, we can identify MODIS measurement biases ( Fig. 2; Sect. 4.1) that are seasonally dependent.
Finally, we calculate monthly-averaged AOD using all MODIS and all AERONET measurements (independent of collocation) to assess MODIS ability to track monthlyaveraged AERONET AOD over the 3+ year study period ( Fig. 7(a); Sect. 4.6). Monthly-averaged AOD for the analysis in Sect. 4.6 is computed using daily-averaged AOD values. Daily-averaged MODIS AOD is the average of Terra and Aqua AOD if measurements from both satellites are available. If AOD on a given day is only retrieved by one satellite, that value is used as the daily-average MODIS AOD. Daily-averaged AERONET AOD is calculated as the average over all AERONET measurements for each day when three or more measurements are made (http://aerone t.gsfc.nasa.gov/new_web/data_description_AOD_V2.html).  Table 1 contains linear regressions for the MODIS/ AERONET inter-comparisons, for Terra and Aqua separately, and their combination. When all Terra & Aqua collocations with QAC = 3 are considered, MODIS AOD shows excellent agreement with AERONET, with high correlation (r = 0.84) and 70.80% of the MODIS AOD retrievals fall within the EE envelope given by Eq. (1). The regression equation is near 1-1, with a slope of 1.06 and a small (~0.03) negative MODIS AOD bias. Monthly-averaged MODIS and AERONET AOD calculated using the collocations (Fig. 2) illustrates that the MODIS AOD bias is fairly uniform (-0.02 to -0.03) for most months over the 3+ year period. A more negative MODIS bias is observed for some warm-season months of 2012 and 2013. Table 1 illustrates that MODIS/AERONET AOD agreement does not degrade when MODIS pixels with QAC < 3 are included in the calculation of box-averaged MODIS AOD. One exception is that the regression slope lies closer to one when only the highest quality pixels (QAC = 3) are used. Similar insensitivity to QAC is observed for Terra and Aqua individually as for their combination. Based on their global MODIS/AERONET inter-comparison, Levy et al. (2010) recommended restricting MODIS AOD usage to QAC = 3 for the highest-quality retrievals and strongly recommended against using QAC = 0 for any quantitative purpose. However, Levy et al. (2010) also acknowledged that "the use of lower confidence data should depend on the trade-offs between an application's tolerance for uncertainty and the spatial coverage requirements". For the rest of this paper, we focus on cases where MODIS QAC = 3 so as to maintain consistency with other published results (Levy et al., 2010;Hyer et al., 2011;More et al., 2013). However, since the MODIS/AERONET agreement seems insensitive to assigned MODIS QAC value, we see the potential for improved MODIS AOD sampling in the Southern Appalachian Mountain region. Levy et al. (2010) suggested that performance of MODIS  Terra and Aqua may differ, in that Terra appeared to have a negative bias since 2004. Hyer et al. (2011 quantified this more fully and found that retrievals of negative AOD were prevalent for low AOD conditions, and that there was a higher percentage of negative AOD retrievals for Terra than for Aqua. Over Appalachian_State, we find that 80% of the 240 Aqua AOD values are contained within the EE envelope (Eq. (1)) with high correlation (r = 0.90). However, only 62.31% of the 260 Terra AOD values are within the EE envelope. Rootmean-squared error (RMSE) is also better for Aqua (0.05) than for Terra (0.07). This is in spite of the fact that the MODIS/AERONET regression slopes and intercepts are slightly better for Terra (m = 1.01; b = -0.02) than for Aqua (m = 1.12; b = -0.03). A large majority of the Terra and Aqua AOD retrievals lying outside the EE envelope occur for low AOD and are biased low (Fig. 3), especially for Terra.

Dependence of MODIS/AERONET Agreement on Satellite and AOD
The pattern of negative MODIS AOD bias under clean conditions is more transparent when stratifying by AERONET AOD (e.g., Levy et al., 2010). We separate "low" and "high" AOD by AERONET AOD < 0.15 and ≥ 0.15, respectively. Combined MODIS Aqua & Terra AOD is poorly correlated with AERONET for low AOD (r = 0.56) and better correlated for high AOD (r = 0.82), as seen in Table 1. When separated, the high AOD cases are similarly correlated for both Terra and Aqua datasets (r = 0.82 and 0.85), but the correlation is much poorer for Terra than Aqua (r = 0.39 and r = 0.75) at low AOD. The linear regression slope difference between the low and high AOD cases is also much smaller for Aqua (m = 1.23 for low AOD versus m = 0.96 for high AOD) than for Terra (m = 0.70 for low AOD versus m = 1.31 for high AOD). In fact, based on applying 95% confidence tests, the slopes and correlation coefficients for the low and high Terra AOD cases are statistically different. The single linear model used in MODIS/AERONET inter-comparisons cannot be applied for Terra in this study. The poor correlation at low AOD, for the combined Aqua and Terra collocations, is almost entirely due to Terra. Levy et al. (2010) found no significant difference between AERONET/Terra agreement and AERONET/Aqua agreement in their global C5 validation study. However, they did report that Terra measured higher (lower) AOD than AERONET over land up until (after) 2004. The Terra AOD drift was attributed to radiance calibration drift, especially in the blue channel. This drift has been reduced for MODIS C6, but the low bias for Terra AOD over land is expected to persist (Levy et al., 2013). Evaluation of C6 data over Appalachian_State will require future study.

