Title: Correlative Analysis of EO-1, Landsat, and Terra Data of the DOE ARM CART Sites: An Investigation of Instrument Performance and Atmospheric Correction
Principal Investigator: Barbara E. Carlson
Institution: NASA/Goddard Space Flight Center Institute for Space Studies
We proposed a detailed analysis of the performance capabilities of the EO-1 sensors using scenes centered on the DOE ARM CART sites. The primary goals of this investigation are to: (1) develop an improved atmospheric correction algorithm; (2) investigate the information content of the EO-1 measurements with respect to atmospheric NO2, O3, water vapor, aerosols, and cloud; and (3) to evaluate the effects of spatial and spectral resolution on both atmospheric correction and the retrieval of atmospheric properties. Algorithm improvements include more accurate treatment of gaseous opacity, continuum absorption, aerosol and cloud extinction, as well as a better separation of atmospheric and surface contributions. The effects of the coarser spatial resolution of LAC on the atmospheric correction provided for ETM+ and ALI will be addressed by using atmospheric corrections derived from the higher spatial resolution, coarser spectral resolution Hyperion measurements. The wide range of correlative, ground-truth data at the CART sites will allow us to examine in detail the range and accuracy of atmospheric retrievals possible using LAC and Hyperion. In particular, the availability of a continuously operating water vapor Raman Lidar and Micro Pulse Lidar provide the data to investigate the extent to which water vapor profile information can be retrieved from LAC and Hyperion data and the expected accuracy of the cirrus atmospheric corrections. Aerosol correction will be addressed using surface-based measurements. These will be compared with the aerosol properties retrieved from MODIS, MISR, LAC, Hyperion, and ALI.
We have been working on the development of an improved atmospheric correction algorithm. We have used as the basic framework for our retrieval algorithm the analysis models that we have developed for the Research Scanning Polarimeter (RSP). Since the RSP makes measurements in nine spectral intervals from 410 to 2250 nm, the spectral range nicely matches that of the EO-1 sensors. The model uses the doubling/adding code to calculate atmospheric scattering. An arbitrary vertical distribution of the scatterers and gases can be specified and the internal radiation fields between any two layers can be calculated as well as the downwelling radiation at the surface and reflected radiation at the top of the atmosphere. This allows the same model to be used to compare surface and satellite data within a consistent framework. The gaseous absorption properties are treated using the correlated k-distribution method for line absorption with continuum absorption from H2O, NO2, and O3 also included. Vector radiative transfer calculations are used at the shorter wavelengths (< 600 nm) to ensure that unexpected errors in atmospheric correction are not caused by the neglect of polarization in the atmospheric correction algorithm while scalar calculations are performed for the longer wavelengths.
In the development of the atmospheric correction algorithm we have paid particular attention to studying the tradeoffs between speed and accuracy for the correction of hyperspectral imagery. For accurate corrections, the spectral registration of the atmospheric correction function must be allowed to vary across the instrument swath. Of particular concern has been the characterization of the spectral smile. Spectral smile is caused by curvature of the image of the slit formed in the focal plane array. Curvature causes the response center of a given sample to vary across the spatial direction of the focal plane array. For the resolution of typical land surface spectral features calibrations need not be orders of magnitude better than the spacing between bands. However, when atmospheric effects must be removed, accuracies must approach 1/100 the width of the bands, which for Hyperion suggests that the needed spectral calibration accuracy is 0.1 nm. Rather than regard the spectral requirements for accurate atmospheric correction as a burden, we have used them to derive a “bootstrap” spectral calibration using major atmospheric absorption features to determine the spectral registration of the hyperspectral measurements. For the Visible Near Infra-Red (VNIR) spectral domain, we have used the oxygen A-band, which is strong, narrow and has a well defined depth, to evaluate the spectral calibration of the VNIR spectrometer. Our “bootstrap” method is straightforward. The data are atmospherically corrected using a particular spectral registration, e.g., the project supplied values. The atmospherically corrected data are fitted with a polynomial (second, or third order polynomials give similar results) and since the underlying surface spectral reflectance is expected to be a smooth function of wavelength, the residual error between the corrected reflectance spectrum and the polynomial fit is regarded as being a function that should be minimized when we have the correct spectral registration. The spectral registration is then varied to find the minimum value of the residual error. This “bootstrap” process is performed separately for each pixel. By performing the spectral calibration over a large number of lines, the uncertainty caused by land surface effects is reduced.
