PUBLICATIONS

Water Resources Research

RESEARCH ARTICLE Hydrologic modeling in dynamic catchments: A data
10.1002/2015WR017192
assimilation approach
Key Points: S. Pathiraja1, L. Marshall1, A. Sharma1, and H. Moradkhani2
 A data assimilation based method for
modeling dynamic catchments is 1
investigated School of Civil and Environmental Engineering, University of New South Wales, Sydney, New South Wales, Australia,
2
 Prior parameters are critical in Department of Civil and Environmental Engineering, Portland State University, Portland, Oregon, USA
determining performance of data
assimilation
 Improved approach for time varying Abstract The transferability of conceptual hydrologic models in time is often limited by both their struc-
parameter estimation presented
tural deficiencies and adopted parameterizations. Adopting a stationary set of model parameters ignores
biases introduced by the data used to derive them, as well as any future changes to catchment conditions.
Supporting Information:
Supporting Information S1
Although time invariance of model parameters is one of the hallmarks of a high quality hydrologic model,
very few (if any) models can achieve this due to their inherent limitations. It is therefore proposed to con-
Correspondence to:
sider parameters as potentially time varying quantities, which can evolve according to signals in hydrologic
S. Pathiraja, observations. In this paper, we investigate the potential for Data Assimilation (DA) to detect known tempo-
[email protected] ral patterns in model parameters from streamflow observations. It is shown that the success of the DA algo-
rithm is strongly dependent on the method used to generate background (or prior) parameter ensembles
Citation: (also referred to as the parameter evolution model). A range of traditional parameter evolution techniques
Pathiraja, S., L. Marshall, A. Sharma, and
H. Moradkhani (2016), Hydrologic
are considered and found to be problematic when multiple parameters with complex time variations are
modeling in dynamic catchments: A estimated simultaneously. Two alternative methods are proposed, the first is a Multilayer approach that
data assimilation approach, Water uses the EnKF to estimate hyperparameters of the temporal structure, based on apriori knowledge of the
Resour. Res., 52, 3350–3372,
doi:10.1002/2015WR017192.
form of nonstationarity. The second is a Locally Linear approach that uses local linear estimation and
requires no assumptions of the form of parameter nonstationarity. Both are shown to provide superior
Received 6 MAR 2015 results in a range of synthetic case studies, when compared to traditional parameter evolution techniques.
Accepted 16 MAR 2016
Accepted article online 21 MAR 2016
Published online 4 MAY 2016

1. Introduction
All hydrologic models are simplified representations of the complex spatially and temporally varying
rainfall-runoff conversion processes. Consequently, significant emphasis is often placed on the model
parameters to account for any unresolved physical processes, scale factors and unknown physiographic
attributes [Gupta et al., 1998; Bardossy, 2007]. It is customary to batch calibrate a hydrologic model against a
particular data period, with the end goal of deriving a globally optimum parameter vector [e.g., Duan et al.,
1994] or stationary distribution of parameters [e.g., Beven and Freer, 2001]. This approach is problematic for
a number of reasons. Firstly, the optimal parameter set(s) are no longer suitable if the catchment conditions
change. Catchments are fundamentally dynamic in nature, often undergoing changes due to natural proc-
esses (e.g., bushfire, changes in surface soil cover due to erosion) or human intervention (e.g., urbanization,
diversions etc.). The effect of catchment dynamics on hydrologic variables such as streamflow has been
widely documented (see for instance Kuczera [1987] and Scott and Van Wyk [1990] in regards to changes in
the runoff regime after bushfires, and Siriwardena et al. [2006] in relation to afforestation/deforestation). Sec-
ond, model parameters often display a strong dependence on the calibration data used to derive them, due
largely to the empirical nature of many conceptual hydrologic models [Ebtehaj et al., 2010]. In particular,
several studies have noted the dependence between model parameters and the dominant climatic regime
of the calibration period, even when the catchment itself has remained relatively stationary [Sorooshian
et al., 1983; Choi and Beven, 2007; Wu and Johnston, 2007]. Coron et al. [2012] also noted the potential for
seasonal variations in parameters, due to changes in the dominant runoff generation mechanisms between
seasons in some catchments. Lastly, the adoption of a time invariant optimal parameter distribution does
not necessarily guarantee consistent model performance even within the calibration period. The chosen
objective function strongly influences the calibration parameters [Thyer et al., 2009; Madsen, 2003; Wang
C 2016. American Geophysical Union.
V et al., 2010; Pathiraja et al., 2012], with emphasis often placed on a particular streamflow characteristic (e.g.,
All Rights Reserved. peak flows). It is near impossible for a time invariant set of model parameters to accurately simulate all

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3350

 Water Resources Research 10.1002/2015WR017192

characteristics of the time series of system states (such as peak flow, base flows, volume of runoff, soil mois-
ture storage fluctuations etc.) [Moussa and Chahinian, 2009; Efstratiadis and Koutsoyiannis, 2010; Westerberg
et al., 2011]. Multiobjective criterion type approaches have shown promise in defining parameter spaces
which provide improved simulations over a range of characteristics, but trade-offs are almost always
required [Madsen, 2003; Efstratiadis and Koutsoyiannis, 2010].
One possible approach to improve the applicability of hydrologic models over long time scales is to allow
the model parameters to evolve with time. Time varying parameter systems have been investigated in fun-
damental systems theory over the last few decades [e.g., Richards, 1983; Schwartz and Ozbay, 1990;
Mohammadpour and Scherer, 2012], the most common being Linear Time Variant systems where the state
space transition matrix is time varying [e.g., Khalil, 1996]. A well-known example of a time varying parameter
system is the variation in aerodynamic coefficients during take-off, cruising and landing for high speed air-
craft [e.g., Tomas-Rodriguez and Banks, 2010]. In hydrologic applications, model parameterizations would
adjust to varying climatic regimes and/or changes in catchment conditions based on signals in the input
and observed data under such a framework. It can be argued that time variations in model parameters are
an artefact of missing system processes. However, it is not always tractable to identify the ‘‘correct’’ system
equations to model such missing processes. Allowing the model parameters to vary provides a mechanism
with which to adjust the model structure to reflect such missing processes, when they become evident in
the observations. Only a few studies have examined strategies for diagnosing parameter nonstationarity in
hydrologic systems. The most common approach in the literature involves split-sample calibration, whereby
the available data set is subdivided and calibration undertaken separately for each period [e.g., Thirel et al.,
2015; Merz et al., 2011; de Vos et al., 2010; Gharari et al., 2013; Seibert and McDonnell, 2010]. Westra et al.
[2014] examined the matter from a different perspective, by assuming one or more parameters vary in time
as a function of selected covariates. Their experiment on the Scott Creek catchment showed an improve-
ment in model predictions when the soil storage parameter in the GR4J model was made to vary based on
seasonal, annual and long term trends. Jeremiah et al. [2013] and Marshall et al. [2006] modeled parameter
variations as a function of the catchment state in a Hierarchical Mixtures of Experts (HME) framework.
Another possible approach, which has rarely been explored, is to extract parameter variation signals from
hydrologic observations using Data Assimilation. Data Assimilation (DA) offers a framework to detect the
potentially time varying nature of model parameters, by adjusting them in real time as observations
become available [Liu and Gupta, 2007; Evensen, 2006]. It allows for the rapid quantification of both model
parameters and prognostic variables in real time using uncertain observations [Evensen, 2006; Liu et al.,
2012; Gelb, 1974]. Prior (a.k.a. background) estimates of a system variable (usually model simulations) are
combined with observation(s), both of which are assumed to contain random errors, to produce an updated
estimate with lower error variance than the background or observation [Evensen, 2006]. The widely used
Kalman Filter (KF) [Kalman, 1960] provides optimal updates for linear dynamical systems with normally dis-
tributed variables and errors [Kalman, 1960; Evensen, 2006]. However, most real world applications do not
satisfy these assumptions. A plethora of DA algorithms have been developed to accommodate less strin-
gent conditions, including Kalman Filter variants (e.g., the Extended Kalman Filter [Jazwinski, 1970], Ensem-
ble Kalman Filter [Evensen, 2006], and the Unscented Kalman Filter [Wan and Van Der Merwe, 2000]) and
Particle Filters (e.g., Particle Filter-SIR [Moradkhani et al., 2005a], Regularised Particle Filter [Musso et al.,
2001], and the Particle Filter-MCMC [Moradkhani et al., 2012]). Although such methods have been widely
applied in the geosciences, they are suboptimal and there is no clear consensus on the most appropriate
methods for use in complex nonlinear systems with nonnormal distributions. The Ensemble Kalman Filter
has been used frequently in hydrologic applications with considerable success, despite its inability to
adequately accommodate nonnormal distributions [see e.g., Komma et al., 2008; Reichle et al., 2002; Weerts
and El Serafy, 2006]. Data Assimilation algorithms in general have predominantly been used in hydrology
for state updating, for instance for characterizing soil moisture [Houser et al., 1998; Walker et al., 2001;
Matgen et al., 2012], rainfall runoff modeling [Aubert et al., 2003; Weerts and El Serafy, 2006] and flood fore-
casting [Vrugt et al., 2006; Noh et al., 2013; Li et al., 2013] (for a detailed review, refer to Liu and Gupta
[2007]). This has in recent years been extended to include parameter estimation within the DA framework,
although it has mainly been limited to deriving stationary parameter distributions [e.g., Maneta and Howitt,
2014; Moradkhani et al., 2005b; Simon and Bertino, 2012; Yang and Delsole, 2009; Aksoy et al., 2006; Xie et al.,
2014; Annan et al., 2005]. One of the few applications of DA for time varying parameter estimation in

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3351

 Water Resources Research 10.1002/2015WR017192

hydrology is the work of Smith et al. [2008]. They examined the time evolution of parameter distributions
from particle filtering as a way of diagnosing model structural inadequacy. Rapid parameter fluctuations in
time were noted, although it is unclear to what extent this is driven by nonuniqueness of parameter solu-
tions and the stochastic nature of the DA algorithm. Vrugt et al. [2013] and Salamon and Feyen [2009] also
applied Particle Filter based algorithms to examine the time evolution of model parameters. However, none
of these studies addressed the question of whether a structured time variation in a parameter can be
detected using DA (e.g., in Vrugt et al. [2013] the true parameters were assumed constant in the synthetic
study using HyMOD). Lin and Beck [2007] used a recursive parameter evolution algorithm to estimate two
state variables and two parameters (one time varying, one time invariant) in a biomass-substrate model.
The algorithm was shown to successfully replicate the temporal pattern of the parameters in the synthetic
case study. However its effectiveness remains to be seen when the dimensionality of time varying parame-
ters is increased and the number of observation variables is reduced.
In this paper, we examine the potential of Data Assimilation for estimating time varying hydrologic model
parameters. It is demonstrated that the parameter evolution model, which is used to generate background
parameter ensembles, is critical in determining the effectiveness of the filter in this regard. A number of
commonly used parameter evolution models are applied within an EnKF framework. They are used to
detect time variations in parameters of the Probability Distributed Model (PDM) [Moore, 2007], a lumped
conceptual hydrologic model. Parameter estimation is undertaken for multiple scenarios with synthetic
data, so as to allow for a robust assessment of algorithm performance. The aim here is to investigate the
efficacy of existing parameter evolution models for a range of parameter temporal patterns. Two alternative
parameter evolution models are proposed which are specifically suited to time varying parameter applica-
tions. The first is a Multilayered approach that uses the EnKF to estimate hyperparameters of the temporal
structure, based on apriori knowledge of the form of nonstationarity. The second is a Locally Linear
approach that uses local linear estimation and requires no assumptions of the form of parameter
nonstationarity.
The remainder of this paper is structured as follows. In section 2, we describe the Joint State-Parameter Esti-
mation framework, including the Dual EnKF algorithm of Moradkhani et al. [2005b] and the various parame-
ter evolution models examined in the study. Section 3 provides details of the experimental setup of the
synthetic study, and the known time variations imposed on model parameters for the various case studies.
Results from all parameter evolution models considered are provided in section 4, along with a discussion
of their efficacy in various scenarios. Finally, we conclude with a summary of the main outcomes of the
study as well as proposed future work.

