Determination of aerosol optical depth over land from satellite remote sensing measurements is extremely complex due to the large variability of aerosol optical properties (i.e. phase function, single scattering albedo etc) as well as the effects of the surface albedo which can vary significantly. In the previous land algorithm, to estimate the VIS ground albedo, a single scene invariant assumption was used to connect the VIS surface albedo to the MIR (2130nm) surface albedos which were directly obtained from the measurements.  Based on these principles, the basic approach for the collection 4 algorithms were

1. Determination of the presence of the dark pixels in the blue (0.47 um) and red (0.66 um) channels using their remotely sensed reflectance in the mid- IR channels (2.1 um ).
2. Estimation of the surface reflectance of the dark pixels in the red and blue channels using the measurements in the mid-IR.
3. Determination of the aerosol type using information on the global aerosol distribution and the ratio between the aerosol path radiance in the red and blue channels.
4. Inversion of the measured radiance at TOA  into the aerosol optical thickness, volume (or mass) concentration and spectral radiative forcing using radiative transfer look-up tables.

            However, it was readily understood that any correlations between the MIR channels and the VIS channels should be a function of the land type and the solar-view geometry. The newest land algorithm for MODIS (C-005) has recently been developed and reprocessing of the MODIS streams is ongoing. This algorithm makes an attempt to improve both the aerosol assumptions as well as the ground albedo models. For the aerosol models, cluster analysis based on large scale Aeronet measurements were used to define a set of most likely aerosol models for each AOD for a given geographic location. To improve on the previous surface models, extensive matchups based on atmospherically cleared MODIS  TOA reflectences using derived Aeronet atmospheres were analyzed as a function of surface classification and sun-view geometry. The results of this matchup were used to develop a more sophisticated algorithm connecting the surface reflectences in the MIR and VIS channels which depends on surface classification and scattering angle. The classification of scene was obtained by investigation of a Modified Vegetation Index  (MVI)  which unlike the Vegetation Index which looks at ratios between the Red and NIR channels, uses the 1270 and 2130 nm channels to avoid errors due to inaccurate atmospheric  retrieval.

            In this paper, we examine the consistency of surface albedo models from high spatial resolution hyperion imagery of the NYC metropolitan area.. In particular, we have examined how the MVI index depends on vegetation and urban regions to see if the urban scene can be classified properly using the MODIS collection 5 algorithm. The pixels were classified into urban and vegetation classes using a supervised classification algorithm after undergoing atmospheric correction based on aeronet AOD measurements which were less than .05 thereby increasing the accuracy of the ground retrieval properties.

RGB Image with Training regions

Segmented Image

            To begin, we have found that the MVI is very insensitive to atmospheric contamination as to be expected in comparison to the traditional VI since the atmospheric effects decrease significantly for the longer channels  However, it was observed that the range of the MVI index for urban scene pixels can be significantly larger than the 0.75 value used by the MODIS C005 algorithm in identifying highly vegetative scenes resulting in misclassification. Furthermore, the  range of the MVI index in the identification of non vegetative pixels  is much larger than the range for the vegetative pixels. This is in contrast to the standard Vegetation Index (VI) where the VI variability between vegetative and urban pixels is nearly the same. This property makes the vegetative index more suitable as a classifier in the presence of random noise but this benefit must be weighed against the sensitivity of the VI to atmospheric corrections.

Finally, we show that a single static relationship of  MVI values for classifying surfaces is not correct and can lead to significant classification errors.  This is shown through an analysis of the surface MODIS BRDF product which allows us to calculate the MVI index as a function of the solar and view geometries. In fact, the variability of the MVI index as a function of the observation angle can be on the order of the width of the MVI classification zones and the variability of this index as a function of angular variability needs to be studied more carefully. Error budgets for the surface models based on our findings will be presented.