Using our EO-1 data collected for the Southern Great Plains (SGP), Oklahoma we find that the spectral offset of pixel 128 band centers is different from the nominal band centers provided with the data by 0.94 nm. We have also examined data obtained for the Coleambally Irrigation Area, NSW and have found that the spectral offset of the pixel 128 band is different from the nominal band centers by 1.04 nm. Our analysis shows the greatest discrepancy between the pre-launch laboratory measurements and our “bootstrap” analysis results occurs at the edge of the spatial field. This is also the area where there is the greatest discrepancy between the analysis of the two different Hyperion data sets (SGP and Coleambally). The shape of the discrepancy suggests that the difference could be a real change in the spectrometer. This is further substantiated by our analysis of the Hyperion data subset that was assembled to better characterize the Hyperion focal plane array.
The spectral bands used for the A-band smile analysis are 742.80, 752.97, 963.14, 773.31, 783.48, and 793.65. The only spectral band in this range that was measured during pre-launch characterization was band 40 to which we assign a nominal center of 753.005 at pixel 128 (based on our analysis of TRW data). This value compares well with the TRW value of 752.97 (i.e., a 0.035 nm discrepancy between our pre-launch analysis and TRW’s nominal band center). It should be noted that the GSFC nominal center for this band is 752.425 nm.
We analyzed the Hyperion data subset using an approach in which the oxygen A-band is normalized by assuming that the surface reflectance is linear in wavelength through the band. The spectral bands on either side of the A-band are used to estimate the surface spectral reflectance. The A-band spectra are normalized (divided by) this surface reflectance under the assumption that for the purposes of this analysis the atmospheric contribution and multiple surface-atmosphere interactions can be neglected. The mean values along with the standard deviations (±1 SD) are used to determine the location of the spectral bands. The spectral band location is determined by dividing the normalized radiances by the direct beam transmission for the altitude of the particular site and fitting the result to a quadratic polynomial. This fit provides some compensation for errors in the initial assumption of linear spectral variation. It is the residual difference between the result and the polynomial fit that is used to determine the location of the A-band. This is based on our expectation that when the RMS value of the residual is minimized the band location is correct, since otherwise the division of the normalized radiance by the direct beam transmission is not a smooth function that can be fitted by a polynomial. The minimum RMS of all residuals as a function of bandshift is therefore determined to be the shift for that Hyperion pixel. The analysis is performed independently for each pixel and is repeated for the mean values of the normalized radiance and their standard deviations. From this analysis, we find that pixel 128 is shifted by 0.76 nm, which is 0.18 nm less than the original estimate based solely on the analysis of the SGP and Coleambally data sets.
We have applied the same “bootstrap” spectral calibration to the Short-Wave Infra-Red (SWIR) spectrometer using the carbon dioxide band at 2000 nm and the water vapor band at 1125 nm. In the case of the Hyperion SWIR spectrometer we have not found a discernable smile, or other pixel dependent spectral registration issue. However, there does appear to be a small shift in the SWIR spectral registration of the on-orbit data compared with the laboratory determined spectral registration provided with the data. Unfortunately, the relatively poor agreement between the calculated transmission values and the Hyperion data in the spectral windows for the 2000 nm CO2 band means that the “bootstrap” approach is less robust in the SWIR than in the VNIR.
We have used our atmospheric correction algorithm to correct several Hyperion images and have compared the retrieved water amounts with ground-based measurements which have shown acceptable agreement. We are currently in the process of constructing a Graphical User Interface (GUI) for our atmospheric correction algorithm, cleaning up and documenting the code. A prototype version of the GUI driven code is available.