An observing system simulation experiment for hydros radiometer-only soil moisture products

Abstract

Based on 1-km land surface model geophysical predictions within the United States Southern Great Plains (Red-Arkansas River basin), an observing system simulation experiment (OSSE) is carried out to assess the impact of land surface heterogeneity, instrument error, and parameter uncertainty on soil moisture products derived from the National Aeronautics and Space Administration Hydrosphere State (Hydros) mission. Simulated retrieved soil moisture products are created using three distinct retrieval algorithms based on the characteristics of passive microwave measurements expected from Hydros. The accuracy of retrieval products is evaluated through comparisons with benchmark soil moisture fields obtained from direct aggregation of the original simulated soil moisture fields. The analysis provides a quantitative description of how land surface heterogeneity, instrument error, and inversion parameter uncertainty impacts propagate through the measurement and retrieval process to degrade the accuracy of Hydros soil moisture products. Results demonstrate that the discrete set of error sources captured by the OSSE induce root mean squared errors of between 2.0% and 4.5% volumetric in soil moisture retrievals within the basin. Algorithm robustness is also evaluated for the case of artificially enhanced vegetation water content (W) values within the basin. For large W(>3 kg·m^-2), a distinct positive bias, attributable to the impact of sub- footprint-scale landcover heterogeneity, is identified in soil moisture retrievals. Prospects for the removal of this bias via a correction strategy for inland water and/or the implementation of an alternative aggregation strategy for surface vegetation and roughness parameters are discussed.

