Validation of a Dual-Pass Microwave Land Data Assimilation System for Estimating Surface Soil...

Validation of a Dual-Pass Microwave Land Data Assimilation Systemfor Estimating Surface Soil Moisture in Semiarid Regions

KUN YANG

Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing, China

TOSHIO KOIKE

Department of Civil Engineering, The University of Tokyo, Tokyo, Japan

ICHIROW KAIHOTSU

Department of Natural Environmental Sciences, Hiroshima University, Hiroshima, Japan

JUN QIN

Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing, China

(Manuscript received 16 June 2008, in final form 1 November 2008)

ABSTRACT

This study examines the capability of a new microwave land data assimilation system (LDAS) for esti-

mating soil moisture in semiarid regions, where soil moisture is very heterogeneous. This system assimilates

the Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) 6.9- and 18.7-GHz

brightness temperatures into a land surface model (LSM), with a radiative transfer model as an observation

operator. To reduce errors caused by uncertainties of system parameters, the LDAS uses a dual-pass as-

similation algorithm, with a calibration pass to estimate major model parameters from satellite data and an

assimilation pass to estimate the near-surface soil moisture. Validation data of soil moisture were collected in

a Mongolian semiarid region. Results show that (i) the LDAS-estimated soil moistures are comparable to

areal averages of in situ measurements, though the measured soil moistures were highly variable from site to

site; (ii) the LSM-simulated soil moistures show less biases when the LSM uses LDAS-calibrated parameter

values instead of default parameter values, indicating that the satellite-based calibration does contribute to

soil moisture estimations; and (iii) compared to the LSM, the LDAS produces more robust and reliable soil

moisture when forcing data become worse. The lower sensitivity of the LDAS output to precipitation is

particularly encouraging for applying this system to regions where precipitation data are prone to errors.

1. Introduction

Soil moisture is a highly variable parameter in semi-

arid regions, where strong coupling between soil mois-

ture and precipitation are suggested by Koster et al.

(2004). Soil moisture estimation is imperative for hy-

drometeorological studies and water resources man-

agement (e.g., Pauwels et al. 2001; Wu and Dickinson

2004; de Goncalves et al. 2006), and how to estimate soil

moisture with good accuracy is still a hot topic of land

hydrology and satellite remote sensing.

Land hydrological modeling is subject to errors in many

factors, including meteorological forcing data, parame-

terization schemes, and model parameters. Accordingly,

individual land models may produce quite diverse re-

sults, and uncertainties of modeled soil moisture are

significant (Entin et al. 1999; Pitman et al. 1999; Schaake

et al. 2004; Yang et al. 2007b). Remote sensing of pas-

sive microwave has been widely used to estimate soil

moisture (e.g., Jackson 1993; Njoku and Entekhabi

1996; Shi et al. 1997; Njoku et al. 2003; Wen et al. 2003,

2005), as it is strongly affected by near-surface soil

moisture but not much by atmospheric state. However,

Corresponding author address: Professor Kun Yang, Institute of

Tibetan Plateau Research, Chinese Academy of Sciences, P.O.

Box 2871, Beijing 100085, China.

E-mail: [email protected]

780 J O U R N A L O F H Y D R O M E T E O R O L O G Y VOLUME 10

DOI: 10.1175/2008JHM1065.1

� 2009 American Meteorological Society

microwave signals are also affected by other parame-

ters, such as soil texture and surface roughness (Engman

1991); these parameters must be specified for estimat-

ing soil moisture from microwave data (Wang 1983;

Wigneron et al. 2003; Njoku and Chan 2006). A land

data assimilation system (LDAS), which combines a

land surface model (LSM) and satellite signals, may

provide a promising means for estimating soil moisture

by assimilating microwave data to correct modeling er-

rors (Reichle et al. 2001; Bach and Mauser 2003; Crow

and Wood 2003; Ni-Meister et al. 2006; Li et al. 2007).

Nonetheless, the assimilation technique does not di-

rectly aim at eliminating uncertainties caused by speci-

fying parameter values in land models and satellite

schemes. The parameter specification is a common and

crucial issue for land modeling, remote sensing, and

data assimilation.

Another common issue is how to make a full scale–

match comparison between satellite-estimated soil

moisture and in situ–observed soil moisture. Soil mois-

ture measured at a single point is often representative

merely on a limited scale, depending on heterogeneity

of soil properties, land cover, and atmospheric condi-

tions. A satellite microwave pixel may cover tens of

kilometers, which can be much larger than the repre-

sentative scale of single-point measurements of soil

moisture. Therefore, a scale–match comparison is es-

sential for evaluating a remote sensing retrieval scheme

or data assimilation algorithm for soil moisture. For

this purpose, some observational networks have been

developed (e.g., Jackson et al. 1999; Cosh et al. 2004;

Bosch et al. 2006; Bindlish et al. 2006; Vivoni et al.

2008).

In response to the aforementioned two issues, this

study evaluates the output of the land data assimilation

system of The University of Tokyo (LDAS-UT) against

measurements from a soil moisture network in a semi-

arid region. LDAS-UT was developed by Yang et al.

(2007a) and its applications, developed at a coordinated

enhanced observing period (CEOP; Koike 2004) Tibet

site, have shown potential for improving the surface

energy budget estimation. The soil moisture network

of interest in this study was established by Kaihotsu

(2005) in Mongolia through the Advanced Earth Ob-

serving Satellite II (ADEOS-II) Mongolian Plateau

Experiment for ground truth (AMPEX) project to col-

lect data for the development and validation of Ad-

vanced Microwave Scanning Radiometer for Earth

Observing System (AMSR-E) soil moisture retrieval

algorithms. AMPEX data have been archived in CEOP

and provide a test bed for remote sensing inversion

schemes and data assimilation algorithms to estimate

areal soil moisture in semiarid regions.

The paper is organized as follows: LDAS-UT is in-

troduced in section 2 [a detailed introduction is pre-

sented in Yang et al. (2007a)], and AMPEX data is

discussed in section 3. Section 4 presents numerical

experiment design, including (i) a control experiment

that is driven with Automatic Weather Station (AWS)

data, with the results analyzed in section 5; and (ii)

several sensitivity studies that are driven with globally

available, low-quality datasets, with the results pre-

sented in section 6. LDAS-UT estimates are evaluated

against observations and its performance is also com-

pared with soil moisture simulations that are not con-

strained by microwave satellite data. Conclusions and

remarks are given in section 7.

