Advances in Water Resources 28 (2005) 135–147

www.elsevier.com/locate/advwatres

Dual state–parameter estimation of hydrological models

using ensemble Kalman filter
a,*
Hamid Moradkhani , Soroosh Sorooshian a, Hoshin V. Gupta b, Paul R. Houser c

a
Department of Civil and Environmental Engineering, University of California, Irvine, CA 92697-2175, USA
b
Department of Hydrology and Water Resources, University of Arizona, Tucson, AZ 85721, USA
c
Hydrological Sciences Branch, NASA-GSFC Code 974, Greenbelt, MD 20771, USA

Received 25 March 2004; received in revised form 20 September 2004; accepted 21 September 2004

Abstract

Hydrologic models are twofold: models for understanding physical processes and models for prediction. This study addresses the
latter, which modelers use to predict, for example, streamflow at some future time given knowledge of the current state of the system
and model parameters. In this respect, good estimates of the parameters and state variables are needed to enable the model to gen-
erate accurate forecasts. In this paper, a dual state–parameter estimation approach is presented based on the Ensemble Kalman Fil-
ter (EnKF) for sequential estimation of both parameters and state variables of a hydrologic model. A systematic approach for
identification of the perturbation factors used for ensemble generation and for selection of ensemble size is discussed. The dual
EnKF methodology introduces a number of novel features: (1) both model states and parameters can be estimated simultaneously;
(2) the algorithm is recursive and therefore does not require storage of all past information, as is the case in the batch calibration
procedures; and (3) the various sources of uncertainties can be properly addressed, including input, output, and parameter uncer-
tainties. The applicability and usefulness of the dual EnKF approach for ensemble streamflow forecasting is demonstrated using a
conceptual rainfall-runoff model.
Ó 2004 Elsevier Ltd. All rights reserved.

Keywords: Streamflow forecasting; Stochastic processes; Data assimilation; Ensemble Kalman filter; Dual estimation; Kernel smoothing

1. Introduction and scope extend the applicability of data assimilation in hydrol-

ogy as emphasized by Troch et al. [39]. However, the
Hydrologic models are defined largely by parameters successful use of data assimilation relies on unbiased
and states, parameters being physical and generally model state prediction, which is largely dependent on
time-invariant descriptions of surface and subsurface accurate parameter estimation. During the past two dec-
characteristics, and states being fluxes and storages of ades, much effort has been directed toward the estima-
water and energy that are propagated in time by the tion of hydrologic model parameters (calibration) to
model physics. In practice, in addition to model simula- improve the forecast accuracy [7,8,11,32]. Conceptual
tion, reliable operation of a watershed system requires a hydrologic models are usually deterministic representa-
continuous correction of the forecast as observational tions, which typically do not contain descriptions of
data become available. This entails the critical need to the various sources of uncertainties. Although it has
been common to translate the inability of a model to
generate accurate streamflow forecasts into parameter
*
Corresponding author. Tel.: +1 949 824 8821; fax: +1 949 884
uncertainty, other sources of uncertainties, such as
8831. model structural error, input, and output measurement
E-mail address: [email protected] (H. Moradkhani). errors, also need to be accounted for [16,17]. Several

0309-1708/$ - see front matter Ó 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.advwatres.2004.09.002
136 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147

authors have studied the uncertainties associated with The development of interactive (dual) state–parameter
parameter estimation, and procedures have been devel- estimation using standard Kalman filter, in the context
oped for the statistical analyses of parameter uncertain- of hydrology, is traced back to [36,37] and later to the
ties [18–20,33,34,42,43]. joint state–parameter by state augmentation technique
The aforementioned calibration procedures generally [3,4] (see Section 3 for detail). Those techniques, how-
minimize long-term prediction error using a historical ever, were limited to linear dynamic systems. For non-
batch of data assuming time-invariant parameters, and linear dynamics, the extended Kalman filter (EKF),
thus make no attempt to include information from which relies on linearization of model using first order
new observations. Batch calibration requires a set of his- approximation of Taylor series, can be used. As re-
torical data to be kept in storage and processed en- ported by Refs. [9,29,30] the EKF can lead to unstable
masse while lacking the flexibility to investigate possible results when the nonlinearity in the system is strong. To
temporal evolution of the model parameters. Thiemann cope with the drawbacks of the EKF, a Monte Carlo-
et al. [35] emphasized another limitation of batch cali- based Kalman filter called ensemble Kalman filter
bration in hydrological prediction of an ungauged wa- (EnKF) was introduced by Evensen [9]. One of the
tershed where the lack of sufficient historical data advantages of the EnKF comparing to the standard
makes the batch method infeasible. These limitations, KF is that the estimation of priori model covariance
as well as an interest in inferring the uncertainty in the (see Section 2.1) is not needed for the updating (analysis)
estimated parameters, motivated Thiemann et al. [35] step although its calculation using the model ensemble is
and Misirli et al. [25] to develop a recursive scheme for straightforward.
model prediction and parameter estimation in the on- The EnKF was originally developed for dynamic
line mode. From another perspective, Kitanidis and state estimation while in this paper its applicability to
Bras [15] stated that adaptive estimation might be suit- static state (parameter) estimation by dual state–para-
able when the forecast lead-time is short in comparison meter estimation strategy is extended and its usefulness
to the response time of the watershed. They explained on streamflow forecasting is examined.
that it would be the case when the error in input is large The organization of the paper is as follow. In Section
while the error in output measurement is small. 2, the general framework for sequential data assimila-
Much of the efforts in simulation-based methods of tion is explained, where the mathematical formulation
hydrologic system analyses have been focused on (1) im- of the EnKF as a special type of Monte Carlo procedure
proved methods for parameter estimation wherein state for state estimation is elaborated. A systematic ap-
variable uncertainties were not explicitly taken into proach for identifying the perturbation factor, as a key
account or (2) improved procedures for estimating feature in the EnKF, and for tackling the uncertainties
time-varying state variables wherein the parameters in forcing data (input) and observation (output) is sug-
were assumed to be known in advance. The commonly gested. In Section 3, the dual EnKF algorithm that deals
used batch calibration techniques only address parame- simultaneously with both model parameters and state
ter uncertainty while uncertainties in input, output and variables is explained and kernel smoothing of parame-
model structure are ignored. The main weakness of such ters is employed for parameter sampling to avoid the
approaches is that they attribute all errors from input, over-dispersion of parameters through random walk.
output and model structure to model parameter uncer- In Section 4, the applicability of dual EnKF on a con-
tainty. Sequential data assimilation procedures have ceptual rainfall-runoff model and the power of this algo-
the potential to overcome this drawback in simulation- rithm in streamflow forecasting is demonstrated.
based methods by explicitly taking into account all the
sources of uncertainty. The Kalman filter [14], a recur-
sive data-processing algorithm, is the most commonly 2. General framework for sequential data assimilation
used sequential data assimilation technique, which re-
sults in optimal estimation for linear dynamic models Over the past decade, a rapid increase in earth system
with Gaussian uncertainties. science data assimilation activities has been witnessed.
Although filtering techniques can address the various Similarly, hydrologic data assimilation techniques have
sources of uncertainties in modeling, the typical pre- garnered a great deal of attention of hydrologists in
sumption of these procedures is that the parameters the sense that by taking advantage of real time observa-
are to be specified in advance and sequential estimation tion, more accurate forecast can be made [22,23,29,
is applied only to the state variables. Because there is no 30,39,40]. The mathematical framework of estimation
guarantee that model behavior does not change over theory provides the tools required to approach a variety
time, the model adjustment through the time variation of data assimilation problems. The basic objective of
of parameters together with state variables is incisive. data assimilation is to characterize the state of a system
Therefore a procedure that can provide the simulta- at some future time from the knowledge of the initial
neous estimate of states and parameters is required. state. The state of a hydrological system, xt, at time t
 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147 137

