CHAPTER 2

Hydrologic Data Assimilation

Jeffrey P. Walker 1 and Paul R. Houser 2

1. Department of Civil and Environmental Engg, The University of Melbourne, Australia

E-mail: [Link]@[Link]

2. Hydrological Sciences Branch, NASA Goddard Space Flight Center, Greenbelt, USA
E-mail: [Link]@[Link]

Key words: Data Assimilation; Quality Control; Validation; Dynamic Assimilation;

Sequential Assimilation; Kalman Filter; Case Studies; Soil Moisture; Downscaling; Snow;
Skin Temperature

2.1 Introduction
Earth observing satellites have revolutionised our understanding and prediction of the
Earth system over the last 30 years, particularly in the meteorologic and oceanographic
sciences. However, historically remote sensing data has not been widely used in hydrology.
This can be attributed to: (i) a lack of dedicated hydrologic remote sensing instruments;
(ii) inadequate retrieval algorithms for deriving global hydrologic information from remote
sensing observations; (iii) a lack of suitable distributed hydrologic models for digesting
remote sensing information; and (iv) an absence of techniques to objectively improve and
constrain hydrologic model predictions using remote sensing data. Three ways that remote
sensing observations have been used in distributed hydrologic models are: (i) as parametric
input data, including soil and land cover properties; (ii) as initial condition data, such as initial
snow water storage; and (iii) as time-varying hydrological state data, such as soil moisture
content, to constrain model predictions. This chapter focuses on the latter.
The historic lack of hydrologic missions and observations has been the result of an
emphasis on meteorologic and oceanographic missions and applications, due to the large
scientific and operational communities that drive those fields. However, significant progress
has been made over the past decade on defining hydrologically-relevant remote sensing
observations through focused ground and airborne field studies. Gradually, satellite-based
hydrologic data are becoming increasingly available, though little progress has been made in
understanding their observation error. Land surface skin temperature and snow cover data
have been available for many years, and satellite precipitation data are becoming available at
increasing space and time resolutions. In addition, land cover and land use maps, vegetation
parameters (albedo, leaf area index and greenness), and snow water equivalent data of
increasing sophistication are becoming available from a number of sensors. Novel
observations such as saturated fraction and changes in soil moisture, evapotranspiration, water
level and velocity (i.e. runoff), and changes in total terrestrial water storage are also under
development. Further, near-surface soil moisture, a parameter shown to play a critical role in
weather, climate, agriculture, flood, and drought processes, is currently available from non-
ideal sensor configuration observations. Moreover, two missions targeted at measuring near-

1
surface soil moisture with ideal sensor configuration are expected before the end of the
decade.
Though remote sensing can make spatially comprehensive measurements of various
components of the hydrologic system, it cannot provide information on the entire system, and
the measurements represent only a snap shot in time. Land surface hydrology process models
may be used to predict the temporal and spatial hydrologic system variations, but these
predictions are often poor, due to model initialisation, parameter and forcing errors, and
inadequate model physics and/or resolution. Figure 2.1 illustrates the hydrologic data
assimilation challenge to optimally merge the spatially comprehensive but limited remote
sensing observations with the complete but typically poor predictions of a hydrologic model
to yield the best possible hydrologic system state estimation, and utilise limited point
measurements to calibrate the model(s) and validate the assimilation results.
Fig. 2.1
While hydrologic data assimilation is still very much in its infancy, a few hydrologic
models have been developed that can use remotely sensed observations. The key is that the
remote sensing observations of interest can be directly related to a prognostic state(s) of the
hydrologic model. Figure 2.2 demonstrates how satellite observations of the near-surface soil
moisture content may be used to constrain the hydrologic model soil moisture prediction
using state-of-the-art hydrologic data assimilation techniques [Walker et al., 2003]. This
example uses actual space-borne near-surface soil moisture observations from a historic
satellite record in a data assimilation framework, and highlights the significant benefit of
using these techniques. However, quantifying the hydrologic model prediction improvement
by assimilating remote sensing data requires targeted field campaigns, and such data are
lacking for these historic satellite records.
Fig. 2.2
Because of its importance, and our increasing ability to observe relevant hydrologic
information remotely, it is expected that the amount of hydrologic remote sensing data will
grow exponentially over the next decade. However, its usefulness will be limited by our
ability to analyse and integrate diverse hydrologic information with hydrologic models.
Quantifying hydrologic process variability will require innovative interpretation of potentially
large hydrologic observation volumes, due to observation type, scale, and error disparities
(Table 2. 1). The effect of variations in instrument type, placement, calibration and accuracy
of both remote sensing and in-situ hydrologic observations must also be quantified. The
complexities of future hydrologic observation scenarios will require systematic methods to
organise and comprehend this information. Therefore a comprehensive hydrologic data
assimilation framework will be a critical component of future hydrologic observation and
modelling systems.

Table 2.1

2.2 History of Hydrologic Data Assimilation

Charney et al. [1969] first suggested combining current and past data in an explicit
dynamical model, using the model’s prognostic equations to provide time continuity and
dynamic coupling amongst the fields. This concept has evolved into a family of techniques
known as data assimilation. In essence, hydrologic data assimilation aims to utilise both our
hydrologic process knowledge as embodied in a hydrologic model, and information that can
be gained from observations. Both model predictions and observations are imperfect and we

2
wish to use both synergistically to obtain a more accurate result. Moreover, both contain
different kinds of information, that when used together, provide an accuracy level that cannot
be obtained when used individually.
For example, a hydrologic model provides both spatial and temporal near-surface and root
zone soil moisture information at the model resolution, including errors resulting from
inadequate model physics, parameters and forcing data. On the other hand, remote sensing
observations contain near-surface soil moisture information at an instant in time, but do not
give the temporal variation or the root zone moisture content. While the remote sensing
observations can be used as initialisation input for models or as independent evaluation,
providing we use a hydrologic model that has been adapted to use remote sensing data as
input, we can use the hydrologic model predictions and remote sensing observations together
to keep the simulation on track through data assimilation [Kostov and Jackson, 1993].
Moreover, large errors in near-surface soil moisture content prediction are unavoidable
because of its highly dynamic nature. Thus when measured soil moisture data are available,
their use to constrain the simulated data should improve the overall estimation of the soil
moisture profile. However, this expectation is based on the assumption that measurement
errors are smaller than simulation errors [Arya et al., 1983].
Data assimilation techniques were pioneered by meteorologists [Daley, 1991] and have
been used very successfully to improve operational weather forecasts for decades. Data
assimilation has also been widely used in oceanography [Bennett, 1992] for improving ocean
dynamics prediction. However, hydrologic data assimilation has just a few case studies
demonstrating its utility. Fortunately, we have been able to jumpstart hydrologic data
assimilation by building on knowledge derived from the meteorologic and oceanographic data
assimilation experience, with significant advancements being made over the past decade.
One of the primary areas of hydrologic data assimilation application has been with soil
moisture content. Other observation types, such as surface temperature, snow, terrestrial
water storage and streamflow, have been used only in more recent applications. The study by
Jackson et al. [1981] was among the first to update soil moisture predictions using near-
surface soil moisture observations. In this application, the soil moisture values in both layers
of the United States Department of Agriculture Hydrograph Laboratory model were
substituted with observed near-surface soil moisture observations as they became available.
The model’s performance improvement was evaluated by annual runoff values. Ottlé and
Vidal-Madjar [1994] used a similar approach but with the assimilation of thermal infrared
derived near-surface soil moisture content.
Another early study based on the direct insertion assimilation method was that of Bernard
et al. [1981]. Here synthetic observations of near-surface soil moisture content were used to
specify the surface boundary condition of a classical one-dimensional soil water diffusion
model, in order to estimate the surface flux. They found that large soil moisture content
variations resulting from rainy periods required special handling of the upper boundary
condition. Prevot et al. [1984] repeated this study with real observations and a similar
approach was used by Bruckler and Witono [1989]. A more popular approach for the
improved estimation of land surface fluxes has been the assimilation of screen-level
measurements of relative humidity and temperature [Bouttier et al., 1993; Viterbo and
Beljaars, 1995]. To date only one study has explored the assimilation of remotely sensed
land surface flux observations [Schuurmans et al., 2003].
The first known study to use an “optimal” assimilation approach is that of Milly [1986]. In
this study, a Kalman filter (a statistical assimilation approach) was used to update a simple
linear reservoir model with near-surface soil moisture observations. It was not until