2. LDAS-UT

LDAS-UT is a dual-pass land data assimilation sys-

tem. It assimilates AMSR-E low-frequency brightness

temperature (Tb) data and its structure is shown in Fig. 1a.

The model operator is a LSM, and the observation oper-

ator is a radiative transfer model (RTM). The LSM pro-

duces near-surface soil moisture (usfc), ground tempera-

ture (Tg), and canopy temperature (Tc), which are then

fed into the RTM to simulate the brightness temperatures.

The difference between simulated Tb (Tbp,est) and ob-

served Tb (Tbp,obs) is then minimized by adjusting either

system parameters or near-surface soil moisture. The al-

gorithm, cost function, model operator, observation op-

erator, and system parameters are introduced below.

a. Dual-pass assimilation algorithm

Estimating brightness temperatures using the RTM

requires the input of several parameters (surface rough-

ness parameters, soil texture, and canopy optical pa-

rameters), in addition to surface variables (usfc, Tg, Tc).

The simulation of these surface variables using the LSM

also requires a number of soil and vegetation parame-

ters. The modeled Tb is thus sensitive to some crucial

parameters used in the LSM and the RTM, which should

be determined before estimating soil moisture. In re-

sponse, we developed the dual-pass assimilation algo-

rithm. Figure 1b shows the schematic of this algorithm

and the detail can be found in Yang et al. (2007a).

Pass 1, the so-called calibration pass, aims at tuning

system parameters; pass 2, the so-called assimilation pass,

estimates soil moisture. The principle behind this algo-

rithm lies in the system state variables responding to

system parameters and to initial near-surface soil mois-

ture at different time scales.

System parameters have a long-term effect on state

variables (such as soil moisture); therefore, a long time

window (several months or longer) is selected to calibrate

JUNE 2009 Y A N G E T A L . 781

the parameters. It should be noted that the parameter

calibration presented in LDAS-UT relies on satellite

microwave data instead of in situ observed soil moisture

and/or surface temperature; thus, it may have a wide

applicability. In principle, pass 1 requires just a single

execution because the optimized parameters only in-

clude static parameters in the LSM and the RTM. It can

be implemented using available high-accuracy data prior

to the real-time assimilation of satellite data for soil

moisture estimation.

By contrast to the parameters, initial near-surface soil

moisture has a short-term effect on the system state var-

iables; therefore, a short time window (;1 day) is selected

to estimate their values by minimizing a cost function.

This pass is similar to a normal variational data assimi-

lation system.

b. Cost function

LDAS-UT assimilates observed brightness tempera-

tures (Tb) of the vertical polarization (V) at a lower

frequency (6.9 GHz) and at a higher frequency (18.7

GHz). As the lower frequency Tb is much more sensi-

tive to near-surface soil moisture than the higher fre-

quency, their difference is correlated with soil wetness.

In pass 1, system parameters are obtained by mini-

mizing a cost function that accounts for the difference

between modeled and observed brightness tempera-

tures for a long-term window (tpass1; scale of months or

longer). The cost function includes an observation error

term and a background error term. The observation

error term is defined as

Fobs 5 �tpass1

t50[(T6.9V

b,est � T6.9Vb,obs)

21 (T18.7V

b,est � T18.7Vb,obs )2], (1)

where the subscript obs denotes the observed value and

est is the modeled value.

The background error term in the cost function is

related to a soil wetness index (SWI) that is defined by

SWI 5 2(T18.7Vb 2 T6.9V

b )/(T18.7Vb 1 T6.9V

b ). A high SWI

value corresponds to a wet surface and a low value to a

dry surface. The background error is taken into account

by adjusting the near-surface soil water content (usfc) at

each observing time, such that the recalculated SWI

value is close to (SWIest,bg 1 SWIobs)/2. Here, SWIest,bg

is the estimated background SWI at the observing time,

and SWIobs is the observed background SWI. Note that

the observation error term is taken into account for the

entire calibration period, but the background error term

is considered at each observing time.

The optimized parameter values in pass 1 are then

transferred into pass 2, where the near-surface soil

moisture is estimated with an assimilation window (tpass2)

of ;1 day. The cost function in pass 2 is defined as

Fobs 5 �tpass2

t50[(T6.9V

b,est � T6.9Vb,obs)

21 (T18.7V

b,est � T18.7Vb,obs )2]

1 [(T6.9Vb0,bg � T6.9V

b0 )21 (T18.7V

b0,bg � T18.7Vb0 )2], (2)

where Tb0,bg and Tb0 are the simulated brightness tem-

perature at the initial time of each assimilation cycle

using the background value of usfc and the renewed

value of usfc, respectively.

c. Model operator

The model operator is the second version of the Simple

Biosphere Model (SiB2), developed by Sellers et al.

(1996). This LSM includes one canopy layer and three soil

layers and describes canopy radiative transfer, aerody-

namic canopy transfer, and conductance–photosynthetic

processes. The model is a typical dual-source model that

FIG. 1. (a) LDAS-UT system structure and (b) schematic of dual-

pass assimilation technique. Here, Tg, Tc, and usfc are the ground

temperature, canopy temperature, and near-surface soil water

content, respectively; Tb is the brightness temperature, F the cost

function, and Dt the data assimilation window; Gp is soil reflectivity;

and tc is the optical thickness of the vegetation. The polarization,

observed value, and estimated value are shown by the subscript p,

obs, and est, respectively (refer to Yang et al. 2007a for details).


parameterizes heat transfer from both the canopy and

the ground. To apply the LSM in arid and semiarid re-

gions, the LSM in LDAS-UT has incorporated an

aerodynamic scheme for sparse canopy (Watanabe and

Kondo 1990) and a heat transfer scheme for bare soil

surfaces (Yang et al. 2008). The LSM is driven with

meteorological data (pressure, wind, temperature, hu-

midity, downward radiation, and precipitation) and it

produces surface state variables and land fluxes.

d. Observation operator

Microwave brightness temperature is given by

Tbp 5 Tg(1� Gp) exp(�tc)

1Tc(1�v)[1� exp(�tc)][1 1 Gp exp(�tc)], (3)

where the subscript p denotes polarization (vertical or

horizontal), Gp is soil reflectivity, tc is the optical thick-

ness of the vegetation, and v is the single-scattering

albedo of the vegetation.