could be conditioned on the observation, y1:t, through EnKF is based upon Monte Carlo or ensemble genera-
the probability density function: tions where the approximation of forecast (a priori)
P ðxt j y 1:t Þ ð1Þ state error covariance matrix is made by propagating
an ensemble of model states using the updated states
Following [13], the generic discrete-time nonlinear sto- (ensemble members) from the previous time step. The
chastic-dynamic system can be expressed in the form of: key point in the performance of the EnKF according
xtþ1 ¼ f ðxt ; ut ; hÞ þ xt ; xt
N ð0; Rmt Þ ð2Þ to [5,29,30] is to generate the ensemble of observations
at each update time by introducing noise drawn from
where xt is an n-dimensional vector representing the sys- a distribution with zero mean and covariance equal to
tem state variables (for example catchment soil moisture the observational error covariance matrix; otherwise
content) at time t. The nonlinear propagator f(.) con- the updated ensemble will possess a very low covariance.
tains the model input vector (deterministic forcing data, A schematic representation of the EnKF is demon-
ut, e.g., mean areal precipitation), and the (possibly) strated in Fig. 1. As seen in Fig. 1, the EnKF propagates
time-invariant model parameter h. The model error is an ensemble of state vectors in parallel such that each
displayed by xt with covariance Rmt and represents all state vector represents one realization of generated
the uncertainties related to model structure and the forc- model replicates. Similar to Eq. (2), the model forecast
ing data. Cohn [6] explained that the model error is gen- is made in the EnKF for each ensemble member as
erally state-dependent even if the operator f(.) is linear. follows:
The state-dependence and even dependence upon the
parameters as part of the uncertain propagator cause xi iþ i i
tþ1 ¼ f ðxt ; ut ; h; tÞ þ xt ; i ¼ 1; . . . ; n ð4Þ
the model error to be unknown. For simplicity, it is
where xi
tþ1 is the ith ensemble member forecast at time
therefore appropriate to represent the model error as a
t + 1 and xiþ
t is the ith updated ensemble member at time
stochastic perturbation in Eq. (2).
t. In addition to representing the additive process noise,
Suppose that a set of scalar observations is taken at
which is common in standard Kalman filtering, the
time t + 1 and that we intend to assimilate the vector
EnKF represents the multiplicative model errors
of observations into the model. The output variables
through forcing data perturbations. The forcing data
of the model are functions of both the model state vari-
perturbations are made by adding the fit noise with
ables and the parameters characterizing the model. The
covariance Rut to the forcing data at each time step:
observation process in general form can be written as:
y tþ1 ¼ hðxtþ1 ; hÞ þ mtþ1 ; mtþ1
N ð0; Rytþ1 Þ ð3Þ uit ¼ ut þ fit ; fit
N ð0; Rut Þ ð5Þ

where propagator h(.) relates the state variables to the Now, we form the expression for the error covariance
measured variables (in our case streamflow) and yields matrix associated with the forecasted (a priori) estimate.
the expected value of the prediction given the model If the true state variables are known, we can use the fol-
states and parameters. All sources of errors in the obser- lowing expectation to estimate the a priori model error
vation are reflected by mt+1, which will be assumed here covariance:
to be Gaussian and independent of model error xt. P  true  true T
tþ1 ¼ E½ðxtþ1  xtþ1 Þðxtþ1  xtþ1 Þ  ð6Þ
However, because the true state is generally unknown, it
2.1. Ensemble Kalman filter (EnKF)-state estimation
is convenient to calculate the ensemble covariance
matrix:
Sequential data assimilation, also known as filtering,
consists of model state estimation at each observation 1
P T
tþ1 ¼ E½X tþ1 X tþ1  ¼ X tþ1 X Ttþ1 ð7Þ
time based only on the observations up to present. In n1
the linear case, this problem is solved by the well-known i  n  
where, X tþ1 Pn¼ ½xitþ1 
xtþ1 ; . . . ; xtþ1 
xtþ1  and
xtþ1 ¼
Kalman filter [14] as an optimal recursive data-process- i 1
E½xtþ1  ¼ n i¼1 xtþ1 .
ing algorithm. In the case of nonlinear dynamics, one The updated (a posteriori) error covariance could be
can linearize the current state vector to use the so-called estimated similarly after updating all of the ensemble
extended Kalman filter (EKF) [13]. The EKF has many members.
well-known drawbacks such as computational demand With the assumption of a priori estimate (forecasted
owing to the error covariance propagation and closure states xi
tþ1 ), we now seek to use the observation yt+1 to
approximation by neglecting the higher order deriva- obtain the posterior estimate (updated states xiþ tþ1 ). A lin-
tives of the model, which correspondingly may produce ear correction equation is used according to standard
instabilities or even divergence [10,13,24]. The ensemble Kalman filter to update forecasted state ensemble
Kalman filter (EnKF) as an alternative to the traditional members:
EKF was first introduced by Evensen [9] and later clar-
ified by Burgers et al. [5] and Van Leewen [41]. The xiþ i i
y itþ1 Þ
tþ1 ¼ xtþ1 þ K tþ1 ðy tþ1  ^ ð8Þ
138 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147