2. Joint State Parameter Estimation with the EnKF

The Ensemble Kalman Filter (EnKF) is a Monte Carlo application of the well-known Kalman Filter [Evensen,
2006]. In the EnKF, distributions of the system variables are replaced with random samples or an ensemble.
This means that unlike the original Kalman Filter [Kalman, 1960] or the Extended Kalman Filter [Jazwinski,
1970], the model error covariance matrix can be approximated by the sample covariance, instead of being
specified apriori and explicitly propagated forward in time [Evensen, 2006]. The sample covariance of the
system variables provide an estimate of the error covariance, assuming unbiasedness. The main limitation
of the EnKF is its suboptimality for nonlinear dynamics with nonnormally distributed errors [Evensen, 2006].
However several studies have demonstrated the usefulness of the EnKF in a variety of highly nonlinear sys-
tems [see e.g., Komma et al., 2008; Reichle et al., 2002; Weerts and El Serafy, 2006]. For further details on the
EnKF, refer to Evensen [2006].
The EnKF and its variants have been successfully applied for estimating both system state variables and
time invariant model parameters [e.g., Maneta and Howitt, 2014; Moradkhani et al., 2005b; Simon and
Bertino, 2012; Yang and Delsole, 2009; Aksoy et al., 2006; Samuel et al., 2014]. Simultaneous estimation of
states and parameters in a DA framework is traditionally undertaken by one of two methods: (1) augment-
ing the state vector with parameter variables so that updating of states and variables occurs concurrently
[e.g., Smith et al., 2013; Franssen and Kinzelbach, 2008]; or (2) Updating parameters and states separately
through two sequential filters, also known as Dual State – Parameter estimation [Moradkhani et al., 2005b;
L€
u et al., 2011; Leisenring and Moradkhani, 2012]. The parameter update equation in the state augmentation

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3352

 Water Resources Research 10.1002/2015WR017192

and dual updating approaches are equivalent when applied with Kalman filtering methods (refer Appendix
A). However, unlike the augmented approach, dual filtering allows biases in the model simulations of sys-
tem states to be removed (at least partially) prior to state updating. This is important, as the successful
implementation of any DA algorithm is reliant on unbiased priors [Dee, 2005]. We therefore adopt the dual
state – parameter estimation approach in this study.
A general framework for undertaking dual State - Parameter estimation using the EnKF was described in
Moradkhani et al. [2005b], although this focused on estimating a stationary distribution of parameters. Here
a methodology is presented for the estimation of time varying parameters in a dual updating framework.
The algorithm is summarized below.

2.1. Dual EnKF

Suppose a dynamical system at any given time t can be described by a vector of states x t and a vector of
associated model parameters ht . The distributions of system variables at time t are represented by an ensem-
   
ble of states x it i51:n and parameters hit i51:n each with n members. A plus (1) superscript indicates an
updated ensemble whereas a minus (-) superscript indicates the background or prior ensemble. Hereafter, we
use the term ‘‘prior’’ to refer to the background or model simulated ensemble and ‘‘update’’ to refer to the
analysis or updated ensemble after assimilation. The Dual EnKF procedure is undertaken as follows:
1. Generate a prior parameter ensemble. At any given time t, begin by generating a prior parameter
ensemble:
 i1 
hi2
t11 5g ht ; z for i51 : n (1)

where g is a parameter evolution model to be chosen by the user, with input parameters z. Parameter
evolution is discussed in more detail in section 2.2.
 
2. Incorporate observation and forcing uncertainty. A set of perturbed observations y it11 i51:n is generated
using the measurement error characteristics. Here we adopt the method of Houtekamer and Mitchell
[1998] and Burgers et al. [1998] for incorporating observation uncertainty into the analysis ensemble. Sto-
chastic perturbations are applied to the observation vector to produce an ensemble of observations using
the assumed observation error statistics:
y it11 5y ot11 1 it11 for i51 : n (2)
 o o

it11  N 0; Ryt11 y
(3)
o o
where y ot11 denotes the raw observation and Ryt11y
denotes the observation error covariance matrix. A simi-
lar approach can be adopted to account for forcing error (refer Moradkhani et al. [2005b]) to generate a set
 
of perturbed forcings uit11 i51:n . Zero forcing error is assumed in this study, as the focus is on the rainfall
to runoff conversion processes and parameter adjustments to compensate for forcing error are undesired.
3. Generate simulated observations using the prior parameters. The updated parameter ensemble is deter-
mined by first calculating simulated observations using the best available estimate of model states and
prior parameters:
 
x^ it11 5f x i1 i2 i
t ; ht11 ; ut11 for i51 : n (4)
i  i i2

y^ t11 5h x^ t11 ; ht11 for i51 : n (5)

where f is the set of model equations, h is an observation operator that converts model states to observed
variables, y^ it11 indicates the ith ensemble member of the model simulation of the observed variable using
the prior parameters, and uit11 is a vector of forcing inputs (e.g., rainfall and evapotranspiration).
4. Perform the Kalman update of the parameters. The prior parameter ensemble is then updated using the
Kalman update equation and the covariance between parameters and simulated observations:
 i 
hi1 i2 h
t11 5ht11 1K t11 y t11 2 y ^ it11 for i51 : n (6)
h i21
^y ^y yo yo
Kht11 5 Rh^y
t11 Rt11 1 Rt11 (7)
 
where Rh^ y
t11 is a matrix of the cross covariance between parameters hi2 t11 i51:n and simulated observed
 i  ^ ^y
variables y^ t11 i51:n and Ryt11 is the covariance matrix of the simulated observations.

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3353

 Water Resources Research 10.1002/2015WR017192

 
5. Generate simulated observations using the updated parameters. The prior state ensemble x i2
t11 is
 i1  i51:n
then generated using the updated model parameter ensemble ht11 i51:n and the model equations:
 i1 i1 
x i2
t11 5f x t ; ht11 ; ut11
i
for i51 : n (8)
i
 i2 i1

y~ t11 5h x t11 ; ht11 for i51 : n (9)

where y~ it11 indicates the simulated observed variable using the updated parameters.
6. Perform the Kalman update of the states. The Kalman equation for correlated measurement and process
noise is used to update states. This is because the standard Kalman equation (equations (6) and (7))
assumes that the errors in the prior and observation are independent. This assumption is no longer valid
in the dual updating approach, as the simulated states have been generated from parameters that were
updated using observations from the same time step. Therefore, updating is undertaken taking into
account the potential correlation between observation and prior noise:
 i 
x i1 i2 x
~ it11
t11 5x t11 1Kt11 y t11 2 y for i51 : n (10)

h i   21
o ~ ~y e~y y o e~y y o T o o
x~y
Kxt11 5 Rt11 1 Ret11
xy
Ryt11 1Rt11 1 Rt11 1 Ryt11
y
(11)

ex it11 5 x i2 ^ it11 ; e~y it11 5 y~ it11 2 y^ it11

t11 2 x (12)
 
where Rx~ y
is a matrix of the cross covariance between states x i2 and simulated observed varia-
 i t11 o
t11 i51:n
 
bles y~ t11 i51:n from Step 5; Ret11 xy
represents the covariance between ex it11 i51:n and the observations;
e~y y o  
Rt11 represents the covariance between the e~y it11 i51:n and the observations; and ðÞT represents the
transpose operator. Note that only the Kalman gain matrix is adjusted, the form of the Kalman update
equation is the same as the standard Kalman equation. For further details on equations (10)–(12), refer to
Appendix B.

2.2. Existing Parameter Evolution Models

The parameter evolution model (equation (1)) is required to generate a suitable prior parameter ensemble
at each time. In the case of time invariant parameters, a logical choice for the prior parameter ensemble,
 i2   
ht11 i51:n would be the updated parameter ensemble from the previous time step, hi1 t i51:n
. However,
this approach can result in a steadily reducing sample variance, since the variance of an update is always
smaller than its background. If the background ensemble variance is small compared to the observation
error variances, the updated ensemble will be dominated by the model simulations, which are seen to have
high confidence because of their small error variance. This slow loss of sensitivity to observations as a result
of reduced background variance is known as filter divergence [Anderson, 2007]. A common method of deal-
ing with this issue is to introduce stochastic noise into the background parameter ensemble [Gordon et al.,
1993]. Here we examine the performance of three existing parameter perturbation techniques when
applied to estimating nonstationary parameters. These are:
1. Standard Kernel Smoother (SKS) [West, 1993]. In this approach, individual ensemble members are drawn
from a truncated multivariate normal distribution with heteroscedastic covariance:
 i1 2 h 
hi2
t11  TMVN ht ; s Rt (13)

where Rht is the covariance matrix of the updated parameter ensemble at time t and s2 is a smoothing
 
parameter which must be tuned by the user. The resulting prior parameter distribution, p h2
t11 is a mix-
 i1 
ture of normals with variance greater than the previous updated ensemble, ht i51:n [Liu and West,
2001]. For this study, the SKS has been modified so that parameters are instead sampled from independ-
ent univarate truncated normal distributions, since parameters are forced to vary independently in the
synthetic case study (refer section 3). The bounded nature of the PDM model parameters necessitates the
use of a truncated normal distribution.
2. Kernel Smoother with location shrinkage (KSLS) [Liu and West, 2001]. This technique has been used in a
number of studies [see e.g., Maneta and Howitt, 2014; Moradkhani et al., 2005b; Xie et al., 2014]. As stated
above and in Liu and West [2001], the Standard Kernel Smoother produces a prior ensemble with
increased variance compared to the updated ensemble from the previous time step. Liu and West [2001]