The soil reflectivity can be calculated using the Q–h

model developed by Wang and Choudhury (1981):

Gp 5 [(1�Q)Rp 1 QRq] exp(�h), (4)

where the subscripts p and q denote vertical and hori-

zontal polarization, respectively; Q and h are empiri-

cally determined surface roughness parameters; and R

is the Fresnel power reflectivity that describes the soil

reflectivity of a smooth surface. The reflectivity depends

on soil moisture usfc.

The model parameters in Eqs. (3) and (4) are fre-

quency dependent and are given by

h 5 (ks)ffiffiffiffiffiffiffiffiffiffiffiffi

0.1 cosup

, (5)

Q 5 Q0(ks)0.795, (6)

tc 5 b9(100l)xwc/cosu, and (7)

v 5 0.000 83/l, (8)

where l (m) is the wavelength; k is the wavenumber

defined as 2p/l; u is the view angle of the satellite sensor;

s (m) is the standard deviation of surface roughness

height; wc (kg m22) is the vegetation water content; and

Q0, b9, and x are empirical coefficients.

e. Optimized parameters

In the LSM and the RTM, some parameters are

crucial to the estimation of soil moisture but have high

spatial variability. They are soil porosity (us), soil hy-

draulic and thermal parameters, surface roughness pa-

rameters (s and Q0), and the vegetation parameter (b9).

Their values are optimized by minimizing the cost

function in pass 1 using the shuffled complex evolution

method (Duan et al. 1993). Two procedures are adopted

to reduce the uncertainties of the parameter estimates.

First, soil thermal and hydraulic parameters are not

directly estimated; instead, they are converted from

soil texture and soil porosity, with the pedotransfer

functions listed below (see references in Yang et al.

2007a):

rscs 5 (0.076 1 0.748rd/rw) 3 106 1 (4.195 3 106)usfc,

(9)

ls 5 ld 1 (lm � ld) exp [0.36(1� 1/usfc)], (10)

lm 5 0.5us [7.7(0.013%sand)2.0(1�0.013%sand)]1�us , (11)

ld 5 (0.135rd 1 64.7)/(2700� 0.947rd), (12)

Ks 5 7.0556 3 10(�6.88410.01533%sand) m s�1, (13)

cs 5�0.01 3 10(1.88�0.01313%sand) m, and (14)

b 5 2.91 1 0.159 3 %clay, (15)

where us is the soil porosity; rscs (J K21 m23) is the soil

heat capacity; ls (W m21 K21) is the thermal conduc-

tivity of the soil; Ks (m s21), cs(m), and b are the hy-

draulic parameters of Clapp and Hornberger (1978);

rd (kg m23) is the bulk density of a dry soil calculated

using rd 5 rs(1 2 us) and rs 5 2650 kg m23; rw 5 1000

kg m23; and %sand and %clay are the percentages of

sand content and clay content in the soil, respectively.

This procedure not only reduces the number of opti-

mized parameters and thus reduces the computational

cost but it also enhances the physical consistency among

the parameter values.

Second, a global dataset is used to provide the default

values for soil texture (%sand and %clay) and an un-

certainty is added to the default values as the upper and

lower bounds of the optimized parameters.

3. AMPEX data

As shown in Fig. 2, the CEOP/Mongolia reference

site was located in Mandal Govi of Mongolia. The site

covers a flat area of 120 km 3 160 km in a semiarid

grassland, where 12 long-term automatic stations for

soil hydrology (ASSH) and six automatic weather sta-

tions (AWSs) were deployed. At ASSH, soil tempera-

ture and moisture were measured at 3 and 10 cm. AWS

measurements included wind, temperature, humidity,

pressure, and precipitation (see measuring height in

Table 1). Among them, data at two AWSs [Tsagaandelger

JUNE 2009 Y A N G E T A L . 783

(TDS) and Choyr (CRS)] were not archived in CEOP

datasets. In this study, near-surface soil moisture data

used for validation are measurements at 3-cm depth of 12

ASSH and four AWS [Bayantsagaan (BTS), Delgertsogt

site (DGS), Deren site (DRS), and Madalgobi site

(MGS)], and meteorological data from the four AWSs are

used to drive the LDAS-UT. Table 2 shows the annual

mean (1 October 2002 to 30 September 2003) of some

meteorological parameters, indicating that the climato-

logical conditions are similar in the experimental area.

AMPEX also collected AMSR-E brightness temper-

ature data for the 12 ASSH stations for the observa-

tional period, which were resampled from the native

resolutions of 56 km for 6.9 GHz and 21 km for 19.7

GHz to a resolution of 0.18.

4. Design of numerical experiments

This paper investigates a control case and several

sensitivity cases. The control case is used to show how

LDAS-UT works to improve the soil moisture estimate,

while the sensitivity cases are used to show the effect of

forcing data on LDAS-UT output. The experiment

design is shown in Table 3.

In the control case, LDAS-UT is applied to the 12

ASSH stations. The forcing data is averaged at the four

AWSs, but downward radiations are taken from Global

TABLE 1. Measured variables and sensor heights above the surface

at AMPEX AWS.

Parameter

Measuring height (m)

MGS DRS DGS BTS

Air temperature 1.6 1.5 1.5 1.5

Relative humidity 1.6 1.5 1.5 1.5

Wind speed 2.5 2.4 2.45 2.4

Precipitation 1.15 1.0 1.05 1.0

Air pressure 1.1 1 1.1 1.1

FIG. 2. Sketch of ADEOS-II AMPEX. CEOP archived data at four AWSs (BTS, DGS, DRS, and MGS) and

12 ASSH.

TABLE 2. Total precipitation (rain, mm), annual mean air tem-

perature (Ta, K), and specific humidity (qa, g kg21) at AMPEX

AWS for the October 2002–September 2003 period. Asterisk (*)

means data for the monsoon season were missing.

