Assessment of second-generation MODIS aerosol retrieval (Collection 005) at Kanpur, India



[1] The second-generation MODIS aerosol retrieval (Collection 005) from EOS- Aqua (2002–2005) was evaluated using ground-based AERONET measurements at Kanpur (26.45°N, 80.35°E) northern India. We found that the aerosol optical depth (AOD) retrievals are more accurate compared to that of previous retrieval (Collection 004). About 70% of the total retrievals at 0.47 μm and 0.55 μm and 60% of the retrievals at 0.66 μm wavelength in the new version fall within the pre-launch uncertainty (Δτ = ±0.05 ± 0.15τ, where τ is AOD) with better correlation (R2 ∼ 0.83) at all three wavelengths. However, MODIS still tends to over-estimate AOD for a few retrievals in the presence of dust aerosols. The error in the fine-dominated AOD was large for most retrievals in the C005. However, the fine-dominated AOD of Collection 004 was better correlated with the equivalent AERONET data. This suggests that the fine-dominated AOD retrievals need to be re-examined further.

1. Introduction

[2] Aerosols are important in the Earth's climate since they play a vital role in radiation budget, cloud processes and air quality. Despite this fact, aerosol radiative forcing is one of the most uncertain components of the climate models because of the lack of a comprehensive database on a global scale [Hansen et al., 1997]. Ground-based measurements of aerosols provide critical information on aerosol physical and optical properties however, they are limited to a small region and aerosols exhibit high spatio-temporal variation. Therefore, these measurements cannot be applied to the remote areas where ground-based aerosol measurements do not exist. Satellite remote sensing of aerosols in this regard play an important role in monitoring and characterizing aerosols regionally as well as globally which is a prime requirement to assess the impact of aerosols on Earth's climate. The Moderate-Resolution Imaging Spectroradiometer (MODIS) sensor on board NASA's Terra (since December 1999) and Aqua (since July 2002) platforms have been retrieving aerosol properties routinely over land and ocean, including aerosol optical depth (AOD) at 0.55 μm and information about aerosol size [Remer et al., 2005]. MODIS is the first sensor to claim accuracy over land surfaces, and from aboard two platforms, provides global coverage twice a day.

[3] Since its launch in early 2000, the MODIS aerosol algorithm has been continually updated and evaluated by the MODIS team. However, the basic structure of the algorithm remained the same. This resulted in a complete set of products called Collection 004 (C004). Several validation studies have compared the C004 aerosol products with those retrieved by ground-based sunphotometers (mainly AERONET) and shown that the AOD retrievals over ocean were well within the pre-launch uncertainty of Δτ = ±0.03 ± 0.05τ, where τ is AOD. Over land, primarily because of spatial heterogeneity and higher uncertainty associated with the surface reflectance, the AOD tended to be over-estimated over bright surfaces and in dust regions [Kaufman et al., 2000; Chin et al., 2004]. Specifically, MODIS C004 systematically over-estimated AOD during periods of large dust loading at Kanpur (26.45°N, 80.35°E), India [Jethva et al., 2007]. Other studies [Chu et al., 2002; Remer et al., 2005] have demonstrated that MODIS tends to under-estimate AOD at higher values of AOD. Jethva et al. [2007] have also noted large positive offset at 0.47 μm during November to February (dominated by fine aerosols) at Kanpur. These and many other studies indicated that there was a need to improve the MODIS algorithm over land in order to reduce the biases.