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3354

 Water Resources Research 10.1002/2015WR017192

argued that this increased variance would result in a loss of information over time. They suggested that
the mean of the Standard Kernel Smoother should be adjusted according to equation (14), so that the
posterior variance from the previous time step would be maintained. For further details of the kernel
smoother with location shrinkage, refer to [Liu and West, 2001].
 1

hi2 i1
t11  TMVN aht 1ð12aÞh t ; h Rt
2 h
for i51 :: n (14)

1  1 h
where h2 512a2 ; a5 3d212d ; d 2 5 ; 1 h t is the updated parameter ensemble mean vector at time t, Rt
is the covariance matrix of the updated parameter ensemble at time t and d is a scalar smoothing parame-
ter, typically between 0.95 and 0.99.
3. Homoscedastic Kernel Smoother (HKS) [e.g., Smith et al., 2008]. This method is similar to the Standard Kernel
Smoother, with the important distinction that the covariance of the sampling distribution is fixed in time.
The parameters are evolved in time using independent truncated normal distributions with standard devi-
ations fixed in time and equal to a fraction of the feasible parameter range:
 i1 
hi2
t11  TMVN ht ; Rh (15)
0 1
ðv:Dh1 Þ2 0 ... 0
B .. .. C
B . ⯗ C
where Rh 5 B 0 . C
B .. .. C
@ ⯗ . . 0 A
2
0 ... 0 ðv:Dhm Þ
Dhj is the width of the feasible parameter range for the jth parameter variable and v is a user defined frac-
tion of the feasible parameter range (between 0 and 1, Smith et al. [2008] adopted 0.2 in their study.)

2.3. A Multilayer EnKF Approach

Here we propose an alternative parameter evolution approach that utilizes apriori knowledge of the form of
parameter nonstationarity. Just as prior states (ie. the model simulations) are determined from model equa-
tions, improved prior parameters could be derived by evolving parameters through a persistence model
instead of an arbitrary random walk. In this approach, the functional form of parameter variation with time
is assumed to be known and incorporated into the EnKF. For instance, suppose a parameter h varies linearly
with time, its functional form would be hðt Þ5 hðt21Þ1 a where a is a time invariant hyperparameter to be
estimated. The dual layer filter described in section 2.1 is extended to three sequential filters, so that the
hyperparameters that describe the time evolution of parameters are estimated through the first level filter.
This is followed by parameter and state estimation in the second and third level filters respectively. This
Multilayer method (ML) is undertaken as follows:
1. Apply the Standard Kernel Smoothing approach (SKS) to generate the background hyperparameter
 
ensemble, wi2
t11 i51:n which is assumed time invariant (refer equation (16)). Note we propose using the
SKS because of the increased susceptibility of the KSLS to filter divergence, which can produce biased
ensemble means following convergence.  
2 w
wi2 i1
t11  TMVN wt ; s Rt for i51 : n (16)
i
2. Estimate parameters h t11 using the background hyperparameters wi2 t11 and assumed parameter persist-
ence equations l: i  

h t11 5l hi1 i2
for i51 : n (17)
t ; wt11

i
3. Calculate model states x it11 using the model equations f , parameters h t11 and forcing data uit11 . Deter-
mine simulated observations y it11 using the observation operator h, and model states x it11 and parame-
i
ters h t11 :
 
x it11 5f x i1 i i
t ; h t11 ; ut11 for i51 : n (18)
 i

y it11 5h x it11 ; h t11 for i51 : n (19)

4. Update the hyperparameters wi2 i

t11 using the perturbed observations y t11 (as discussed in section 2.1) and
i
the simulated observations y t11 in the standard Kalman update equation:

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3355

 Water Resources Research 10.1002/2015WR017192

w  i 
wi1 i2
t11 5wt11 1K t11 y t11 2 y  it11 for i51 : n (20)
h o
i21
wy  y
K wt11 5 Rt11 Ryt11 1 Ryt11 (21)
 
where Rwy
t11 is a matrix of the cross covariance between hyperparameters wi2 t11 i51:n and simulated
 i   y
observed variables y t11 i51:n ; and Ryt11 is the covariance matrix of the simulated observations.
5. Generate the background parameters hi2 i1
t11 using the updated hyperparameters wt11 and the parameter
persistence equations l:  i1 
hi2
t11 5l ht ; wt11
i1
for i51 : n (22)
 i 
6. Generate model simulations of the observed variables y^ t11 i51:n using the background parameter
 
ensemble, hi2 t11 i51:n as per Step 2 in section 2.1.
7. Update the model parameters using the Kalman update equation for correlated process and measure-
ment noise (equations (23) and (24)). Equations (23) and (24) are used in place of equations (6) and (7) as
the errors in the simulated observed variable and observations are no longer independent.
 i 
hi1 i2 h
t11 5ht11 1K t11 y t11 2 y ^ it11 for i51 : n (23)
h i   21
eh y o ^ ^y e^y y o e^y y o T o o
Kht11 5 Rh^ y
t11 1 Rt11 Ryt11 1Rt11 1 Rt11 1 Ryt11
y
(24)

eh it11 5 hi2 i i
y it11 2 y it11
t11 2 h t11 ; e^y t11 5^ (25)
 
where Rh^ y
t11 is a matrix of the cross covariance between parameters hi2 t11 i51:n and simulated observed
 i  o  
variables y^ t11 i51:n from Step M6; Ret11 hy
represents the covariance between eh it11 i51:n and the observa-
e^y y o  
tions; Rt11 represents the covariance between e^y it11 i51:n and the observations; and ðÞT represents the
transpose operator.
8. Update states as per Steps 5 and 6 in section 2.1.

2.4. A Locally Linear Approach

We also propose a second parameter evolution model, which unlike the Multilayer EnKF, requires no assump-
tions about the form of parameter nonstationarity. This method involves linearly extrapolating the ensemble
mean of the updated parameters from the previous two time steps to propose a background ensemble, as
demonstrated in Figure 1. If the change in updated parameters over a single time step exceeds a predefined
threshold value, the gradient of the linear fit is determined based on the preceding updated parameters. This
minimizes both overfitting and avoids parameter drift due to isolated large updates. Parameter values and
the linearized rate of change of parameters at time t 5 0 are required to initialize the algorithm.
The locally linear approach (LL) is undertaken as follows:
1. Apply the
n SKS oto generate an ensemble of background parameters for any given time
t 11; hiint; t11 :
i51:n  
hiint; t11  TMVN hi1 2 h
t ; s Rt for i51 : n (26)

where s2 is a smoothing parameter which must be tuned by the user. For this study, the parameters are
instead sampled from independent univarate truncated normal distributions, since parameters are forced
to vary independently in the synthetic case study (refer section 3).
2. Calculate the gradient of the linear function fitted to the updated parameter ensemble mean from time
t-1 and time t. Repeat for time t-2 and time t-1:
   
h1 1
t 2 ht21 h1 1
t21 2 ht22
mt 5 and mt21 5 (27)
Dt Dt21

where Dt 5t2ðt21Þ and Dt21 5ðt21Þ2ðt22Þ:

3. Define the gradient for linearly extrapolating the updated parameter ensemble mean at time t11:
(
mt ; jmt j  mmax
mt11 5 (28)
mt21 ; jmt j > mmax

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3356

 Water Resources Research 10.1002/2015WR017192

where mmax is a user defined positive which indicates the maximum allowable jump over a single time
step.
4. Shift the parameter ensemble from Step 1, hiint; t11 , so that h2 1
t11 5 ht 1mt11 : Dt11 , that is, calculate:
 
hi2 i
t11 5 hint; t11 1 h1
t 1mt11 : Dt11 2 hint;t11 for i51 : n (29)

3. Synthetic Case Studies With the Probability Distributed Model

The ability of the Dual EnKF to estimate nonstationary model parameters was investigated through an
application to the lumped conceptual hydrologic model, PDM [Moore, 2007]. The PDM consists of three
buckets representing the main storage and routing components of the catchment, ie. soil moisture, surface
runoff routing and groundwater flow. In this version, soil storage capacity is assumed to be Pareto distrib-
uted with shape parameter b and maximum point soil storage depth cmax . Excess rainfall is partitioned into
a surface runoff store and groundwater store through the splitting parameter a. Outflow from these two lin-
ear routing tanks is controlled by parameters ks (for the surface runoff store) and kb (for the groundwater
store). A schematic of the model is shown in Figure 2.
State and parameter estimation was undertaken by assimilating streamflow observations into the PDM at a
daily time step for approximately 3 years. In order to properly assess algorithm performance, the true sys-
tem states and model parameters must be known apriori. Synthetic daily rainfall, potential evapotranspira-
tion (PET) and model parameters were therefore specified for use in the PDM to develop the ‘‘true’’ system
variables. Catchment averaged rainfall derived from the 5km x 5km gridded daily rainfall SILO database
[Jeffrey et al., 2001] for the Snowy Creek Catchment in Victoria, Australia was used as the basis for synthetic
rainfall. Extended periods of no rainfall were replaced with rainfall events from wet years, to minimize peri-
ods of zero observability associated with dry conditions, and to avoid the need for likelihood functions
designed to operate for such ephemeral catchments [Smith et al., 2010]. Zero (or small) streamflow values
contain no information about the model parameters, meaning that extended dry periods create additional
challenges for the time varying parameter estimation problem. Daily PET data were developed by uniformly
disaggregating monthly average PET data from evapotranspiration maps published by the Australian
Bureau of Meteorology as part of their Climate Atlas Series (see http://www.bom.gov.au/climate/how/new-
products/IDCetatlas.shtml). Mean annual PET for the synthetic time series is approximately 1000 mm/yr
whilst mean annual rainfall is 1800 mm/yr. Multiple time varying parameter scenarios were developed in
order to assess algorithm performance for a range of situations, as detailed in Table 1. These include param-
eter combinations with a strong signal in the streamflow (Scenarios 1 and 2) and a weaker signal in the
streamflow (Scenarios 3 and 4). Signal strength is defined by the magnitude of runoff residuals when
parameters are assumed constant and equal to their initial values (refer supporting information Figure S1).
We considered a sinusoidal variation in the model parameters, following the work of Paik et al. [2005], Coron
et al. [2012] and Ye et al. [1997] who noted the potential for seasonal variations in hydrologic model param-
eters; and also He et al. [2011] where snow melt factors are treated as sinusoidally varying. Linear trends in
model parameters corresponding to a steady gradual change in land cover (e.g., afforestation/deforestation
or urbanization) [Croke et al., 2004] were also considered. Finally, parameter sets with different forms of non-
stationarity (Scenario 3) and differing rates of change (Scenario 2) were also devised to assess the ability to
detect temporally correlated parameter fluctuations.