3
Entekhabi et al. [1994] that this approach was extended, when synthetically-derived vertical
and horizontal polarised passive microwave and thermal infrared observations were
assimilated into a one-dimensional soil moisture and temperature diffusion model using the
Kalman filter. This synthetic study was further extended by Walker et al. [2001a] to more
realistic observation times and Walker et al. [2001b] to a field application. Since then there
has been a plethora of one-dimensional Kalman filter and variational assimilation studies.
Georgakakos and Baumer [1996] were one of the first to use the Kalman filter to update a
hydrologic basin model with near-surface soil moisture measurements. Results showed that
even when the observations carried substantial measurement errors, estimation of soil
moisture profiles and total soil moisture storage was possible with an error that was smaller
than that achieved without the use of remotely sensed data. Walker et al. [2002] were also
among the first to use a three-dimensional Kalman filter based assimilation in a small
catchment distributed hydrologic model. Houser et al. [1998] was the first detailed study of
several alternative assimilation approaches, including direct insertion, statistical correction,
Newtonian nudging and optimal interpolation. Both the Newtonian nudging and optimal
interpolation approaches, pathological cases of the Kalman filter, showed the greatest
improvement.

2.3 Summary of Data Assimilation

The data assimilation challenge is: given a (noisy) model of the system dynamics, find the
best estimates of system states X from (noisy) observations Z. Most current approaches to
this problem are derived from either the direct observer (i.e. Kalman filter) or dynamic
observer (i.e. variational through time) techniques. Figure 2.3 illustrates schematically the
key differences between these two approaches to data assimilation. To help the reader
through the large amount of jargon typically associated with data assimilation, a list of
terminology has been provided (Table 2.2).

Fig. 2.3

Table 2.2

2.3.1 Direct Observer Assimilation

The direct observer techniques sequentially update the model forecast, using the
difference between observation Z and model predicted observation Ẑ , known as the
‘innovation’, whenever observations are available. The predicted observation is calculated
from the model predicted or ‘background’ states, indicated by the superscript b. The
correction added to the background state vector is the innovation multiplied by a weighting
factor K known as the ‘gain’ (sometimes called the Kalman gain). The gain represents the
relative uncertainty in the observation and model variances, and is a number between 0 and 1
in the scalar situation. The resulting estimate of the state vector is known as the ‘analysis’, as
indicated by the superscript a.

k k (
ˆ a = Xb + K Z − Z
X k
ˆ
k ) (2.1)

The subscript k refers to the time of the update. If the uncertainty of the predicted observation
(as calculated from the background states and their uncertainty) is large relative to the
uncertainty of the actual observation, then the analysis state vector takes on values that will
yield the actual observation. Conversely, if the uncertainty of the predicted observation is

4
small relative to the uncertainty of the actual observation, then the analysis state vector is
unchanged from the original background value. The commonly used direct observer methods
are:

1. Direct Insertion
2. Statistical Correction
3. Successive Correction
4. Optimal Interpolation/Statistical Interpolation
5. Analysis Correction
6. Nudging
7. 3D Variational
8. Kalman Filter and variants

While approaches like direct insertion, nudging and optimal interpolation are computationally
efficient and easy to implement, the updates do not account for observation uncertainty or
utilise system dynamics in estimating model background state uncertainty, and information on
estimation uncertainty is limited. The Kalman filter, while computationally demanding in its
pure form, can be adapted for near-real-time application and provides information on
estimation uncertainty. However, it has only limited capability to deal with model errors, and
necessary linearisation approximations can lead to unstable solutions. The ensemble Kalman
filter, while it can be computationally demanding (depending on the size of the ensemble) is
well suited for near-real-time applications, is robust, very flexible and easy to use, and is able
to accommodate a wide range of model error descriptions.

2.3.2 Dynamic Observer Assimilation

The dynamic observer techniques find the best fit between the forecast model state and the
observations, subject to the initial state vector uncertainty Σ and observation uncertainty R, by
minimising over space and time an objective function J such as

( ) ( ) ( ) ( )
N −1
J = 1 X 0 − X b0 Σ b0 X 0 − Xb0 + 1 ∑ Z k − Z
−1
T
ˆ T R −1 Z − Z
ˆ , (2.2)
2 2 0 k k k k

where the superscript b refers to the initial or ‘background’ estimate of the state vector, the
subscript refers to time, and N is the number of time steps. To minimise the objective
function over time, an assimilation time ‘window’ is defined and an ‘adjoint’ model is
typically used to find the derivatives of the objective function with respect to the initial model
state vector X0. The adjoint is simply a mathematical operator that allows one to determine
the sensitivity of the objective function to changes in the solution of the state equations by a
single forward and backward pass over the assimilation window. While an adjoint is not
strictly required (i.e. a number of forward passes can be used to numerically approximate the
objective function derivatives with respect to each state), it makes the problem
computationally tractable. The dynamic observer techniques can be considered simply as an
optimisation or calibration problem, where the state vector – not the model parameters – at the
beginning of each assimilation window is “calibrated” to the observations over that time
period. The dynamic observer techniques can be formulated with:

1. Strong Constraint (variational)

2. Weak Constraint (dual variational or representer methods)

5
Strong constraint is where the model is assumed perfect, as in equation 2.2, while weak
constraint is where errors in the model formulation are taken into account as process noise.
This is achieved by including an additional term in equation 2.2 so that

( ) ( ) ( ) ( )
N −1 N −1
J = 1 X 0 − X b0 Σ b0 X 0 − X b0 + 1 ∑ Z k − Z ˆ + 1 ∑ w T Q −1w , (2.3)
−1
T
ˆ T R −1 Z − Z
2 2 0 k k k k 2 0 k k k

where w is the model error vector and Q is the model error variance-covariance matrix.
Dynamic observer methods are well suited for smoothing problems, but provide
information on estimation accuracy only at considerable computational cost. Moreover,
adjoints are not available for many existing hydrologic models, and the development of robust
adjoint models is difficult due to the nonlinear nature of hydrologic processes.