[4] A new algorithm to retrieve aerosol properties over land from MODIS-measured top-of-atmosphere reflected spectral radiances, called Collection 005 (C005), was implemented recently [Levy et al., 2007]. The entire old algorithm (C004) was restructured to incorporate the following major changes in the new algorithm: 1) the new algorithm uses updated aerosol models based on Dubovik et al. [2002] and derived using ‘subjective cluster analysis’ of AERONET Level-2 data which yields three fine-dominated models and one spheroid dust model with new geographic distribution of fine-dominated models, 2) look-up-table (LUT) calculations are based on the combination of MIEV code [Wiscombe, 1981] and RT3 radiative transfer code [Evans and Stephens, 1991] that accounts for the effects of polarization, 3) the new algorithm no longer assumes that the 2.1 μm channel is transparent to aerosols and 4) the SWIR relationships used to estimate surface reflectance in the visible channels (ρ0.66s, ρ0.47s) from the reflectance in 2.1 μm channel (ρ2.1s) are parameterized which includes dependency on “green-ness” of the target (NDVIswir, based on 1.24 μm and 2.1 μm) and scattering angles (θ),

equation image

[5] In the C004, the ratios of reflectance were assumed to be constant in space and time. Many previous MODIS validation studies found positive offset of about 0.1 between MODIS and AERONET-derived AOD, indicating that the surface reflectance was under-estimated. Levy et al. [2005] demonstrated that higher values of VIS/SWIR surface ratio improved the continuity of the MODIS over-land aerosol products along the coastline of the DelMarVa Peninsula. It is well known that the land surface exhibits high spatio-temporal characteristics and has strong Bi-directional Reflectance Distribution Functions (BRDF). The VIS/SWIR surface ratios are functions of viewing geometry as well as the surface type and often differs from the ratios assumed by the C004 [Remer et al., 2001; Gatebe et al., 2001]. The above studies led to the introduction of parameterized surface reflectance relationships where surface reflectance in visible channels was not just a function of reflectance at 2.1 μm but also depends on scattering angle and “green-ness” of the pixel (equation (1)). These relationships have been derived using atmospherically corrected surface reflectances from co-located MODIS and AERONET data.

[6] In the present study, we have compared the reprocessed Aqua/MODIS Collection 005 (C005) aerosol optical depth, fine-dominated AOD and fine-dominated model weighting (FMW) obtained from, with the equivalent parameters derived from the ground-based Aerosol Robotic Network (AERONET) [Holben et al., 1998] at Kanpur, India. Kanpur is an urban site situated in the center of the Indo-Gangetic plane in the northern India. This region is surrounded by the tall Himalaya in the north and mountains (up to 1 km height) in the south, thus acts as a basin. Due to high population density and economic growth, this region has become one of the most aerosol loaded regions in India. The unique feature about this region is that it encounters different types of aerosols depending on season, air mass and sources. While high aerosol optical depth (0.6–1.2) attributed to the dust loading measured by AERONET during the pre-monsoon (April to June), the region is dominated by fine mode aerosols (radius < 0.6 μm) spread uniformly with AOD in the range 0.4 to 0.6 in the winter (November to February) [Jethva et al., 2005]. The AERONET sunphotometer has been deployed at Kanpur since year 2001 which provides more than 5 years of ground truth that is required for the validation. These make Kanpur a good test bed for the validation of MODIS aerosol retrieval. Under cloud free conditions, the AERONET-derived spectral AOD are accurate to within ±0.02 [Holben et al., 1998; Eck et al., 1999]. We have used a spatio-temporal window technique [Ichoku et al., 2002] in which the MODIS Level-2 products (10 km at nadir) were averaged spatially in a grid box of size 0.5° × 0.5° centered at sunphotometer site and compared with the temporally averaged (within ±30 minutes of MODIS overpass) AERONET Level-2 (cloud screened and quality assured) parameters.