Fig. 1. Ensemble Kalman filter (EnKF) schematic. xi iþ

t : Forecasted state ensemble member and xt : updated state ensemble member.

where y itþ1 is the ith trajectory of the observation repli- ance of the state variables xi y itþ1 . The
tþ1 and prediction ^
cates generated by adding the noise of gitþ1 , with covari- above form of Kalman gain is its modified version of
ance Rytþ1 , to the actual observation: the standard Kalman gain represented as (K tþ1 ¼
1
P T  T 
tþ1 H ½HP tþ1 H þ Rtþ1  ) where P tþ1 is defined in (6),
y itþ1 ¼ y tþ1 þ gitþ1 ; gitþ1
N ð0; Rytþ1 Þ ð9Þ
H is the observation transition operator after lineariza-
This is one of the features of the EnKF in which obser- tion of observation model (2) and Rt+1 is the same as
vations (Eq. (9)) are treated as random variables by gen- Rytþ1 defined by (9). One of the advantages of EnKF is
erating an observation ensemble with mean equal to the that the estimation of P tþ1 is not needed (although pos-
actual observation at each time and a predefined covari- sible as explained in (7)), whereas its estimation in the
ance. One may want to consider an alternate strategy in standard KF is necessary.
updating step (Eq. (8)) by using the ensemble square root
filter (EnSRF) [47] such that the perturbation of observa- 2.2. Identification of hyper-parameters and estimation
tion is not needed. Whitaker and Hamill [47] justified the of ensemble size
applicability of EnSRF to the linear observation models
while, in this study we are interested in filtering the non- In general, the performance of most ensemble fore-
linear model dynamics; hence we develop our strategy casts (EF) is influenced by the quality of the ensemble
according to the version of the EnKF that treats the generation method, the forecast model and also the
observation as a random variable and perturbation of analysis scheme. A large number of procedures exist to
observation is required [5,29,30]. Therefore, in the next evaluate the ensemble forecasts [12,27,38,48]. The key
section, we will elaborate on a systematic procedure to feature of the EnKF, however, is the perturbation of
tune the magnitude of the forcing data and observation forcing data to generate replicates of the model state
covariances in order to generate a reliable ensemble while variables, and then the correction of the forecasted
ensemble size can be determined correspondingly. ensemble members through the analysis (update) step
Similarly, ^y itþ1 is the ith predictive variable at time (Eq. (8)). A question may arise on how to perturb the
t + 1: system to construct a reliable ensemble where the spread
^y itþ1 ¼ hðxi ð10Þ of the ensemble is within a meaningful range. Another
tþ1 ; hÞ
issue in EF is the efficiency of the procedure, which is
In Eq. (8), Kt+1 is the Kalman gain matrix which, in highly related to the ensemble size. As shown by Eqs.
adaptation to the ensemble based approach can easily (5) and (9), the perturbation of forcing data and obser-
be proven to be as: vation (input and output) are made by adding noise to
1 the variable of interest. A fundamental limitation here
K tþ1 ¼ Rxy yy y
tþ1 ½Rtþ1 þ Rtþ1  ð11Þ
is connected with the identifiability of the noise variance
where Ryy
tþ1is the forecast error covariance matrix of such that it can tackle the uncertainty in input and out-
the prediction ^y itþ1 , and Rxy
tþ1 is the forecast cross covari- put. Stochastic noises are assumed to be Gaussian with
 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147 139

predetermined variances, which are assumed to be het- where n and T are the ensemble size and period of anal-
eroscedastic (variance changing) [31]. Therefore the var- ysis respectively.
iance of the noises introduced to the input and output If the actual observation is statistically indistinguish-
variables (Eqs. (5) and (9)) are proportional to the mag- able from n ensemble members the expected value of the
nitude of the variables as follows: RMSE ratio Ra, as explained in [1,27] should be
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Rut ¼ c  ut ð12Þ ðn þ 1Þ
E½Ra ¼ ð17Þ
2n
Rytþ1 ¼ q  y tþ1 ð13Þ
The ratio of Ra to E[Ra] is referred to as the Normalized
Here, we call the proportionality factors of c and q as RMSE Ratio (NRR),
hyper-parameters. Ra
The magnitudes of these unknown hyper-parameters NRR ¼ ð18Þ
E½Ra
determine the ensemble spread. A modelÕs failure to
properly fit the observations is a measure of model In order to evaluate the ensemble performance, the
error, for which an obvious approximation is a compar- normalized RMSE ratio is used, while NRR > 1 indi-
ison between the spread of an ensemble and the ensem- cates that the ensemble has too little spread, and
ble mean forecast error. The deficiency in spread might NRR < 1 is an indication of an ensemble with too much
be a measure of the uncertainty associated with the spread. Ideal ensemble generation should produce a
ensemble mean. Anderson [1] discussed a simple proce- NRR value close to unity. As seen in Eqs. (13) and
dure based on [27] for evaluating the similarity of truth (14), hyper-parameters control the ensemble spread
versus randomly selected members of the ensemble. through the perturbation variance. Tuning of these
According to this method, the ratio of the time-averaged new parameters results to the meaningful ensemble gen-
RMSE of the ensemble mean, R1, to the mean RMSE of eration while input and output errors are taken into
the ensemble members R2, is calculated: account.
R1 2.3. Tuning of hyper-parameters
Ra ¼ ð14Þ
R2
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
#2ffi To demonstrate the EnKF hyper-parameter tuning
T u
u" !
1 X X n procedure presented above, the conceptual Hydrologic
R1 ¼ t 1 ^y i  y it ð15Þ
T n i¼1 t MODel (HyMOD) described by Refs. [2,44] (see Fig.
t¼1
2) was used. HyMOD originates in the probability dis-
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi tributed moisture model (PDM) [26], an extension of
u T some of the lumped storage models developed in
1X n u
t1
X
R2 ¼ ð^y i  y it Þ2 ð16Þ 1960s, and later to the case of multiple storages
n i¼1 T t¼1 t representing a spatial distribution of different storage

Fig. 2. Hydrologic MODel (HyMOD) conceptualization.