Full text

Page 1

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 43, NO. 6, JUNE 20051289
An Observing System Simulation Experiment for
Hydros Radiometer-Only Soil Moisture Products
Wade T. Crow, Member, IEEE, Steven Tsz K. Chan, Senior Member, IEEE, Dara Entekhabi, Senior Member, IEEE,
Paul R. Houser, Ann Y. Hsu, Thomas J. Jackson, Fellow, IEEE, Eni G. Njoku, Fellow, IEEE,
Peggy E. O’Neill, Senior Member, IEEE, Jiancheng Shi, Senior Member, IEEE, and Xiwu Zhan
Abstract—Based on 1-km land surface model geophysical
predictions within the United States Southern Great Plains
(Red-Arkansas River basin), an observing system simulation ex-
periment (OSSE) is carried out to assess the impact of land surface
heterogeneity, instrument error, and parameter uncertainty on
soil moisture products derived from the National Aeronautics
and Space Administration Hydrosphere State (Hydros) mission.
Simulated retrieved soil moisture products are created using three
distinct retrieval algorithms based on the characteristics of passive
microwave measurements expected from Hydros. The accuracy
of retrieval products is evaluated through comparisons with
benchmark soil moisture fields obtained from direct aggregation
of the original simulated soil moisture fields. The analysis provides
a quantitative description of how land surface heterogeneity,
instrument error, and inversion parameter uncertainty impacts
propagate through the measurement and retrieval process to
degrade the accuracy of Hydros soil moisture products. Results
demonstrate that the discrete set of error sources captured by
the OSSE induce root mean squared errors of between 2.0%
and 4.5% volumetric in soil moisture retrievals within the basin.
Algorithm robustness is also evaluated for the case of artificially
enhanced vegetation water content ?
For large
?? kg m
to the impact of sub-footprint-scale landcover heterogeneity, is
identified in soil moisture retrievals. Prospects for the removal
of this bias via a correction strategy for inland water and/or the
implementation of an alternative aggregation strategy for surface
vegetation and roughness parameters are discussed.
? values within the basin.
??, a distinct positive bias, attributable
IndexTerms—Microwaveremotesensing,observingsystemsim-
ulation experiment, soil moisture, spaceborne radiometry.
Manuscript received July 16, 2004; revised January 19, 2005. This work
was supported by the National Aeronautics and Space Administration (NASA)
Earth System Science Pathfinder (ESSP) Program Risk-Mitigation Phase
funding to the Hydros Science Team. The high-resolution land surface mod-
eling approach and forcing dataset used in this analysis were developed under
the direction of E. Wood at Princeton University through support from NASA
under Grants NAG8-1517 and NAG5-6494 and from the National Oceanic
and Atmospheric Administration under Grant NA96GP0413. The research
described in this paper was carried out in part at the Jet Propulsion Laboratory,
California Institute of Technology, under contract to the National Aeronautics
and Space Administration.
W. T. Crow, A. Y. Hsu, and T. J. Jackson are with the Hydrology and
Remote Sensing Laboratory, U.S. Department of Agriculture, Agricultural
Research Service, BARC-West, Beltsville, MD 20705 USA (e-mail: wcrow@
hydrolab.arsusda.gov).
S. T. K. Chan and E. G. Njoku are with the Jet Propulsion Laboratory,
California Institute of Technology, Pasadena, CA 91109 USA.
D.EntekhabiiswiththeDepartmentofCivilandEnvironmentalEngineering,
Massachusetts Institute of Technology, Cambridge, MA 02139 USA.
P. R. Houser, P. E. O’Neill, and X. Zhan are with the Hydrological Sciences
Branch, NASA Goddard Space Flight Center, Greenbelt, MD 20771 USA.
J. Shi is with the Institute for Computational Earth System Sciences, Univer-
sity of California, Santa Barbara, CA 93106 USA.
Digital Object Identifier 10.1109/TGRS.2005.845645
I. INTRODUCTION
A
reliability of soil moisture retrievals in densely vegetated areas
and the global extent over which retrievals will be possible. Er-
rors in retrieved soil moisture can originate from a variety of
sources within the measurement and retrieval process. In addi-
tion to instrument error, two key contributors to retrieval error
are the masking of the soil microwave signal by vegetation [1],
[2]andtheinterplaybetweennonlinearretrievalphysicsandthe
relatively poor spatial resolution of passive microwave space-
borne sensors [3]–[6]. Quantification of these errors requires
the realistic specification of land surface soil moisture hetero-
geneityandspatialvegetationpatterns.Sincedetailedsoilmois-
ture patterns are currently difficult to obtain from direct obser-
vation,anattractivealternativeistheapplicationofanobserving
system simulation experiment (OSSE) in which simulated land
surface states are propagated through the sensor measurement
andretrievalprocesstoinvestigateandconstrainexpectedlevels
of retrieval error [7], [8].
This analysis describes an OSSE performed for soil moisture
products to be derived from the NASA Hydrosphere State (Hy-
dros) mission. The Hydros mission is an Earth System Science
Pathfinder (ESSP) mission selected in 2002 by NASA for fur-
ther development and currently scheduled for launch in 2010
[9].Hydrosisdesignedtoprovideglobalmapsoftheearth’ssoil
moisture and freeze/thaw state every 2–3 days for weather and
climate prediction, water, energy, and carbon cycle studies, and
natural hazards monitoring. Hydros utilizes a unique active and
passive L-band (1.2–1.4 GHz) microwave concept to simulta-
neously measure microwave emission and backscatter from the
surface across a wide spatial swath [9], [10]. The Hydros an-
tenna is an approximately 6-m diameter deployable lightweight
mesh reflector that provides footprint sizes of approximately
40 km for the radiometer and 30 km for the radar. The radar
resolution is enhanced to 1–3 km using synthetic aperture pro-
cessing. The key derived products are soil moisture at 40-km
resolution for hydroclimatology obtained from the radiometer
measurements, soil moisture at 10-km resolution for hydrome-
teorologyobtainedbycombining theradarand radiometermea-
surements in a joint retrieval algorithm, and freeze/thaw state at
3-km resolution for terrestrial carbon flux dynamics studies ob-
tained from the radar measurements.
For the OSSE described in this paper, we focus on the 40-km
soil moisture product, and use a baseline set of passive mi-
N IMPORTANT issue in the development of a dedicated
spacebornesoilmoisturesensorhasbeenconcernoverthe
0196-2892/$20.00 © 2005 IEEE