MGS DRS DGS BTS Mean

Rain 200.2 149.6 188.0 176.8 178.7

Ta 275.1 274.9 274.3 273.1 274.4

qa 3.1 1.8* 3.3 3.2 3.2


Land Data Assimilation System (GLDAS; Rodell et al.

2004) 18 3 18 data, because they were not measured

through AMPEX. The forcing data are applied to all the

stations. The input microwave data are the station-

dependent 0.18 gridded brightness temperatures of

AMSR-E 6.9- and 18.7-GHz vertical polarization. Re-

sults of this case will be analyzed in section 5.

There are three sensitivity case studies and the first

two cases each have three runs driven by different

forcing data. First, LDAS-UT is driven with three in-

dependent datasets, that is, AWS data, GLDAS output

for CEOP, and Japan Meteorological Agency (JMA)

model output for CEOP (hereafter AWS, GLDAS, and

JMA, respectively). AWS is identical to that used in the

control case, GLDAS has a resolution of 18 3 18, and

JMA has a resolution of about 60 km. The latter two

provide their data at the grid nearest to the AMPEX

area. GLDAS precipitation and radiation are satellite-

derived or satellite/gauge merged, and other meteoro-

logical data are derived by atmospheric assimilation;

thus, the quality of the GLDAS data should be gener-

ally better than JMA output. GLDAS wind speed is not

archived in CEOP, and thus AWS wind speed is inser-

ted into GLDAS data. This case study contributes to the

sensitivity analysis of LDAS-UT output to sources of

TABLE 3. Numerical experiment design. For sensitivity studies,

case 1 includes three runs driven by AWS, GLDAS, and JMA,

respectively; and case 2 also includes three runs driven by AWS,

except precipitation data are from AWS, GLDAS, and JMA, re-

spectively. Microwave data are the station-dependent data (0.18) in

the control case and the station-averaged data in the sensitivity

cases.

Case Forcing data Microwave data

Control AWS Station dependent

Sensitivity

Case 1 AWS; GLDAS; JMA Station average

Case 2 AWS; AWS 1 GLDAS rain;

AWS 1 JMA rain

Station average

Case 3 AWS 1 no rain Station average

FIG. 3. Daily and accumulated precipitations of AWS observation, GLDAS, and JMA output in

the AMPEX area.

JUNE 2009 Y A N G E T A L . 785

forcing data. Second, LDAS-UT is driven with AWS

data, except that precipitation is taken from one of

AWS, GLDAS, and JMA. This case contributes to un-

certainty analyses of LDAS-UT output in response to

precipitation data, whose quality is most questionable in

many regions. Third, LDAS-UT is driven by AWS

forcing, except that precipitation is set to zero. This case

investigates the role of microwave signal in the dual-

pass LDAS. These cases will be analyzed in section 6.

Note that LDAS-UT in all sensitivity cases assimilates

the station-averaged data rather than the 0.18 station-

dependent data.

Figure 3 shows daily and accumulated precipitations

of AWS, GLDAS, and JMA datasets in the AMPEX

area. JMA shows positive biases and GLDAS shows

negative biases in both precipitation frequencies and

accumulated amount. Figure 4 shows the monthly-mean

diurnal variations of other meteorological data. Both

GLDAS and JMA show general observed patterns, but

biases in temperature and humidity are evident.

Other ancillary data (leaf area index and default soil

texture) are directly taken from global datasets. The

default values of the soil texture and soil parameters are

sourced from 18 3 18 International Satellite Land Sur-

face Climatology Project (ISLSCP) Initiative II soil data

(Global Soil Data Task Group 2000), and values of

vegetation parameters (classification and coverage) are

derived from 18 3 18 ISLSCP Initiative II vegetation

data (Loveland et al. 2000). Leaf area index data is

sourced from Moderate Resolution Imaging Spectror-

adiometer (MODIS) 0.258 3 0.258 gridded 8-day leaf

area index products (Knyazikhin et al. 1999).

Note that the period of interest in this study is from

May to September 2003 to avoid modeling errors due to

soil freezing and thawing processes, which have not

been parameterized in the system.

5. Control experiment

LDAS-UT is applied to each ASSH station, driven by

the averaged AWS data and assimilating AMSR-E 0.18

gridded data. The accuracy of LDAS-UT estimates is

analyzed against measurements as follows.

Figure 5 compares the observed and estimated soil

moisture for five stations (A3, C2, E4, G6, and H7; see

their positions in Fig. 2), where soil moisture data were

FIG. 4. Monthly-mean diurnal variations of meteorological data of AWS observation, GLDAS,

and JMA output in the AMPEX area.


continuously recorded. It shows that the estimates

and observations agree well for some stations (e.g., A3

and C2) but does not for the other stations (e.g., G6 and

H7). This is not surprising, because the footprint of

AMSR-E 6.9-GHz data is about 56 km, while observed

soil moisture values at 3-cm depth of the 16 stations are

quite diverse, as indicated in Fig. 6. The comparisons in

Fig. 5, therefore, do not match scale. As soil moisture

was spatially heterogeneous to a large extent in this

region, a comparison of areal averages would be better

than a point-to-footprint comparison. As shown in Fig. 7a,

the LDAS-UT estimates are agreeable with the obser-

vations after spatial averaging. Some observed dramatic

changes in soil moisture are also produced by LDAS-

UT, indicating that the dual-pass LDAS has successfully

estimated the temporal variations of soil moisture in this

region. By contrast, Fig. 7b shows that the LSM simu-

lation with default parameter values overestimates soil

moisture in general.

To explain how the two passes (i.e., calibration pass

and assimilation pass) of LDAS-UT work to improve

soil moisture estimations, the LSM was run again with

the LDAS-UT–calibrated parameters. Figure 8 shows

FIG. 5. Comparisons of daily mean near-surface soil volumetric water content between observations

and estimations of LDAS-UT for five stations. The day of the year is shown on the x axis.

