GEOPHYSICAL RESEARCH LETTERS, VOL. 32, L15403, doi:10.
1029/2005GL023623, 2005
Observation operators for the direct assimilation of TRMM microwave
imager retrieved soil moisture
M. Drusch
Data Assimilation Section, European Centre for Medium-Range Weather Forecasts, Reading, UK
E. F. Wood and H. Gao
Department of Environmental Engineering, Princeton University, Princeton, New Jersey, USA
Received 25 May 2005; revised 30 June 2005; accepted 6 July 2005; published 9 August 2005.
[1] Each soil moisture data set is characterized by its on linear estimation theory, systematic differences between
specific mean value, variability, and dynamical range. For the modeled first guess and the observations prevent a
the assimilation of soil moisture observations into numerical statistically optimal analysis and therefore have to be
models observation operators have to be developed, which corrected [Dee and da Silva, 1998].
reduce systematic differences. In this study, cumulative [4] In the case of soil moisture analyses, systematic
distribution function (CDF) matching is used to derive differences between the model and the satellite data are
observation operators for TRMM Microwave Imager (TMI) due to: (1) The terrestrial water balance equation depends
derived soil moisture for the southern US, modeled soil on the temporal change of soil moisture. Consequently,
moisture fields from the European Centre for Medium- every model has its own soil moisture climatology with a
Range Weather Forecasts (ECMWF), and model output dynamical range defined through the wilting point and field
from the Variable Infiltration Capacity model (VIC). It is capacity [Koster and Milly, 1997]. (2) Microwave observa-
found that the transferability of these observation operators tions in the frequency range from 5 GHz to 11 GHz
in space and time strongly depends on the geographical represent the upper most soil layer with a depth of up to
region. In the Central US, where the assimilation of soil 0.02 m. In contrast, most operational hydrological models
moisture is most promising, the observation operator have a surface layer of between 0.05 to 0.10 m depth. The
exhibits little variability in time. The temporal variability thin surface layer monitored by the satellite instrument will
in the observation operator can result in substantial show a larger dynamical range and a different probability
differences between the modeled field and the density function of soil moisture representing a faster dry
observation. In Numerical Weather Prediction (NWP) down compared to models with a relatively thicker layer.
applications, where the model tends to be updated on a Observation operators will be needed to remove the sys-
regular basis, dynamic observation operators will be tematic differences between satellite derived soil moisture
necessary to assimilate soil moisture. For climate studies and the modeled first guess.
or re-analysis projects long time series are required to define [5] In this context, Reichle and Koster [2004] introduced
an observation operator, which correctly reproduces the concept of Cumulative Distribution Function (CDF)
interannual variability. Citation: Drusch, M., E. F. Wood, matching. Satellite derived soil moisture from the Scanning
and H. Gao (2005), Observation operators for the direct Multifrequency Microwave Radiometer (SMMR) was com-
assimilation of TRMM microwave imager retrieved soil pared against the NASA land catchment model and it was
moisture, Geophys. Res. Lett., 32, L15403, doi:10.1029/ found that local CDF matching efficiently removed the bias.
2005GL023623. The present study focuses on the observation operator,
which is obtained through CDF matching, its transferability
in time and space, and its dependency on the modeled first
1. Introduction guess (i.e. the model). Soil moisture from the operational
[2] Soil moisture is a key variable for hydrological and numerical weather prediction model at the European Centre
meteorological processes in the soil and over land [e.g., for Medium-Range Weather Forecasts (ECMWF), the Var-
Koster et al., 2004]. Global soil moisture data sets can either iable Infiltration Capacity (VIC) model [Liang et al., 1994],
be obtained from numerical models or from satellite obser- Oklahoma Mesonet in situ measurements, and the TRMM
vations. The most advanced satellite products currently Microwave Imager (TMI) Soil Moisture Pathfinder Data Set
available originate from observations in the low frequency have been compared in this study.
