![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
Neural network retrieval of soil moisture: application to SMOS |
VerfasserIn |
Nemesio Rodriguez-Fernandez, Philippe Richaume, Filipe Aires, Catherine Prigent, Yann Kerr, Jana Kolassa, Carlos Jimenez, François Cabot, Ali Mahmoodi |
Konferenz |
EGU General Assembly 2014
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 16 (2014) |
Datensatznummer |
250096307
|
Publikation (Nr.) |
EGU/EGU2014-11805.pdf |
|
|
|
Zusammenfassung |
We present an efficient statistical soil moisture (SM) retrieval method using SMOS
brightness temperatures (BTs) complemented with MODIS NDVI and ASCAT
backscattering data. The method is based on a feed-forward neural network (hereafter NN)
trained with SM from ECMWF model predictions or from the SMOS operational
algorithm.
The best compromise to retrieve SM with NNs from SMOS brightness temperatures in a
large fraction of the swath (~ 670 km) is to use incidence angles from 25 to 60 degrees (in 7
bins of 5 deg width) for both H and V polarizations. The correlation coefficient (R) of the SM
retrieved by the NN and the reference SM dataset (ECMWF or SMOS L3) is 0.8. The
correlation coefficient increases to 0.91 when adding as input MODIS NDVI, ECOCLIMAP
sand and clay fractions and one of the following data: (i) active microwaves observations
(ASCAT backscattering coefficient at 40 deg incidence angle), (ii) ECMWF soil
temperature. Finally, the correlation coefficient increases to R=0.94 when using a
normalization index computed locally for each latitude-longitude point with the
maximum and minimum BTs and the associated SM values from the local time
series.
Global maps of SM obtained with NNs reproduce well the spatial structures present in the
reference SM datasets, implying that the NN works well for a wide range of ecosystems and
physical conditions. In addition, the results of the NNs have been evaluated at selected
locations for which in situ measurements are available such as the USDA-ARS watersheds
(USA), the OzNet network (AUS) and USDA-NRCS SCAN network (USA). The time series
of SM obtained with NNs reproduce the temporal behavior measured with in situ sensors. For
well known sites where the in situ measurement is representative of a 40 km scale like the
Little Washita watershed, the NN models show a very high correlation of (R =
0.8-0.9) and a low standard deviation of 0.02-0.04 m3/m3 with respect to the in situ
measurements. When comparing with all the in situ stations, the average correlation
coefficients and bias of NN SM with respect to in situ measurements are comparable
to those of ECMWF and SMOS L3 SM (R = 0.6). The standard deviation of the
difference (STTD) of those products with respect to in situ measurements is lower for
NN SM, in particular for the NN models that use local information on the extreme
BTs and associated SM values, for which average STDD is 0.03 m3/m3, twice as
low as the average STDD values obtained with ECMWF and L3 SM (0.05-0.07
m3/m3).
In conclusion, SM obtained using NN give results of comparable or better quality to other
SM products. In addition, NNs are an efficient method to obtain SM from SMOS data (one
year of SMOS observations can be inverted in less than 60 seconds). These results have been
obtained in the framework of the SMOS+NN project funded by ESA and they open
interesting perspectives such as a near real time processor and data assimilation in weather
prediction models. |
|
|
|
|
|