2. Aerosol Optical Depth Comparison

[7] Figure 1 compares the AOD from MODIS and AERONET at 0.47 μm, 0.55 μm and 0.66 μm wavelengths. Different markers have been used for different seasons indicated as legends in the plot. We refer to winter as November to February, pre-monsoon as March to June and monsoon as July to October. The solid lines are pre-launch uncertainty in AOD retrieval: Δτ = ±0.05 ± 0.15τ, where τ is AOD [Remer et al., 2005] and dashed lines are linear regression fit for all data points with regression equation given at top of each plot. The co-location procedure resulted in a total of 236 matched points from Aqua for the period 2002–2005. The associated root-mean-square difference, correlation coefficients and % pixels falling within the predicted uncertainty for C005 (in bold) and C004 are listed in Table 1. We find that the new retrievals are more accurate and better correlated (R2 ∼ 0.83) with AERONET measurements and have less error than the C004 retrieval at all three wavelengths. Consequently, the number of retrievals that fall within the pre-launch uncertainty have also increased. We have noted that the magnitude of AOD retrievals in C005 were lower than that of C004 for Kanpur in the pre-monsoon season for about 86% of the pixels at 0.66 μm and higher in the winter for about 77% of the pixels at 0.47 μm. The reasons for the improvement in the AOD retrievals could be the introduction of the parameterized reflectance relationship (equation (1)) and new aerosol models based on Dubovik et al. [2002]. Using the C004 retrieval at Kanpur, Jethva et al. [2007] have correlated the error in AOD with the apparent reflectance at 2.1 μm for the pre-monsoon season when surface becomes bright (0.15 < ρ2.1 < 0.25, ρ2.1 is the apparent reflectance at 2.1 μm) and dust is a dominant aerosol type. This is because the ratios of reflectance in the C004 were only a function of reflectance at 2.1 μm and assumed to be constant in space and time which seems to be not appropriate for the bright surfaces. The new algorithm however, accounts for the surface type and scattering angle dependence while deriving the surface reflectance in the visible channels. At the same time, the new algorithm no longer assumes that the 2.1 μm channel is fully transparent to aerosols which is definitely true for the dust aerosols. These improvements may have resulted in the more accurate surface reflectance in visible channels and consequently better AOD retrieval. The single-scattering albedo at 0.66 μm assumed by the V5.2 over the Indian region is 0.90 which is very close to the mean value of 0.90 ± 0.03 and 0.88 ± 0.04 derived by AERONET Version 1 [Dubovik and King, 2000] and Version 2 respectively.

Figure 1.

Scatter-plot of aerosol optical depth between MODIS and AERONET at 0.47, 0.55 and 0.66 μm wavelengths. A total of 236 matched points are used from Aqua/MODIS 2002–2005. The dashed solid line is a linear regression fit for all data points and solid lines represent pre-launch uncertainty over land: Δτ = ±0.05 ± 0.15τ, where τ is aerosol optical depth. The regression equation at each wavelength is given at the top of the plot.

Table 1. Root-Mean-Square Error, Correlation Coefficients and % Pixels Fall Within the Predicted Uncertainty at Three Wavelengths, i.e. 0.47, 0.55 and 0.66 μm for MODIS Old Algorithm (Collection 004) and New Algorithm (Collection 005) for Period 2002–2005 at Kanpur
Channels, μmRoot-Mean-Square DifferenceCorrelation Coefficient% Pixels

3. Fine AOD Comparison

[8] The AOD contributed by the fine-dominated model (such as urban/industrial or biomass burning aerosol) called fine-dominated AOD and the ratio of fine-dominated AOD to the total AOD at 0.55 μm is called Fine Model Weighting (FMW). Each MODIS aerosol model (three spherical fine-dominated models and one spheroid dust model) is multi-modal and contains fine and coarse modes. On the other hand, AERONET sky retrievals designate fine mode fraction to be the volume contribution of aerosols below a radius < 0.6 μm in the case of Dubovik and King [2000] retrieval whereas O'Neill et al. [2003] method is based on the fitting an effective optical size distribution to the spectral AOD derived from direct sun measurements. Hence, the FMW derived by MODIS is not the same quantity as that obtained from the AERONET. However, the definitions of FMW are similar enough so that they should be correlated with each other.

