Journal of Hydrology 548 (2017) 291–304

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

A probabilistic drought forecasting framework: A combined dynamical

and statistical approach
Hongxiang Yan a, Hamid Moradkhani b,⇑, Mahkameh Zarekarizi b
a
Hydrology Technical Group, Pacific Northwest National Laboratory, Richland, WA, United States
b
Remote Sensing and Water Resources Lab, Department of Civil and Environmental Engineering, Portland State University, United States

a r t i c l e i n f o a b s t r a c t

Article history: In order to improve drought forecasting skill, this study develops a probabilistic drought forecasting
Received 5 October 2016 framework comprised of dynamical and statistical modeling components. The novelty of this study is
Received in revised form 9 January 2017 to seek the use of data assimilation to quantify initial condition uncertainty with the Monte Carlo ensem-
Accepted 4 March 2017
ble members, rather than relying entirely on the hydrologic model or land surface model to generate a
Available online 6 March 2017
This manuscript was handled by K.
single deterministic initial condition, as currently implemented in the operational drought forecasting
Georgakakos, Editor-in-Chief, with the systems. Next, the initial condition uncertainty is quantified through data assimilation and coupled with
assistance of Ashish Sharma, Associate a newly developed probabilistic drought forecasting model using a copula function. The initial condition
Editor at each forecast start date are sampled from the data assimilation ensembles for forecast initialization.
Finally, seasonal drought forecasting products are generated with the updated initial conditions. This
Keywords: study introduces the theory behind the proposed drought forecasting system, with an application in
Probabilistic drought forecasting Columbia River Basin, Pacific Northwest, United States. Results from both synthetic and real case studies
Data assimilation suggest that the proposed drought forecasting system significantly improves the seasonal drought fore-
Initial condition uncertainty casting skills and can facilitate the state drought preparation and declaration, at least three months
Particle Markov chain Monte Carlo before the official state drought declaration.
Published by Elsevier B.V.

1. Introduction drought is necessary. A drought forecasting system which provides

forecasts of potential drought conditions in a timely manner is crit-
Drought is a naturally occurring climate phenomenon driven by ical to stakeholders and decision-makers. In the U.S., there are few
extreme macroclimatic variability originating from atmospheric federal drought forecasting systems, including the National Ocea-
interactions and feedback between the atmosphere, the oceans, nic and Atmospheric Administration (NOAA) Climate Prediction
and the land surface (Ahmadalipour et al., 2017; Keyantash and Center’s (CPC’s) Seasonal Drought Outlook (SDO) (Steinemann,
Dracup, 2002; Mishra and Singh, 2010; Yan, 2016). Contrary to 2006) and the National Integrated Drought Information System’s
other natural disasters such as floods, hurricanes, and tornados, (NIDIS’s) Drought Early Warning Systems (DEWS). Despite these
droughts develop slowly over large spatial extents and last for efforts, the World Climate Research Program still recognizes that
multiple years. It is usually difficult to identify droughts until sev- seasonal drought forecasting is one of the major research gaps in
ere damage has already occurred (Luo and Wood, 2007; Mishra hydroclimatology (WCRP, 2010).
and Singh, 2011). Moreover, drought is the most costly extreme Drought forecasting generally relies on drought indicators com-
event among all natural disasters (NCDC, 2012; Ross and Lott, puted by either dynamical model simulations or statistical meth-
2003; Sheffield et al., 2014). For instance, Howitt et al. (2014) esti- ods (DeChant and Moradkhani, 2014; Kumar et al., 2014b; Mao
mated that the recent California (CA) extreme drought resulted in et al., 2015; Wood and Lettenmaier, 2008). One option to generate
an economic loss of $2.7 billion in 2014; Washington Department hydrologic prediction is the Ensemble Streamflow Prediction (ESP)
of Agriculture (2015) calculated that the economic loss of the 2015 framework (Day, 1985). The ESP is generated by deriving the
state drought was more than $335 million. hydrologic modeling using resampled historical meteorological
In order to mitigate the extensive negative socio-economic variables and future drought condition is estimated accordingly.
impacts, comprehensive preparation and effective response to However, the sampled forcing data is not necessarily representa-
tive of the non-stationary climate (Milly et al., 2008; Yan and
⇑ Corresponding author. Edwards, 2013). Another option is to use dynamical forecasting
E-mail address: hamidm@pdx.edu (H. Moradkhani). products, such as the Climate Forecast System Version 2 (CFSv2)

http://dx.doi.org/10.1016/j.jhydrol.2017.03.004
0022-1694/Published by Elsevier B.V.
292 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304