140 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147

capacities in a watershed. Boyle [2] described HyMOD which has been a test basin in numerous studies
as a rainfall excess model through a nonlinear tank con- [2,11,32,35,42]. The data consist of potential evapo-
nected with two series of linear tanks (three identical transpiration, ET (mm/d), mean areal precipitation, P
quick-flow tanks) in parallel to a slow-flow tank repre- (mm/d) as forcing data, and streamflow (m3/s) as
senting the groundwater flow. From the definitions of observation.
state variables given in Section 2, state variables in this According to Eq. (19), NRR was calculated for a
system are S: storage in the nonlinear tank representing range of observation perturbation factor (observation
the watershed soil moisture content, x1, x2 and x3: the hyper-parameter) and forcing data perturbation (input
quick-flow tank storages representing the temporary hyper-parameter) with 50 ensemble members, where
(short-time) detentions, e.g., depression storages, and four combinations of input and observation noise mag-
x4: the slow-flow tank storage (subsurface storage). Cor- nitudes were investigated (Fig. 3). The light shaded area
respondingly parameters of this model are Cmax, as the in each subplot shows the acceptable bound for NRR
maximum storage capacity within the watershed, bexp, and a range of acceptable hyper-parameters. As seen,
the degree of spatial variability of the soil moisture NRR is more sensitive to the observation hyper-para-
capacity within the watershed, a, a factor for partition- meter than the input hyper-parameter, implying that
ing the flow between two series of tanks, Rq and Rs as accurate estimation using HyMOD model is highly
the residence time parameters of quick-flow and slow- dependent on the observation replicate generation.
flow tanks, respectively. To further investigate the variation of NRR with re-
The above procedure was applied to tune the hyper- spect to ensemble size, NRR was examined for a range
parameters and ensemble size for streamflow forecasting of observation hyper-parameter (q = 5–25%) and input
of the Leaf River watershed, a humid watershed with an hyper-parameter (c = 10%) (Fig. 4). As seen in case A
area 1944 km2 area located north of Colins, Mississippi of Fig. 3, the minimum noise in input and observation

(A) (B)
1.091.07
1.05

0.93
0.91
21 21
0.69
0.79
0.85
1.03

0.91
0.89
1.01
0.99
0.97

0.81

0.77
0.95
0.93

0.87

0.83

0.79
0.81
18 18

0.83
15 15 0.99
1.01

0.97

0.85
0.87
0.95

0.89
0.71

12 12
1.11

0.75
0.73

1.05
1.11

1.07
1.09

1.03

9 9
1.13
Input hyper-parameter, γ [%]

6 6

3 3
3 6 9 12 15 18 21 24 3 6 9 12 15 18 21 24

(C) (D)
1.07

0.99
1.03

1.01

1.0
1.01
0.95
0.93
1.0

1.0

21 21
1.09
1.11

1.05

5
0.97
3
7

18 18

15 15
1.01

0.99
0.97
1.09
1.11

0.91

12 12

9 9
1.05

1.
1.07

6 6
0.

01
1.03

1.0

99
07

3
1.

3 3
3 6 9 12 15 18 21 24 3 6 9 12 15 18 21 24
Observation hyper- parameter, ρ [%]

Fig. 3. NRR space with respect to hyper-parameters with 50 ensemble members for different noise scenarios, (A) input noise = 5%, observation
noise = 5%, (B) input noise = 20%, observation noise = 5%, (C) input noise = 5%, observation noise = 20%, and (D) input noise = 20%, observation
noise = 20%.
 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147 141

1.3 1.3
Obs. Hyper-parameter = 5% (A) (B)
10%
1.2 (ρ) 15% 1.2
20%
25%
1.1 1.1

1 1

0.9 0.9
Normalized Rmse Ratio [–]

0.8 0.8

0.7 0.7
20 30 40 50 60 70 80 20 30 40 50 60 70 80

1.3 1.3
(C) (D)
1.2 1.2

1.1 1.1

1 1

0.9 0.9

0.8 0.8

0.7 0.7
20 30 40 50 60 70 80 20 30 40 50 60 70 80
Ensemble Size Ensemble Size

Fig. 4. Normalized RMSE ratio (NRR) against ensemble size for input hyper-parameter (input perturbation factor, c = 10%) and different
observation hyper-parameter (observation perturbation factors, q = 5–25%). Four input and output error combinations: (A) Input noise = 5%,
observation noise = 5%, (B) input noise = 20%, observation noise = 5%, (C) input noise = 5%, observation noise = 20%, and (D) input noise =
20%, observation noise = 20%. Dash lines demonstrate the upper and lower bound of acceptable NRR (0.99–1.01) for this study.