Page 2

1290IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 43, NO. 6, JUNE 2005
crowave retrieval algorithms to illustrate Hydros retrieval capa-
bilities and examine error propagation in the algorithms. The
synthetic experiment is driven by realistically heterogeneous
land surface geophysical variables generated from a distributed
land surface model. These states are used to derive a set of sim-
ulated brightness temperatures which are degraded (i.e., spa-
tially aggregatedandrandomly perturbedwithsimulated instru-
mentnoise)tosimulateHydrosmeasurementsandtheninverted
back into soil moisture products using various retrieval algo-
rithms. Comparison of these retrievals to the original soil mois-
ture field reveals how the performance of soil moisture retrieval
algorithmswillbeimpactedbyvegetationdensity,measurement
resolution, inversion parameter uncertainty, and land cover het-
erogeneity. The intent of this analysis is not to capture the full
rangeoferrorsourcesimpactingtheaccuracyofspacebornesoil
moistureretrievals.Rather,itistostudytheerrorpropagationof
adiscretenumberoferrorsourcesthroughtheentireHydrosob-
servation and retrieval system and examine prospects for reme-
diation strategies to reduce the impact of these errors on Hydros
soil moisture products. OSSE-type experiments provide a con-
trolled-synthetic environment within which such issues can be
addressed prior to sensor launch. The comparison of results in
this paper obtained from different retrieval algorithms is not in-
tended as a selection test for the algorithms but as a means to in-
vestigate the sensitivities of the different algorithm types to the
error sources studied in this OSSE. Continuing evaluations of
these and other algorithmswill be performed in theyears ahead,
using a variety of simulation and observational datasets, for fur-
ther algorithm optimization and selection prior to the launch
of Hydros. This particular study focuses exclusively on 40-km
Hydros soil moisture products derived from radiometer obser-
vations and passive-only soil moisture algorithms. Future work
will focus on higher resolution Hydros products derived from
radar observations.
II. HYDROS INSTRUMENT DESIGN
The Hydros instrument is designed for a 670-km circular,
Sun-synchronous orbit, with equator crossings at 6 A.M. and
6 P.M. local time. The instrument combines radar and ra-
diometer subsystems that share a single feedhorn and parabolic
mesh reflector. The radar operates with vertical (VV), hori-
zontal (HH), and horizontal–vertical (HV) transmit–receive
polarizations, and uses separate transmit frequencies for the H
(1.26 GHz) and V (1.29 GHz) polarizations. The radiometer
operates with V, H and U (third Stokes parameter) polarizations
at 1.41 GHz. The reflector is offset from nadir and rotates
about the nadir axis at 14.6 rpm, providing a conically scanning
antenna beam with a surface incidence angle of approximately
40 . The reflector diameter is roughly 6 m, providing a ra-
diometer footprint of approximately 40 km defined by the
one-way 3-dB beamwidth. The two-way 3-dB beamwidth de-
fines the real-aperture radar footprint of approximately 30 km.
The swath width of 1000 km provides global coverage within
3 days at the equator and 2 days at boreal latitudes (
or S). To obtain the desired 3-km spatial resolution, the radar
employs range and Doppler discrimination. Relevant Hydros
instrument characteristics are listed in Table I. Discussion of
N
TABLE I
HYDROS INSTRUMENT CHARACTERISTICS
theinstrument measurementerrors and samplingcharacteristics
is deferred to Section VI. Details of the Hydros mission and
instrument design are provided in [9] and [10].
III. OSSE DESIGN
The OSSE is designed to simulate Hydros sensor and orbital
characteristics. The simulation involves four elements: 1) a land
surface model (LSM); 2) a forward microwave emission model
(MEM); 3) an orbit and sensor model (OSM); and 4) an inverse
retrieval model (RM). A flowchart of the forward and inverse
modeling is shown in Fig. 1. The LSM is used to generate a
simulated 2-D spatial scene as a 1-km grid of geophysical pa-
rameters covering a sufficiently large region of the U.S. to in-
corporate a diversity of land cover types and representative dy-
namic ranges of soil moisture, temperature, and other surface
characteristics.
Brightness temperatures
frequencies, polarizations and look angle via the MEM. Inputs
to the MEM are provided by 1-km gridded geophysical vari-
ables—predicted by the LSM—and observed land surface char-
acteristics. The OSM applies representative orbit sampling and
antenna spatial resolution characteristics to the simulated mea-
surements, and adds expected instrument noise and relative cal-
ibration error. It outputs simulated Hydros radiometer measure-
ments at 40 km which are run through a series of RMs to obtain
40-km soil moisture estimates. These retrieved soil moistures
are compared with the original soil moisture “truth” fields (after
degrading these to 40 km) to obtain the error characteristics
of the radiometer-based soil moisture retrievals. Consequently,
the OSSE can conceptually be divided up into two parts. The
are computed at the Hydros

