The atmospheric correction, i.e. the conversion from top-of-atmosphere (TOA) radiance to reflectance images, is one of the most important steps in the pre-processing of remote sensing data. The CHRIS Atmospheric Correction Tool converts from TOA radiance to surface reflectance images in an automated manner by means of a modular approach which involves the characterization of atmospheric and instrumental parameters. The algorithm is implemented so that it must be applied after noise removal and cloud screening. The interface between atmospheric and geometric correction modules is not established at this stage of the project, so the correction of topographic effects with a digital elevation model are not considered in the current version of the Atmospheric Correction Tool.
The algorithm implemented in the Atmospheric Correction Tool is described in detail in Guanter et al. (2005a,b). In the most general processing, the following steps are carried out:
Update of spectral characterisation (CHRIS modes 1 & 5)
Spectral polishing - removal of systematic errors in surface reflectance (CHRIS modes 1 & 5)
The different steps to be performed along the atmospheric correction process depend on the CHRIS acquisition mode. In particular, CWV is only derived for CHRIS acquisition modes 1, 3 and 5, as no sufficient sampling of water vapor absorption features is provided by modes 2 and 4. On the other hand, spectral calibration and an optional spectral polishing are only performed on modes 1 and 5, which are the ones providing the necessary high spectral resolution around sharp spectral features.
The different processing modules composing the CHRIS Atmospheric Correction Tool assume clear-sky conditions. Cloudy pixels are screened out from the processing using the cloud mask provided by the cloud screening module and the Cloud product threshold to be selected by the user as a processing parameter. This indicates the limit value of the probabilistic cloud mask discriminating cloudy and clear pixels. This threshold is not necessary if only a binary cloud mask was generated by the cloud screening module (no unmixing applied).
The MODerate resolution TRANsmittance (MODTRAN4) atmospheric radiative transfer code (Berk et al., 2003) was used for the generation of a Look-Up Table (LUT) which provides the atmospheric parameters used by the different modules from multidimensional linear interpolation. MODTRAN4 has been selected for its good parameterisation of both scattering and absorption atmospheric processes, as interposed by an algorithm dealing with simultaneous aerosol and water vapor retrieval. The LUT depends on 6 free input parameters: view zenith angle, solar zenith angle, relative azimuth angle, surface elevation, aerosol optical thickness at 550 nm (AOT550) and CWV. The atmospheric vertical profile is given by the default midlatitude summer atmosphere, the aerosol type is fixed to the continental model, and the ozone concentration is fixed to 7.08 g·m−2. The LUT was generated using the New Kurucz extraterrestrial solar irradiance data base in MODTRAN4.
Update of spectral characterisation
The first step in the Atmospheric Correction Tool for modes 1 and 5 is the evaluation of potential problems in the instrument spectral calibration. The lower spectral resolution and sampling of the acquisition modes 2, 3 and 4 leads to the smaller sensitivity to spectral calibration errors in these modes. In the case of pushbroom imaging spectrometers such as CHRIS, errors in the instrument spectral calibration are generally associated to the knowledge of the channels center wavelength, and are mostly caused by spectrometer shifts (systematic bias in channel position) and spectral smile (non-linear variation of channel position within the across-track direction of the image). The spectral response function and bandwidth are expected to have lower deviations from the laboratory calibration. Errors in surface reflectance appear around atmospheric absorption features when the band configuration applied in the resampling of the atmospheric parameters during atmospheric correction differs from the actual one. The fundamental basis for the determination of the real wavelength position is the removal of those errors. The method for spectral characterisation in the Atmospheric Correction Tool (Guanter et al, 2006) performs an iterative search for the channel position which generates the smoothest surface reflectance spectrum after atmospheric correction. The O2 absorption feature centered at 760 nm is used as reference for the CHRIS spectral calibration. The across-track direction is represented by a set of radiance spectra derived by averaging all the spectra in the along-track direction into one single pixel for each image column. The channel spectral position is evaluated independently for each one of them in order to characterise spectral shift and smile at the 760 nm region. However, in view of the low peak-to-peak smile detected in CHRIS (Guanter et al, 2006), and in order to speed-up the processing, atmospheric correction is not performed column-wise accounting for the exact channel position at each column, but only a mean value characterising the overall spectral shift from the nominal channel positions is used to update the later ones. The spectral shift in the 760 nm is assumed to be representative of the entire CHRIS spectral range.
Different AOT retrieval methods are applied to land and water modes. For the retrieval of AOT over land targets, the technique already described in Guanter et al. (2005) is applied. The total aerosol loading is parameterised by the AOT at 550 nm. No attempt to derive information on the aerosol model is made, but it is assumed that the aerosol optical properties are sufficiently described by the rural aerosol model implemented in MODTRAN4. It is also assumed that the AOT is constant all over the imaged area. After cloud screening, the lowest radiance value in each spectral band within the image is found. The resulting spectrum is employed as a reference dark target. It provides the highest limit for the aerosol content: an iterative procedure looks for the AOT550 value leading to the atmospheric path radiance which is closest to the radiance in the dark spectrum, not allowing path radiance to be higher than the dark spectrum in any of the visible bands (from 410 to 680 nm). The next step is refining that initial AOT550 estimation with a more sophisticated method involving the inversion of TOA radiances in combinations of green vegetation and bare soil pixels. This is performed only over those pixels which are classified as land pixels. AOT550 is then retrieved from 5 pixels with high spectral contrast inside this window, by means of a multiparameter inversion of the TOA spectral radiances in those pixels.
