Thermal Land-surface Variables From METEOSAT-IR Data
279
TOVS Initial Guess Retrieval (TIGR) library (globally distributed, quality checked
and representative radiosondes). The data were chosen from moderate northern
latitudes and the elevation, the scan-angle, and the surface temperature were varied
over an appropriate range. The trained network was verified using data sets
generated for atmospheric profiles from ECMWF data from 1996 over Europe.
1.2 The Radiative Transfer Model Modtran-3
MODTRAN-3 simulates the transport of radiation in the visible, the infra-red, and
the microwave spectral range with a resolution of lcm- I . For a given atmospheric
situation and with some additional parameters MODTRAN-3 can calculate the
radiation density at the satellite sensor, e.g. for calculations in the terrestrial infrared spectral range also surface emissivity and surface temperature have to be
supplied, while the path through the atmosphere is derived from surface location and
viewing angle. The measurement in a satellite channel and its corresponding
brightness temperature can be determined by applying the sensors response function
to the calculated radiation densities. For a known surface emissi vity the atmospheric
correction is then given with an accuracy of 0.7 K by the difference between the
known surface temperature and the calculated brightness temperature
(Chinnaswamy, 1999). This forms the basis of the "single channel method"
developed by Reutter et ai. (1994), which was improved by developing dedicated
spatial and temporal interpolation schemes (Schroedter et aI., 2001; Schadlich et aI.,
2001). For mid-latitude summer and winter and with an average surface emissivity
of E = 0.975 ± 0.025 the temperature error associated with this emissivity is
estimated to be ca. ± 1.4 K and the total accuracy of the method is estimated to be ca.
±2.1 K. Unfortunately, MODTRAN-3 only manages about 90 forward calculations
per minute on a Sun Ultra-Sparc. Therefore, the method is not suitable to calculate
vast amounts of corrections, e.g. for several months or years of METEOSA T data.
1.3 Feed-forward Neural Networks
Feed-forward networks are the most commonly used type of neural network (NN).
Two reasons for their popularity are that they are well suited for pattern matching
task and that well established training algorithms for this type of network exist. In
its most simple case a feed-forward network consists of two layers, the input and the
output layer, but it usually also has at least one hidden layer. In feed-forward
networks the information is fed into the input layer from where it is passed via the
hidden layer(s) to the output layer. The output of each neuron is specifically
weighted for all connections to the neurons of the following layer where it serves as
input and is processed further - there is no horizontal spreading of information
279
TOVS Initial Guess Retrieval (TIGR) library (globally distributed, quality checked
and representative radiosondes). The data were chosen from moderate northern
latitudes and the elevation, the scan-angle, and the surface temperature were varied
over an appropriate range. The trained network was verified using data sets
generated for atmospheric profiles from ECMWF data from 1996 over Europe.
1.2 The Radiative Transfer Model Modtran-3
MODTRAN-3 simulates the transport of radiation in the visible, the infra-red, and
the microwave spectral range with a resolution of lcm- I . For a given atmospheric
situation and with some additional parameters MODTRAN-3 can calculate the
radiation density at the satellite sensor, e.g. for calculations in the terrestrial infrared spectral range also surface emissivity and surface temperature have to be
supplied, while the path through the atmosphere is derived from surface location and
viewing angle. The measurement in a satellite channel and its corresponding
brightness temperature can be determined by applying the sensors response function
to the calculated radiation densities. For a known surface emissi vity the atmospheric
correction is then given with an accuracy of 0.7 K by the difference between the
known surface temperature and the calculated brightness temperature
(Chinnaswamy, 1999). This forms the basis of the "single channel method"
developed by Reutter et ai. (1994), which was improved by developing dedicated
spatial and temporal interpolation schemes (Schroedter et aI., 2001; Schadlich et aI.,
2001). For mid-latitude summer and winter and with an average surface emissivity
of E = 0.975 ± 0.025 the temperature error associated with this emissivity is
estimated to be ca. ± 1.4 K and the total accuracy of the method is estimated to be ca.
±2.1 K. Unfortunately, MODTRAN-3 only manages about 90 forward calculations
per minute on a Sun Ultra-Sparc. Therefore, the method is not suitable to calculate
vast amounts of corrections, e.g. for several months or years of METEOSA T data.
1.3 Feed-forward Neural Networks
Feed-forward networks are the most commonly used type of neural network (NN).
Two reasons for their popularity are that they are well suited for pattern matching
task and that well established training algorithms for this type of network exist. In
its most simple case a feed-forward network consists of two layers, the input and the
output layer, but it usually also has at least one hidden layer. In feed-forward
networks the information is fed into the input layer from where it is passed via the
hidden layer(s) to the output layer. The output of each neuron is specifically
weighted for all connections to the neurons of the following layer where it serves as
input and is processed further - there is no horizontal spreading of information