microwave region ranging from 5 GHz to 11 GHz [e.g.,
Gao et al., 2005; Wagner and Scipal, 2000; Njoku et al.,
2003]. 2. Soil Moisture Data Sets for the Southern US
[3] In recent years a number of studies addressed the [6] The TMI Soil Moisture Pathfinder Data Set for the
assimilation of satellite derived soil moisture or brightness southern US as described by Gao et al. [2005] was derived
temperatures into numerical land surface models [e.g., from the 10.65 GHz observations at horizontal polarization,
Reichle and Koster, 2005]. In any assimilation system based surface and vegetation temperature fields from VIC, and
geophysical data sets from the North American Land Data
Copyright 2005 by the American Geophysical Union. Assimilation System (NLDAS) database [Mitchell et al.,
0094-8276/05/2005GL023623 2004]. In this study, the daily averaged, quality checked
L15403 1 of 4
19448007, 2005, 15, Downloaded from [Link] by Universita Di Brescia, Wiley Online Library on [01/09/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
L15403 DRUSCH ET AL.: OBSERVATION OPERATORS L15403
Figure 1. Time series of volumetric soil moisture for Figure 3. Observation operators as derived from observed
the Oklahoma state area for the period from 1 June to soil moisture (TMI: thick lines, in situ: thin lines) and the
31 October, 2002. modeled data sets (ECMWF: solid lines, VIC: dashed lines).
The individual elements of the ranked data arrays are
displayed as squares (TMI/ECMWF), triangles (TMI/VIC),
Level 3 data set at 0.125° resolution has been used. The data
diamonds (in situ/ECMWF), and stars (in situ/VIC). The
were re-sampled to the reduced Gaussian model grid at
data represent the 2002 time period shown in Figure 1.
T511 spectral resolution (40 km), which represents the
resolution of ECMWF’s current operational forecast system.
The ECMWF soil moisture fields for the top layer (7 cm climatology characterized by a specific statistical distribu-
depth) were taken from the operational analysis at 12 UTC. tion. The most striking features are differences in the mean
For a description of the land surface scheme and the volumetric soil moisture content with VIC showing the
Optimal Interpolation soil moisture analysis the reader is highest values and the observations exhibiting the lowest
referred to van den Hurk et al. [2000] and Douville et al. values. Both models have a comparable dynamical range of
[2000], respectively. surface soil moisture, which is slightly smaller when com-
[7] For the Oklahoma state area, two additional data sets pared to in situ observations. In general, the two model-
have been analysed. In situ observations from the Oklahoma derived data sets and the in situ observations are in good
Mesonet were used to compute daily means for the period agreement. The TMI-derived soil moisture data set is
from 1 June to 31 October, 2002. Spatial averages for the significantly different since it represents a much thinner
Oklahoma state were derived from 72 individual sites. The layer: The dynamical range is larger and the dry down
observations have been taken at a depth of 5 cm; details on occurs much faster. Both effects result in a non-Gaussian
this data set can be found under ([Link] frequency distribution, which is shifted towards low soil
For the corresponding period, modeled soil moisture fields moisture values and is characterised by a significant low
from the NLDAS/VIC runs were analysed. The horizontal bias compared to the model data sets and the in situ
resolution for the VIC model runs is 0.125°, the depth of the observations. Obviously, an appropriate observation opera-
topsoil layer is 10 cm. The VIC model output was also tor is needed to correctly assimilate the TMI data into one of
spatially averaged to represent the Oklahoma state area. the models. The same is true, but to a lesser extent, for the in
[8] Time series of surface soil moisture for the Oklahoma situ observations.
state area are shown in Figure 1. Every data set has its own
3. CDF Matching for the Oklahoma Area
[9] Cumulative Distribution Function matching techni-
ques have been widely used in a range of geophysical
applications. Reichle and Koster [2004] have been using
the concept for bias reduction in satellite retrieved surface
soil moisture. Conceptually, the satellite derived data have
to be scaled so that the CDFs of both data sets match
(Figure 2) and:
cdfm x1 ¼ cdfs ð xÞ ð1Þ
where cdfs and cdfm are the CDFs of the satellite derived soil
moisture and the corresponding model data, respectively,
and x and x1 are the original and transformed satellite data
[Reichle and Koster, 2004]. Technically, the two data sets
(modeled soil moisture and observations) have to be ranked.