2.3.3 Features of Data Assimilation

The potential benefit of data assimilation for hydrologic science is tremendous and can be
summarised as follows [adapted from Rood et al., 1994]:
• Organises the data by objectively interpolating from the observation space to the
model space. The raw observations are organised and given dynamical consistency
with the model equations, thereby enhancing their usefulness.
• Supplements the data by constraining the model's physical equations with
parsimonious observations, which can be used to estimate unobserved quantities. This
allows the progress of research that would be impossible without assimilation, because
it allows for a more complete understanding of the true state of a hydrologic system
(see Figure 2.4a).
• Complements the data by propagating information into regions of sparse observations
using either observed spatial and temporal correlations, or the physical relationships
included in the model (see Figure 2.4b).
• Quality controls the data through comparison of observations with previous forecasts
to identify and eliminate spurious data. By performing this comparison repeatedly, it
is possible to calibrate observing systems and identify biases or changes in observation
system performance.
• Validates and improves the hydrological models by continuous model confrontation
with real data. This helps to identify model weaknesses, such as systematic errors,
and correct them.

Fig. 2.4 a, b

2.3.4 Quality Control for Data Assimilation

One of the major components of any data assimilation system is quality control of the input
data stream. Quality control is a pre-assimilation rejection or correction of questionable or
bad observations, which begins where the remote sensing product quality control activities
leave off. The observation data from remote sensing systems contain errors that can be
classified into two types:
1. Natural Error (including instrument and representativeness error)
2. Gross Error (including improperly calibrated instruments, incorrect registration or
coding of observations, and telecommunication error)

6
These errors can be either random or spatially and/or temporally correlated with each other;
inversion techniques and instrument biases can be correlated in time and space, and
calibrations of remote sensing instruments can drift. To address these problems a number of
quality control operations are performed.
The quality control process consists of a set of algorithms which examine each data item,
individually or jointly, in the context of additional information. Their primary purpose is to
determine which of the data are likely to contain unknown (incorrigible) gross errors, and
which are not. Quality control proceeds in a three step process: (i) test for potential problem
observations; (ii) attempt to correct the problem observation; and (iii) decide the fate of the
observation (data rejection). The quality control algorithms can be categorised as follows:
• Quality control flags are used to check the data for inconsistencies noted during the
measurement, transmission, pre/post processing and archiving stages.
• Consistency or sanity checks see if the observation absolute value or time rate of
change is physically realistic. This check filters such things as observations outside
the expected range, unit conversion problems, etc.
• Buddy checks compare the observation with comparable nearby (space and time)
observations of the same type and reject the questioned observation if it exceeds a
predefined level of difference.
• Background checks examine if the observation is changing similarly to the model
prediction. If it is not, and the user has some reasonable confidence in the model, the
observation may be questioned.

2.3.5 Validation Using Data Assimilation

The continuous confrontation of model predictions with observations in a data assimilation
system presents a rich opportunity to better understand physical processes and observation
quality in a structured, iterative, and open-ended learning process. Inconsistencies between
observations and predictions are easily identified in a data assimilation system, providing a
basis for observational quality control and validation. Systematic differences between
observations and model predictions can identify systematic error. This methodology clearly
illustrates the importance of a good quality forecast and an analysis that is reasonably faithful
to the observations. If the hydrologic model makes reasonably good predictions, then the
analysis must only make small changes to an accurate background field.
The validation of observations in a data assimilation system is centred on: (i) comparisons
of new observations with the model forecast and the data assimilation analysis; and (ii)
interpretation of the forecast error covariances. The data assimilation validation algorithms
can be categorised as follows:
• Innovation evaluation compares the observation with the model prediction as either a
single point in time or change over time; large or obvious deviations from the model
prediction are probably wrong. Means, standard deviations, and time evolution of
observed minus predicted fields are examined with the goal of detecting abrupt
changes.
• Analysis residual evaluation compares the observation with the data assimilation
analysis. Examination of the means, standard deviations, and time evolution of
observed minus predicted fields will help to diagnose systematic or abrupt observation
system changes. This technique is useful to diagnose the performance of the analysis,

7
 and if the observations are being used effectively [Hollingsworth and Lonnberg,
1989].
• Observation withholding is a stringent method for validation in an assimilation system
where some of the observational data are withheld from the analysis procedure in
data-dense regions. This allows the analysis to be validated against the withheld
observations.
• Εrror propagation is undertaken and changes in the regional distribution or absolute
value of these errors could indicate observational problems.
• Model and observation bias is generally assumed to be zero and uncorrelated in space.
These assumptions work reasonably well for in situ observations, but satellite
observations are usually biased by inaccurate algorithms, and their errors are usually
horizontally correlated because the same sensor is making all the observations. With
recent work by Dee and Todling [2000] the bias of the model and observations can be
continuously estimated and corrected for. Evaluation of these bias estimates in space
and time may lead to additional insights on the observational characteristics.

2.4 Direct Observer Assimilation Methods

Land surface hydrology process models are typically nonlinear, and can be considered to
forecast the system state vector X at time k+1 as a function of the system state vector estimate
at the previous time step k and a forcing vector U. The model state forecast is subject to a
model error vector w, which represents errors in the model forcing data, initial conditions,
parameters and physics. The state equation is given by

X k +1 = ak (X k , U k ) + w k , (2.4)

where a is a nonlinear operator. This equation can be linearised to obtain the ‘tangent linear
model’ as

X k +1 = A k X k + B k U k + w k . (2.5)

The state space equation is subject to the initial state vector

X 0 = X(t0 ) + e 0 , (2.6)

with error vector e. The observation equation is given by

ˆ = h (X ) + v ,
Z (2.7)
k k k k

where h is a nonlinear operator. This equation can also be linearised as

ˆ = HX + Y + v ,
Z (2.8)
k k k k

with error vector v.

8
 The key assumptions of this assimilation approach are that the error terms w, v and e are
uncorrelated (white) through time and have Gaussian distributions as represented by their
covariance matrices Q, Σ and R, respectively. That is

( )
E (w k ) = 0 E w k w Tk = Q k
E (e 0 ) = 0 E (e e ) = Σ
T
0 0
b
0 , (2.9)
E (v k ) = 0 E (v v ) = R
k
T
k k

where E is the expectation operator. The assumption that observation and model errors are
unbiased relative to each other and the “truth” is the most restrictive assumption, most
commonly violated assumption, and most detrimental assumption in terms of predictive
performance.
One key question in the direct observer data assimilation technique, and the fundamental
difference between the various methods, is the choice of the gain matrix K. Ultimately Kk
should be chosen such that X ak approaches the expectation of Xk as k approaches infinity.
This can be achieved by choosing K as the optimal least squares estimator or Best Linear
Unbiased Estimator (BLUE) analysis obtained as a solution to the variational optimisation
problem posed in equation 2.2. That is, choosing K such that objective function J is a
minimum. This can be shown analytically to produce [Bouttier and Courtier, 1999]

(
K = Σ b H T HΣ b H T + R )
−1
, (2.10)

where HΣ b H T = Rˆ is the covariance matrix of the predicted observation Ẑ . Thus, on

assimilation interval k∈[0,N], the analysis X aN given by the Kalman filter should be equal to
the converged solution obtained by the adjoint method at time k equal to N.
From application of standard error propagation theory on the correction equation it can
also be shown that the updated uncertainty of the states is given by

Σ a = (I − KH )Σ b (I − KH ) + KRK T ,
T
(2.11)

where I is the identity matrix. Equations 2.1, 2.5, 2.8, 2.10 and 2.11 form the basis of the
Kalman filter approach [Kalman, 1960] to data assimilation.
Apart from the assumption that errors are unbiased and normally distributed, the difficulty
associated with applying these equations is an estimate of the background variance-covariance
matrix Σb, and that to find the analysis Xa one must compute Σ b H T (H Σ b H T + R )− 1 (Z − Zˆ b ),
which is computationally expensive. As a result, approximations to these equations and/or
alternative methodologies of solving the key equations are sought. Ultimately, it is
approximations to K that are typically made.