3.1. Specification of EnKF components

We now describe the specification of the components of the EnKF, such as initialization and error character-
ization. In order to characterize the observation uncertainty, perturbed observation ensembles were gener-
ated following the method of Houtekamer and Mitchell [1998] and Burgers et al. [1998], using known error
variances. A heteroscedastic error variance was considered for generating the synthetic observations, and
defined as a proportion of the true streamflow, so that larger flows have greater errors than low flows. In
this study, d was chosen to be 0.1, and the error characteristics assumed known. A truncated normal distri-
bution was utilized to ensure positive streamflow observations:

qiobs ðtÞ5 qobs ðtÞ1 eiq where eiq  TNð0; d x qtrue ðt ÞÞ for i51 : n (30)

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3357

 Water Resources Research 10.1002/2015WR017192

Initial parameter ensembles were gen-

erated by sampling from a truncated
normal distribution with mean equal
to the true parameter value at t 5 0
(see Table 2). Perfect initial conditions
were adopted as the purpose of the
study is to assess the performance of
the filter in detecting parameter varia-
tions over time, rather than to correct
for initial condition error. For applica-
tions involving real data, this would
reflect a situation where model param-
eters have been well-calibrated over a
stationary period and used to initialize
Figure 1. Schematic of the Locally Linear Approach for proposing background the algorithm. Hyperparameters in the
parameters. The circles indicate the parameter ensemble mean. The mean of the Multilayer approach are assumed time-
background parameters for the current time (h2 t11 ) is estimated by linearly extrap- invariant, therefore initial hyperpara-
olating from the updated parameter ensemble mean from the previous time step.
The gradient (m) is calculated from the updated parameters from the previous meter ensembles were developed by
two time steps (t and t21). A background ensemble for time t11 is generated sampling from a uniform distribution
using the Standard Kernel Smoother, and the entire ensemble shifted so that its with bounds determined based on the
mean is equal to h2t11 .
feasible parameter range and initial
parameters. All hyperparameters were
updated simultaneously. Linearly varying parameters have only one hyperparameter, a (ht 5 ht21 1a),
 
whilst sinusoidally varying parameters have four hyperparameters, b; c; d and e (ht 5bsin 2p c t1d 1e). Ini-
tial states were also generated In order to minimize sampling error, an ensemble size of 100 members was
adopted, based on the findings of Moradkhani et al. [2005b] and Aksoy et al. [2006]. Due to the stochastic-
dynamic nature of the algorithm, ensemble statistics were calculated over 50 realizations of the dual filter-
ing process. For the sake of comparison, the same set of perturbed observations and initial ensembles were
used when evaluating the different parameter evolution models.
The difficulties in estimating time varying parameters with increasing dimensionality have been noted in
previous studies [Yang and Delsole, 2009]. In order to examine the effects of increased dimensionality on
algorithm performance, dual filtering for the scenarios described in Table 1 was undertaken for two case
studies. The first involves updating only the routing parameters and states of the PDM (the surface runoff
routing coefficient ks ; surface storage state Ss , groundwater routing coefficient kb , groundwater storage
state Sg , and the excess runoff splitting parameter a) in the EnKF, with the soil storage state and parameters
assumed known. The second involves estimating all model states and parameters within the Dual EnKF
framework.
All the parameter evolution models evaluated are parametric in that they require the specification of a tun-
ing parameter (ie. s2 for the SKS, Multilayer and Locally Linear methods; v for the HKS; and d for the KSLS).
An objective method for specifying such tuning parameters is through Variance Variable Multipliers
[Leisenring and Moradkhani, 2012]. This approach considers the tuning parameters as dynamic variables,
which are adjusted in time based on the absolute bias and the 95th percentile uncertainty bound. The abso-
lute bias refers to the absolute difference between forecast and observed streamflow from previous time
steps, whilst the 95th percentile uncertainty bound ðubt Þ refers to the distance between the ensemble
mean and 5th or 95th percentile of the forecast:
8
< y^t 2 y^5 if yt < y^t
t
ubt 5 (31)
: ^95 ^
yt 2 yt if yt > y^t

where y^t 5 ensemble mean forecast streamflow at time t, y^t5 5 5th percentile forecast streamflow at time t,
^ 5 95th percentile forecast streamflow at time t.
yt95
An initial tuning parameter value is supplied and increased at any given time if the absolute bias is greater
than the 95th percentile uncertainty bound (indicating the ensemble variance is too small because the

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3358

 Water Resources Research 10.1002/2015WR017192

Figure 2. PDM Schematic (adapted from Moore [2007]).

observation lies outside the forecast ensemble) and decreased if the absolute bias is less than the 95th per-
centile uncertainty bound (indicating the ensemble variance is too large). For further details on Variable Var-
iance Multipliers, refer to Leisenring and Moradkhani [2012]. Initial tuning parameter values of 0.03 for the
SKS and LL methods and 0.002 for the ML method were adopted within the Variable Variance Multiplier
framework. The remaining methods (KSLS and HKS) are not amenable to the VVM approach, hence, trial
experiments with a range of tuning parameters were undertaken. For the KSLS, a range of values for d
between 0.95 and 0.99 (as recommended in Liu and West [2001]) were evaluated and found to produce sim-
ilar parameter ensembles trajectories. In the HKS, v 5 0.2 was tested following Smith et al. [2008], however it
was found that smaller values were more suitable (refer section 4).

4. Results and Discussion

Clear differences between the estimation performance of each of the parameter evolution models were
seen when detecting temporal patterns in parameters. The results demonstrate that the parameter evolu-
tion model, i.e., the mechanism for defining the background (or prior) parameter ensemble, has a significant
impact on the detection of temporal patterns in parameters. A summary of the parameter estimation per-
formance for the various scenarios is provided in Figure 3. The Nash Sutcliffe Efficiency (NSE) of the parame-
ter ensemble mean and the percentage of time the parameter ensemble mean is within acceptable limits
(denoted PTW here) are provided. Both statistics are provided as averages over the estimated parameters.
Acceptable limits for each parameter are defined for time t as ht 1 rdp where ht is the true parameter value
at time t, dp is the feasible parameter range (Table 2) and r is a fraction varying between 5% and 10%.
Results are provided for each scenario for Case Study 1 (estimating routing parameters only) in the top
quadrant and for Case Study 2 (estimating full model parameters) in the bottom quadrant. Average NSE val-
ues closer to 1 and Average PTW scores closer to 100% indicate better performance.
When calculating the average Nash Sutcliffe Efficiency over all the model parameters, the Multilayer approach
gave higher average NSE values in comparison to the other parameter evolution models, regardless of the

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3359

 Water Resources Research 10.1002/2015WR017192

Table 1. True Parameters and Impact on Streamflow for the Various Scenarios
Scenario 1—High Impact on Q, Similar Scenario 2—High Impact on Q, Different Scenario 3—Low Impact on Q, Different Per- Scenario 4—Low Impact on Q, Similar
Parameter Variation Rates of Change sistence Types Parameter Variation
b bðt11Þ5bðtÞ2231024 As per Scenario 1 As per Scenario 1 As per Scenario 1
22
cmax cmax ðt11Þ5cmax ðt Þ16:08310
 As per Scenario
 2pt 1 As per Scenario 1 As per Scenario 1
2pt
ks ks ðtÞ5 0:1sin 300 10 10:5 ks ðt Þ5 0:1sin 200 20:2p 10:5 As per Scenario 1 As per Scenario 1
kb 2pt
kb ðt Þ5 0:05sin 300 10 10:85 2pt
kb ðtÞ5 0:05sin 150 10:4p 10:85 kb ðt11Þ5kb ðtÞ2131024 As per Scenario
 2pt 1
a aðtÞ5 0:15sin 3002pt
1p 10:7 aðtÞ5 0:15sin 300 2pt
10:2p 10:7 aðt11Þ5aðt Þ2231024 aðt Þ5 0:15sin 300 10 10:7

parameter scenario or the case study investigated (Figure 3, left). The Locally Linear method provides good
quality estimates for all scenarios involving the updating of routing parameters only (average NSE is 0.55).
Both methods yield superior parameter estimation performance compared to the existing parameter evolu-
tion models (SKS, KSLS and HKS). However, the performance of the Locally Linear method degrades when the
parameter dimensionality is increased, producing results similar to the SKS in the full model case study. We
now examine the various parameter evolution models and test scenarios in further detail.

4.1. Estimation of Routing Parameters

We begin by considering the first case study, where the soil storage parameters are assumed known and
only the routing parameters of the PDM are estimated for Scenarios 1–4. The results demonstrate the
potential of both the Multilayer approach and Locally Linear method even when the information content in
observations is low (Scenarios 3 and 4) and when model parameters vary at different rates with different
structures (Scenario 2 and 3 respectively). Across Scenarios 1–4, the Multilayer approach and the Locally Lin-
ear Method provide the highest NSE and PTW scores (refer Figure 3). The Locally Linear Method, which is
less restrictive in terms of its assumptions, provides parameter estimates with comparable quality to the
Multilayer approach for all scenarios in this case study. For the scenario with the most complex combination
of time varying parameters (Scenario 2), the long term temporal structure is mostly well represented by the
Locally Linear method (see Figure 4), with the Multilayer approach demonstrating the most superior per-
formance. Both the Multilayer and Locally Linear methods were also found to correctly identify stationary
parameters when applied to scenarios with a combination of time varying and time invariant parameters.
The existing parameter evolution models (KSLS, SKS and HKS) have varying degrees of success in the rout-
ing parameter estimation case study. The KSLS parameter evolution model performs poorly for all scenarios,
and is unsuitable for estimating parameters which vary over lengthy time scales. Representative results
from the KSLS for Scenario 1 (ie. parameter variations have a strong impact on streamflow) demonstrate
that the shrinkage of kernel locations encourages filter divergence (Figure 5) That is, the ensemble variance
is continuously reduced over time, such that the filter becomes insensitive to observations. Background
parameters are subsequently seen as having high confidence and are no longer updated, leading to conver-
gence to a stationary distribution. This result is to some extent expected, given that the KSLS has been
designed specifically for the detection of time invariant parameters [Liu and West, 2001].
The predecessor to the KSLS, the SKS, is more suited to hydrologic parameter estimation than the KSLS
where parameter changes occur over long time scales (see Figure 4). Unlike the KSLS, the SKS is less suscep-
tible to filter divergence because of its in-built variance inflation. The background parameter ensemble
 i2 
h from the SKS has larger variance than the updated ensemble from the previous time step
 t11  i51:n  
hi1
t i51:n
, unlike in the KSLS where the shrinkage of locations retains the variance of hi1
t i51:n
[Liu and

Table 2. Sampling Distributions of Initial Parameters and Their Feasible Rangesa

Feasible Range

Parameter Description Units Initial Sampling Distribution Lower Bound Upper Bound
b Pareto-distributed soil storage shape N(b(t 5 0), 0.001) 0.10 1.50
parameter
cmax Maximum point soil storage depth (mm) N(cmax (t 5 0), 10) 150 350
ks Surface Runoff Routing Coefficient N(ks (t 5 0), 0.008) 0.00 0.69
kb Groundwater Routing Coefficient N(kb (t 5 0), 0.001) 0.70 0.99
a Excess Runoff Splitting Parameter N(a(t 5 0), 0.001) 0.50 0.99
a
For all parameters, the mean of the initial sampling distribution is equal to the true initial parameter.

