278
F. Giittsche, F.-S. Olesen
thermodynamic land surface temperatures (Gottsche and Olesen, 2002).
The spatial interpolation from a limited number of locations with
known atmospheric profiles (ECMWF grid cells) to each pixel (Shepard
algorithm; Schroedter et aI., 2001).
The temporal interpolation from 4 times per day, for which the
atmospheric situation is known, to all 48 METEOSAT slots (Schadlich
et aI., 2001).
A model consisting of a cosine (daytime) and an exponential decay
(night -time) is fitted to 10 day or monthly composites of cloud free data
in order to derive thermal surface parameters, e.g. mllllmum
temperature, diurnal amplitude, and the time of the maximum
temperature (Gottsche and Olesen, 20(1).
The merging of METEOSA T and A VHRR IR data to time series of 2
km spatial and 30 minutes temporal resolution.
All the above components were developed at the IMK. The two most recent
components, the atmospheric corrections using neural networks and the
model for the derivation of thermal surface parameters, are described in more
detail below. In combination with work carried out by the FU-Berlin this will
result in one of the most complete evaluations of long term satellite data.
1 Neural Network Versus Modtran
1.1 Principal Considerations
The atmospheric influence on the radiance measured by a satellite can be calculated
with a radiative-transfer model, e.g. MODTRAN, provided one has sufficient
knowledge about the state of the atmosphere (moisture- and temperature-profiles)
and the surface (emissivity). These calculations are very expensive in terms of
computing time and are, therefore, not well suited to calculate corrections for large
quantities of data, i.e. one year for Europe. On the other hand, faster split-window
methods require two or more channels to estimate the atmospheric influence. Until
recently these were only available on sun-synchronous satellites. At moderate
latitudes these satellites provide a maximum of 4 samples with a variable viewing
geometry per day. In order to exploit the long time series (20 years) of METE OS AT
data with 30 minutes temporal resolution the possibility of atmospherically
correcting single channel IR data using neural networks is investigated. The neural
networks are developed and trained using the Stuttgart Neural Network Simulator
(SNNS) and the evolutionary algorithm "Evolutionarer Netzwerk Optimierer
(ENZO)". The training and validation data sets consist of MODTRAN-3
calculations for a representative selection of atmospheric profiles taken from the
F. Giittsche, F.-S. Olesen
thermodynamic land surface temperatures (Gottsche and Olesen, 2002).
The spatial interpolation from a limited number of locations with
known atmospheric profiles (ECMWF grid cells) to each pixel (Shepard
algorithm; Schroedter et aI., 2001).
The temporal interpolation from 4 times per day, for which the
atmospheric situation is known, to all 48 METEOSAT slots (Schadlich
et aI., 2001).
A model consisting of a cosine (daytime) and an exponential decay
(night -time) is fitted to 10 day or monthly composites of cloud free data
in order to derive thermal surface parameters, e.g. mllllmum
temperature, diurnal amplitude, and the time of the maximum
temperature (Gottsche and Olesen, 20(1).
The merging of METEOSA T and A VHRR IR data to time series of 2
km spatial and 30 minutes temporal resolution.
All the above components were developed at the IMK. The two most recent
components, the atmospheric corrections using neural networks and the
model for the derivation of thermal surface parameters, are described in more
detail below. In combination with work carried out by the FU-Berlin this will
result in one of the most complete evaluations of long term satellite data.
1 Neural Network Versus Modtran
1.1 Principal Considerations
The atmospheric influence on the radiance measured by a satellite can be calculated
with a radiative-transfer model, e.g. MODTRAN, provided one has sufficient
knowledge about the state of the atmosphere (moisture- and temperature-profiles)
and the surface (emissivity). These calculations are very expensive in terms of
computing time and are, therefore, not well suited to calculate corrections for large
quantities of data, i.e. one year for Europe. On the other hand, faster split-window
methods require two or more channels to estimate the atmospheric influence. Until
recently these were only available on sun-synchronous satellites. At moderate
latitudes these satellites provide a maximum of 4 samples with a variable viewing
geometry per day. In order to exploit the long time series (20 years) of METE OS AT
data with 30 minutes temporal resolution the possibility of atmospherically
correcting single channel IR data using neural networks is investigated. The neural
networks are developed and trained using the Stuttgart Neural Network Simulator
(SNNS) and the evolutionary algorithm "Evolutionarer Netzwerk Optimierer
(ENZO)". The training and validation data sets consist of MODTRAN-3
calculations for a representative selection of atmospheric profiles taken from the