FIG. 6. Observed daily mean near-surface soil moisture varia-

tions at AMPEX’s 12 ASSH and four AWSs during 30 Apr to

30 Sep 2003.

JUNE 2009 Y A N G E T A L . 787

comparisons between the simulation with default pa-

rameters (LSM-def), the one with calibrated parame-

ters (LSM-opt), and LDAS-UT output. As indicated by

the error metrics [mean bias error (MBE) and root-

mean-square error (RMSE)] in each panel of Fig. 8,

LDAS-UT has the best accuracy, and LSM-opt is inferior

to LDAS-UT but better than LSM-def. Therefore, the

improvement of soil moisture estimations in LDAS-UT

is realized through both the parameter calibration and

the data assimilation.

6. Sensitivity studies

Areal mean soil moisture can be obtained by assimi-

lating microwave data at each station and then averag-

ing the estimated soil moistures, as in the control case. It

can also be obtained by directly assimilating the station-

averaged microwave data. Figure 9 shows that these two

estimates of areal mean soil moisture are almost iden-

tical. This result is consistent with previous studies

(Galantowicz et al. 2000; Burke and Simmonds 2003),

which indicated that regional mean soil moisture was little

affected by subpixel heterogeneity of soil moisture and

soil properties. In all sensitivity studies, LDAS-UT as-

similates the station-averaged microwave data to obtain

areal mean soil moisture, to save the computational time.

a. Case 1: Sensitivity to forcing data sources

LDAS-UT and the LSM are driven with AWS,

GLDAS, and JMA data. Figure 10 shows the compar-

ison of daily mean soil water content between the ob-

servations, LDAS-UT output, and the LSM output,

corresponding to three sets of forcing. It shows a general

worse tendency of soil moisture estimates for both

LDAS-UT and the LSM, when the forcing data was

FIG. 7. Comparisons of AMPEX daily mean station-averaged

near-surface soil volumetric water content with (a) LDAS-UT

output and (b) LSM simulation with default parameter values

(LSM-def) during 30 Apr to 30 Sep 2003.

FIG. 8. Comparisons of station-averaged near-surface soil

moisture between (a) LSM simulation with default parameters, (b)

LSM simulation with parameters calibrated in LDAS-UT, and (c)

LDAS-UT estimations. Shown are the station-averaged observed

value (mean), MBE, RMSE, and correlation coefficient R.


changed from AWS and GLDAS to JMA. However,

LDAS-UT produces better estimates than the LSM

does in any of these three runs, as also indicated by the

comparison of the mean biases and root-mean-square

errors in Table 4.

b. Case 2: Sensitivity to precipitation

This sensitivity study includes three runs of both

LDAS-UT and the LSM, each of which is driven with

AWS data, except precipitation data is sourced from

one of AWS, GLDAS, and JMA. Unlike what Crosson

and Laymon (2002) did (they only changed the precip-

itation amount), these three sets of data have totally

different precipitation amounts and frequency, and thus

they might be more reasonable to show the effects of

different sources of precipitation data.

Figure 11 shows the temporal variations of near-surface

soil moisture estimated from these three runs. The LSM

output seems comparably sensitive to positive biases in

JMA precipitation and negative biases in GLDAS

precipitation (see Fig. 3 for JMA and GLDAS data),

while the LDAS-UT output seems more sensitive to

positive biases in precipitation than to negative biases

in precipitation. This asymmetry of the sensitivity in

LDAS-UT output can be related to the nonlinear re-

lationship between soil moisture and brightness tem-

perature. Bright temperature changes more slowly with

soil moisture in a drier range than that in a wetter range.

Accordingly, the error in the simulated brightness tem-

perature is less, given a negative bias in precipitation

data, than that given a positive bias; thus, the assimila-

tion system would take a longer time to reduce the er-

rors in the latter case, and the estimated soil moisture

may be more sensitive to positive biases in precipitation.

Even so, much lessened is the sensitivity of the LDAS-

estimated soil water content to both positively and

negatively biased precipitation, compared to the LSM

estimates. This result demonstrates that LDAS-UT can

provide more robust estimates of soil moisture than a

land surface model.

c. Case 3: Role of microwave signal

This special case completely removes precipitation

from forcing data, and thus soil moisture estimates is

FIG. 9. Comparing areal-average near-surface soil moisture

values retrieved with the brightness temperature of the individual

stations to those retrieved with the brightness temperature aver-

aged over these stations.

FIG. 10. Comparison of daily mean near-surface soil water content between observation and (top) LDAS-UT and (bottom)

LSM output, driven with meteorological data by AWS observation, GLDAS, and JMA output, respectively.

JUNE 2009 Y A N G E T A L . 789

mainly controlled by the satellite data. This case is still

different from common microwave remote sensing,

because the surface roughness parameters for moisture

retrieval is calibrated by the dual-pass assimilation al-

gorithm rather than tuned with in situ observations or

specified as default values.

Figure 12 shows LDAS-UT–estimated soil moisture

compared against observations. Surprisingly, LDAS-

UT produces fairly good estimates of soil moisture,

though precipitation values have been set to zero. This,

in turn, indicates that soil moisture can be estimated

with good accuracy from 6.9- and 18.7-GHz Tb, given

proper parameter values; therefore, the smaller uncer-

tainties of LDAS-UT in sections 6a,b can be attributed

to the constraint of the microwave signals.

d. Evaluation of optimized parameter values

Table 5 shows the observed values and the optimized

values of soil parameters. All the assimilation cases

have produced similar parameter values. The soil po-

rosity value (us) and soil water potential at saturation

(cs) in all the cases are quite comparable to the obser-

vation averaged over the individual sites, but the esti-

mates of the soil hydraulic conductivity (Ks) are one

order of magnitude lower than the observed one. The

estimates of the pore size distribution index (b) are also

much higher than the observed one. Therefore, the

optimized values of Ks and b are tuned values rather

than physical ones. The results suggest that the near-

surface soil moisture retrieved from the satellite data is

not sufficient to physically estimate all the soil param-

eters, but it is possible to estimate the most sensitive

parameters, such as the soil porosity.