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3360

 Water Resources Research 10.1002/2015WR017192

Figure 3. Summary performance metrics from using a suite of parameter evolution models in an EnKF framework for estimating time varying parameters. A range of time varying param-
eters were examined (Scenarios 1–4, refer Table 1) in two separate Case Studies 1) Updating routing parameters only and 2) Updating all model parameters. The NSE of the parameter
ensemble mean (averaged over all parameters) is shown in the left panel. The right panel shows the percentage of time the ensemble mean is within acceptable limits (averaged over
all parameters). Acceptable limits for each parameter are defined for time t as ht 1rd p where ht is the true parameter value at time t, d p is the feasible parameter range (Table 2) and r
is a fraction varying between 5% and 10%. HKS 5 Homoscedastic Kernel Smoother, KSLS 5 Kernel Smoother with Location Shrinkage, SKS5 Standard Kernel Smoother, LL 5 Locally
Linear, ML 5 Multilayer.

West, 2001]. Whilst the SKS performs reasonably well for Scenario 1 where parameter variations strongly
impact streamflow response (on average, parameters are within their acceptable limits approximately 60%
of the time), its estimation of a and kb in particular for Scenario 2 and 4 is poor (parameters within accepta-
ble limits for only 30–40% of the time). Scenario 2 in particular shows that even if the parameter variation
signal in streamflow is fairly strong, the SKS struggles when multiple parameters vary at different rates (refer
Figure 4).
Similar to the KSLS, parameter trajectories from the HKS fail to replicate the true parameter variations in all
scenarios. Figure 5 shows the ensemble mean trajectories for the 3 year simulation period with v 5 0.2 (as
adopted in Smith et al. [2008]) for a representative case, Scenario 1. Rapid temporal fluctuations in the
ensemble mean (as much as 0.2 over a time step) as well as the width of the ensemble are indicative of
overdispersion due to the perturbation process. Smaller values of v led to reduced short-term parameter
fluctuations, but were still unable to detect long-term patterns (refer Figure 5b shown for v 5 0.005).
Although the HKS is fairly similar to the SKS in its structure, it would seem that the use of a time invariant
variance for parameter resampling leads to either overdispersion or underdispersion at certain times,
degrading the overall performance of the filter.

4.2. Estimation of All Model Parameters

We now consider the case where all the parameters of the PDM are estimated simultaneously by the filter,
that is, the Pareto distribution shape parameter b and maximum point soil storage depth cmax are also esti-
mated. As was the case with the first case study, the Multilayer approach produces high quality estimates
for all parameters, across all scenarios, with average NSE and PTW values in excess of 0.85 and 85% respec-
tively. The linear trend in the soil storage parameters b and cmax was captured well by the SKS, Multilayer

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3361

 Water Resources Research 10.1002/2015WR017192

Figure 4. Updated parameter ensemble for Case Study 1 (Routing parameter updating only), Scenario 2. The darker grey areas indicate the middle 90% of the ensemble, whilst the
lighter grey areas indicate the middle 50% of the ensemble. The blue line indicates the ensemble mean whilst the red line indicates the synthetic true parameter. Results are provided
for (a) the Multilayer Method (ML), (b) the Locally Linear Method (LL), and (c) the Standard Kernel Smoother (SKS).

and Locally Linear methods in all scenarios (refer Figure 6). For both the HKS and the KSLS, inclusion of the
soil storage parameters produced similar poor results to the case where only the routing parameters were
estimated. However, for the remaining parameter evolution models, estimation of the routing parameters
(a, ks and kb ) was degraded in comparison to the case where b and cmax were assumed known. Inclusion of
the soil storage parameters only slightly degraded the quality of routing parameter estimates (in terms of
capturing the overall trend) for Scenario 3, where kb and a both vary linearly in time (see Figure 6) . However
in terms of parameter bias, the NSE reduced significantly for the routing parameters (in particular, for the
Locally Linear method, see Table 3), and the sinusoidal temporal structure in Scenarios 1, 2 and 4 was less
pronounced in the ensemble mean (see e.g., Figure 7). This reduction in estimation quality is attributed to
the sensitivity of these parameters to errors in the excess runoff, particularly a and kb . Biases in the soil stor-
age state accumulate over time and are not fully compensated for in the state updating step due to small
errors in the soil storage parameters. This leads to errors in the excess runoff which degrade estimates of
the less observable routing parameters, kb and a. This is particularly problematic when kb and a have com-
plex time varying structures (e.g., high frequency sinusoidal), as the impacts of parameter fluctuations on
the streamflow may be drowned out by errors in the excess runoff. Despite the reduced parameter estima-
tion quality, the Locally Linear method still leads to superior estimation of streamflow compared to the
Standard Kernel Smoother across the scenarios.
The full model parameter case study demonstrates the efficacy of the Multilayer approach, particularly in
situations where several parameters are time varying and there is reduced information content in the obser-
vations. However, it is limited to situations where the form of nonstationarity can be defined apriori and the
hyperparameters which govern the nonstationarity are time invariant. The Locally Linear approach is a suita-
ble alternative which has modest apriori knowledge requirements (initial parameter value and initial rate of
change). It has been shown to provide parameter estimates of comparable quality to the Multilayer
approach in low dimensional systems, even when the impact of parameter variation on the observations is

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3362

 Water Resources Research 10.1002/2015WR017192

Figure 5. Updated parameter ensemble for Case Study 1 (Routing parameter updating only), Scenario 1. The darker grey areas indicate
the middle 90% of the ensemble, whilst the lighter grey areas indicate the middle 50% of the ensemble. The blue line indicates the ensem-
ble mean whilst the red line indicates the synthetic true parameter. Results are shown for (a) the Kernel Smoother with Location Shrinkage
and (b) the Homoscedastic Kernel Smoother.

weak. However, the Local Linear approach is less resistant to periods of poor observability, and was shown
to perform poorly when the dimensionality and nonlinearity was increased. It remains to be seen whether
similar dimensionality issues arise when the Local Linear method is applied to other model structures. Issues
of observability and identifiability in parameter estimation using DA will be investigated in more detail in a
follow up paper.

Table 3. Percentage Change in NSE of the Routing Parameters When All Model Parameters are Estimated, Compared to When Only the
Routing Parameters are Estimateda
Percentage
 change in NSE

NSEfull model 2 NSErouting only
jNSErouting only j

Scenario Parameter Locally Linear Method Standard Kernel Smoother Multilayer Method
Scenario 1 ks 228% 243% 0%
kb 2252% 242% 215%
a 242% 286% 24%
Scenario 2 ks 270% 281% 2%
kb 2485% 2552% 12%
a 2134% 258% 21%
Scenario 3 ks 219% 285% 24%
kb 2115% 281% 215%
a 242% 240% 215%
Scenario 4 ks 217% 262% 29%
kb 2938% 25% 229%
a 2114% 5% 218%
Average over all parameters and scenarios: 2188% 294% 26%
a
Negative values indicate a reduction in the NSE score when all model parameters are estimated simultaneously. The quality of the
parameter estimates is almost always degraded when the dimensionality is increased.

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3363

 Water Resources Research 10.1002/2015WR017192

Figure 6. Updated parameter ensemble for Case Study 2 (Full model parameter updating), Scenario 3. The darker grey areas indicate the middle 90% of the ensemble, whilst the lighter
grey areas indicate the middle 50% of the ensemble. The blue line indicates the ensemble mean whilst the red line indicates the synthetic true parameter. Results are provided for (a)
the Multilayer Method (ML), (b) the Locally Linear Method (LL), and (c) the Standard Kernel Smoother (SKS).

4.3. Robustness of the Multilayer Method

Additional testing was undertaken to determine the robustness of the Multilayer method to errors in the
assumed parameter dynamics. Any given parameter h was assumed to follow a more general structure:

2pt
hðtÞ5 asin 1c 1d1et (32)
b

where a; b; c; d and e are the 5 hyperparameters updated in the first stage of the filter for each model
parameter and t 5 time. The algorithm was found to be fairly robust to changes in assumptions about the
temporal structure of parameters, with only one parameter experiencing a reduction in performance due
to poor identifiability. For the first case study, results were largely similar to the case where there are no
ambiguities in the dynamical structure of parameters (refer supporting information Figure S2 for

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3364

 Water Resources Research 10.1002/2015WR017192

Figure 7. Updated parameter ensemble for Case Study 2 (Full model parameter updating), Scenario 4. The darker grey areas indicate the middle 90% of the ensemble, whilst the lighter
grey areas indicate the middle 50% of the ensemble. The blue line indicates the ensemble mean whilst the red line indicates the synthetic true parameter. Results are provided for (a)
the Multilayer Method (ML), (b) the Locally Linear Method (LL), and (c) the Standard Kernel Smoother (SKS).

representative results). Only a minor degradation in performance was seen for Case Study 2. All parame-
ters except for cmax were well represented, with performance similar to that of the perfect parameter
structure case (refer supporting information Figure S3). The reduction in cmax performance is attributed to
the weak correlation between its hyperparameters and the observations. This means that introducing fur-
ther unknowns in temporal structure reduces the ability of the filter to detect changes in this parameter.
Ultimately, the method was found to be fairly robust to incorrect assumptions in parameter temporal
structure, but its performance in this case may reduce as dimensionality and nonlinearity increase and
parameter identifiability reduces.