products (Yuan et al., 2011) and the North American Multimodel dynamical hydrologic modeling, the DA implementation, and the
Ensemble (NMME) climate forecasts (Yuan, 2016). While dynami- statistical modeling using copula functions to generate the proba-
cal forecast provides future climate information, the precipitation bilistic forecasts. Section 3 applies the proposed forecasting system
forecasts are subject to high uncertainty and exhibit low skill with a case study in the Columbia River Basin, Pacific Northwest.
beyond one month lead time (Hayes et al., 2005; Lavers et al., Finally, section 4 provides the conclusion.
2009; Yuan et al., 2013). As an alternative, a recently developed
probabilistic drought forecasting approach built on copula func-
tions (Madadgar and Moradkhani, 2013) could be used. Recent 2. Probabilistic drought forecasting system framework
studies have shown that such a statistical approach could lead to
good seasonal drought forecasting skill (Chen et al., 2015; The framework of the proposed probabilistic drought forecast-
Madadgar and Moradkhani, 2016, 2014; Mishra and Desai, 2005). ing system is described here. The system is composed of three main
Recent works show that seasonal hydrologic forecast skill is components: (1) hydrologic modeling, (2) data assimilation, and (3)
mostly controlled by hydrologic initial condition (DeChant and a copula-based probabilistic drought forecasting model. The hydro-
Moradkhani, 2015, 2011; Koster et al., 2010; Li et al., 2009; logic model is first calibrated for the study region. Then the copula-
Shukla and Lettenmaier, 2011; Wood et al., 2016; Wood and based probabilistic drought forecasting model is developed based
Lettenmaier, 2008; Yossef et al., 2013; Yuan et al., 2016). For on the long-term open-loop (OL) simulations that is simulation
instance, the initial condition refers to the root-zone soil moisture with no assimilation, of hydrologic observations. Next, the DA sys-
for agricultural drought forecasting. Initial condition uncertainty tem is used to assimilate the remotely sensed and in-situ observa-
arises from the chaotic properties of the Earth system and is tions to improve hydrologic simulation and characterize the initial
unavoidable due to inability to accurately observe land surface condition uncertainty by estimating the probability density func-
states (Stainforth et al., 2005). In the western U.S., the accurate tion (PDF) of the initial conditions. Last, the initial conditions sam-
estimation of initial condition uncertainty is particularly important pled from the PDF are used in the copula-based probabilistic
due to ongoing coverage of drought in California and Pacific North- drought forecasting model to generate seasonal drought forecasts.
west. However, the initial condition uncertainty is not considered Fig. 1 displays the proposed drought forecasting framework.
in the aforementioned drought forecasting systems, where a single
deterministic hydrologic initial condition is utilized. Therefore, in 2.1. Hydrologic modeling
order to improve drought forecasting skill, this study seeks the
use of ensemble data assimilation (DA) to improve land initializa- The Precipitation-Runoff Modeling System (PRMS) (Leavesley
tion by quantifying the initial condition uncertainty (DeChant and et al., 1983) is used in this study to model the land surface states.
Moradkhani, 2011; Yan et al., 2015; Yan and Moradkhani, 2016). The PRMS is a modular deterministic, distributed-parameter, and
In summary, the scope of this study is to develop a probabilistic physical-process watershed model (Markstrom et al., 2008). The
drought forecasting system, with the use of DA and a copula-based land surface hydrologic process simulated by PRMS includes the
probabilistic drought forecasting method. It is hypothesized that evapotranspiration, runoff, infiltration, interflow, snowpack, and
an accurate quantification of uncertainty in hydrologic initial con- soil moisture. Instead of delineating a watershed into uniform grid
dition would lead to a better drought forecasting skill. The paper is cells, the PRMS partitions a watershed into hydrologic response
organized as follows. Section 2 describes the framework of the pro- units (HRUs) that are based on the physical attributes of the water-
posed probabilistic drought forecasting system, including the shed such as land-surface elevation, slope and aspect, vegetation

Meteorological Hydrologic
Probabilisc
Forcing Modeling
Inial
Condion

Remotely
DA
Sensed Data

Probabilisc Mulvariate
Sampled Inial
Drought Copula
Condion
Forecast Probabilisc
Drought
Forecasng
Model