Table 6 shows the estimates of the surface roughness

parameters s, Q0, and vegetation parameter b9. The

estimates of parameters Q0 and b9 are relatively stable

in various cases, but the value of s, which represents the

standard deviation of surface roughness height, nearly

decreases with respect to the increase of the precipitation

amount in forcing data, except in the case of ‘‘AWS 1

JMA rain.’’ Brightness temperature is a decreasing

function of near-surface soil moisture and an increasing

function of surface roughness. For a given value of

brightness temperature, near-surface soil moisture is thus

an increasing function of the surface roughness. In the

cases where precipitation data are not used or lower than

the observation, the calibration pass would produce

an increase in the surface roughness parameter s to

make the estimated near-surface soil moisture increase.

Therefore, the estimated parameter values in Table 6 are

well accordant with the theoretical analysis. The excep-

tional case (AWS 1 JMA rain) does not produce a lower

value of s (0.252 mm) than other cases, though it uses the

maximum precipitation amount. Possibly, other forcing

data also exert effects on the parameter estimations, and

this needs further investigation in future studies.

7. Conclusions and remarks

Based on data collected at 16 stations in a Mongolian

semiarid region, this paper compares the capability of

soil moisture estimation by a dual-pass (calibration 1

assimilation) microwave land data assimilation system

(LDAS-UT) and a land surface model as well as their

sensitivity to forcing data. Major findings from our study

are listed below:

1) LDAS-UT produces good estimates of spatially av-

eraged near-surface soil moisture in this semiarid

FIG. 11. Daily mean near-surface soil water content estimates of

(top) LDAS-UT and (bottom) LSM. Both include three runs, each

of which is driven with observed forcing data, except for precipi-

tation data taken from one of observations, GLDAS, and JMA.

FIG. 12. Comparison of daily mean near-surface soil water

content between observation and LDAS-UT driven with AWS

data, but precipitation is set to zero.

TABLE 4. MBE and RMSE in near-surface volumetric soil water

content estimated by LDAS-UT and LSM. Both are driven with

AWS, GLDAS, and JMA forcing data. LSM-def denotes a LSM

simulation with default parameters, and R is the correlation

coefficient.

LDAS-UT LSM-def

Forcing AWS GLDAS JMA AWS GLDAS JMA

MBE 0.010 20.005 0.011 0.033 0.016 0.037

RMSE 0.027 0.027 0.037 0.040 0.041 0.053

R 0.868 0.736 0.617 0.908 0.527 0.593


region. This performance is attributed to both sat-

ellite-based calibration pass and assimilation pass.

We also confirm that it is risky to evaluate micro-

wave-derived soil moisture based on measurements

at a single site or a sparse network because of the

spatial heterogeneity of soil moisture.

2) Compared to the land surface model, LDAS-UT–

estimated soil moisture is not sensitive to forcing

data. This indicates that the constraint of microwave

signal helps lessen uncertainties in land hydrological

modeling.

3) Particularly, LDAS-UT produces more robust and

reliable soil moisture than a land surface model does

when forced by different sources of precipitation

data. This result is encouraging and would contribute

to hydrometeorological studies for remote regions—

such as the Tibet Plateau and the Mongolia Plateau—

where globally available datasets of precipitation

are usually prone to errors due to the lack of

observations.

Previous studies have shown that specifying effective

model parameters are crucial for land hydrological

models, and many of them have presented sophisticated

methods for parameter calibrations; however, they

usually have limited applications because their calibra-

tions are based on in situ data (e.g., Merlin et al. 2006;

Judge et al. 2008). This paper presents the potential to

estimate the effect of parameter values using satellite

data rather than in situ data. This is particularly im-

portant for the establishment of an operational hydro-

logical model, a remote sensing retrieval scheme, or a

land data assimilation system; therefore, this strategy

should be pursued in future studies, even though the

estimated parameter values might be effective values

rather than physical values.

Finally, it should be mentioned that this study mainly

addresses how the parameter calibration may improve

estimates of soil moisture, but it is possible to further

improve the performance of LDAS-UT by introducing

advanced assimilation methods, such as an ensemble

Kalman filter or particle filters, and this issue should be

addressed in future studies.

Acknowledgments. This work was supported by the

100-talent program of Chinese Academy of Sciences

and the EU’s Seventh Research Framework Programme

‘‘Coordinated Asia–European Long-Term Observing

TABLE 5. The observed and estimated soil parameter values [see Eqs. (11) and (13)–(15) for the meaning of the parameters] in the three

sensitivity cases introduced in Table 3.

us Sand% Clay% b cs (m) Ks (1026 m s21)

Observed

MGS* 0.334 — — 3.37 — 18.7

MGS** 0.388 — — 3.62 20.12 51.9

DRS* 0.393 — — 3.06 — 44.0

DRS** 0.428 — — 3.00 20.11 88.6

BTS** 0.456 — — — — 49.3

E4* 0.358 — — 2.93 — 21.3

G6* 0.341 — — 2.70 — 8.9

C2* 0.299 — — 3.10 — 19.3

C4* 0.395 — — 2.82 — 28.5

Avg 0.377 — — 3.08 20.115 36.7

LDAS-UT

AWS 0.368 47 28 7.34 20.18 4.9

GLDAS 0.355 47 33 8.14 20.18 4.9

JMA 0.348 38 30 7.75 20.24 3.5

AWS 1 no rain 0.369 46 32 8.07 20.19 4.6

AWS 1 GLDAS rain 0.361 47 32 7.99 20.19 4.8

AWS 1 JMA rain 0.349 43 32 8.05 20.21 4.1

* Data from Table 1 of Yamanaka et al. (2005).

** Data from Table I-2 and Table I-3 of appendix 1 in Kaihotsu (2005).

TABLE 6. Estimates of surface roughness parameters and veg-

etation parameter [see Eqs. (5)–(7) for the meaning of the pa-

rameters] in the sensitivity cases, which are arranged in ascending

order of total precipitation.