4.4. Improving the Locally Linear Approach—Local Linear Regression?

The Locally Linear approach demonstrated strong potential for the estimation of time varying parameters
in the first case study, with minimal apriori knowledge required. The method generates background

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3365

 Water Resources Research 10.1002/2015WR017192

parameters at any given time by using

the rate of change of the updated
parameters from the previous two
time steps. A natural question might
be whether the Locally Linear
approach could be improved by con-
sidering a longer historical time win-
dow of updated parameters than
simply the previous two time steps to
propose the background. This was
examined by applying Local Linear
Regression (hereafter LLR) to the full
time history of updated parameters,
thereby producing a temporally
smoothed time series. The commonly
Figure 8. Schematic of the Locally Linear Approach combined with local linear
regression for proposing background parameters. Initially, local linear regression
used Epanechnikov kernel [Epanechni-
is undertaken on the updated parameter ensemble means from previous time kov, 1969] was adopted as the weight-
steps (refer to the blue line). The mean of the background parameters for the ing function and a number of
current time {h2 t11 } is then estimated by linear extrapolation from the regressed
values.
bandwidth sizes examined. Back-
ground parameters were then pro-
posed in a similar fashion to the Locally Linear approach, except that the smoothed parameters from the
previous two time steps were used for extrapolation (refer Figure 8).
Incorporating LLR into the Locally Linear approach produced temporally smoother parameter trajectories
than when it was not used. However, its ability to replicate the temporal structure of parameters was signifi-
cantly reduced compared to the nonsmoothed Local Linear approach.
Figure 9 shows the parameter ensemble when combining the Locally Linear approach with LLR for Case
Study 1 - Scenario 1. Bandwidth values of 10, 20 and 5 days were adopted for all parameters based on trial
runs. It can be seen that the filter produces a delayed response near the crests and troughs where the
parameters experience the most rapid change in gradient, (see for instance near days 800–1000 for kb ) .

Figure 9. Parameter ensemble mean trajectories for the Locally Linear Method with Local Linear Regression (LLR) and the Locally Linear
Method (LL) for Case Study 1 (Routing parameter updating only), Scenario 1.

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3366

 Water Resources Research 10.1002/2015WR017192

This behavior occurs due to the LLR smoothing out regions of parameter updates of duration less than the
bandwidth, which may indicate a critical change in the parameter. This is demonstrated here where even a
relatively small bandwidth of 5 days is used.

5. Conclusions
The accuracy of hydrologic predictions from conceptual rainfall runoff models is strongly dependent on the
calibration parameters used to derive them. Consideration of a unique parameter set or stationary distribution
of parameters is problematic in that 1) it ignores potential changes in the catchment that may have occurred
outside the calibration period (e.g., changing land use); 2) they have a tendency to be highly dependent on
the dominant climatic regime of the calibration period [Sorooshian et al., 1983; Choi and Beven, 2007; Wu and
Johnston, 2007]; and 3) it is difficult to adequately simulate all features of hydrologic variables (such as peak
flows, low flows, volume of runoff) due to model structural deficiencies [Moussa and Chahinian, 2009; Efstratia-
dis and Koutsoyiannis, 2010; Westerberg et al., 2011]. A framework which relaxes the need for a time invariant
parameter set has the potential to improve the transferability of hydrologic models in time, particularly for
nonstationary catchments. This paper examined the potential for Data Assimilation (DA) to detect time varia-
tions in model parameters from hydrologic observations, namely streamflow. It serves as an important first
step in examining whether DA can be used to detect a structured time variation in multiple model parameters
simultaneously. Joint state-parameter estimation with the Ensemble Kalman Filter was undertaken for a range
of synthetic experiments using the conceptual rainfall runoff model, the Probability Distributed Model (PDM).
It was demonstrated that the success of the DA algorithm is strongly dependent on the mechanism for gener-
ating the background (or prior) parameter ensembles, an issue which has not been investigated in previous
studies [Smith et al., 2008; Vrugt et al., 2013]. Existing techniques such as the Kernel Smoother with Location
Shrinkage [Liu and West, 2001] and the Standard Kernel Smoother [West, 1993] are shown to be poorly suited
for time varying parameter estimation from streamflow observations, particularly when multiple parameters
vary in a complex, uncorrelated fashion. Two alternative parameter evolution models are proposed which are
specifically suited to time varying parameter applications. The first is a Multilayer approach which uses the
EnKF to estimate hyperparameters of the temporal structure, based on apriori knowledge of the form of non-
stationarity. The second is a Locally Linear approach that uses local linear estimation to propose temporal
changes and requires no assumptions of the form of parameter nonstationarity. The results demonstrate that
both methods more accurately capture temporal variations in parameters compared to the traditional param-
eter evolution models. This indicates that incorporating information about temporal structure into the back-
ground, in addition to ensuring sufficient variance, can improve the detection of time varying parameters.
Expert knowledge of how catchment nonstationary affects model parameters (e.g., whether parameters will
vary linearly or cyclically with time to reflect changes in catchment conditions) will significantly improve filter
performance when changing processes affect several model parameters. This is demonstrated by the superior
performance of the Multilayer method compared to other approaches examined. The method was shown to
be fairly robust to inaccurate assumptions of parameter dynamics, meaning that reasonable performance can
still be obtained by considering all the likely modes of parameter variability together. This however has the
potential to degrade the detection of less identifiable parameters. The Locally Linear method requires only
knowledge of the initial parameters and their likely rate of change. It was shown to provide parameter esti-
mates of comparable quality to the Multilayer approach in low dimensional systems, indicating that it can be
a suitable alternative where changes in catchment conditions lead to time variations in only a few model
parameters. Small increases in dimensionality and nonlinearity have the potential to degrade the perform-
ance of the Locally Linear method, indicating a limit to which several time varying parameters can be
estimated from streamflow observations alone, without the addition of further external information (as
in the case of the Multilayer method). The optimum number of parameters that could be satisfactorily
estimated by the Locally Linear method would be model dependent, and also depend on the signal
strength in the observed variables. An application of the proposed methods to a real data case study will
be investigated in a follow up publication. This will allow assessment of the extent to which time varying
parameters can compensate for missing processes and the resulting improvement in streamflow predic-
tion. Future work will also examine the performance of the proposed parameter evolution models with
other DA algorithms such as the Ensemble Kalman Smoother (EnKS) and Particle Filter, along with a
range of model structures.

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3367

 Water Resources Research 10.1002/2015WR017192

Appendix A
Here we demonstrate that the parameter update equation in the state augmentation approach is equiva-
lent to the dual update procedure when applied with Kalman filtering methods. For the purposes of dem-
onstration, we will assume our system is one dimensional (the same logic applies for higher dimesions):
x t11 5 xt11 ; y t11 5 yt11 ; ht11 5 ht11

We start with the state augmentation approach for parameter updating. The augmented state vector is
defined as:
" #
xt11
zt11 5 (A1)
ht11

The augmented state is updated using the Kalman update equation:

i 
zi1 i2
y it11 for i51 : n
t11 5zt11 1 K t11 yt11 2 ^ (A2)
h i21
^y ^y yo yo
Kt11 5 Rz^
y
t11 Rt11 1 Rt11 (A3)

That is,
" # " # 2 
   3
i1
xt11 i2
xt11 1 cov ; ^y i2 i2
xt11
t11 i51:n
i
i51:n 
5 1  ^^ 4  i2   i2   5 yt11 2 ^y t11 for i51 : n
i
(A4)
yo yo
hi1
t11 hi2
t11
yy
Rt11 1 Rt11 cov ht11 i51:n ; ^y t11 i51:n

Since
(" i2
#) ! 2  i2   i2  3
 2  xt11  i2  cov xt11 i51:n
; ^
y t11 i51:n
Rz^
y
t11 5cov zt11 ; ^
y t11 5cov ; ^y t11 i51:n 54  i2   i2  5 (A5)
hi2
t11 i51:n cov ht11 i51:n ; ^
y t11 i51:n

Now extract the parameter update equation (i.e., second element of vector):
1  i2    
i 
hi1 i2
t11 5ht11 1

^ ^y o yo
 cov ht11 i51:n ; ^y i2 y it11
t11 i51:n yt11 2 ^ for i51 : n (A6)
Ryt11 1 Ryt11
h 
h^y ^y ^y yo yo
5hi2
t11 1Rt11 Rt11 1 Rt11 21 yt11
i
2 ^y it11  for i51 : n (A7)

This is the equivalent to the parameter update equation presented in the Dual approach (refer equations
(6) and (7)).

Appendix B
Here we provide further details on the state updating step in the dual state-parameter estimation frame-
work (i.e., equations (10–12)). As noted earlier, the observations have been utilized to generate the back-
ground model states (by way of the updated model parameters). There is therefore potentially some
correlation between the model simulation noise (also referred to as process noise) and the observation (or
measurement) noise. In such cases, the Kalman update equation remains unchanged, but the Kalman gain
matrix now includes terms accounting for the covariance between process and measurement noise [e.g.,
Simon, 2006]:


21
K5 Pt HT 1Ct HPt HT 1HC1 Ct T HT 1Rt (B1)

where x t11 5 Fx t 1Gut11 1xt ; y t 5 Hx t 1tt ; cov ðxt ; xt Þ5Pt 5 background or process error covariance,
cov ðtt ; tt Þ5Rt 5 observation error covariance, cov ðxt ; tt Þ5Ct 5 covariance between background and
observation errors, H 5 linearized observation operator, F and G represent linearized model dynamics.
In an EnKF framework, error covariances are estimated by the sample covariances of the monte carlo
ensembles (assuming unbiasedness). For instance, in standard state updating using the EnKF, the following
estimate is used:

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3368

 Water Resources Research 10.1002/2015WR017192

Pt HTt 5cov ðxt ; xt ÞHTt 5cov ðxt ; Hxt Þ

(B2)
5Rxy
t

5 covariance between model simulated states and simulated observed variables.

The same approach can therefore be used to estimate Ct , with Ct given by the cross covariance between
the simulated model states using the updated parameters and the observations. This is problematic for a
number of reasons. Firstly, consider the two ensembles of model states available prior to state updating:

 
x^ it11 5f x i1 i2
t ; ht11 ; ut11 for i51 : n (B3)

 i1 i1 
x i2
t11 5f x t ; ht11 ; ut11 for i51 : n (B4)
   
where x i2
t11 i51:n represents the model simulated states using the prior parameters and x^ it11 i51:n repre-
sents the model simulated states using the updated parameters. To simplify the notation, we define the
following:
 
X5 x i2t11 i51:n
 
^X 5 x^ i
t11 i51:n
 
Y5 y it11 i51:n

X is potentially correlated with the Y (the observation errors at time t11), whilst X^ is not, as the observa-
tions have not been used in any capacity to generate these states. Now X and X^ are highly correlated with
 
each other. Due to sampling issues, it is possible that corr X^ ; Y is nonzero, and since X and X^ are highly
correlated, this nonzero correlation will inflate the estimate of Ct11 . A more appropriate estimate for Ct11 is
o
hence given by Ret11
xy
:

o       
Ct11  Ret11
xy
5cov X2 X^ ; Y 5 cov ex it11 i51:n ; y it11 i51:n (B5)

where ex it11 5 x i2 ^ it11

t11 2 x

To see why this is the case, note that equation (B5) can be written as:
o    
Ret11
xy
5cov X2 X^ ; Y 5cov ðX; Y Þ2cov X^ ; Y (B6)

Equation (B6) shows that any spurious correlations between X^ and Y are removed from cov ðX; Y Þ, and that
o    
Ret11
xy
is equivalent to cov ðX; Y Þ if corr X^ ; Y is indeed zero (ie. Ct11 5 cov ðX; Y Þ if corr X^ ; Y 5 0). It was
found from experiments in this study that such spurious correlations have the potential to overestimate
Ct11 , leading to degraded parameter and state estimates.
Now the term HCt11 is simply the covariance between the process noise in observation space and the
observations. Using the estimate of Ct11 from equation (B5), we have:
e~y y o    
Rt11  HCt11 5 Hcov X2 X^ ; Y 5 cov HX2 HX^ ; Y
     (B7)
5cov e~y it11 i51:n ; y it11 i51:n

where e~y it11 5 y~ it11 2 y^ it11 .