Dynamical Modeling Stascal Modeling Dataset

Data Assimilaon Modular Inial Condion Uncertainty Sampled Inial Condion

Fig. 1. The framework of the proposed dynamical-statistical probabilistic drought forecasting system. For each forecasting initialization date, the data assimilation (DA)
technique is used to account for the initial condition uncertainty, in terms of a probability density function (PDF). An updated initial condition is then sampled from the PDF
and input into the multivariate copula drought forecasting model to generate the probabilistic drought forecasts.
 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304 293

type, soil type, and spatiotemporal climate patterns (Markstrom sampling importance resampling (PF-SIR) (Moradkhani et al.,
et al., 2015). The physical attributes and hydrologic response of 2005) and uses the PF-SIR to design efficient high-dimensional pro-
each HRU are assumed to be homogeneous. The meteorological posal distributions for MCMC algorithm. Despite the reasonable
forcing data for PRMS are precipitation, minimum temperature performance of the EnKF, the PF was found to be more robust since
and maximum temperature. Excess runoff is routed to the outlet it can relax the Gaussian error assumption, preserve water balance,
through the cascade and Muskingum routing methods. and provide a more complete representation of state posterior dis-
The PRMS version 4.0.1 (PRMS-IV) released on March 15, 2015 tribution for a nonlinear non-Gaussian hydrologic system
was used in this study. The PRMS-IV takes soil moisture into (DeChant and Moradkhani, 2012; Dong et al., 2015; Plaza et al.,
account in three reservoir systems: the preferential-flow reservoir, 2012; Yan and Moradkhani, 2016).
the capillary reservoir, and the gravity reservoir. The preferential-
flow reservoir represents the water content between
preferential-flow threshold and total soil saturation; it is available 2.2.1. PF-SIR algorithm
for fast interflow and Dunnian surface runoff. The capillary reser- Following Moradkhani (2008), the state-space models that
voir represents the moisture content between wilting point and describe the generic non-linear earth system are as follows:
field capacity; it is only available for evapotranspiration and not
for drainage. The gravity reservoir is limited to the water content
yt ¼ hðxt Þ þ v t ð1Þ
between field capacity and preferential-flow threshold. The water
content in this reservoir is available for slow interflow, groundwa- xt ¼ f ðxt1 ; ut ; hÞ þ wt ð2Þ
ter recharge, and Dunnian surface runoff (Markstrom et al., 2015).
where xt 2 Rn is a vector of the uncertain state variables at current
2.2. Data assimilation algorithm time step, yt 2 Rm is a vector of observation data, ut is the uncertain
forcing data, h 2 Rd is a vector of model parameters, hðÞ is a non-
The ensemble DA system is used to quantify the initial condi- linear function that relates the states xt to the observations yt , wt
tion uncertainty. The recently developed particle filter Markov represents the model error, and v t indicates the observation error.
chain Monte Carlo (PMCMC) (Moradkhani et al., 2012) is used in The errors wt and v t are assumed to be white noise with mean zero
this study. The PMCMC is an extension of the particle filter with and covariance Q t and Rt , respectively.

Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun 1-Jan

Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun 1-Feb

Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun
1-Mar

Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun
1-Arp

SEASONAL INTER-ANNUAL 1st SEASONAL 2nd SEASONAL 3rd SEASONAL

FORECASTING FORECASTING OUTLOOK OUTLOOK OUTLOOK
Predicon for this Predicon for next Predicon for Predicon for Predicon for
season based on half a year based next season next season next season
last season. on last half a year. given last season. given last season given last
Performed on 1- Performed on 1- Performed on 1- + last month. season + last
Jan, 1-Apr, 1-Jul, Jan, 1-Apr, 1-Jul, Jan, 1-Apr, 1-Jul, Performed on 1- two months.
and 1-Oct and 1-Oct and 1-Oct Feb, 1-May, 1- Performed on
Aug, 1-Nov 1-Mar, 1-Jun,
1-Sep, 1-Dec

Fig. 2. The seasonal and inter-annual drought forecasting framework.

294 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304

where wit is the prior particle weights, and pðyt jxt Þ can be com-
According to Moradkhani et al. (2005), the posterior distribu- i

tion of the state variables xt given a realization of the observations puted from the likelihood Lðyt jxt Þ. Generally, a Gaussian distribution
i

y1:t is written as follows: is used to estimate Lðyt jxit Þ:

pðyt jxt Þpðxt jy1:t1 Þ  

1 1 i T 1 i
pðxt jy1:t Þ ¼ pðxt jy1:t1 ; yt Þ ¼
pðyt jy1:t1 Þ Lðyt jxit Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp  ð v Þ R ð v Þ ð9Þ
ð2pÞm jRt j 2 t t t

pðyt jxt Þpðxt jy1:t1 Þ

¼R ð3Þ
pðyt jxt Þpðxt jy1:t1 Þdxt where v it ¼ yt  hðxit Þ is the residual. The estimate of the normaliza-
tion factor pðyt jy1:t1 Þ in Eq. (6) is given by:
Z
pðxt jy1:t1 Þ ¼ pðxt ; xt1 jy1:t1 Þdxt1 X
N X
N