2.4.1 Direct Insertion

One of the earliest and most simplistic approaches to data assimilation is direct insertion.
As the name suggests, the forecast model states are directly replaced with the observations by
essentially assuming that K = 1. This approach makes the explicit assumption that the model
is wrong (has no useful information) and that the observations are right, which both
disregards important information provided by the model and preserves observation errors. A

9
further key disadvantage of this approach is that model physics are solely relied upon to
propagate the information to unobserved parts of the system [Houser et al., 1998; Walker et
al., 2001a].

2.4.2 Statistical Correction

A derivative of the direct insertion approach is the statistical correction approach, which
adjusts the mean and variance of the model states to match those of the observations. This
approach assumes the model pattern is correct but contains a non-uniform bias. First the
predicted observations are scaled by the ratio of observation field standard deviation to
predicted field standard deviation. Second, the scaled predicted observation field is given a
block shift by the difference between the means of the predicted observation field and
observation field [Houser et al., 1998]. This approach also relies upon the model physics to
propagate the information to unobserved parts of the system.

2.4.3 Successive Correction

This is an iterative type approach that uses weights W to smooth observations into the
model states, by modifying the states at all grid points within a specified radius of influence r
of each observation s [Bratseth, 1986]. Any weighting system can be used, but the Cressman
weights given by

⎧ r 2 − d ij2
⎪ d ij < r
Wijs = ⎨ r 2 + d ij2 (2.12)
⎪ 0 d ij ≥ r
⎩

are commonly used, where dij is the distance between grid point i,j and the observation. In
practice the approach is usually applied consecutively to each observation s from 1 to sf as

(
X ks +1 = X ks + Wks Z k − Z
ˆ ,
k ) (2.13)

~s ~s
and then setting X ak = X k f . This is equivalent to using K k = W f in equation 2.1 where W f
s

is calculated from
~
W 1 = W1
M
~ . (2.14)
~
( ~
W s+1 = W s +1 I + HW s + W s )
M

This approach assumes that the observations are more accurate than model forecasts, with the
observations fitted as closely as is consistent. Moreover, it is ineffective in data sparse
regions [Nichols, 2001].

2.4.4 Optimal Interpolation

The optimal interpolation (OI) approach, sometimes referred to as statistical interpolation,
approximates the “optimal” solution from equation 2.10 by choosing

10
 ~ ~
(
K = Σ b H T HΣ b H T + R )
−1
(2.15)

~
where Σ b is an approximated background covariance matrix with a “fixed” structure for all
time steps, and is often given by prescribed variances and a correlation function given only by
distance [Lorenc, 1981].

2.4.5 Analysis Correction

This is a modification to the successive correction approach that is applied consecutively to
each observation s from 1 to sf as [Lorenc et al., 1991]

(
X ks +1 = X ks + Wk Vk Z ks − Z
ˆ ,
k ) (2.16)

where the observation vector Zs is also successively updated by

(
Z ks+1 = Z ks − Vk Z ks − Z
ˆ ,
k ) (2.17)

and the weight matrices W and V given by

~
Wk = Σ bk H Tk R −k 1 (2.18a)
(
Vk = I + H k Wk )
−1
. (2.18b)

In practice Zs is not updated and Vk is approximated to avoid inversion. The result of these
assumptions is an update equation equivalent to that for optimal interpolation [Nichols, 1991].

2.4.6 Nudging
The nudging approach approximates the gain matrix by the empirical function

( )
K ≈ G W T ΘW (W I ) ,
−1
(2.19)

where G is a nudging factor that gives the magnitude of the nudging term and has a value
from 0 to 1, Θ is an observational quality factor with a value from 0 to 1, I is the identity
matrix and W is a temporal and spatial weighting function also with a value from 0 to 1. The
function W is given by wxywzwt, where wxy is a horizontal weighting function (i.e. Cressman),
wz is a similar vertical weighting function, and wt is a temporal weighting function. Each of
these temporal/spatial weighting functions has a value from 0 to 1 [Stauffer and Seaman,
1990].

2.4.7 3D Variational
This approach directly solves the iterative minimisation problem given by equations 2.2 or
2.3 for N = 1 [Parrish and Derber, 1992]. The same approximation for the background
covariance matrix as in the optimal interpolation approach is typically used. The solution
gives an analysis which is similar in nature to the direct insertion approach.

11
2.4.8 Kalman Filter
The family of Kalman filter data assimilation approaches calculate the gain matrix in
equation 2.10 by directly forecasting the background covariance matrix. In the traditional
Kalman filter (KF) approach this is achieved by application of standard error propagation
theory on the (tangent) linear model in equation 2.5. (The only difference between the
Kalman filter and the extended Kalman filter is that the forecast model is linearised using a
Taylor’s series expansion; the same forecast and update equations are used for each.) The
state covariance forecast equation is

Σ bk +1 = A k Σ bk A Tk + Q k , (2.20)

where A is the linear operator from equation 2.5 and Q is the model error covariance matrix
given in equation 2.9. Thus, the (extended) Kalman filter requires propagation of the state
covariances along with the states. While the approach gives an optimal analysis for the
assumed statistics, the initial state error covariance matrix Σ0 and more seriously the model
error covariance matrix Q are difficult to define, and often assumed ad hoc.
The standard extended Kalman filter (EKF) approach assumes an explicit model, which is
limiting in terms of computational runtime as a result of the small step size necessary to
satisfy stability criteria. However, it is also possible to apply the same update and state
covariance forecast equations to an implicit formulation, such as the Crank-Nicholson scheme

Φ1 X k +1 + Ω1 = Φ 2 X k + Ω 2 , (2.21)

by making the substitutions that

A k = Φ1−1Φ 2 (2.22a)
B k U k = Φ [Ω 2 − Ω1 ] ,
−1
1 (2.22b)

but the inverse and multiplication required to calculate A is costly for large systems [Walker
et al., 2001b].
The standard extended Kalman filter update and state covariance forecast equations can
also be applied directly with a nonlinear state forecast model. This is achieved by numerically
approximating the Jacobians A and H as required by

∂Xbk +1
Ak = (2.23a)
∂Xbk
∂Z k
Hk = . (2.23b)
∂Xbk

However, the cost of doing this is n+1 times the standard model run time, where n is the
number of state variables to be updated by the assimilation. Note that only states with
significant correlation to the observation need be included in the state covariance forecast and
update [Walker and Houser, 2001].
A further approach to estimating the state covariance matrix is the ensemble Kalman filter
(EnKF). As the name suggests, the covariances are calculated from an ensemble of state

12
forecasts using the Monte Carlo approach rather than a single discrete forecast of covariances.
In this case m ensembles of n model predicted states X are stored as x using different initial
conditions and forcing [Turner et al., 2004], different parameters and/or models, different
model error (additive/multiplicative/etc.), etc., in order to get a representative spread of state
forecasts amongst the ensemble members. While this is quite straight forward, the question of
what model error w to apply, and how, is still a major unknown. Moreover, special care is
required when m is less than the number of observations n.
Using this approach, the background state covariance matrix is basically calculated as