Acknowledgments References
This study was funded by the
Aksoy, A., F. Zhang, and J. Nielsen-Gammon (2006), Ensemble-based simultaneous state and parameter estimation in a two-dimensional
Australian Research Council as part of
sea-breeze model, Mon. Weather Rev., 134, 2951–2970.
the Discovery Project DP140102394.
Anderson, J. L. (2007), An adaptive covariance inflation error correction algorithm for ensemble filters, Tellus, Ser. A, 59(2), 210–224, doi:
Marshall is additionally supported
10.1111/j.1600-0870.2006.00216.x.
through a Future Fellowship
Annan, J., J. Hargreaves, N. Edwards, and R. Marsh(2005), Parameter estimation in an intermediate complexity earth system model using
FT120100269. Data utilized in this
an ensemble Kalman filter, Ocean Modell., 8(1-2), 135–154, doi:10.1016/j.ocemod.2003.12.004.
study can be made available from the
Aubert, D., C. Loumagne, and L. Oudin(2003), Sequential assimilation of soil moisture and streamflow data in a conceptual rainfall–runoff
authors upon request.
model, J. Hydrol., 280(1-4), 145–161, doi:10.1016/S0022-1694(03)00229-4.

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3369

 Water Resources Research 10.1002/2015WR017192

Bardossy, A. (2007), Calibration of hydrological model parameters for ungauged catchments, Hydrol. Earth Syst. Sci., 11, 703–710.
Beven, K., and J. Freer (2001), Equifinality, data assimilation, and uncertainty estimation in mechanistic modelling of complex environmen-
tal systems using the GLUE methodology, J. Hydrol., 249, 11–29.
Burgers, G., P. J. van Leeuwen, and G. Evensen (1998), On the analysis scheme in the ensemble Kalman filter, Mon. Weather Rev., 126,
1719–1724.
Choi, H. T., and K. Beven(2007), Multi-period and multi-criteria model conditioning to reduce prediction uncertainty in an application of
TOPMODEL within the GLUE framework, J. Hydrol., 332, 316–336.
Coron, L., V. Andr eassian, C. Perrin, J. Lerat, J. Vaze, M. Bourqui, and F. Hendrickx (2012), Crash testing hydrological models in contrasted cli-
mate conditions: An experiment on 216 Australian catchments. Water Resour. Res., 48, W05552, doi:10.1029/2011WR011721.
Croke, B. F.W., W. S. Merritt, and A. J. Jakeman (2004), A dynamic model for predicting hydrologic response to land cover changes in
gauged and ungauged catchments, J. Hydrol., 291(1-2), 115–131, doi:10.1016/j.jhydrol.2003.12.012.
Dee, D. P. (2005), Bias and data assimilation, Q. J. R. Meteorol. Soc., 131(613), 3323–3343, doi:10.1256/qj.05.137.
de Vos, N. J., T. H. M. Rientjes and H. V. Gupta (2010), Diagnostic evaluation of conceptual rainfall-runoff models using temporal clustering,
Hydrol. Processes, 24, 2840–2850, doi:10.1002/hyp.7698.
Duan, Q., S. Sorooshian, and V. K. Gupta (1994), Optimal use of the SCE-UA global optimization method for calibrating watershed models,
J. Hydrol., 158, 265–284.
Ebtehaj, M., H. Moradkhani, and H. V. Gupta (2010), Improving robustness of hydrologic parameter estimation by the use of moving block
bootstrap resampling, Water Resour. Res., 46, W07515, doi:10.1029/2009WR007981.
Efstratiadis, A., and D. Koutsoyiannis(2010), One decade of multi-objective calibration approaches in hydrological modelling: A review,
Hydrol. Sci. J., 55(1), 58–78, doi:10.1080/02626660903526292.
Epanechnikov, V. A. (1969), Non-parametric estimation of a multivariate probability density, Theory Probab. Appl., 14(1), 153–158, doi:
10.1137/1114019.
Evensen, G. (2006), Data Assimilation: The Ensemble Kalman Filter, Springer, N. Y.
Franssen, H. H. J., and W. Kinzelbach(2008), Real-time groundwater flow modeling with the Ensemble Kalman Filter: Joint estimation of
states and parameters and the filter inbreeding problem, Water Resour. Res., 44, W09408, doi:10.1029/2007WR006505.
Gelb, A. (1974), Applied Optimal Estimation, MIT Press, Cambridge, U. K.
Gharari, S., M. Hrachowitz, F. Fenicia, and H. H. G. Savenije (2013), An approach to identify time consistent model parameters: Sub-period
calibration, Hydrol. Earth Syst. Sci., 17(1), 149–161, doi:10.5194/hess-17-149-2013.
Gordon, N. J., D. J. Salmond, and A. F. M. Smith (1993), Novel approach to non-linear/non-Gaussian Bayesian state estimation, IEE Proc., Part
F, 140, 107–113.
Gupta, H. V., S. Sorooshian, and P. O. Yapo (1998), Toward improved calibration of hydrologic models: Multiple and noncommensurable
measures of information, Water Resour. Res., 34(4), 751–763, doi:10.1029/97WR03495.
He, M., T. S. Hogue, K. J. Franz, S. A. Margulis, and J. A. Vrugt (2011), Characterizing parameter sensitivity and uncertainty for a snow model
across hydroclimatic regimes, Adv. Water Resour., 34(1), 114–127, doi:10.1016/j.advwatres.2010.10.002.
Houser, P. R., W. J. Shuttleworth, J. S. Famiglietti, H. V. Gupta, K. H. Syed, and D. C. Goodrich (1998), Integration of soil moisture remote sens-
ing and hydrological modeling using data assimilation, Water Resour. Res., 34(12), 3405–3420.
Houtekamer, P. L., and H. L. Mitchell (1998), Data assimilation using an ensemble Kalman filter technique, Mon. Weather Rev., 126, 796–811.
Jazwinski, A. H. (1970), Stochastic Processes and Filtering Theory, Elsevier, N. Y.
Jeffrey, S. J., J. O. Carter, K. B. Moodie, and A. R. Beswick (2001), Using spatial interpolation to construct a comprehensive archive of
Australian climate data, Environ. Modell. Software, 16(4), 309–330.
Jeremiah, E., L. Marshall, S. A. Sisson, and A. Sharma(2013), Specifying a hierarchical mixture of experts for hydrologic modeling: Gating
function variable selection, Water Resour. Res., 49, 2926–2939, doi:10.1002/wrcr.20150.
Kalman, R. E. (1960), A new approach to linear filtering and prediction problems, J. Basic Eng., 82 (1960), 35–45.
Khalil, H. K. (1996), Nonlinear Systems, 143 pp., Simon & Schuster, Upper Saddle River, N. J.
Komma, J., G. Bl€ oschl, and C. Reszler (2008), Soil moisture updating by Ensemble Kalman Filtering in real-time flood forecasting, J. Hydrol.,
357(3-4), 228–242, doi:10.1016/j.jhydrol.2008.05.020.
Kuczera, G. (1987), Prediction of water yield reductions following a bushfire in ash-mixed species eucalyptus forest, J. Hydrol., 94, 215–236.
Leisenring, M., and H. Moradkhani (2012), Analysing the uncertainty of suspended sediment load prediction using sequential data assimila-
tion, J. Hydrol., 468–469, 268–282.
Li, Y., D. Ryu, A. W. Western, and Q. J. Wang (2013), Assimilation of stream discharge for flood forecasting: The benefits of accounting for
routing time lags, Water Resour. Res., 49, 1887–1900, doi:10.1002/wrcr.20169.
Lin, Z., and M. B. Beck (2007), On the identification of model structure in hydrological and environmental systems, Water Resour. Res., 43,
W02402, doi:10.1029/2005WR004796.
Liu, J., and M. West (2001), Combined parameter and state estimation in simulation based filtering, in Sequential Monte Carlo Methods in
Practice, edited by A. Doucet, N. de Freitas, and N. Gordon, pp. 197–223, Springer, N. Y.
Liu, Y., and H. V. Gupta (2007), Uncertainty in hydrologic modeling: Toward an integrated data assimilation framework, Water Resour. Res.,
43,W07401, doi:10.1029/2006WR005756.
Liu, Y., A. H. Weerts, M. Clark, H.-J. Hendricks Franssen, S. Kumar, H. Moradkhani, and P. Restrepo (2012), Advancing data assimilation in
operational hydrologic forecasting: Progresses, challenges, and emerging opportunities, Hydrol. Earth Syst. Sci. Discuss., 16, 3863–3887.
doi:10.5194/hessd-9-3415-2012.
L€
u, H., Z. Yu, Y. Zhu, S. Drake, Z. Hao, and E. A. Sudicky (2011), Dual state-parameter estimation of root zone soil moisture by optimal
parameter estimation and extended Kalman filter data assimilation, Adv. Water Resour., 34(3), 395–406, doi:10.1016/
j.advwatres.2010.12.005
Madsen, H. (2003), Parameter estimation in distributed hydrological catchment modelling using automatic calibration with multiple objec-
tives, Adv. Water Resour., 26(2), 205–216, doi:10.1016/S0309-1708(02)00092-1.
Maneta, M. P., and R. Howitt(2014), Stochastic calibration and learning in nonstationary hydroeconomic models, Water Resour. Res., 50,
3976–3993, doi:10.1002/2013WR015196.
Marshall, L., A. Sharma, and D. J. Nott (2006), Modelling the catchment via mixtures: Issues of model specification and validation, Water
Resour. Res., 42, W11409, doi:10.1029/2005WR004613.
Matgen, P., F. Fenicia, S. Heitz, D. Plaza, R. de Keyser, V. R. N. Pauwels, and H. Savenije (2012), Can ASCAT-derived soil wetness indices
reduce predictive uncertainty in well-gauged areas? A comparison with in situ observed soil moisture in an assimilation application,
Adv. Water Resour., 44, 49–65, doi:10.1016/j.advwatres.2012.03.022.