(each 5%), the minimum ensemble size of 40 and obser- guarantee that model behavior does not change over
vation hyper-parameter, q = 10% will keep the NRR time; therefore model adjustment over time may be re-
within the acceptable range (0.99–1.01). Although for quired. Additionally, due to the multiplicative nature
q = 15% with the ensemble size of 30, NRR still lies of errors in forcing data and observation, it is prudent
within the acceptable range, the reliability on the ensem- to assemble the parameter adaption in the state evolu-
ble size for this case is low because by increasing the tion and forecasting system [40]. The need for real time
ensemble size the NRR drops suddenly, implying that state–parameter estimation of hydrological models is
the ensemble spread becomes too large. By increasing not free from empiricism and has been reported in sev-
the input noise in case B, and keeping the observation eral studies [3,4,36,37,40].
noise same as A, minor changes in NRR with respect Section 2.1 illustrated that recursive state estimation
to observation hyper-parameter q and ensemble size in a stochastic-dynamic system is carried out such that
are seen, whereas in case C, by just increasing the obser- the parameters are assumed to be a time-invariant sys-
vation noise and having the input noise the same as case tem description. In this section, we consider the com-
A, more significant changes in the output hyper-para- bined estimation problem, in which both model state
meter and required ensemble size is seen. variables and parameters are estimated simultaneously
The above procedure to quantify the input and out- given erroneous forcing data and observations. One ap-
put perturbation factors and their impact on ensemble proach for combined estimation is provided by joint
generation has been carried out separately from the estimation where state and parameter vectors are con-
uncertainty associated with state and parameter esti- catenated into a single joint state vector (state augmen-
mates. In the following section, an interactive procedure tation) [3,4,29,36,37,40]. The drawback of such a
in the context of EnKF to provide a probabilistic esti- strategy is that, by increasing the number of unknown
mate of states and parameters is developed. model states and parameters, the degree of freedom in
the system increases and makes the estimation unstable
and intractable especially in the nonlinear dynamic
3. Dual state–parameter estimation with EnKF models. An alternative approach to joint estimation is
dual estimation, designed as two interactive filters moti-
Although the parameters of a hydrologic model can vated either by the need to estimate state from the model
be estimated in a batch-processing scheme, there is no (parameters) or by the need to estimate the model from
142 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147

state. Examples of dual estimation are the dual extended meter sampling is the over-dispersion of parameter sam-
Kalman filter (EKF) developed by Refs. [28,45] for esti- ples and loss of information between time points when
mating neural networks model signal (state) and weights the parameters are considered to be fixed. In other
(parameter). The dual EKF requires separate state-space words, loss of information results in posterior distribu-
representation for the state and parameters through two tion of parameters that are too diffuse when compared
parallel filters. To extend the applicability of the EnKF to the posteriors of fixed parameters [21]. One remedy
to simultaneous state–parameter estimation, we need to to this problem is the Kernel smoothing of parameter
treat the parameters similar to state variables with a dif- samples introduced by West [46]. Suppose that the hit ,
ference that parameter evolution is set up artificially, and their weights wit , i = 1, . . . , n denote a random mea-
i.e., it is assumed that the parameters follow a random sure fhit ; wit g to characterize the discrete Monte Carlo
walk; therefore, in EnKF, parameter samples can be approximation to posterior density of parameters as
made as follows: P(ht+1jy1, . . . , yt) with mean
ht and the variance
matrix Vt. West [46] explained that the smooth kernel
hi iþ i
tþ1 ¼ ht þ st ; sit
N ð0; Rht Þ ð19Þ
density form of the Monte Carlo approximation to
Using the artificially forecasted parameters and forcing P(ht+1jy1, . . . , yt) could be a mixture of Gausian densities
data replicates, a model state ensemble and predictions with mean mit and variance h2Vt, weighted by sample
are made, respectively: weights wit :
xi iþ i i X
n
tþ1 ¼ f ðxt ; ut ; htþ1 Þ ð20Þ
P ðhtþ1 j y 1 ; . . . ; y t Þ
wit N ðhtþ1 j mit ; h2 V t Þ ð28Þ
i i¼1
^y itþ1 ¼ hðxi
tþ1 ; htþ1 Þ ð21Þ
where, h is the smoothing or variance reduction para-
Updating the parameter ensemble members is made meter. The standard kernel method considers that
according to the standard Kalman filter equation: mit ¼ hit , however, this results to an over-dispersed kernel
hiþ i h i
y itþ1 Þ density relative to posterior samples. West [46] and later
tþ1 ¼ htþ1 þ K tþ1 ðy tþ1  ^ ð22Þ
Liu [21] suggested that this flaw can be corrected by
here, K htþ1
is the Kalman gain for correcting the param- shrinkage of kernel locations:
eter trajectories and is obtained by: pffiffiffiffiffiffiffiffiffiffiffiffiffi
1 mit ¼ ahit þ ð1  aÞ
ht with a ¼ 1  h2 ð29Þ
K htþ1 ¼ Rhy yy y
tþ1 ½Rtþ1 þ Rtþ1  ð23Þ
If the Monte Carlo approximation to posterior density
here Rhy
tþ1is the cross covariance of parameter ensemble P(ht+1jy1, . . . , yt) has mean
ht and variance matrix Vt,
and prediction ensemble. the parameter evolution in Eq. (19) with independent
Now using the updated parameter, the new model
perturbation sit , which was assumed to be independent
state trajectories (state forecasts) and prediction trajec-
of ht, has the correct mean
ht but variance matrix
tories are generated:
V t þ Rht . This problem was reported as loss of informa-
iþ
xi iþ i
tþ1 ¼ f ðxt ; ut ; htþ1 Þ ð24Þ tion in [21,46]. Therefore Liu [21] showed that the artifi-
cial evolution needs to be modified by considering the
iþ
^y itþ1 ¼ hðxi
tþ1 ; htþ1 Þ ð25Þ correlations between ht and the perturbation sit . By
doing so the conditional evolution density of parameters
Model states ensemble is similarly updated as follows:
is written as follows:
xiþ i x i
y itþ1 Þ
tþ1 ¼ xtþ1 þ K tþ1 ðy tþ1  ^ ð26Þ
P ðhtþ1 j ht Þ
N ðhitþ1 j ahit þ ð1  aÞ
h; h2 V t Þ ð30Þ
where K xtþ1
is the Kalman gain for correcting the state
where, a ¼ 3d1
2d
and d is a factor in (0 1], which is typi-
trajectories and is obtained by:
cally around 0.95–0.99. For more information about
K xtþ1 ¼ Rxy yy y 1 the derivation of the conditional density and the para-
tþ1 ½Rtþ1 þ Rtþ1  ð27Þ
meters associated with it, please refer to Liu [21].
where, K xtþ1 is the cross covariance of states ensemble A dual state–parameter estimation flowchart using
and prediction ensemble. EnKF with kernel smoothing of parameters is shown
in Fig. 5.
3.1. Kernel smoothing of parameter samples