Σ b
=
(x b
k )(
− x bk x bk − x bk ) T

. (2.24)
m −1
k

This could then be used in equation 2.10 directly, except some smart maths is typically used
so only matrices of size (n x m) are required [Evensen, 1994; Houtekamer and Mitchell,
1998]. Thus, Σb is never calculated explicitly. Here the analysis equation is presented as

X ak = X bk + B Tk b k , (2.25)

where

B Tk = Σ bk H Tk (2.26a)
(
b k = H k Σ bk H Tk + R k ) (Z
−1
k −Z
ˆ .
k ) (2.26b)

By rearranging equation 2.26a and letting y = Z + ζ, where ζ is a zero mean random

observation error term with covariance matrix R, b is solved for each ensemble from

(H Σ H k
b
k
T
k )
+ Rk bk = y k − Z (
ˆ ,
k ) (2.27)

where

q k q Tk
H k Σ bk H Tk = (2.28)
m −1

and

(
q k = H k x bk − x bk = zˆ k − zˆ k . ) ( ) (2.29)

The vector ẑ is the predicted observation vector for each of the respective ensemble
members. In this case it is not necessary to solve for H either, and the updates are made
individually to each of the ensemble members. Finally, B can be estimated from

B Tk =
(x − x bk T
b
k
qk .
) (2.30)
m −1

Reichle et al. [2001a] applied the ensemble Kalman filter to the soil moisture estimation
problem and found it to perform as well as the numerical Jacobian approximation approach to

13
the extended Kalman filter, with the distinct advantage that the error covariance propagation
is better behaved in the presence of large model nonlinearities. This was the case even when
using only the same number of ensembles as required by the numerical approach to the
extended Kalman filter i.e. n+1.

2.5 Dynamic Observer Assimilation Methods

In its pure form, the ‘variational’ (otherwise known as Gauss-Markov) dynamic observer
assimilation methods use an adjoint to efficiently compute the derivatives of the objective
function J with respect to each of the initial state vector values X0. This adjoint approach is
derived by defining the Lagrangian L as the adjoining of the model to the model response
using Lagrange multipliers λ

N −1
L = J + ∑ λ Tk +1 [X k +1 − ak (X k , U k )] , (2.31)
0

where ideally the second term is zero. Thus the Lagrange multiplier is chosen such that
∇L = 0 and λN = 0, yielding (i.e. backward pass)

(
λ k = A Tk λ k +1 − H Tk R k−1 Z k − Z )
ˆ .
k (2.32)

The derivative of the objective function is given from the Lagrange multiplier at time zero by
− λ T0 [Castelli et al., 1999; Reichle et al., 2001a]. Note that AT, the adjoint operator, is from
the tangent linear model in equation 2.5, and effectively needs to be saved during the forward
pass [Bouttier and Courtier, 1999]. Solution to the variational problem is then achieved by
minimisation and iteration. In practical application the number of iterations is usually
constrained to a small amount.
While ‘adjoint compilers’ are available (see [Link] for automatic
conversion of the nonlinear forecast model into a tangent linear model, application of these is
not straight forward. It is best to derive the adjoint at the same time as the model is
developed.

2.6 Case Studies

Significant advances in hydrologic data assimilation have been made over the past decade
from which we have selected a few case studies to demonstrate the utility of hydrologic data
assimilation.

2.6.1 Case Study 1: Soil Moisture Assimilation

A one-dimensional Kalman filter soil moisture assimilation strategy was developed by
Walker and Houser [2001] that provides a framework to constrain model predicted soil
moisture with observations, using covariances that represent their respective uncertainty. A
one-dimensional extended Kalman filter was used because of its computational efficiency and
the fact that horizontal correlations between soil moisture prognostic variables of adjacent
catchments at the scales of interest to climate modelling are likely only through the large-
scale correlation of atmospheric forcing. A set of numerical experiments was undertaken for
North America to illustrate the effectiveness of the Kalman filter assimilation scheme in

14
providing an accurate soil moisture estimate. When assimilating surface soil moisture once
every 3 days, the scheme was generally able to retrieve the “true” profile soil moisture after
only 1 month, and the predicted evapotranspiration and runoff fluxes were significantly
improved.
Walker and Houser [2004] used this same EKF framework to address soil moisture
satellite mission accuracy, repeat time and spatial resolution requirements through a
numerical twin study. Simulated soil moisture profile retrievals were made by assimilating
near-surface soil moisture observations with various accuracy, repeat time and spatial
resolution. It was found that: (i) near-surface soil moisture observation error must be less
than the model forecast error required for a specific application and must be better than 5%v/v
accurate to positively impact soil moisture forecasts, (ii) daily near-surface soil moisture
observations achieved the best soil moisture and evapotranspiration forecasts, (iii) near-
surface soil moisture observations should have a spatial resolution of around half the model
resolution, and (iv) satisfying the spatial resolution and accuracy requirements was much
more important than repeat time. This kind of study is important for planning future
observation systems, but it must be recognised that observation requirements are also highly
application specific; for example, flood forecasting and precision agriculture requirements
will likely have different requirements than climate modelling and policy planning, as they
operate at different scales.
Walker and Houser [2002] was the first known study to use space-borne measurements of
near-surface soil moisture content to estimate the spatial and temporal variation of soil
moisture content at the continent scale by the process of data assimilation. Near-surface soil
moisture measurements from the 6.6 GHz (C-band) channels of the Scanning Multi-channel
Microwave Radiometer (SMMR) were assimilated into a land surface model over North
America using a Kalman filter to correct for soil moisture estimation errors. Comparison with
the limited ground-based point measurements of soil moisture content found a net
improvement when near-surface soil moisture observations were assimilated. Walker et al.
[2003] used a similar approach to estimate the spatial and temporal variation of soil moisture
content across Australia (Figure 2.2). Unfortunately the lack of appropriate soil moisture
evaluation data and mismatch in scale between model output and available data made it
difficult to draw any conclusive statements regarding improvements in soil moisture
predictions. There was however an obvious increase in correlation between soil moisture
predictions and NDVI data when SMMR surface soil moisture data were assimilated. This
provides some encouragement for pursuing assimilation experiments using the new AMSR-E
data, and the collection of more appropriate ground-based soil moisture data for validation
purposes.

2.6.2 Case Study 2: Downscaling With Data Assimilation

In a data assimilation framework, it may be possible to effectively increase the resolution
of observations by making use of forecasts that include higher resolution meteorological, land
cover, and soil texture information. Figure 2.5 demonstrates how low resolution brightness
temperature observations have been used to estimate soil moisture at high resolution through
variational assimilation into a land surface model [Reichle et al., 2001b]. Spatial structures at
scales well below the scale of the observations were resolved satisfactorily. This means that
brightness images with a resolution of a few tens of kilometres are useful, even if the
estimation scale of interest is of the order of a few kilometres, provided that fine-scale
information is available on the meteorological forcing, land cover, and texture.

