For this reason, better statistical modeling of surfaces based on land class can be used as a better constraint of a realistic surface albedo and ultimately improve accuracy and retrieval frequency. Mathematically, this problem can be stated as the optimal selection of aerosol parameters whose effective surface reflectance spectrum is not an outlier in the spectral space determined from the surface reflectantance parameterization used. To efficiently carry out this approach, we use land class-based empirical orthogonal functions to project the higher simensional spectal space into a small number of principle components.. We have found that this is a good compression method for surface reflectances since we find that for most land classes can explained using two principal components while three components can explain ~ 90% of the variance in urban areas.
Prior to our analysis, we also made a statistical study of the factors that are best representative of biases that are observed in the matchup between AERONET and MODIS. To remove aerosol classification factors or satellite geometry, we explored the biases for 10 AERONET pair sites where an urban and non-urban AERONET site are examined and biases between them and the MODIS retreivals can be explored in detail. Analyis indicates that overestimating bias and reduction of correlation is observed in the urban site in comparison to the non-urban companion.
To generate realistic spectral albedos for different land classes, we used the MODIS BRDF 43A1 product as a training dataset to obtain land-class dependent EOF at any satellite geometry. Look-up tables were created to store the covariance matrix of the Ross-Li BRDF parameters in each land class and for each month of the year which is then used for a given satellite/view geometry to calculate the EOF from the LUT covariance matrices. The existing MODIS aerosol parameter space (defined by the optical depth and fine coarse ratio) is then searched for the surface reflectance spectrum that is closest (using Mahalanobis distance) to a linear mixture of EOF basis functions. The operational MODIS aerosol models were used; however, the use of our method allowed us to increase the number of bands used in the aerosol retrieval to three to six making it possible to extend the aerosol models to more complex cases in future.
We tested our algorithm by comparing the results of our EOF algorithm run at 500 m resolution and spatially averaged to 3km in a 20 km box for multiple urban Aeronet sites. The same procedure was applied to the operational MODIS 3 km DT algorithm. The main result is a significant increase in the R2 value for the unbiased regression line, which indicates that the algorithm performs much better than the operational algorithm in urban environments. Although operational MODIS 3km retrievals at individual urban locations may be fitted to a full (including bias) linear regression model, the bias coefficients are quite variable from site to site and also depend on season so that a global fitting of the urban retrievals will result in larger errors than those obtained using our EOF approach. Most importantly, this method inherently bypasses the operational filters on the surface models allowing for more frequent retrievals. At 3km resolution, for urban sites, the increase in retreival frequency > 60% while for 1.5 km, the improvement is > 110 %. Application for GOES-R and VIIRS will also be discussed as well as the potential improvements for PM25 retieval