Consecutively, the differences in soil moisture between the
Figure 2. CDFs for the Oklahoma region (region 5) corresponding elements of each ranked data set have to be
computed from TMI derived soil moisture and ECMWF computed. The observation operator, which removes the
top layer soil moisture for the period from 1 March to systematic differences between both data sets, is a
31 October 2002. The arrows illustrate how the TMI data polynomial fit to the ranked observed soil moisture values
are transferred into the adjusted data set. and the corresponding differences.
2 of 4
19448007, 2005, 15, Downloaded from [Link] by Universita Di Brescia, Wiley Online Library on [01/09/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
L15403 DRUSCH ET AL.: OBSERVATION OPERATORS L15403
80° w.L. and 34° n.L. to 37° n.L.. Region no. 5 represents
the Oklahoma state area reasonably well. For each area the
observation operators for TMI/ECMWF have been derived
based on time series from 1 March to 31 October 2002
(Figure 5). The observation operators differ significantly
for the individual regions. The mountain regions 1 to 4
show comparably small temporal variations in modeled
and observed soil moisture with most of the TMI soil
moisture values in the range from 10% to 20% and
differences for modeled values varying from 8% to 12%.
The Central US regions 5 to 7 form a group with similar
Figure 4. Time series of volumetric soil moisture for the observation operators, which have the same characteristics
Oklahoma state area for the period from 1 June to 31 found for the Oklahoma state area. Region 8 exhibits a
October 2002, for the original TMI data set (light blue), the smaller dynamical range of TMI observations and the
ECMWF surface layer product (black), and the transformed corresponding observation operator is comparable to the
TMI product (blue). one obtained for regions 5 to 7, albeit by a shift of 4%.
[14] For the same areas, observation operators have been
derived from the corresponding 2001 period (not shown).
Qualitatively, the spatial variability for the eight regions is
[10] Based on the data shown in Figure 1 observation the same as for the 2002 period. In general, the differences
operators for the modeled and observed soil moisture data in the observation operators between 2002 and 2001 are
sets have been derived through least square fits of a third small for the central US regions 5 to 7 and the east (region 8).
order polynomial (Figure 3). A general feature to notice is For the western US, the differences are more pronounced.
that the observation operators are non-linear for the TMI Although the general shape of the observation operators are
observations. For the in situ observations, linear transfor- comparable, biases of 2% to 3% were found.
mations would be a reasonable approximation for the largest [15] It is interesting to quantify the systematic and
part of the observed soil moisture range. It also becomes random errors introduced through the use of non-ideal
clear that the observation type defines the shape of the observation operators, which have been derived for a
observation operator through the specific statistical proper- different spatial region or time. Table 1 summarizes selected
ties of its frequency distribution. The two different models, results for three southern US areas and the 2001 and 2002
which were forced with different input data sets but are seasons described previously. For each region, the original
characterized with comparable soil layer depths, show TMI data have been compared against the ECMWF model
different biases. However, the shape of the observation field (first line). In addition, the consistency test is shown to
operator is very similar (Figure 3) for the ECMWF model demonstrate the applicability of CDF matching for various
and VIC. regions with different soil moisture climatologies (second
[11] As a self consistency test, the observation operator line). These values have to be compared with the results
was applied to the 2002 TMI time series (Figure 4). It can obtained for an observation operator from a different region
be noted that the bias between the ECMWF and the TMI (third line) and from a different period (fourth line).
data sets has been reduced from 11.67% to 0.35%. In
addition, the agreement between the dynamical ranges of
both data sets is very high. CDF matching is limited through
the polynomial fit so that there can not be perfect agree-
ment. In general, the random difference between both data
sets will change. For the first part of the time series from
1 June to 1 October 2002, the explained variance r2 is
decreased from 0.18 to 0.01. For the second part values
of 0.69 and 0.66 have been obtained for the original and
transformed data set, respectively.