[9] Figure 2 shows the scatter-plot of fine-dominated AOD (Figure 2, left) and FMW (Figure 2, right) at 0.55 μm wavelength derived from AERONET (Dubovik retrieval, Version 2) and MODIS (C004 and C005). The results are a bit surprising. The fine-dominated AOD and FMW derived by the new MODIS algorithm are nearly zero in most cases whereas fine-dominated AOD of C004 is better correlated with the AERONET data. The FMW of C004 has a binary distribution (1 or 0) and rarely picks an intermediate value. However, the comparison on a monthly scale showed that MODIS C004 was able to capture the seasonal variation of aerosol type at Kanpur [Jethva et al., 2005]. The MODIS-derived fine-dominated AOD was also compared with that retrieved using O'Neill's method (AERONET, Level 1.5) but there was no significant difference noted from the Dubovik retrieval. Figure 3 shows the spatial and seasonal distribution of monthly mean FMW derived from Aqua/MODIS monthly C004 product (Figure 3, top) and C005 product (Figure 3, bottom) over the Indian subcontinent. The FMW value was less than 0.2 over large part of India in all three seasons this means that the new MODIS algorithm has used the dust aerosol model in most cases for the AOD retrieval. This is in contrast with the C004 in which MODIS had used both fine-dominated models and dust model depending on the season and aerosol type in the AOD retrieval. The large error in the FMW retrieval at Kanpur could be attributed to the non-uniqueness of the inversion problem where the AOD retrievals are more accurate but the choice of aerosol model is somehow inappropriate. The new MODIS algorithm chooses one value of out of thirteen values of η, where η is the FMW, which provides the best spectral fit in radiance measured by MODIS at top-of-atmosphere. In the present case, the spheroid dust model assumed in the V5.2 provides the best spectral fit in radiance measured by MODIS at top-of-atmosphere (zero error at 0.47 μm and minimum error at 0.66 μm) in most cases. We believe that the new algorithm missed out other values of η which might have slightly larger error in fitting spectral radiances measured by MODIS. Unlike the previous version of algorithm (C004) in which the selection of the dynamic aerosol model was based on the ratio of path radiance in 0.66 μm and 0.47 μm, the new algorithm does not use the spectral dependence of aerosol path radiance which is a useful information in selecting non-dust and/or dust model for the retrieval.

Figure 2.

Comparison of MODIS-derived fine-dominated AOD (left) and fine model weighting (right) with the AERONET inversions at Kanpur for period 2002–2005. The open circles and closed circles denote Aqua/MODIS Collection 004 retrievals (old version) and Collection 005 retrievals (new version) respectively. The number of points used (N) and correlation coefficients (R2) are given separately for both the dataset.

Figure 3.

Spatial and seasonal distributions of monthly mean fine-dominated model weighting derived from Aqua/MODIS monthly (top) C004 product and (bottom) C005 product over the Indian subcontinent.

4. Conclusions

[10] The performance of the new MODIS aerosol algorithm (V5.2) was tested using ground-based AERONET data at Kanpur, India. Compared to the previous version (C004), the aerosol optical depth retrievals were more accurate at all wavelengths and well correlated with AERONET data. This could be due to the improved estimation of visible surface reflectance using parameterized reflectance relationship and updated new aerosol models used in the algorithm. However, MODIS still tends to over-estimate aerosol optical depth for few retrievals in the presence of dust which requires further investigation. The error in the fine-dominated AOD was too large for most retrievals in the C005 data whereas fine-dominated AOD of C004 showed better correlation with the equivalent AERONET data. This suggests that the new retrieval of fine-dominated AOD and fine-dominated model weighting over the Indian region need to be re-examined.


[11] We would like to thank Level 1 and Atmosphere Archive and Distribution System (LAADS) and MODIS team for their online support of MODIS Level-2 data. We also thank B. N. Holben and R. P. Singh and their staff for establishing and maintaining the AERONET site at Kanpur whose data was used in this investigation.