The artificial parameter evolution at each time step 4. Streamflow forecasting by applying dual EnKF on
by adding small random perturbation provides a new HyMOD model
parameter set in simulation and has been performed
by many authors, from which [36] is one of the earliest The applicability and usefulness of the dual EnKF on
in hydrologic application. The drawback of such para- state–parameter estimation of HyMOD for one-day
 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147 143

Fig. 5. Dual state–parameter estimation flowchart using the ensemble Kalman filter by kernel smoothing of parameters.

ahead streamflow forecasting in the Leaf River basin and 320. The starting point in the parameter space is
were investigated. The system was initialized by defining sampled from the uniform distribution, then forecast-
the prior uncertainty range associated with the parame- updates of all state variables and parameters are made
ters in Table 1. With regard to state variables, storages simultaneously using the dual estimation. Owing to the
in the linear tanks have no threshold, and storage in stochastic-dynamic nature of the problem, it is required
nonlinear tank limited to the minimum and maximum to run the model for the sufficient number of parameter
bound defined for the nonlinear tank parameters which samples to examine the time evolution of predictive
is found from model formulation [2] to be between 60 uncertainties.

Table 1
Prior uncertainty associated with parameters in HyMOD model
Parameter Description Minimum Maximum
Rq Residence time for quick-flow tanks 0.20 0.70
Rs Residence time for slow-flow tank 0.01 0.10
a Partitioning factor between tanks 0.60 0.99
bexp Spatial variability of soil moisture capacity 0.10 1.50
Cmax Maximum storage capacity of watershed 150.00 350.00
144 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147

For this experiment, 500 starting points were sampled Table 2

in the parameter space, and dual ensemble filtering with Expected value of the parameter sets in HyMOD estimated by different
algorithms
ensemble size of 50 (resulted from the tuning of hyper-
parameters) was performed from each starting point. Parameter Estimation algorithm
Fig. 6 displays the time evolution of HyMOD parame- SCE-UA SCEM-UA BaRE Dual EnKF
ters for the water years of 1950–1953. Shaded areas in Rq 0.465 0.46 0.478 0.463
this figure represent the evolution of confidence intervals Rs 0.01 0.0027 0.0295 0.0127
obtained from 500 trajectories, while each trajectory is a 0.861 0.84 0.667 0.82
bexp 0.251 0.38 0.15 0.406
the mean of 50 ensemble members. As seen all parame- Cmax 282.51 257.33 181.91 258.34
ters converge smoothly to the certain region in parame-
ter space where the uncertainty bounds stabilize.
It also appears from Fig. 6 that quick-flow tank
parameter Rq is the most identifiable parameter by column is a Bayesian recursive estimation technique,
showing the fastest convergence with a minimum degree which also investigates the time evolution of parameter
of uncertainty comparing to the others. In contrast, the probabilities.
maximum storage capacity of the watershed displayed Although the dual EnkF result is comparable with
by Cmax is less identifiable than the others and shows other algorithms and, to a higher extent, with the batch
the slowest convergence. It is apparent from the model calibration schemes, it has some advantages over the
configuration (Fig. 2) that Cmax and bexp are in high above models such as:
interaction such that one compensates for another, that
is, the uncertainty in Cmax can be compensated with suit- (1) The capability of dual EnKF in interactive parame-
able degree of convexity or concavity of the nonlinear ter and state estimation in which the updated
reservoir represented by bexp to provide the most accu- parameters at each time step are used to update
rate excess rainfall possible for parallel tanks. As a com- the model state. As an example, the estimation of
parison, the expected values of parameters obtained one of the state variables as the storage of nonlinear
using different algorithms are shown in Table 2. The last tank, conceptually representing the watershed soil
three columns in Table 2 give the algorithms developed moisture storage, is demonstrated in Fig. 7. This is
at University of Arizona from which the SCE-UA [7,8] an unobservable quantity and the accuracy of its
and SCEM-UA [42] are the global optimization algo- estimation is translated through the accuracy of
rithms suitable for the batch calibration of hydrologic streamflow forecasting as the observable and pre-
model parameters. BaRE algorithm [25,35] in the last dictable variable.

Fig. 6. Time evolution of the HyMOD parameters for 3 years of dual ensemble filtering (water years of 1950–1953). Shaded areas correspond to 95,
75, 68 and 10 percentile confidence intervals.
 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147 145

Precipitation [mm/d]
20

40

60

80
0 30 60 90 120 150 180 210 240 270 300 330 360
Time [Day]
Soil Moisture Storage [mm]

250

200

150

100

50

0
30 60 90 120 150 180 210 240 270 300 330 360
Time [Day]

Fig. 7. Soil moisture storage variation (storage in the nonlinear tank of the HyMOD) in the Leaf River watershed (water year 1952–1953) with 95
percentile confidence interval. Dash line is the mean value of the ensemble.

(2) In addition to parameter uncertainty as the only in the recession limb of hydrograph which can be seen
source of uncertainty affecting the performance of in lower subplot in Fig. 8. This persistent bias can be ex-
the estimation considered in the above-mentioned plained as the role of model structural error that has not
procedures, the dual EnKF undertakes other been considered in this study which could be included in
sources of errors. future studies.
(3) Dual EnKF as a recursive procedure does not
require keeping all of the data in storage; thus, by
availability of observation at any time, the variables 5. Summary and conclusion
in the system can be adjusted for better conformity
with the observation. Hydrologic models are still far from perfect, and
(4) Examining the stable uncertainty bounds in Fig. 6 hydrologists need to put the models in better compliance
determined by dual EnKF reveals that the parame- with observations prior to use in forecasting. Batch cal-
ters do not converge to single points and, therefore, ibration procedures as the most commonly used tech-
degeneracy of parameter samples does not happen. niques in hydrology, and even the recursive calibration
This is the drawback that BaRE algorithm [25,35] schemes concern primarily the estimation of parameters
suffers from, that is the parameter uncertainty and the identification of uncertainties associated with
bounds vanish in a short amount of time after the them. However, more general algorithms that account
recursive estimation starts. for the simultaneous interactions of model states and
parameters are encouraging while different sources of er-
In keeping up with previous studies [25,42], the per- rors are considered. In this study, an integrated and
formance of the dual EnKF in streamflow forecasting algorithmic framework for dual state–parameter estima-
is demonstrated in Fig. 8. The hydrograph simulation tion using EnKF was presented, which leads to the
is the result of sequential dual estimation by assimilating ensemble streamflow forecasting. Perturbation of input
streamflow everyday for the water year of 1952–1953 in and output to generate and modify the ensemble of
the Leaf River basin using the HyMOD. The forecasting model variables and to determine the ensemble size are
results with 95% confidence intervals are derived from key features of the EnKF, and identification of the mag-
model output ensemble at each time step. As seen the nitude of perturbation in a systematic framework is de-
ensemble mean of daily streamflow forecasting is in very sired and elaborated in this study.
good agreement with the observations, implying that the In the hydrologic model (HyMOD) used for this
dual EnKF is a reliable and effective approach for study, the analysis certainly indicated the feasibility of
streamflow forecasting. The uncertainty bound also cov- sequential ensemble filtering that incorporates parame-
ers the observation in a consistent manner, despite the ters of the model in addition to state variables. In es-
small negative bias in the rising limb and positive bias sence, the dual EnKF use the ensemble of model
146 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147