15
 Fig. 2.5
The downscaling properties of data assimilation may also be able to help overcome the
limitations of passive radiometric remote sensing (higher accuracy but lower spatial
resolution) and active radar remote sensing (higher spatial resolution but lower accuracy) for
sensing soil moisture. Zhan et al. [2004] tested this hypothesis by conducting an Observation
System Simulation Experiment (OSSE), where the feasibility of retrieving surface soil
moisture at a medium spatial scale (10 km) from both coarse scale (40 km) radiometer
brightness temperature and fine scale (3 km) radar backscatter cross-section observations was
evaluated using the extended Kalman filter. In this case the background field was an
inversion of the passive data. Compared with the results from traditional soil moisture
inversion algorithms, the combined active-passive extended Kalman filter retrieval algorithm
significantly reduced soil moisture error at the medium scale. There is the potential to further
enhance this downscaling approach by additionally using the information contained in
overlapping observations.

2.6.3 Case Study 3: Snow Assimilation

Because of snow’s high albedo, thermal properties, feedback to the atmosphere and being a
medium-term water store, improved snow state estimation has the potential to greatly increase
climatological and hydrological prediction accuracy. An analysis scheme to assimilate
observed snow water equivalent into a land surface model has been developed [Sun et al.,
2004]. Using a set of numerical ‘twin’ experiments, the scheme is shown to be successful in
retrieving the snow states (snow water equivalent, snow depth and snow temperature) from
observations of snow water equivalent alone. The study illustrates that by assimilating
remotely sensed snow water equivalent observations, the errors in forecast snow states from
poor initial conditions can be removed (Figure 2.6), and the prediction of runoff and
atmospheric fluxes improved. A comparison between monthly-averaged runoff and
atmospheric fluxes showed negligible differences between the assimilation and truth
simulations. Moreover, the assimilation significantly improved both upward shortwave and
longwave radiation, and runoff predictions, as compared to no assimilation.
Fig. 2.6
Snow has several properties that make it uniquely challenging to assimilate, as follows.
First, snow cover and depth observations provide an incomplete description of the multi-layer
snow water equivalent, temperature and density states used in most physical snow models. For
example, snow cover observations provide a binary snow presence description without snow
quantity information. This is generally incompatible with data assimilation schemes that act
on snow states. Rodell et al. [2002] overcame this problem by adding an arbitrarily thin layer
of snow to model elements that have no snow when snow cover was observed. Second, snow
is a highly transient model state that disappears and is not predicted for long periods of time
during the year. This is generally incompatible with modern data assimilation techniques that
seek to propagate error covariances; when there is no snow state prediction (i.e. when the
temperature is above freezing), then there can also be no error propagation. This problem was
overcome by Sun et al., [2004] by simply reinitializing the error covariances when snow
reappeared, but may not be so easily overcome in an ensemble Kalman filter context where
ensemble members may vary significantly. Finally, in the presence of temperature bias, snow
assimilation may have an undesirable water budget impact. Cosgrove et al. [2004] show that
large water balance errors occur when imperfect snow melting processes interact with the
direct insertion of perfect snow observations. Constraining these snow melt biases is important
for achieving optimal assimilation results, and is an important topic for future research.

16
2.6.4 Case Study 4: Skin Temperature Assimilation
The land surface skin temperature state is a principle control on land-atmosphere fluxes of
water and energy. It is closely related to soil water states, and is easily observable from space
and aircraft infrared sensors in cloud-free conditions. The usefulness of skin temperature in
land data assimilation studies is limited by its very short memory (on the order of minutes)
due to the very small heat storage it represents. Radakovich et al. [2001] have demonstrated
skin temperature data assimilation in a land surface model (Figure 2.7) using three-hourly
observations from the International Satellite Cloud Climatology Project (ISCCP).
Incremental and semi-diurnal bias correction techniques based on Dee and da Silva [1998]
were developed to account for biased skin temperature forecasts. The assimilation of ISCCP-
derived surface skin temperature significantly reduced the bias and standard deviation
between model predictions and the NCEP reanalysis [Kalnay et al., 1996]. However, the
assimilation of ISCCP-derived surface skin temperature has a substantial impact on the sensible
heat flux, due to an enhanced gradient between the surface and 2m air temperatures. If the near-surface
air temperature were interactive, as in a coupled land-atmosphere model, then it would respond to this
enhanced flux rather then maintaining the artificial temperature gradient.
Fig. 2.7

2.7 Summary
Hydrologic data assimilation is an objective method to estimate the hydrologic system
states from irregularly distributed observations. These methods integrate observations into
numerical prediction models to develop physically consistent estimates that better describe
the hydrologic system state than the raw observations alone. This process is extremely
valuable for providing initial conditions for hydrologic system prediction and/or correcting
hydrologic system prediction, and for increasing our understanding and improving
parameterisation of hydrologic system behaviour through various diagnostic research studies.
Hydrologic data assimilation is still in its infancy, with many open areas of research.
Development of hydrologic data assimilation theory and methods is needed to: (i) better
quantify and use model and observation errors; (ii) create model independent data
assimilation algorithms that can account for the typical non-linear nature of hydrologic
models; (iii) optimise data assimilation computational efficiency for use in large operational
hydrological applications; (iv) use forward models to enable the assimilation of remote
sensing radiances directly; (v) link model calibration and data assimilation to optimally use
available observation information; (vi) create multivariate hydrologic assimilation methods to
use multiple observations with complementary information; (vii) quantify the potential of data
assimilation downscaling; and (viii) create methods to extract the primary information content
from observations with redundant or overlaying information. Further, the regular provision of
snow, soil moisture, and surface temperature observations with improved knowledge of
observation errors in time and space are essential to advance hydrologic data assimilation.
Hydrologic models must also be improved to: (i) provide more “observable” land model
states, parameters, and fluxes; (ii) include advanced processes such as river runoff and
routing, vegetation and carbon dynamics, and groundwater interaction to enable the
assimilation of emerging remote sensing products; (iii) have valid and easily updated adjoints,
and (iv) have knowledge of their prediction errors in time and space. The assimilation of
additional types of hydrologic observations, such as streamflow, vegetation dynamics,
evapotranspiration, and groundwater or total water storage must be developed.

17
 As with most current data assimilation efforts, we describe data assimilation procedures
that are implemented in uncoupled models. However, it is well known that the high-
resolution time and space complexity of hydrologic phenomena have significant interaction
with atmospheric, biogeochemical, and oceanic processes. Scale truncation errors, unrealistic
physics formulations, and inadequate coupling between hydrology and the overlying
atmosphere can feedback to cause serious systematic hydrologic errors. Hydrologic balances
cannot be adequately described by current uncoupled hydrologic data systems, because large
analysis increments that compensate for errors in coupling processes (e.g. precipitation) result
in important non-physical contributions to the energy and water budgets. Improved coupled
process models with improved feedback processes, better observations, and comprehensive
methods for coupled assimilation are needed to achieve the goal of fully coupled data
assimilation systems that should produce the best and most physically consistent estimates of
the Earth system.