Page 3

CROW et al.: OBSERVING SYSTEM SIMULATION EXPERIMENT FOR HYDROS1291
Fig. 1.
brightness temperature imagery. Key components include the land surface
modeling (LSM), microwave emission modeling (MEM), orbital and sensor
modeling (OSM), and soil moisture retrieval modeling (RM).
Illustration of OSSE procedure using simulated soil moisture and
“forward” portion of the simulation—described in Sections IV
and V—which consists of the generation, aggregation, and per-
turbation of synthetic Hydros observations, and the “retrieval”
portion—described in Section VI—which entails the inversion
of synthetic microwave observations into Hydros soil moisture
products.
IV. LAND SURFACE MODELING
The starting point for the simulations is the generation of
high-resolution (1-km) geophysical fields using a land surface
model (LSM) applied within the United States Southern Great
Plains. The LSM used in the computations is the TOPLATS hy-
drological model [11]. The spatial domain is the 575000-km
Red-Arkansas River Basin in the south-central United States
(seeFig.2).Theapproachtothelandsurfacemodelinganddata-
base generation is based on [7] and [12]. The simulations over
the spatial domain are forced by the following static datasets: 1)
a U.S. Geologic Survey land cover database1at 1 km derived
from Advanced Very High-Resolution Radiometer (AVHRR)
overpasses between April 1992 and March 1993 using the clas-
sification methodology presented in [13] and [14]; 2) a soil tex-
ture database at 1 km derived from merged State Soil Geo-
graphic (Penn State University) products;23) a USGS digital
elevation model at 1 km;3and 4) a normalized difference veg-
1See http://www.nationalatlas.gov/landcvm.html.
2See http://www.essc.psu.edu/.
3See http://edcdaac.usgs.gov/gtopo30/gtopo30.asp.
Fig. 2.Location of OSSE domain within the continental United States.
etation index (NDVI) database4at 1 km derived from AVHRR
data and based on composite imagery from June 1995. The land
cover and soil texture classifications are listed in Tables II and
III.
TOPLATS is run with three water and energy balance layers.
The top water balance layer extends from the surface to a depth
of 5 cm, and the second layer from 5 cm to the top of the water
table (depth to water table is also a dynamic state variable). The
top layer soil moisture value is taken to be an integrated value
over the top 5 cm. The energy balance discretization has a node
atthesurface,anodeat5cmandthirdnodeat50cm.The50-cm
node is prescribed to vary seasonally. Input forcings are de-
rived from 4-km Weather Surveillance Radar precipitation im-
agery, 1-km Geostationary Operational Environmental Satellite
(GOES) solar radiation imagery, and spatially interpolated Na-
tional Climate Data Center surface airways meteorology data.
Model outputs are generated at hourly time steps and saved in
the output database every 12 h at 6 A.M. and 6 P.M. local times
to simulate the Hydros overpass times. The outputs generated
by the model include: 1) vertically integrated 0- to 5-cm soil
moisture expressed as volumetric soil moisture; 2) surface (soil
or vegetation) skin temperature; and 3) soil temperature at 5 cm
below the surface. Model validation against ground measure-
ments from the Southern Great Plains Atmospheric Radiation
Measurement Cloud and Radiation Test bed (ARM-CART) site
is described in [12]. The time period for the LSM simulations
is from April 1, 1994 to July 31, 1994. Basin-averaged rain-
fall and simulated soil moisture conditions for this time period
are plotted in Fig. 3. However, the actual OSSE is conducted
only for the subset of this period between May 26, 1994 to June
28,1994thatdemonstratesapronouncedtransitionbetweenwet
and dry surface conditions.
4See http://edc.usgs.gov/products/landcover/ndvi.html.