Precipitation [mm/d]
0
(A)
20
40
60
80
0 30 60 90 120 150 180 210 240 270 300 330 360
Time [Day]

Streamflow [m 3/s] 400 (B)

300

200

100

0
0 30 60 90 120 150 180 210 240 270 300 330 360
Residual [m /s]

40
(C)
3

20
0
-20
-40
0 30 60 90 120 150 180 210 240 270 300 330 360
Time [Day]

Fig. 8. Ensemble streamflow forecasting by dual state–parameter estimation of HyMOD for the Leaf River watershed (water year 1952–1953), (A)
forcing data (precipitation), (B) daily assimilation of observed streamflow (solid dots), streamflow forecast (shaded area with 95% uncertainty
bound), and mean value of the ensemble streamflow forecast (solid line), and (C) model residuals (observed––ensemble mean forecast).

trajectories in an interactive parameter–state space Finally, it is safe to say that dual EnKF provides a
and provides the confidence interval of parameter–state more flexible approach compared to other estimation
estimation. Because the traditional random walk of procedures explained in Section 4. It is a suitable tech-
parameters may result in over-dispersion/information nique for nonlinear models and together with kernel
loss and consequently the collapsing the parameter var- smoothing of parameter samples is a robust and effective
iance, the kernel smoothing of parameters can be algorithm that can tackle input, output and parameter
employed. uncertainties properly.
Using the dual technique, the time evolution of
parameter uncertainties and one of the state variables
were demonstrated. The one-day ahead streamflow fore-
casting in the Leaf River watershed using the estimated Acknowledgments
states and parameters was performed, and result seemed
to be very consistent with observation. From a filtering Partial financial support for this research was pro-
point of view, this study offers the following features vided by SAHRA (center for Sustainability of semi-Arid
which do not exist in nonensemble methods: Hydrology and Riparian Areas) under the NSF-STC,
Agreement EAR-9876800, and by the Hydrologic Labo-
(1) It allows incorporating a wide range of uncertainties ratory of the National Weather Service under Agree-
to the model. ment NA87WHO582.
(2) It provides a quantitative basis for probabilistic rep-
resentation of estimates.
(3) The measure of confidence interval becomes possi- References
ble using such ensemble technique.
(4) It provides a flexible and reliable strategy to deal [1] Anderson JL. An ensemble adjustment filter for data assimilation.
Monthly Weather Rev 2001;129:2884–903.
with nonlinear dynamic models; the problem that [2] Boyle DP. Multicriteria calibration of hydrological models. PhD
could not be overcome entirely by even extended Dissertation, Dep of Hydrol and Water Resour, Univ of Arizona,
Kalman filter. Tucson; 2000.
(5) It employs the kernel smoothing procedure that pro- [3] Bras RL, Restrepo-Posada P. Real time automatic parameter
tects the parameter sampling from over-dispersion calibration in conceptual runoff forecasting models. In: Proceed-
ings of the Third International Symposium on Stochastic
or loss of information while doing the forecasting. Hydraulics, 1980. p. 61–70.
The concept of random walk of parameters was [4] Bras RL, Rodriguez-Iturbe I. Random functions and hydrol-
always suffering from this drawback. ogy. Reading, MA, USA: Addison Wesley; 1985. 559 pp.
 H. Moradkhani et al. / Advances in Water Resources 28 (2005) 135–147 147