Please, note that only the top threshold of AOT550 is calculated by the current version of the aerosol retrieval module for CHRIS modes 1, 3, 4 and 5. The calculated top AOT value is labeled as 'Maximum AOT550' in the MPH file. An improved version of the aerosol retrieval algorithm over land including the inversion of vegetation and soil pixels will be available in future releases of the software.
In the case of inland water pixels, the particular performance of CHRIS in the mode 2 configuration (optimised for the observation of water bodies) causes that a different AOT retrieval approach must be considered. The use of land pixels is avoided for mode 2 data, due to the saturation usually found for surface albedos higher than 20-25%. The procedure described above for the land modes to set the maximum AOT value is used to calculate the final AOT value in mode 2. The AOT550 leading to the path radiance spectrum equal to or lower than the TOA radiance from the darkest pixels in the in the 435-690 nm range is selected. The first band, centered in 410 nm, is avoided because of the high noise levels detected. The final AOT550 is expected to be closer to the real value in the case of dark water bodies, from which most of the TOA signal is due to atmospheric scattering. In the case that the user can provide a reliable AOT550 value from external sources, it can be entered in the Aerosol optical thickness at 550 nm processing parameter field. The AOT retrieval module would not be applied in this case.
Water vapor retrieval is based on a band-fitting approach making use of the water vapor absorption centred at 940 nm. Surface reflectance outside and inside the left wing of that absorption feature is assumed to be linear with wavelength for the band-fitting inversion. A real band-fitting technique is only applied to modes 1 and 5, which are the ones providing a complete sampling of the 940 nm water vapor absorption feature. In the case of the mode 3, only two bands, centered about 900 and 910 nm, are affected by that absorption. However, the inversion of these two bands following the procedure described before has shown to be sufficient for the accurate water vapor retrieval. More accurate CWV values may be obtained if an initial value is provided to the algorithm (processing parameter Initial water vapour column (g cm-2)). The user can also select a faster processing option (deactivating Generate water vapour map) in which water vapor retrieval is only performed on a selected subset of pixels, rather than per-pixel, the retrieved mean value been applied to the complete image. However, thanks to the efficient JAVA implementation of the code, the per-pixel CWV retrieval is recommended as default.
Surface reflectance retrieval
Reflectance images are derived from TOA radiance after AOT and CWV retrieval. AOT and pixel-wise CWV are used as inputs to the MODTRAN4 LUT for the calculation of the atmospheric parameters to be used for the radiance-to-reflectance inversion. The resulting reflectance images can be further processed to remove the image blurring caused by those photons reflected by the target environment and scattered by the atmosphere particles into the sensor’s line-of-sight. This effect is called adjacency effect, because the apparent signal at the TOA for a given pixel comes also from the adjacent ones. The simple formulation proposed by Vermote et al. (1997) is followed here. It is based on the idea of weighting the strength of the adjacency effect by the ratio of diffuse to direct ground-to-sensor transmittance. This approach may be too simplistic in some cases (i.e. large view or illumination zenith angles, large aerosol loading) so the application of adjacency correction is performed only under user request (processing parameter Perform adjacency correction).
Systematic errors in the form of spikes and dips may appear in the surface reflectance product. Out of absorption regions these errors are mostly due to problems in the instrument gain coefficients (radiometric calibration), but inside atmospheric absorptions they can also be associated to inaccurate radiative transfer simulations. The fact that they are systematic allows the correlation with an error-free reference reflectance for the correction. The inversion of artificial endmembers is used to provide the reference reflectance patterns. Smooth spectra are calculated by spectral unmixing of spiky spectra. Up to 50 reference spectra are extracted from the reflectance images. The selection is made from a previous classification of the area based on three Normalized Difference Vegetation Index (NDVI) categories: vegetation, bare soil and mixtures. A recalibration coefficient for each CHRIS channel is calculated by correlating the 50 pairs of spectral reflectance from the spiky and smooth spectra. The reflectance image is multiplied by these recalibration coefficients, yielding the final reflectances. It must be remarked that the spectral polishing can lead to errors in reflectance for images where the endmembers cannot reproduce the actual reflectance patterns. For this reason, the recalibration process is carried out under user request only (processing parameter Perform spectral polishing).
A method for the surface reflectance retrieval from PROBA/CHRIS data over land: application to ESA SPARC campaigns L. Guanter, L. Alonso, J. Moreno, IEEE Transactions on Geoscience and Remote Sensing, 43, 2908-2917, 2005a. http://dx.doi.org/10.1109/TGRS.2005.857915
First results from the PROBA/CHRIS hyperspectral/multiangular satellite system over land and water targets L. Guanter, L. Alonso, J. Moreno, IEEE Geoscience and Remote Sensing Letters, 2, 250-254, 2005b. http://dx.doi.org/10.1109/LGRS.2005.851542
MODTRAN4 Version 3 Revision 1 User’s Manual A. Berk, G. P. Anderson, P. K. Acharya, M. L. Hoke, J. H. Chetwynd, L. S. Bernstein, E. P. Shettle, M. W. Matthew and S. M. Adler-Golden, Technical report, Air Force Research Laboratory, Hanscom Air Force Base, MA, USA, 2003.
Atmospheric correction of visible to middle infrared EOS-MODIS data over land surface: Background, operational algorithm and validation. E. F. Vermote, N. El-Saleous, C. O. Justice, Y. J. Kaufman, J. L. Privette, L. Remer, J. C. Roger, and D. Tanré, Journal of Geophysical Research, 102:17131–17141, 1997. http://modis.gsfc.nasa.gov/EOS/SCITEAM/ARTICLES/MST_A0199.pdf