Page 4

1292IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 43, NO. 6, JUNE 2005
TABLE II
LAND COVER CLASSIFICATIONS AND REPRESENTATIVE ROUGHNESS AND
VEGETATION PARAMETERS (DESCRIBED IN TEXT)
TABLE III
SOIL TEXTURE CLASSES
V. MICROWAVE EMISSION MODELING
BasedontheLSMoutputfields,amicrowaveemissionmodel
(MEM) is used to simulate radiometer observations of the geo-
physical scene. The parameterizations of the MEM represent
tradeoffs between the need to adequately represent the key ef-
fects of surface characteristics on microwave signatures at the
spatial scale of interest (40 km), and the need for a sufficiently
simple representation for application to satellite retrieval algo-
rithms. The MEM incorporates the effects of dynamic features,
e.g., surface soil moisture and soil temperature, and static fea-
tures (on the simulation time scale) such as soil texture, soil
surface roughness, land-cover and vegetation type, and vegeta-
tion water content. Effects of soil freezing, thawing, and snow
cover are not considered in the simulations, and effects of at-
mospheric variability are assumed negligible at L-band for non-
raining conditions. Computations are performed at the 40 look
Fig.3.
moisture simulations within the OSSE domain. Dashed vertical lines represent
the starting and stopping times of the OSSE simulation.
TimeseriesofareallyaveragedprecipitationandLSM(TOPLATS)soil
angle and at 1.41-GHz V and H polarizations. Heterogeneity is
taken into account by aggregating from the 1-km scale up to the
40-km product scale for Hydros radiometer-based soil moisture
retrievals.
The microwave emission model is based on a layered single-
scattering soil–vegetation model commonly used for passive
microwave sensing at L-band [1]. The modeled brightness tem-
perature
includes components from the soil and vegetation
and is expressed as
(1)
The subscript
fective temperature,
nadir vegetation opacity,
albedo, and
to the emissivity by
at the Hydros look angle of
that vegetation multiple scattering and reflection at the vegeta-
tion–air interface are negligible. Specular reflection is assumed
at the soil surface, but with reflectivity modified for surface
roughness. The parameters
and
vegetation volume (leaves, stalks, branches, trunks) modeled as
a homogeneous medium. Nadir vegetation opacity is related to
the total columnar vegetation water content
refers to polarization ( or ),
is the vegetation temperature,
is the vegetation single scattering
is the soil reflectivity. The reflectivity is related
, and
is the soil ef-
is the
, , andare values
. Equation (1) assumes
represent the composite
kg mby
(2)
The coefficient
the predominantly vertical orientation of vegetation stem and
depends on vegetation type [2], [15]. Due to