[5] Burgers GP, Van Leeuwen J, Evensen G. Analysis scheme in the [29] Reichle RH, McLaughlin D, Entekhabi D. Hydrologic data
ensemble Kalman filter. Monthly Weather Rev 1998;126:1719–24. assimilation with the ensemble Kalman filter. Monthly Weather
[6] Cohn SE. An introduction to estimation theory. J Meteo Soc Jpn Rev 2002;130:103–14.
1997;75(1B):147–1787. [30] Reichle RH, Walker JP, Koster RD, Houser PR. Extended versus
[7] Duan Q, Sorooshian S, Gupta VK. Effective and efficient global ensemble Kalman filtering for land data assimilation. J Hydromet
optimization for conceptual rainfall-runoff models. Water Resour 2002;3:728–40.
Res 1992;28(4):1015–31. [31] Sorooshian S, Dracup JA. Stochastic parameter estimation
[8] Duan Q, Gupta VK, Sorooshian S. A shuffled complex evolution procedures for hydrologic rainfall-runoff models: correlated
approach for effective and efficient global minimization. J Optim and heteroscedastic error cases. Water Resour Res 1980;16(2):
Theory Appl 1993;76(3):501–21. 430–42.
[9] Evensen G. Sequential data assimilation with a nonlinear quasi- [32] Sorooshian S, Duan Q, Gupta VK. Calibration of rainfall-runoff
geostrophic model using Monte Carlo methods to forecast error models: application of global optimization to the soil moisture
statistics. J Geophys Res 1994;99:10143–62. accounting model. Water Resour Res 1993;29(4):1185–94.
[10] Gelb A, Kasper JF, Nash RA, Price CF, Sutherland AA. Applied [33] Spear RC, Hornberger GM. Eutrophication in Peel Inlet, 2,
optimal estimation. The MIT Press; 1974. 374 pp. Identification of critical uncertainties via generalized sensitivity
[11] Gupta VK, Sorooshian S, Yapo PO. Towards improved calibra- analysis. Water Resour Res 1980;14:43–9.
tion of hydrological models: multiple and noncomensurable [34] Spear RC, Grieb TM, Shang N. Parameter uncertainty and
measures of information. Water Resour Res 1998;34:751–63. interaction in complex environmental models. Water Resour Res
[12] Hamill TM. Notes and correspondence on ‘‘Interpretation of rank 1994;30(11):3159–69.
histograms for verifying ensemble forecasts’’. Monthly Weather [35] Thiemann M, Trosset M, Gupta H, Sorooshian S. Bayesian
Rev 2001;129:550–60. recursive parameter estimation for hydrologic models. Water
[13] Jazwinski AH. Stochastic processes and filtering theory. San Resour Res 2001;37(10):2521–35.
Diego: Academic; 1970. [36] Todini E, Szollosi-Nagy A, Wood EF. Adaptive state–parameter
[14] Kalman R. New approach to linear filtering and prediction estimation algorithm for real time hydrologic forecasting; a case
problems. Trans AMSE J Basin Eng 1960;82D:35–45. study. In: IISA/WMO Workshop on the Recent Developments in
[15] Kitanidis PK, Bras RL. Adaptive filtering through detection of Real Time Forecasting/Control of Water Resources Systems,
isolated transient errors in rainfall-runoff models. Water Resour Laxemburg (Austria), 1976.
Res 1980;16(4):740–8. [37] Todini, E. Mutually interactive state–parameter (MISP) estima-
[16] Kitanidis PK, Bras RL. Real-time forecasting with conceptual tion, Applications of Kalman filters to hydrology, hydraulics and
hydrologic models, 1. Analysis of uncertainty. Water Resour Res water resources. In: Chiu C-L, editor. Proceedings of AGU
1980;16(6):1025–33. Chapman Conference, 1978.
[17] Kitanidis PK, Bras RL. Real-time forecasting with conceptual [38] Toth Z, Talagrand O, Candille G, Zhu Y. Probability and
hydrologic models, 2. Applications and results. Water Resour Res ensemble forecasts. In: Jolliffe IT, Stephenson DB, editors.
1980;16(6):1034–44. Environmental forecast verification: a practitionerÕs guide in
[18] Kuczera G. On validity of first-order prediction limits for atmospheric science, John Wiley, in press.
conceptual hydrological models. J Hydrol 1998;103:229–47. [39] Troch PA, Paniconi C, McLaughlin D. Catchment-scale hydro-
[19] Kuczera G, Mroczkowski M. Assessment of hydrological para- logical modeling and data assimilation. Adv Water Res
meter uncertainty and the worth of multiresponse data. Water 2003;26:131–5.
Resour Res 1998;34(6):1481–9. [40] Young PC. Advances in real-time flood forecasting. Philos Trans
[20] Kuczera G, Parent E. Monte Carlo assessment of parameter Roy Soc Lond 2002;360:1433–50.
uncertainty in conceptual catchment models: The Metropolis [41] Van Leeuwen PJ. Comment on data assimilation using an
algorithm. J Hydrol 1998;211:69–85. ensemble Kalman filter technique. Monthly Weather Rev
[21] Liu F. Bayesian time series: analysis methods using simulation 1999;127:1374–7.
based computations. PhD Dissertation, Institutes of Statistics and [42] Vrugt JA, Gupta HV, Bouten W, Sorooshian S. A shuffled
Decision Sciences, Duke University, 2000. complex evolution metropolis algorithm for optimization and
[22] McLaughlin D. Recent developments in hydrologic data assim- uncertainty assessment of hydrologic model parameters. Water
ilation. Rev Geophys 1995(Suppl):977–84. Resour Res 2003;39(8):1–14.
[23] McLaughlin D. An integrated approach to hydrologic data [43] Vrugt JA, Gupta HV, Sorooshian S, Wagener T, Bouten W.
assimilation: interpolation, smoothing, and filtering. Adv Water Confronting parameter uncertainty in hydrologic modeling:
Res 2002;25:1275–86. application of the SCEM-UA algorithm to the Sacramento Soil
[24] Miller RN, Ghil M, Guathiez F. Advanced data assimilation in Moisture Accounting Model. J Hydrol, in review.
strongly nonlinear dynamical systems. J Atmos Sci 1994;51(8): [44] Wagener T, Boyle DP, Lees MJ, Wheater HS, Gupta HV,
1037–56. Sorooshian S. A framework for development and application
[25] Misirli F, Gupta HV, Sorooshian S, Thiemann M. Bayesian of hydrological models. Hydrol Earth Syst Sci 2001;5(1):
recursive estimation of parameter and output uncertainty for 13–26.
watershed models: Calibration of Watershed Models. In: Duan Q, [45] Wan EA, Nelson TA. Dual Kalman filtering methods for
et al., editor. Water Sci Appl Ser, vol 6, 2003. p. 113–24. nonlinear prediction, estimation and smoothing. Adv Neural
[26] Moore RJ. The probability-distributed principle and runoff Informat Process Syst 1997;9.
production at point and basin scales. Hydrol Sci J 1985;30(2): [46] West M. Mixture models, Monte Carlo, Bayesian updating and
273–97. dynamic models. Comput Sci Statist 1993;24:325–33.
[27] Murphy JM. The impact of ensemble forecast on predictability. [47] Whitaker JS, Hamill TM. Ensemble data assimilation without
Quart J Roy Meteor Soc 1988;114:89–125. perturbed observations. Monthly Weather Rev 2002;130:
[28] Nelson AT. Nonlinear estimation and modeling of noisy time 1913–24.
series by dual Kalman filtering methods. PhD Dissertation, [48] Wilks DS. Forecast verification, Statistical methods in the
Oregon Graduate Institute, 1999. atmospheric sciences. Academic; 1995. p. 233–83.