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3370

 Water Resources Research 10.1002/2015WR017192

Merz, R., J. Parajka, and G. Bl€ oschl (2011), Time stability of catchment model parameters: Implications for climate impact analyses, Water
Resour. Res., 47, W02531, doi:10.1029/2010WR009505.
Mohammadpour, J., and C. W. Scherer (2012), Control of Linear Parameter Varying Systems With Applications, 3 pp., Springer, N. Y.
Moore, R. J. (2007), The PDM rainfall-runoff model, Hydrol. Earth Syst. Sci., 11(1), 483–499, doi:10.5194/hess-11-483-2007.
Moradkhani, H., K.-L. Hsu, H. Gupta, and S. Sorooshian(2005a), Uncertainty assessment of hydrologic model states and parameters: Sequen-
tial data assimilation using the particle filter, Water Resour. Res., 41, W05012, doi:10.1029/2004WR003604.
Moradkhani, H., S. Sorooshian, H. V. Gupta, and P. R. Houser (2005b), Dual state–parameter estimation of hydrological models using ensem-
ble Kalman filter, Adv. Water Resour., 28(2), 135–147, doi:10.1016/j.advwatres.2004.09.002.
Moradkhani, H., C. M. DeChant, and S. Sorooshian(2012), Evolution of ensemble data assimilation for uncertainty quantification using the
particle filter-Markov chain Monte Carlo method, Water Resour. Res., 48, W12520, doi:10.1029/2012WR012144.
Moussa, R., and N. Chahinian(2009), Comparison of different multi-objective calibration criteria using a conceptual rainfall-runoff model of
flood events, Hydrol. Earth Syst. Sci., 13, 519–535.
Musso, C., N. Oudjane, and F. LeGland (2001), Improving regularised particle filters, in Sequential Monte Carlo Methods in Practice, edited by
A. Doucet, J. F. G. de Freitas, and N. J. Gordon, pp. 247–271, Springer, N. Y.
Noh, S. J. I. N., Y. Tachikawa, M. Shiiba, and S. Kim (2013), Sequential data assimilation for streamflow forecasting using a distrib-
uted hydrologic model: Particle filtering and ensemble Kalman filtering, in Floods: From Risk to Opportunity, IAHS Publ., vol. 357,
341–349.
Paik, K., J. H. Kim, H. S. Kim, and D. R. Lee (2005), A conceptual rainfall-runoff model considering seasonal variation, Hydrol. Processes,
19(19), 3837–3850, doi:10.1002/hyp.5984.
Pathiraja, S., S. Westra, and A. Sharma (2012), Why continuous simulation? The role of antecedent moisture in design flood estimation,
Water Resour. Res., 48, W06534, doi:10.1029/2011WR010997.
Reichle, R. H., D. B. McLaughlin, and D. Entekhabi (2002), Hydrologic Data Assimilation with the Ensemble Kalman Filter, Mon. Weather Rev.,
130(1), 103–114, doi:10.1175/1520-0493(2002)130 < 0103:HDAWTE>2.0.CO;2.
Richards, J. A. (1983), Analysis of Periodically Time-Varying Systems, Springer, N. Y.
Salamon, P., and L. Feyen (2009), Assessing parameter, precipitation, and predictive uncertainty in a distributed hydrological model using
sequential data assimilation with the particle filter, J. Hydrol., 376(3-4), 428–442, doi:10.1016/j.jhydrol.2009.07.051.
Samuel, J., P. Coulibaly, G. Dumedah, and H. Moradkhani (2014), Assessing model state variation in hydrologic data assimilation, J. Hydrol.,
513, 127–141, doi:10.1016/j.jhydrol.2014.03.048.
Schwartz, C. A., and H. Ozbay (1990), An identification procedure for linear continuous time systems with jump parameters, in Proceedings
of the International Symposium MTNS-89 Volume 1 on Realization and Modelling in Systems Theory, edited M. A. Kaashoek, J. H. van
Schuppen, and A. C. M. Ran, pp. 471-480, Birkha€ user Boston, Berlin.
Scott, D. F., and D. B. Van Wyk (1990), The effects of wildfire on soil wettability and hydrological behaviour of an afforested catchment,
J. Hydrol., 121, 239–256.
Seibert, J., and J. J. McDonnell (2010), Land-cover impacts on streamflow: A change-detection modelling approach that incorporates
parameter uncertainty, Hydrol. Sci. J., 55(3), 316–332, doi:10.1080/02626661003683264.
Simon, D. (2006), Optimal State Estimation, Kalman, H Infinity, and Non-linear Approaches, 1st ed. Wiley-Interscience.
Simon, E., and L. Bertino (2012), Gaussian anamorphosis extension of the DEnKF for combined state parameter estimation: Application to a
1D ocean ecosystem model, J. Mar. Syst., 89(1), 1–18, doi:10.1016/j.jmarsys.2011.07.007.
Siriwardena, L., B. L. Finlayson, and T. A. McMahon (2006), The impact of land use change on catchment hydrology in large catchments:
The Comet River, Central Queensland, Australia, J. Hydrol., 326(1-4), 199–214, doi:10.1016/j.jhydrol.2005.10.030.
Smith, P. J., K. J. Beven, and J. A. Tawn (2008), Detection of structural inadequacy in process-based hydrological models: A particle-filtering
approach, Water Resour. Res., 44, W01410, doi:10.1029/2006WR005205.
Smith, P. J., G. D. Thornhill, S. L. Dance, A. S. Lawless, D. C. Mason, and N. K. Nichols (2013), Data assimilation for state and parameter esti-
mation: Application to morphodynamic modelling, Q. J. R. Meteorol. Soc., 139(671), 314–327, doi:10.1002/qj.1944.
Smith, T., A. Sharma, L. Marshall, R. Mehrotra, and S. Sisson (2010), Development of a formal likelihood function for improved Bayesian
inference of ephemeral catchments, Water Resour. Res., 46, W12551, doi:10.1029/2010WR009514.
Sorooshian, S., V. K. Gupta, and J. F. Fulton (1983), Evaluation of maximum likelihood parameter estimation techniques for and length on
model credibility, Water Resour. Res., 19(1), 251–259.
Thirel, G., V. Andr eassian, C. Perrin, J. N. Audouy, L. Berthet, P. Edwards, and J. Vaze (2015), Hydrology under change: An evaluation protocol
to investigate how hydrological models deal with changing catchments, Hydrol. Sci. J., 60(7–8), 1184–1199. doi:10.1080/
02626667.2014.967248.
Thyer, M., B. Renard, D. Kavetski, G. Kuczera, S. W. Franks, and R. Srikanthan (2009), Critical evaluation of parameter consistency and predic-
tive uncertainty in hydrological modelling: A case study using Bayesian total error analysis, Water Resour. Res., 45, W00B14, doi:10.1029/
2008WR006825.
Tomas-Rodriguez, M., and S. P. Banks (2010), Linear, Time-varying Approximations to Nonlinear Dynamical Systems: With Applications in Con-
trol and Optimization, Lect. Notes Control Inf. Sci., Springer-Verlag, Berlin Heidelberg, pp. 75–76.
Vrugt, J. A., H. V. Gupta, B. O. Nuallain, and W. Bouten (2006), Real-time data assimilation for operational ensemble streamflow forecasting,
J. Hydrometeorol., 7, 548–565.
Vrugt, J. A., C. J. F. ter Braak, C. G. H. Diks, and G. Schoups (2013), Hydrologic data assimilation using particle Markov chain Monte Carlo sim-
ulation: Theory, concepts and applications, Adv. Water Resour., 51, 457–478, doi:10.1016/j.advwatres.2012.04.002.
Walker, J. P., G. R. Willgoose, and J. D. Kalma (2001), One-dimensional soil moisture profile retrieval by assimilation of near-surface observa-
tions: A comparison of retrieval algorithms, Adv. Water Resour., 24(6), 631–650.
Wan, E. A., and R. Van Der Merwe (2000), The Unscented Kalman Filter for Nonlinear Estimation, in Adaptive Systems for Signal Processing,
Communications, and Control Symposium 2000, AS-SPCC, IEEE, Lake Louise, Alberta, Canada, pp. 153–158.
Wang, Y. C., P. S. Yu, and T. C. Yang (2010), Comparison of genetic algorithms and shuffled complex evolution approach for calibrating dis-
tributed rainfall-runoff model, Hydrol. Processes, 24(8), 1015–1026.
Weerts, A. H., and G. Y. H. El Serafy (2006), Particle filtering and ensemble Kalman filtering for state updating with hydrological conceptual
rainfall-runoff models, Water Resour. Res., 42, W09403, doi:10.1029/2005WR004093.
West, M. (1993), Mixture Models, Monte Carlo, Bayesian updating and dynamic models, in Proceedings of the 24th Symposium on the Inter-
face on Computing Science and Statistics, pp. 325-333, Interface Found. of North Am., Fairfax Station, Va.
Westerberg, I. K., J. L. Guerrero, P. M. Younger, K. J. Beven, J. Seibert, S. Halldin, and C.-Y. Xu (2011), Calibration of hydrological models using
flow-duration curves, Hydrol. Earth Syst. Sci., 15(7), 2205–2227, doi:10.5194/hess-15-2205-2011.

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3371

 Water Resources Research 10.1002/2015WR017192

Westra, S., M. Thyer, M. Leonard, D. Kavetski, and M. Lambert (2014), A strategy for diagnosing and interpreting hydrological model nonsta-
tionarity, Water Resour. Res., 50, 5090–5113, doi:10.1002/2013WR014719.
Wu, K., and C. A. Johnston (2007), Hydrologic response to climate variability in a Great Lakes Watershed: A case study with the SWAT
model, J. Hydrol., 337, 187–199.
Xie, X., S. Meng, S. Liang, and Y. Yao (2014), Improving streamflow predictions at ungauged locations with real-time updating: Application
of an EnKF-based state-parameter estimation strategy, Hydrol. Earth Syst. Sci., 18(10), 3923–3936, doi:10.5194/hess-18-3923-2014.
Yang, X., and T. Delsole (2009), Using the ensemble Kalman filter to estimate multiplicative model parameters, Tellus, Ser. A, 61(5), 601–609,
doi:10.1111/j.1600-0870.2009.00407.x.
Ye, W., B. C. Bates, N. R. Viney, M. Sivapalan, and A. J. Jakeman (1997), Performance of conceptual rainfall-runoff models in low-yielding
ephemeral catchments, Water Resour. Res., 33(1), 153–166.

PATHIRAJA ET AL. HYDROLOGIC MODELING IN DYNAMIC CATCHMENTS 3372