2.8 References
Arya, L. M., Richter, J. C., and Paris, J. F., (1983). Estimating profile water storage from
surface zone soil moisture measurements under bare field conditions. Wat. Resour. Res.,
19(2): 403-412.
Bennett, A. F., (1992.) Inverse methods in physical oceanography. Cambridge University
Press, 346 pp.
Bernard, R., Vauclin, M., and Vidal-Madjar, D., (1981). Possible use of active microwave
remote sensing data for prediction of regional evaporation by numerical simulation of soil
water movement in the unsaturated zone. Wat. Resour. Res., 17(6): 1603-1610.
Bouttier, F., and Courtier P., (1999). Data assimilation concepts and methods. ECMWF
training course notes.
Bouttier, F., Mahfouf, J. -F., and Noilhan, J., (1993). Sequential assimilation of soil moisture
from atmospheric low-level parameters. Part I: Sensitivity and calibration studies, J. Appl.
Meteorol., 32(8): 1335-1351.
Bratseth, A. M., (1986). Statistical interpolation by means of successive corrections. Tellus
38A: 439-447.
Bruckler, L., and Witono, H. (1989). Use of remotely sensed soil moisture content as
boundary conditions in soil-atmosphere water transport modeling: 2. Estimating soil water
balance. Wat. Resour. Res., 25(12): 2437-2447.
Castelli, F., Entekhabi, D., and Caporali, E., (1999). Estimation of surface heat flux and an
index of soil moisture using adjoint-state surface energy balance. Wat. Resour. Res., 35(10):
3115-3125.
Charney, J. G., Halem, M., and Jastrow, R., (1969). Use of incomplete historical data to infer
the present state of the atmosphere. J. Atmos. Sci. 26: 1160-1163.
Cosgrove, B. A., Houser P. R., and Toll D. L., (2004). Impact of surface forcing biases on
snow assimilation. In preparation.
Daley, R., (1991). Atmospheric data analysis. Cambridge University Press, 460 pp.
Dee, D. P., and da Silva, A., (1998). Data assimilation in the presence of forecast bias. Q. J.
R. Meteorol. Soc., 124: 269-295.

18
Dee, D. P., and Todling, R., (2000). Data assimilation in the presence of forecast bias: The
GEOS moisture analysis. Mon. Wea. Rev., 128: 3268-3282.
Entekhabi, D., Nakamura H., and Njoku, E. G., (1994). Solving the inverse problem for soil
moisture and temperature profiles by sequential assimilation of multifrequency remotely
sensed observations. IEEE Trans. Geosci. Rem. Sens., 32: 438-448.
Evensen, G., (1994). Sequential data assimilation with a nonlinear quasi-geostrophic model
using Monte Carlo methods to forecast error statistics. J. Geophys. Res., 99(C5): 10143-
10162.
Georgakakos, K. P., and Baumer, O. W., (1996). Measurement and utilization of on-site soil
moisture data. J. Hydrol., 184: 131-152.
Hollingsworth, A. and Lonnberg, P., (1989). The verification of objective analyses:
Diagnostics of analysis system performance. Meteor. Atmos. Phys., 40: 3-27.
Houser, P. R., et al., (1998). Integration of soil moisture remote sensing and hydrologic
modeling using data assimilation. Wat. Resour. Res., 34(12): 3405-3420.
Houtekamer, P. L., and Mitchell, H. L., (1998). Data assimilation using a Ensemble Kalman
filter techniques, Mon. Weather Rev., 126: 796-811.
Jackson, T. J., et al., (1981). Soil moisture updating and microwave remote sensing for
hydrological simulation. Hydrol. Sci. B., 26(3): 305-319.
Kalman, R. E., (1960). A new approach to linear filtering and prediction problems. Trans.
ASME, Ser. D, J. Basic Eng. 82: 35-45.
Kalnay, E., et al., (1996) The NCEP/NCAR 40-year reanalysis project. Bullet. Amer.
Meteorol. Soc., 77: 437-471.
Kostov, K. G., and Jackson, T. J., (1993). Estimating profile soil moisture from surface layer
measurements – A review. In: Proc. The International Society for Optical Engineering,
Vol. 1941, Orlando, Florida, 125-136.
Lorenc, A., (1981). A global three-dimensional multivariate statistical interpolation scheme.
Mon. Wea. Rev., 109: 701-721.
Lorenc, A. C., Bell, R. S., and Macpherson, B. (1991). The Meteorological Office analysis
correction data assimilation scheme. Quart. J. Roy. Meteor. Soc., 117, 59-89
Milly, P. C. D., (1986). Integrated remote sensing modelling of soil moisture: sampling
frequency, response time, and accuracy of estimates. Integrated Design of Hydrological
Networks - Proceedings of the Budapest Symposium, 158: 201-211.
Nichols, N. K., (2001). State estimation using measured data in dynamic system models,
Lecture notes for the Oxford/RAL Spring School in Quantitative Earth Observation.
Ottlé, C., and Vidal-Madjar, D., (1994). Assimilation of soil moisture inferred from infrared
remote sensing in a hydrological model over the HAPEX-MOBILHY Region. J. Hydrol.,
158: 241-264.
Parish, D., and Derber, J., (1992). The National Meteorological Center's spectral statistical
interpolation analysis system. Mon. Wea. Rev., 120: 1747-1763.
Prevot, L., et al., (1984). Evaporation from a bare soil evaluated using a soil water transfer
model and remotely sensed surface soil moisture data. Wat. Resour. Res., 20(2): 311-316.

19
Radakovich, J. D., Houser, P. R., da Silva, A., and Bosilovich, M. G., (2001). Results from
global land-surface data assimilation methods. Proceedings AMS 5th Symposium on
Integrated Observing Systems, Albuquerque, NM, 14-19 January, 132-134.
Reichle, R. H., and McLaughlin, D. B., (2001a). Variational data assimilation of microwave
radiobrightness observations for land surface hydrologic applications. IEEE Trans. Geosci.
Rem. Sens., 39(8): 1708-1718.
Reichle, R. H., Entekhabi, D., and McLaughlin, D. B., (2001b). Downscaling of
radiobrightness measurements for soil moisture estimation: A four-dimensional variational
data assimilation approach, Wat. Resour. Res., 37(9): 2353-2364.
Rodell, M., (2002). Use of MODIS-derived snow fields in the Global Land Data Assimilation
System. Proceedings GAPP Mississippi River Climate and Hydrology Conference, New
Orleans.
Rood, R. B., Cohn, S. E., and Coy, L., (1994.) Data assimilation for EOS: The value of
assimilated data, Part 1. The Earth Observer, 6(1): 23-25.
Schuurmans, J. M., et al., (2003). Assimilation of remotely sensed latent heat flux in a
distributed hydrological model. Adv. Water Resour., 26(2): 151-159.
Stauffer, D. R., and Seaman, N. L., (1990). Use of four-dimensional data assimilation in a
limited-area mesoscale model. Part I: Experiments with synoptic-scale data. Mon. Wea.
Rev., 118: 1250-1277.
Sun, C., Walker, J. P., and Houser, P. R., (2004). A methodology for snow data assimilation
in a land surface model. J. Geophys. Res.-Atmos., 109.
Turner, M. R. J., Walker, J. P., and Oke, P. R., (2004). Ensemble member generation for
sequential data assimilation. In preparation.
Viterbo, P., and Beljaars, A., (1995). An improved land surface parameterization scheme in
the ECMWF model and its validation. J. Climate, 8: 2716–2748.
Walker J. P., and Houser, P. R., (2001). A methodology for initialising soil moisture in a
global climate model: assimilation of near-surface soil moisture observations, J. Geophys.
Res.-Atmos., 106(D11): 11761-11774.
Walker, J. P., Willgoose, G. R., and Kalma, J. D., (2001a). One-dimensional soil moisture
profile retrieval by assimilation of near-surface observations: A comparison of retrieval
algorithms. Adv. Water Resour., 24(6): 631-650.
Walker, J. P., Willgoose, G. R. and Kalma, J. D., (2001b). One-dimensional soil moisture
profile retrieval by assimilation of near-surface measurements: A simplified soil moisture
model and field application. J. Hydromet., 2(4): 356-373.
Walker, J. P., Willgoose, G. R. and Kalma, J. D., (2002). Three-dimensional soil moisture
profile retrieval by assimilation of near-surface measurements: Simplified Kalman filter
covariance forecasting and field application. Wat. Resour. Res., 38(12): 1301.
Walker, J. P., and Houser, P. R., (2002). Soil moisture estimation using remote sensing.
Proceedings 27th Hydrology and Water Resources Symposium. The Institute of Engineers
Australia, Melbourne, Australia, 20 - 23 May.
Walker, J. P., and Houser, P. R., (2004). Requirements of a global near-surface soil moisture
satellite mission: Accuracy, repeat time, and spatial resolution. Adv. Water Resour., 27:
785-801.

