| Meris Case-2 Regional Algorithm Specification | |
Details of the algorithms are provided in two ATBDs, for atmospheric correction (not yet released) and for the retrieval of water constituents. This gives only a brief overview.
The atmospheric correction is based on the principle idea that the path radiances as well as the
transmittances of the first 9 MERIS bands (412 -708 nm) can be derived from the top of atmosphere
directional radiance reflectance of 4 MERIS bands in the near infrared spectral range (708 - 870
nm). We assume that the variability of the path radiances and transmittances of all bands, for a
standard atmosphere, depend on the aerosol types and concentrations in different layers of the
atmosphere, on thin cirrus clouds, and on the specular reflectance of scattered light at the sea
surface (sky light glitter) and on the specular reflectance of direct sun light at the sea surface (sun
glint), both of which are wind dependent. By knowing the reflectance in the 4 NIR bands it is
possible to determine the path radiance and transmittance of first 9 MERIS bands, which are used
for water remote sensing. Since the absorption of pure water in the NIR bands is very high, the
contribution of other water constituents to the total absorption can be neglected.
In turbid water the 4 NIR bands might also be affected by the scattering of particles in water. Thus
an algorithm as we provide here for case 2 water has to take this effect into account.
The techniques we have used to determine the path radiances and transmittances from these 4 NIR
bands is a neural network. It is trained by simulated top of standard atmosphere directional radiance reflectances.
Input to the neural network are the reflectances in the 4 NIR bands as well as the solar and viewing
zenith angel and the difference between solar and viewing azimuth angle. Output of the NN are the
9 path radiances and transmittances of the first 9 MERIS bands. This information can then be
simply used to compute the directional water leaving radiance reflectances for these 9 bands.
The NN is applied to the reflectances top of a standard atmosphere (TOSA). These reflectances are
derived from the Top of Atmosphere (TOA) radiance reflectances by taking the difference between
the standard and actual surface pressure into account to correct for the scattering of air molecules
(Rayleigh scattering) of this correction layer. In the same way the difference between the standard
and actual ozone concentration is used in this correction layer (details s. ATBD Atmospheric
correction). The correction for these two deviations from a standard atmosphere is performed by an
analytical algorithm. This algorithm converts TOA radiance reflectances into TOSA radiance
reflectances, which are thin input to the NN.
Furthermore, we have included in this correction procedure an optional (s. parameter file)
correction for the spectral deviations of the individual pixels of each camera due to the remaining
misalignment of the cameras (in other papers called "smile correction"). This misalignment has
mainly an effect on the path radiance calculation (scattering by air molecules) and on the solar
irradiance at top of atmosphere.
A further, experimental, add on is a special neural network for taking the influence of the
polarisation of light by air molecules into account for the atmospheric correction. Also this
procedure can be optionally switched on by a parameter.
Output products of the atmospheric correction are the TOSA radiance reflectances, the path
radiances, the transmittances and the water leaving radiances. The inclusion of all of these variables
in the output product file can be selected in the parameter file band by band individually.
Although it is possible to perform the atmospheric correction and retrieval of water constituents in
one step with one neural network, we have separated these tasks in order not to involve a specific
bio-optical model in the atmospheric correction, so that the use of only the atmospherically
corrected data is independent from the optical properties of the water constituents as assumed in a
bio-optical model.
The retrieval of the water constituents from the directional water leaving radiance reflectances is
performed also with a neural network.
In this version of the processor we provide two alternatives, which can be selected by the parameter
file. In both cases the networks are trained with simulated directional radiance reflectances. Details
of the bio-optical model, the simulations and training procedure can be found in Doerffer & Schiller
(2006) and in the ATBD.
The first version is a combination of a backward and forward neural network. Input to the backward
network are the directional water leaving radiance reflectances of 8 MERIS bands (412 - 708 nm,
excluding the fluorescence band at 681 nm). Output of the network are 3 IOPs, i.e. (1) scattering of
all particles, (2) absorption of phytoplankton pigments and (3) absorption of Gelbstoff and bleached
particles, all three at MERIS band 2 (442 nm). These coefficients are also input to a forward neural
network which computes the water leaving radiance reflectances. The computed reflectances are
then compared with the measured ones (derived after atmospheric correction). In the case that the
sum of the squared deviations between both spectra surmounts a threshold of 4 a warning flag is
triggered. Furthermore the values of the agreement is also provided as a selectable product.
The advantage of this alternative is that it is rather fast.
The second selectable alternative is based on an optimization procedure. It uses only the forward
neural network. By modifying the three IOPs in the optimization loop an optimum agreement
between the simulated and measured spectrum of the directional water leaving radiance
reflectance is searched. If the deviation is below a certain minimum or if the maximum number of
iterations is reached, the three IOPs are provided as the result. Furthermore, the remaining Chi2
deviation is also provided.
Comparisons between both alternatives have shown that in most cases the remaining deviation is
significantly smaller when the optimization procedure is used instead of the backward neural
network. Also the results show less noise. However, this alternative requires significantly more
computation time.