Forcing data Rain (mm) s (mm) Q0 b9

AWS 1 no rain 0 0.315 0.94 3.73

GLDAS 135 0.214 0.98 3.70

AWS 1 GLDAS rain 135 0.209 0.99 3.94

AWS 173 0.124 0.97 3.98

JMA 238 0.093 0.80 3.87

AWS 1 JMA rain 238 0.252 0.71 3.95

JUNE 2009 Y A N G E T A L . 791

System of Qinghai–Tibetan Plateau Hydrometeoro-

logical Processes and the Asian–Monsoon System with

Ground Satellite Image Data and Numerical Simula-

tions’’ (CEOP–AEGIS). AMPEX was implemented in

the framework of NASDA–JRA’s ‘‘Ground Truth for

Evaluation of Soil Moisture and Geophysical/Vegetation

Parameters Related to Ground Surface Conditions with

AMSR and GLI in the Mongolian Plateau’’ (PI: Prof.

Ichirow Kaihotsu of the University of Hiroshima). The

Mongolian partnership is with the Institute of Meteo-

rology and Hydrology, and the National Agency for

Meteorology, Hydrology and Environment Monitoring

of Mongolia. The authors thank the editor and the re-

viewers for their contributions to the analysis in this

manuscript.

REFERENCES

Bach, H., and W. Mauser, 2003: Methods and examples for remote

sensing data assimilation in land surface process modeling.

IEEE Trans. Geosci. Remote Sens., 41, 1629–1637.

Bindlish, R., T. J. Jackson, A. J. Gasiewski, M. Klein, and E. G.

Njoku, 2006: Soil moisture mapping and AMSR-E validation

using the PSR in SMEX02. Remote Sens. Environ., 103,127–139.

Bosch, D. D., V. Lakshmi, T. J. Jackson, M. Choi, and J. M. Jacobs,

2006: Large scale measurements of soil moisture for valida-

tion of remotely sensed data: Georgia soil moisture experi-

ment of 2003. J. Hydrol., 323, 120–137.

Burke, E. J., and L. P. Simmonds, 2003: Effects of sub-pixel het-

erogeneity on the retrieval of soil moisture from passive mi-

crowave radiometry. Int. J. Remote Sens., 24, 2085–2104.

Clapp, R. B., and G. M. Hornberger, 1978: Empirical equations

for some hydraulic properties. Water Resour. Res., 14,

601–604.

Cosh, M. H., T. J. Jackson, R. Bindlish, and J. H. Prueger, 2004:

Watershed scale temporal and spatial stability of soil moisture

and its role in validating satellite estimates. Remote Sens.

Environ., 92, 427–435.

Crosson, W. L., C. A. Laymon, R. Inguva, and M. P. Schamschula,

2002: Assimilating remote sensing data in a surface flux–soil

moisture model. Hydrol. Processes, 16, 1645–1662.

Crow, W. T., and E. F. Wood, 2003: The assimilation of remotely

sensed soil brightness temperature imagery into a land surface

model using ensemble Kalman filtering: A case study based on

ESTAR measurements during SGP97. Adv. Water Resour.,

26, 137–149.

de Goncalves, L. G. G., W. J. Shuttleworth, S. C. Chou, Y. Xue,

P. R. Houser, D. L. Toll, J. Marengo, and M. Rodell, 2006:

Impact of different initial soil moisture fields on Eta model

weather forecasts for South America. J. Geophys. Res., 111,

D17102, doi:10.1029/2005JD006309.

Duan, Q., V. K. Gupta, and S. Sorooshian, 1993: Shuffled complex

evolution approach for effective and efficient global minimi-

zation. J. Optim. Theory Appl., 76, 501–521.

Engman, E. T., 1991: Applications of microwave remote sensing of

soil moisture for water resources and agriculture. Remote

Sens. Environ., 35, 213–226.

Entin, J. K., A. Robock, K. Y. Vinnikov, V. Zabelin, S. Liu,

A. Namkhai, and T. Adyasuren, 1999: Evaluation of global

soil wetness project soil moisture simulations. J. Meteor. Soc.

Japan, 77, 183–198.

Galantowicz, J. F., D. Entekhabi, and E. G. Njoku, 2000: Esti-

mation of soil-type heterogeneity effects in the retrieval of soil

moisture from radiobrightness. IEEE Trans. Geosci. Remote

Sens., 38, 312–315.

Global Soil Data Task Group, cited 2000: Global gridded surfaces

of selected soil characteristics (IGBPDIS). [Available online

at http://www.daac.ornl.gov/.]

Jackson, T. J., 1993: Measuring surface soil moisture using passive

microwave remote sensing. Hydrol. Processes, 7, 139–152.

——, D. M. Le Vine, A. Y. Hsu, A. Oldak, P. J. Starks, C. T. Swift,

J. D. Isham, and M. Haken, 1999: Soil moisture mapping at

regional scales using microwave radiometry: The Southern

Great Plains hydrology experiment. IEEE Trans. Geosci.

Remote Sens., 37, 2136–2151.

Judge, J., A. W. England, J. R. Metcalfe, D. McNichol, and B. E.

Goodison, 2008: Calibration of an integrated land surface

process and radiobrightness (LSP/R) model during summer-

time. Adv. Water Resour., 31, 189–202.

Kaihotsu, I., 2005: Ground truth for evaluation of soil moisture and

geophysical/vegetation parameters related to ground surface

conditions with AMSR and GLI in the Mongolian Plateau.

Japanese Aerospace Exploration Agency Rep. JAXA-RM-

04-017, 113 pp.

Knyazikhin, Y., and Coauthors, 1999: MODIS leaf area index

(LAI) and fraction of photosynthetically active radiation ab-

sorbed by vegetation (FPAR) product, (MOD15). Algorithm

theoretical basis document version 4.0. [Available online at

http://modis.gsfc.nasa.gov/data/atbd/atbd_mod15.pdf.]

Koike, T., 2004: The coordinated enhanced observing period—An

initial step for integrated global water cycle observation.

WMO Bull., 53, 1–8.

Koster, R. D., and Coauthors, 2004: Regions of strong coupling be-

tween soil moisture and precipitation. Science, 305, 1138–1140.

Li, X., and Coauthors, 2007: Development of a Chinese land data

assimilation system: Its progress and prospects. Prog. Nat.

Sci., 17, 881–892.