20
Walker, J. P., Ursino, N., Grayson, R. B., and Houser, P. R., (2003). Australian root zone soil
moisture: Assimilation of remote sensing observations, In: D. Post (Ed.), Proceedings of the
International Congress on Modelling and Simulation (MODSIM). Modelling and
Simulation Society of Australia and New Zealand, Inc., Townsville, Australia, 14-17 July,
1: 380-385.
Zhan, X., et al., (2004). Retrieving medium resolution surface soil moisture from coarse
resolution radiometer and fine resolution radar observations using the Kalman filter. In
preparation.

21
Table 2.1: Characteristics of hydrologic observations potentially available within the next
decade.
Hydrologic Remote Time Space Accuracy Considerations Example Sensors
Quantity Sensing Scale Scale
Technique
Thermal Hourly 4km Tropical convective clouds GOES
infrared 1day 1km only MODIS, AVHRR
15days 60m Landsat, ASTER
Precipitation Passive 3hour 10km Land calibration problems TRMM, SSMI,
microwave AMSR-E, GPM
Active Daily 10m Land calibration problems TRMM, GPM
microwave
Passive 1-3days 25-50km Limited to sparse AMSR-E, SMOS,
microwave vegetation, low Hydros
Surface soil
topographic relief
moisture
Active 3days 3km Significant noise from
microwave 30days 10m vegetation and roughness ERS, JERS, RadarSat
Surface skin Thermal 1hour 4km Soil/vegetation average, GOES
temperature infrared 1day 1km cloud contamination MODIS, AVHRR
15days 60m Landsat, ASTER
Snow cover Visible/ 1hour 4km Cloud contamination, GOES
thermal 1day 500m- vegetation masking, bright MODIS, AVHRR
infrared 15days 1km soil problems Landsat, ASTER
30-60m
Passive 1-3days 10km Limited depth penetration AMSR-E
Snow water microwave
equivalent Active 30days 100m Limited spatial coverage SnoSat or CLPP
microwave (proposed missions)
Laser 10days 100m Cloud penetration ICESAT
Water level/
problems
velocity
Radar 30days 1km Limited to large rivers TOPEX/POSEIDON
Total water Gravity 30days 1000km Bulk water storage change GRACE
storage changes
changes
Thermal 1hour 4km Significant assumptions GOES
Evaporation infrared 1day 1km MODIS, AVHRR
15days 60m Landsat, ASTER

22
Table 2.2: Commonly used data assimilation terminology.
State condition of a physical system ie. soil moisture
Prognostic a model state required to propagate the model forward in time
Diagnostic a model state/flux diagnosed from the prognostic states – not required
to propagate the model
Observation measurement of a model diagnostic or prognostic
Covariance matrix describes the standard deviations & correlations
Prediction model estimate of states or covariances
Update correction to a model prediction using observations
Background prediction prior to an update
Analysis prediction after an update
Innovation observation-prediction
Gain matrix correction factor applied to the innovation
Tangent linear model linearised (using Taylor’s series expansion) version of a non-linear
model
Adjoint operator allowing the model to be run backwards in time

23
 Remote Sensing
Satellite

Near-Surface Moisture
Skin Temp
Logger etc Land Surface Model
θ (z) = f [qs, Dψ (θ), ψ(θ)]
Stream “Gauge” Point
Measurement

Figure 2.1: Schematic of the hydrologic data assimilation challenge.

24
Figure 2.2: Satellite observations of near-surface soil moisture content made by the scanning
multifrequency microwave radiometer (SMMR) are used to constrain hydrologic model
predictions of soil moisture throughout the root zone using data assimilation.

25
 x
x
State Value

x
x
x
observation
model
background
x analysis

Time
a)

Window 1 Window 2

x
x
State Value

x
x
observation
model
background
x analysis
Time
b)

Figure 2.3: Schematic of the a) direct observer and b) dynamic observer assimilation
approaches.

26
 a)

Depth
Depth

Matric Head Matric Head

b)
Model O bservation Model with 4DDA

Tombstone, AZ

0% 20%

Figure 2.4: Example of how data assimilation supplements data and complements
observations: (a) Numerical experiment results demonstrating how near-surface soil moisture
measurements are used to retrieve the unobserved root zone soil moisture state using (left
panel) direct insertion and (right panel) a statistical assimilation approach [Walker et al.,
2001a]; (b) Six Push Broom Microwave Radiometer (PBMR) images gathered over the
USDA-ARS Walnut Gulch Experimental Watershed in southeast Arizona were assimilated
into the TOPLATS hydrological model using several alternative assimilation procedures
[Houser et al., 1998]. The observations were found to contain horizontal correlations with
length scales of several tens of kilometres, thus allowing soil moisture information to be
advected beyond the area of the observations.

27
 TRUE PRIOR EST (1:4) EST (1:16)
day of year 169.7

160

120 0.6
y [km]

0.4
40
0.2
0
day of year 174.66

160

120 0.6
y [km]

0.4
40
0.2
0
day of year 178.66

160

120 0.6
y [km]

0.4
40
0.2
0
0 40 80 0 40 80 0 40 80 0 40 80
x [km] x [km] x [km] x [km]

Figure 2.5: True and prior surface soil saturation (in the first two columns) at three different
times. Corresponding estimates for downscaling ratios of (1:4) and (1:16) are shown in the
third and fourth columns. The resolution of the observations used to compute the downscaled
estimates are indicated with solid black grid lines [Reichle, 2001b].

28
 SWE
Snow Depth
Snow Temp
Snow Fraction a) Truth b) Assimilation c) Control

Figure 2.6: Comparison of snow simulations on January 5, 1987 over North America for
snow water equivalent (in mm, top row); snow depth (in mm, second row); average snow
temperature (in °C, third row); and areal snow fraction (bottom row) from a) truth run (using
spin up initial condition), b) assimilation run (with degraded initial condition and assimilation
of daily total snow water equivalent observations), and c) control run (with degraded initial
condition) [Sun et al., 2004].

29
Figure 2.7: Differences between simulated and reanalysis (top left), assimilated and reanalysis (bottom
left) mean skin temperature (K), and the resulting differences between simulated and reanalysis (top
right), and assimilated and reanalysis (bottom right) mean sensible heat fluxes (Wm-2) for September
through November 1992. Global terrestrial mean bias and standard deviation (SD) for September
through November are also noted [Radakovich et al., 2001].

30