[12] In summary, CDF matching is an appropriate method
to remove systematic differences. It is applicable also to non
linear relationships and data sets with non-Gaussian distri-
butions. For data assimilation applications rms errors are
less critical since advanced analysis systems are designed to
handle random errors in a statistically optimal way.
4. Spatial and Temporal Transferability of TMI
Observation Operators
[13] For operational applications it is important to know Figure 5. Observation operators calculated from 2002 data
if the observation operator is transferable in space and time. (2002/3/1 to 2002/10/31) for eight individual regions in the
To address the spatial variability of the observation operator southern US. The values in brackets define the longitudinal
the southern US has been divided in eight sub-regions of 5° extent of each region, the northern and southern boundaries
longitude and 3° latitude, covering the area of 120° w.L. to for each region are 37° and 34°, respectively.
3 of 4
19448007, 2005, 15, Downloaded from [Link] by Universita Di Brescia, Wiley Online Library on [01/09/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
L15403 DRUSCH ET AL.: OBSERVATION OPERATORS L15403
Table 1. Performance of Observation Operators for the 2002 [19] The length of the temporal sampling was found to
Period and Three Selected Regions (1, 5, and 8; First Two have an impact on the observation operator in the western
Columns)a US and the eastern region. It would be desirable to study
Comparison this effect in more detail but operational numerical weather
TMI Obs. Obs. Operator Against ECMWF prediction (NWP) models are frequently up dated. Conse-
Region Time Region Time Bias r2 quently, the statistical properties of the modeled first guess
1 2002 – – 6.79 0.34 will not be constant for long time periods. For operational
1 2002 1 2002 0.00 0.32 NWP applications, it might be necessary to introduce
1 2002 2 2002 0.89 0.34 temporally changing observation operators. Practically, the
1 2002 1 2001 1.47 0.36 CDF matching would have to be done regularly based on
5 2002 – – 11.1 0.58
5 2002 5 2002 0.01 0.56
the latest data available. Since the interannual variability of
5 2002 6 2002 0.47 0.57 vegetation characteristics, used in the soil moisture retrieval
5 2002 5 2001 0.08 0.58 and the modeled first guess, are poorly known a temporally
8 2002 – – 15.9 0.16 varying bias correction scheme might be required anyway.
8 2002 8 2002 0.00 0.16 The work presented gives valuable information on the
8 2002 5 2002 3.25 0.15
8 2002 8 2001 0.26 0.16 design of future observation operators for the direct
a
For each region, the first row shows the bias and r2 between observation
operational assimilation of satellite retrieved soil moisture
and modeled field without the application of an observation operator. The products.
consecutive rows show the consistency test, results obtained for an
observation operator from a different region, and results obtained for an [20] Acknowledgments. The authors acknowledge the availability of
observation operator from a different period. data from the NLDAS project, of which the second author is a co-
investigator. The authors also acknowledge and appreciate the soil moisture
data made available by the Oklahoma Mesonet system. We would like to
[16] The general conclusion from these comparisons is thank P. Viterbo and E. Andersson (both ECMWF) for their helpful
comments on the manuscript and A. Bowen (ECMWF) for her help with
that the transferability strongly depends on the region of the preparation of the final figures.
interest. For the western US the temporal sampling seems to
be more important than the spatial sampling, for the eastern References
US it seems to be the other way around. The most Dee, D. P., and A. da Silva (1998), Data assimilation in the presence of
insensitive region is the central US area. This is encourag- forecast bias, Q. J. R. Meteorol. Soc., 124, 269 – 295.
ing, because (1) the microwave observations are most Douville, H., P. Viterbo, J.-F. Mahfouf, and A. C. M. Beljaars (2000),
Evaluation of the optimum interpolation and nudging techniques for soil
sensitive to soil moisture in sparsely vegetated areas, and moisture analysis using FIFE data, Mon. Weather Rev., 128, 1733 – 1756.