Loveland, T. R., B. C. Reed, J. F. Brown, D. O. Ohlen, J. Zhu,

L. Yang, and J. W. Merchant, 2000: Development of a global

land cover characteristics database and IGBP DISCover from

1 km AVHRR data. Int. J. Remote Sens., 21, 1303–1330.

Merlin, O., A. Chehbouni, Y. H. Kerr, and D. C. Goodrich, 2006:

A downscaling method for distributing surface soil moisture

within a microwave pixel: Application to the Monsoon ‘90

data. Remote Sens. Environ., 101, 379–389.

Ni-Meister, W., P. R. Houser, and J. P. Walker, 2006: Soil moisture

initialization for climate prediction: Assimilation of scanning

multifrequency microwave radiometer soil moisture data

into a land surface model. J. Geophys. Res., 111, D20102,

doi:10.1029/2006JD007190.

Njoku, E. G., and D. Entekhabi, 1996: Passive microwave remote

sensing of soil moisture. J. Hydrol., 184, 101–129.

——, and S. K. Chan, 2006: Vegetation and surface roughness

effects on AMSR-E land observations. Remote Sens. Envi-

ron., 100, 190–199.

——, T. J. Jackson, V. Lakshmi, T. K. Chan, and S. V. Nghiem,

2003: Soil moisture retrieval from AMSR-E. IEEE Trans.

Geosci. Remote Sens., 41, 215–229.

Pauwels, V. R. N., R. Hoeben, N. E. C. Verhoest, and F. P. De Troch,

2001: The importance of the spatial patterns of remotely sensed

soil moisture in the improvement of discharge predictions for

small-scale basins through data. J. Hydrol., 251, 88–102.


Pitman, A. J., and Coauthors, 1999: Key results and implications

from phase 1(c) of the project for intercomparison of land–

surface parameterization schemes. Climate Dyn., 15, 673–684.

Reichle, R. H., D. B. McLaughlin, and D. Entekhabi, 2001: Vari-

ational data assimilation of microwave radiobrightness ob-

servations for land surface hydrology applications. IEEE

Trans. Geosci. Remote Sens., 39, 1708–1718.

Rodell, M., and Coauthors, 2004: The global land data assimilation

system. Bull. Amer. Meteor. Soc., 85, 381–394.

Schaake, J. C., and Coauthors, 2004: An intercomparison of soil

moisture fields in the North American Land Data Assimilation

System (NLDAS). J. Geophys. Res., 109, D01S90, doi:10.1029/

2002JD003309.

Sellers, P. J., and Coauthors, 1996: A revised land surface pa-

rameterization (SiB2) for atmospheric GCMs. Part I: Model

formulation. J. Climate, 9, 676–705.

Shi, J. C., J. Wang, A. Y. Hsu, P. E. O’Neill, and E. T. Engman,

1997: Estimation of bare surface soil moisture and surface

roughness parameter using L-band SAR image data. IEEE

Trans. Geosci. Remote Sens., 35, 1254–1266.

Vivoni, E. R., M. Gebremichael, C. J. Watts, R. Bindlish, and T. J.

Jackson, 2008: Comparison of ground-based and remotely-

sensed surface soil moisture estimates over complex terrain

during SMEX04. Remote Sens. Environ., 112, 314–325.

Wang, J. R., 1983: Passive microwave sensing of soil moisture

content: The effects of soil bulk density and surface rough-

ness. Remote Sens. Environ., 13, 329–344.

——, and B. J. Choudhury, 1981: Remote sensing of soil moisture

content, over bare fields at 1.4 GHz frequency. J. Geophys.

Res., 86, 5277–5282.

Watanabe, T., and J. Kondo, 1990: The influence of canopy

structure and density upon the mixing length within and above

vegetation. J. Meteor. Soc. Japan, 68, 227–235.

Wen, J., Z. Su, and Y. Ma, 2003: Determination of land surface

temperature and soil moisture from tropic rainfall measuring

mission/microwave imager remote sensing data. J. Geophys.

Res., 108 (D2), 4038, doi:10.1029/2002JD002176.

——, T. J. Jackson, R. Bindlish, A. Y. Hsu, and Z. B. Su, 2005:

Retrieval of soil moisture and vegetation water content using

SSM/I data over a corn and soybean region. J. Hydrometeor.,

6, 854–863.

Wigneron, J.-P., J.-C. Calvet, T. Pellarin, A. A. Van de Griend,

M. Berger, and P. Ferrazzoli, 2003: Retrieving near-surface

soil moisture from microwave radiometric observations:

Current status and future plans. Remote Sens. Environ., 85,489–506.

Wu, W., and R. E. Dickinson, 2004: Time scales of layered soil

moisture memory in the context of land–atmosphere inter-

action. J. Climate, 17, 2752–2764.

Yamanaka, T., I. Kaihotsu, D. Oyunbaatar, and T. Ganbold, 2005:

Regional-scale variability of the surface soil moisture re-

vealed by the AMPEX monitoring network. Ground truth for

evaluation of soil moisture and geophysical/vegetation pa-

rameters related to ground surface conditions with AMSR

and GLI in the Mongolian Plateau, Japanese Aerospace Ex-

ploration Agency Rep. JAXA-RM-04-017, 33–42.

Yang, K., T. Watanabe, T. Koike, X. Li, H. Fujii, K. Tamagawa,

Y. Ma, and H. Ishikawa, 2007a: Auto-calibration system de-

veloped to assimilate AMSR-E data into a land surface model

for estimating soil moisture and the surface energy budget.

J. Meteor. Soc. Japan, 85A, 229–242.

——, and Coauthors, 2007b: Initial CEOP-based review of the

prediction skill of operational general circulation models and

land surface models. J. Meteor. Soc. Japan, 85A, 99–116.

——, and Coauthors, 2008: Turbulent flux transfer over bare

soil surfaces: Characteristics and parameterization. J. Appl.

Meteor. Climatol., 47, 276–290.

JUNE 2009 Y A N G E T A L . 793

Validation of a Dual-Pass Microwave Land Data Assimilation System for Estimating Surface Soil...

Documents

Transcript of Validation of a Dual-Pass Microwave Land Data Assimilation System for Estimating Surface Soil...