Page 5

CROW et al.: OBSERVING SYSTEM SIMULATION EXPERIMENT FOR HYDROS1293
trunk structure,
types [16]. The vegetation water content
overthe1-kmpixelsincenoattemptismadetomodelfractional
vegetation cover within a 1-km pixel. In addition, the polariza-
tion dependence of
is ignored.
Roughness influences the sensor response through the root
mean squared (RMS) surface height , the horizontal correla-
tionlength ,andtheroughnessspectrum.Theroughnesseffects
can be computed using a theoretical model such as the integral
equation model [17] or empirically using the
AtL-bandthepolarizationmixingparameter
roughness effect using this model can be approximated as
tends to be larger thanfor most vegetation
is the average value
model [18].
issmall,andthe
(3)
The parameter
surface height , and
smooth soil surface, determined by the soil dielectric constant
using the Fresnel equations.
At the 6 A.M. overpass time, the vegetation temperature
is assumed to be equal to the soil surface skin temperature
The soil microwave effective temperature
the weighted average of the surface skin temperature and the
5-cm temperature
. The weighting is actually a function of
soil moisture, since the penetration depth varies with moisture,
but, for simplicity, the direct average is used here
is assumed to be linearly related to the RMS
is the reflectivity of the equivalent
.
is estimated as
(4)
For a water surface the observed brightness temperature is
(5)
where
water at temperature
is neglected.
In general, the key assumptions implicit in (1)–(5) (e.g., the
neglect of multiple vegetation scattering and reflection, the ne-
glect of polarization differences in
between
and, and the linear relationship between
) have been supported by airborne and tower campaigns over
grass, shrubland, and agricultural crop land cover types [1],
[2], [20]. However, much less evidence is available to support
theirsuitabilityoverforestedregions.Consequently,predictions
made by the MEM are likely more accurate over lightly vege-
tated western and central portions of the Red-Arkansas River
Basin than over forested areas along the eastern edge of the
OSSE domain.
is the Fresnel smooth surface reflectivity for fresh
[19]. Wind-induced surface roughness
, the linear relationship
and
A. Soil Dielectric Mixing Model
The real part of the soil dielectric constant, , and the Fresnel
smoothsurfacereflectivities,
tions of soil moisture for the distribution of soil textures (sand
and clay fractions,
and) found over the basin. The compu-
tation is done using the empirical relations of Dobson et al. [21]
based on the data of Hallikainen et al. [22]. Parameters required
in these models include the soil bulk density and the specific
and ,arecomputedasfunc-
density of solid soil particles. A value of 2.66 g cm
sumed for soil specific density. Bulk density values were calcu-
lated from specific density and porosity values estimated from
andusing the regression relationship presented in [23].
was as-
B. Parameterization and Dataset Generation
Due to large uncertainties in the correct field-scale emission
parameterization for many land surface types, it is difficult to
unambiguously specify vegetation and surface parameters re-
quired by (1). While emission parameters are selected based on
best available published information and constitutive relation-
ships, some level of uncertainty and error is inevitable. How-
ever, it is important to note that the purpose of this study is not
toperfectlymimicactuallandsurfaceemissionforthestudype-
riod, but rather to investigate—as accurately as possible—key
aspects of error propagation in Hydros soil moisture retrievals
given our current level of understanding concerning large-scale
patterns in land surface emission parameters. Therefore, a rea-
sonable level of error in the parameters described in this section
is both expected and tolerable.
Vegetationandsurfaceroughnessparametersassignedtoland
cover types are listed in Table II. Surface roughness parameters
andare assigned to the different land cover classes in the
basin using values for categories of bare soils, crops, and grass-
land given in [20]. A factor of 0.1 cm
mate scaling between
(cm) and
values are intended to be representative of mean magnitudes for
the different land cover classes. For example, cropped surfaces
areassignedhighermeanroughnessthanpastureoruncultivated
surfaces. Due to a lack of reliable published information con-
cerning
values for trees, they are set to a baseline value of
1 cm. Nonpolarized values of the vegetation parameters
arealsoassignedbasedonlandcovertypes,usingrepresentative
values for herbaceous and woody vegetation listed in [1]. Based
loosely on polarization results presented in [16], values for ver-
tically polarized (horizontally polarized) -parameters are esti-
mated by increasing (decreasing) the baseline nonpolarized
values by 10% for grasses/crop/shrub landcovers and 20% for
forested areas (Table II). The canopy vegetation water content
is obtained from the NDVI dataset using the relationship
[20]
provides an approxi-
at L-band [20], [24]. Listed
and
NDVI NDVI (6)
This relationship was derived for grassland/crop conditions
using NDVI observations derived from 30-m Landsat Thematic
Mapper data over south-central Oklahoma. Here we extrapolate
its use to all land cover types using 1-km AVHRR observation
of the Red-Arkansas River basin between May 13 and July 7,
1995. While this is clearly stretching (6) somewhat beyond its
original formulation, the resulting 1-km
bestavailablerepresentationofspatialpatternsof
required for the analysis. For NDVI
values less than zero, in which case
the NDVI is sensitive to vegetation greenness, its signal comes
primarilyfromthefoliarcanopy and notthewoodycomponents
of the vegetation. In contrast, the microwave signals include
contributions from the water content of the trunks and branches
fields represent the
variability
, (6) predicts
is set to zero. Since