(2) the sensitivity of the atmosphere to soil moisture Gao, H., E. F. Wood, T. Jackson, M. Drusch, and R. Bindlish (2005), Using
changes is comparably high in this area [Koster et al., TRMM/TMI to retrieve soil moisture over the southern United States
2004]. from 1998 to 2002, J. Hydrometeorl., in press.
Koster, R. D., and P. C. D. Milly (1997), The interplay between transpira-
tion and runoff formulations in land surface schemes used with atmo-
spheric models, J. Clim., 10, 1578 – 1591.
5. Summary and Discussion Koster, R. D., and the GLACE Team (2004), Regions of strong coupling
between soil moisture and precipitation, Science, 305, 1138 – 1140.
[17] Since model soil moisture and satellite derived soil Liang, X., D. P. Lettenmaier, E. F. Wood, and S. J. Burges (1994), Hydro-
moisture represent different layer thickness and each soil logically based model of land-surface water and energy fluxes for general
moisture data set has its own soil moisture climate, obser- circulation models, J. Geophys. Res., 99, 14,415 – 14,428.
Mitchell, K. E., et al. (2004), The multi-institution North American Land
vation operators are needed for data assimilation applica- Data Assimilation System (NLDAS): Utilizing multiple GCIP products
tions. CDF matching has been used to derive observation and partners in a continental distributed hydrological modeling system,
operators for the direct assimilation of satellite derived soil J. Geophys. Res., 109, D07S90, doi:10.1029/2003JD003823.
moisture. It was shown that two different hydrological Njoku, E. G., T. J. Jackson, V. Lakshmi, T. Chan, and S. V. Nghiem (2003),
Soil moisture retrieval from AMSR-E, IEEE Trans. Geosci. Remote
models, which were forced with different meteorological Sens., 41, 215 – 229.
data, require similar observation operators. In general, CDF Reichle, R., and R. D. Koster (2004), Bias reduction in short records
matching is an appropriate technique to reduce systematic of satellite soil moisture, Geophys. Res. Lett., 31, L19501, doi:10.1029/
2004GL020938.
differences between the first guess and the observation Reichle, R., and R. D. Koster (2005), Global assimilation of satellite surface
without attributing the biases to a source. soil moisture retrievals into the NASA catchment land surface model,
[18] However, the transferability of observation operators Geophys. Res. Lett., 32, L02404, doi:10.1029/2004GL021700.
van den Hurk, B. J. J. M., P. Viterbo, A. C. M. Beljaars, and A. K. Betts
depends on the region and the temporal sampling. It was (2000), Offline validation of the ERA40 surface scheme, ECMWF Tech.
shown by Reichle and Koster [2004] that a moving window Memo. 295, Eur. Cent. for Medium-Range Weather Forecasts, Reading,
of 2° gives reliable results. For individual regions, e.g. the UK.
central US, it might even be possible to extend the sampling Wagner, W., and K. Scipal (2000), Large-scale soil moisture mapping in
western Africa using the ERS scatterometer, IEEE Trans. Geosci. Remote
area to larger domains, which is beneficial for global Sens., 38, 1777 – 1782.
operational applications. In addition, a large spatial sam-
pling window allows to decrease the sampling time without
reducing the actual sample size. This will help to define M. Drusch, Data Assimilation Section, ECMWF, Shinfield Park, Reading
RG2 9AX, UK. ([Link]@[Link])
observation operators for short time series, e.g. the first H. Gao and E. F. Wood, Department of Environmental Engineering,
year(s) after a satellite has been launched. Princeton University, Princeton, NJ 08544, USA.
4 of 4