Thermal Land-surface Variables From METEOSAT-IR Data
291
verified using profiles extracted from ECMWF analyses. The mean temperature
error of ENZO-NN (Table 2) for the verification data (ECMWF) was 0.31 K as
compared to MOOTRAN-3. For the validation data (TIGR profiles not used for
training) the mean temperature error was 0.25 K. The biggest advantage of the
neural network is speed: it is approximately 5000 times faster than MOOTRAN-3.
It is planned to use emissivity as a further input of the neural network. The
neural network is expected to achieve an accuracy comparable to MOOTRAN-3 for
an average emissivity of 0.975 with an error of ± 0.025. For the standard profiles
'mid-latitude summer' and 'mid-latitude winter' the temperature error associated
with this emissivity is estimated to be 1.4 K. The combined temperature error due
to MOOTRAN-3 and the above emissivity error is estimated to be 2.1 K (Schadlich
et al., 2001). If a more accurate method to determine emissivity can be found (Dash
et al., 2001), e.g. by using METEOSAT Second Generation (MSG), the error can
be reduced accordingly.
It was shown that diurnal temperature cycles can be modelled in a stable and
meaningful way. The modelling yields residual temperature, temperature amplitude,
time of the maximum, and attenuation during the night. Low residual temperature
anomalies are primarily caused by elevation effects, whereas, high residual
temperature anomalies can be observed for water-bodies. The temperature amplitude
proved to be the most discriminating parameter for surface characteristics, as it
reflects the surface moisture and bio-mass. In order to yield a continuous spatial
coverage for change detection, monthly composites of the IR data were generated.
Cloud screening of the data combined with subsequent composition and modelling
ensures that the thermal parameters are practically free from synoptic artefacts.
Good results were achieved using median and maximum value composites, which
describe the typical and the hottest situations during the composite interval,
respectively. The NOVI and the corrected temperature amplitude T a - oT were
compared. For many areas Ta - oT is strongly correlated to the NOVI. A deviating
behaviour can be a hint for different land use (Gottsche and Olesen, 2001).
The next step will be the calculation of parameters for several years in order to
allow change detection. The necessary amount of computing time as well as the
amount of data that must be handled are manageable due to the composites. The
fitting algorithm proved to be stable and robust in respect to undetected clouds and
data gaps and only little user-interaction is required. Therefore, this is a promising
method to add IR information to change detection, which is until now based on
visible data alone.
The data evaluation oflong time series (years) of METEOSAT and A VHRR IR
data is feasible and the data as well as the required algorithms are available at the
FU-Berlin and at the Forschungszentrum Karlsruhe - IMK. MSG, which is
scheduled to become operational at the end of 2002, combines the capabilities of
A VHRR and METEOSA T and will allow to expand the existing time series of
satellite data. The principles underlying the developed algorithms, in particular the
291
verified using profiles extracted from ECMWF analyses. The mean temperature
error of ENZO-NN (Table 2) for the verification data (ECMWF) was 0.31 K as
compared to MOOTRAN-3. For the validation data (TIGR profiles not used for
training) the mean temperature error was 0.25 K. The biggest advantage of the
neural network is speed: it is approximately 5000 times faster than MOOTRAN-3.
It is planned to use emissivity as a further input of the neural network. The
neural network is expected to achieve an accuracy comparable to MOOTRAN-3 for
an average emissivity of 0.975 with an error of ± 0.025. For the standard profiles
'mid-latitude summer' and 'mid-latitude winter' the temperature error associated
with this emissivity is estimated to be 1.4 K. The combined temperature error due
to MOOTRAN-3 and the above emissivity error is estimated to be 2.1 K (Schadlich
et al., 2001). If a more accurate method to determine emissivity can be found (Dash
et al., 2001), e.g. by using METEOSAT Second Generation (MSG), the error can
be reduced accordingly.
It was shown that diurnal temperature cycles can be modelled in a stable and
meaningful way. The modelling yields residual temperature, temperature amplitude,
time of the maximum, and attenuation during the night. Low residual temperature
anomalies are primarily caused by elevation effects, whereas, high residual
temperature anomalies can be observed for water-bodies. The temperature amplitude
proved to be the most discriminating parameter for surface characteristics, as it
reflects the surface moisture and bio-mass. In order to yield a continuous spatial
coverage for change detection, monthly composites of the IR data were generated.
Cloud screening of the data combined with subsequent composition and modelling
ensures that the thermal parameters are practically free from synoptic artefacts.
Good results were achieved using median and maximum value composites, which
describe the typical and the hottest situations during the composite interval,
respectively. The NOVI and the corrected temperature amplitude T a - oT were
compared. For many areas Ta - oT is strongly correlated to the NOVI. A deviating
behaviour can be a hint for different land use (Gottsche and Olesen, 2001).
The next step will be the calculation of parameters for several years in order to
allow change detection. The necessary amount of computing time as well as the
amount of data that must be handled are manageable due to the composites. The
fitting algorithm proved to be stable and robust in respect to undetected clouds and
data gaps and only little user-interaction is required. Therefore, this is a promising
method to add IR information to change detection, which is until now based on
visible data alone.
The data evaluation oflong time series (years) of METEOSAT and A VHRR IR
data is feasible and the data as well as the required algorithms are available at the
FU-Berlin and at the Forschungszentrum Karlsruhe - IMK. MSG, which is
scheduled to become operational at the end of 2002, combines the capabilities of
A VHRR and METEOSA T and will allow to expand the existing time series of
satellite data. The principles underlying the developed algorithms, in particular the