Dependence of MODIS/AERONET Agreement on DT Surface Assumptions
Monthly-averaged MODIS surface reflectance at 2113 nm (e.g., scene reflectance) and NDVI swir (from MYD09A1 data) are shown in Figs. 4(a) and 4(b) to examine possible roles of MODIS surface albedo assumptions on MODIS AOD accuracy. Scene reflectance of ~0.06-0.08 and NDVI swir ~0.60-0.70 during May-September are consistent with active, dark green vegetation in the heavily forested Southern Appalachian Mountain region. Scene reflectance of ~0.10-0.12 and NDVI swir values of ~0.30 during November-March are the result of somewhat brighter, less green vegetation during the dormant season. April and October represent transition months. Levy et al. (2010) reported that MODIS C5 DT AOD agreed best with AERONET when scene reflectance was 0.10-0.12 and when NDVI swir ~0.4. AOD was overestimated for brighter surfaces (by ~0.02 for scene reflectance of0.17) and underestimated over darker surfaces (by ~0.02  for scene reflectance of less than ~0.07). Retrieved AOD was biased high for less green surfaces (NDVI swir < 0.25) and biased low (by ~0.03) for greener surfaces (NDVI swir about 0.60-0.70). From these global generalities, one can use Figs. 4(a) and 4(b) to estimate the expected bias to the MODIS -retrieved AOD over our site. We estimate a MODIS AOD bias of -0.05 during May-September (-0.03 due to NDVI swir and -0.02 due to scene reflectance), which lies close to the observed monthly-averaged MODIS bias of -0.03 to -0.04 for a majority of these months (Fig. 2). For November-April, the values of scene reflectance and NDVI swir are closer to optimal for dark-target retrieval, so that biases (of Fig. 2) are smaller. Surface assumption errors are thus consistent with the negative AOD bias during May-September but not during November-April. Changes made as part of MODIS C6 are expected to increase AOD over vegetated surfaces including much of the Eastern US by ~0.02, due to correcting a C5 processing software bug in the assumed relationship between VISvs2113nm surface reflectance and NDVI swir over land (Levy et al., 2013).

Dependence of MODIS/AERONET Agreement on DT Aerosol Property Assumptions
Systematic biases in MODIS AOD also result from incorrect assumptions of aerosol type (optical properties) used in the LUT to retrieve AOD (Kaufman et al., 1997). These may be errors of size distribution (and resulting effect on spectral dependence of AOD), and/or errors of absorption characteristics (characterized by SSA). Levy et al. (2010) showed that the AOD retrieval error (MODIS-AERONET) tended to be smaller when the aerosol was dominated by fine-mode particles, as indicated by larger values of Ångström Exponent (AE). Since the range of monthly mean AERONET AE observed in our study (Fig. 4(c)) is indicative of being dominated by fine-mode particles, there is no expected systematic bias due to aerosol size. Based on studies such as Ichoku et al. (2002), the MODIS retrieval is expected to have a negative (positive) MODIS AOD bias where the algorithm overestimates (underestimates) aerosol SSA, The MODIS C5 LUT assigns a weakly-absorbing fine-mode dominated aerosol type (SSA ~0.95 at 553 nm) to the southeastern US during all seasons (Levy et al., 2007b). This assumption can be compared with in-situ measurements (dried; RH < 40%) of SSA at AppalAIR's NOAA-ESRL site. SSA ranges from 0.88 during the winter to 0.94 during the summer (Fig. 4(d)), After accounting for scattering hygroscopic factors, also measured at the site (not shown), SSA during the summertime (under typical RH of 70-80% in boundary layer), is closer to 0.95 as assumed by MODIS. However, a moderately absorbing fine-mode aerosol type (SSA ~0.90) is a more suitable choice for the LUT during September-March ( Fig. 4(d)). For C6, Levy et al. (2013) updated the retrieval to assume the moderately absorbing choice during the winter months, and we will evaluate the C6 results in a future study. Regardless, AOD in the winter tends to be low enough (Figs. 2 and 7(a)), such that errors in the aerosol model should not be the primary contributor to retrieval errors. It is possible, however, that assumption of SSA contributes to the observed ~-0.02 to -0.03 bias during September-March.

Diurnal Representativeness of MODIS AOD Measurements
Previous studies (e.g., Kaufman et al., 2000;Zhang et al., 2012) have considered whether AOD measured at the time of MODIS overpass is representative of daily-averaged AOD, and hence suitable for long-term climate or daily air-quality applications. On average (over 50-70 globallyspaced sites), MODIS overpass time is representative of the daily mean AOD ) However, diurnal variability is important for some sites (Zhang et al., 2012). Fig. 5 shows AERONET AOD at Appalachian_State, averaged over the MODIS Terra and Aqua overpass hours (local solar hours 10 and 13, respectively), compared to daily-averaged AERONET AOD in different seasons. Correlation coefficients are between 0.90-0.97 for all seasons and slopes are between 0.92-1.12 (Fig. 5), indicating that that daily measurements made by Terra and Aqua should be representative of the daily average AOD over Appalachian_ State. Yet, during all seasons, there is a diurnal cycle of AOD ( Fig. 6(a)), with dawn and/or dusk maxima and little AOD variability (~0.02 or less) during the intervening hours. Ångström exponent also demonstrates little diurnal variability ( Fig. 6(b)). Sherman et al. (2015) reported minimal diurnal variability for lower tropospheric aerosol light scattering coefficient and scattering Ångström exponent at the collocated NOAA-ESRL site.

Fig. 5.
Linear regressions of AERONET AOD (at 550 nm) averaged over the Terra and Aqua overpass hours (local standard hours 10 and 13, respectively) versus daily-averaged AERONET AOD. The thick solid black line is the best-fit line y = mx + b. The thin dotted black line represents the one-to-one line y = x. Lower and upper bounds listed for the linear regression parameters encompass 95% confidence intervals.