Z pðyt jy1:t1 Þ  t pðyt jxt Þ ¼

wi i
t Lðyt jxt Þ
wi i
ð10Þ
i¼1 i¼1
¼ pðxt jxt1 Þpðxt1 jy1:t1 Þdxt1 ð4Þ
To obtain approximate samples from pðxt jy1:t Þ, a resampling
where pðyt jxt Þ is the likelihood, pðxt jy1:t1 Þ is the prior distribution, operation is necessary to reduce the weight degeneration problem
and pðyt jy1:t1 Þ is the normalization factor. The marginal likelihood where all but one of the importance weights are close to zero. The
pðy1:t Þ can be computed as: SIR algorithm suggests resampling the particles with a probability
Y greater than the uniform probability. After resampling, all the par-
pðy1:t Þ ¼ pðy1 Þ pðyt jy1:t1 Þ ð5Þ ticle weights are set equal to 1=N. Based on the above descriptions,
the PF-SIR algorithm can thus be summarized as follows:
where the normalization factor pðyt jy1:t1 Þ is:
Z Z At time t ¼ 1,
pðyt jy1:t1 Þ ¼ pðyt ; xt jy1:t1 Þdxt ¼ pðyt jxt Þpðxt jy1:t1 Þdxt ð6Þ
 Initialize the system by sampling the model states
In practice, Eq. (3) does not have an analytical solution except xit ; i ¼ 1; . . . ; N from a uniform distribution
for few special cases (e.g., the linear system with Gaussian assump-  Assign the particle weights uniformly wiþ t ¼ wtþ1 ¼ 1=N;
i

tion). Instead, the posterior distribution pðxt jy1:t Þ is approximated i ¼ 1; . . . ; N

using a set of Monte Carlo (MC) random samples as:  At time t P 2,
 Propagate the N model states forward in time xit through model
X
N operator
pðxt jy1:t Þ  t dðxt  xt Þ
wiþ i
ð7Þ pðyt jxit Þ
t ¼ w t PN
 Update the particle weights wiþ i
i¼1 wi pðyt jxit Þ
i¼1 t

 Resample to obtain N new equally-weighted particles xit with

where wiþt is the posterior weight of the i th particle, d is the Dirac
t ¼ 1=N; i ¼ 1; . . . ; N
wiþ
delta function, and N is the ensemble size. The normalized weights
are calculated as follows:
2.2.2. PMCMC algorithm
pðyt jxit Þ
t ¼ wt XN
wiþ ð8Þ
i
The PMCMC is an extension of the particle filter with sampling
t pðyt jxt Þ
wi i
i¼1
importance resampling (PF-SIR) and uses the PF-SIR to design effi-
cient high-dimensional proposal distributions pðxt jy1:t Þ for MCMC
algorithm. The PMCMC has the following three steps:

Fig. 3. The location of the Columbia River Basin and the Kling-Gupta efficiency (KGE) values for the 146 No Regulation No Irrigation (NRNI) streamflow gauges.
 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304 295

 
(1) Initialization (j ¼ 0): run PF-SIR targeting pðxt jy1:t Þ, sample pðy1:t Þ
min 1; ð11Þ
X t ð0Þ  pðxt jy1:t Þ and let pðy1:t Þð0Þ denote the corresponding pðy1:t Þðj  1Þ
marginal likelihood estimate.
and set X t ðjÞ ¼ X t and pðy1:t ÞðjÞ ¼ pðy1:t Þ ; otherwise set
(2) Iteration (j P 1): sample X t  pðxt jy1:t Þ again and let pðy1:t Þ
X t ðjÞ ¼ X t ðj  1Þ and pðy1:t ÞðjÞ ¼ pðy1:t Þðj  1Þ. The marginal likeli-
denote the corresponding marginal likelihood estimate.
hood is estimated based on the Eqs. (5) and (10).
(3) Calculate the acceptance ratio as:

2.3. Copula-based probabilistic drought forecasting model

A newly developed probabilistic drought forecasting method

based on copula functions (Madadgar and Moradkhani, 2014,
2013) is used here which is coupled with the initial condition
uncertainty through DA. Since drought indicators are usually sta-
tistically dependent on their past status, they can be expressed
within the Bayesian networks, which are capable of describing
the conditional dependencies of two random variables. The core
of this forecasting model is to apply a multivariate distribution
function to forecast future drought conditions given the current
or past drought status.
Copulas are multivariate distribution functions on the n -
dimensional unit cube. The variables of copula functions are uni-
formly distributed with ½0; 1. Following Sklar’s theorem (Sklar,
1959), a multivariate distribution Pðx1 ; . . . xn Þ can be expressed by
copula functions as follows:
Pðx1 ; . . . xn Þ ¼ C½Pðx1 Þ; . . . ; Pðxn Þ ¼ Cðu1 ; . . . ; un Þ ð12Þ
where C is the cumulative distribution function (CDF) of copula, and
Pðxi Þ is the marginal distribution of xi being uniform with ½0; 1,
which is denoted by ui . If the C½Pðx1 Þ; . . . ; Pðxn Þ is absolutely contin-
uous, its PDF can be written as:
@ n Cðu1 ; . . . ; un Þ
cðu1 ; . . . ; un Þ ¼ ð13Þ
@u1 . . . @un
Using the PDF of the copula, the joint PDF of pðx1 ; . . . xn Þ can be
expressed as:
Y
n
pðx1 ; . . . xn Þ ¼ cðu1 ; . . . ; un Þ pðxi Þ ð14Þ
i¼1

For the drought forecasting application, only two random vari-

Fig. 4. The normalized information contribution (NIC) value. The positive value ables are modeling (n ¼ 2Þ. As a result, the conditional probability
indicates that the DA improves soil moisture prediction against OL; negative value distribution of xtþ1 given xt is written as:
indicates the degradation over the OL.

Fig. 5. Comparison of the basin-averaged daily root-zone soil moisture (m3/m3) by the open-loop (OL) and data assimilation (DA) for fall 2014 and winter/spring/summer
2015 across the CRB. The error bars show the 95% prediction intervals.
296 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304

pðxtþ1 ; xt Þ Using Eq. (16), the probabilistic forecast of drought states in

pðxtþ1 jxt Þ ¼ ð15Þ
pðxt Þ time t þ 1 given the drought condition at time t can be examined.
The mode of the pðxtþ1 jxt Þ is correspondent to the maximum like-
Replacing the joint PDF pðxtþ1 ; xt Þ in Eq. (15) with Eq. (14), the
lihood estimation (MLE) of xtþ1 . For seasonal drought forecasting,
conditional probability of Eq. (15) can be revised as:
the xtþ1 represents the drought condition in the coming season,
pðxtþ1 ; xt Þ cðutþ1 ; ut Þpðxtþ1 Þpðxt Þ and xt indicates the initial drought condition in the current season.
pðxtþ1 jxt Þ ¼ ¼
pðxt Þ pðxt Þ Following Madadgar and Moradkhani (2013), to build the condi-
¼ cðutþ1 ; ut Þpðxtþ1 Þ ð16Þ tional PDF of Eq. (16), the Monte Carlo (MC) approach is used to
sample from the copula density function cðutþ1 ; ut Þ. Different

Fig. 6. Seasonal probabilistic drought forecasting for both OL and DA for different seasons in 2015 given the drought status in each of the previous seasons.
 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304 297

copula functions can be used to join the marginal distributions of the root-zone soil moisture initial condition uncertainty with the
correlated and dependent variables. Among various copula fami- assimilation of satellite surface soil moisture for improving agricul-
lies, the Archimedean and elliptical families are usually used in tural drought forecasting. Similar to what has been explained here,
hydrological applications (Madadgar and Moradkhani, 2016). The other remotely sensed or in-situ data, e.g. total water storage and
Gaussian and t copulas from the elliptical family, and the Gumbel precipitation, can also be assimilated in the same way to improve
and Clayton from the Archimedean family, can be tested with the forecasting of hydrological and meteorological droughts.
parameters estimated by the inference function for margins
(IFM) method (Joe, 1997). To select the appropriate copula function
among the various copula functions, a goodness-of-fit (GOF) test
can be used, such as the parametric bootstrapping GOF test 3.1. Study area
(Genest and Rémillard, 2008).
Based on the descriptions in Sections 2.1–2.3, the proposed The Columbia River Basin (CRB), located in Pacific Northwest
dynamical-statistical drought forecasting system can thus be sum- (PNW), covers about 674,500 km2 of U.S. (85%) and Canada
marized as follows: (15%) (Beechie and Imaki, 2014). In the U.S., the CRB spans seven
At each drought forecasting initialization date t, states, including Washington, Oregon, Idaho, Montana, Utah,
Wyoming, and Nevada (Fig. 3). The CRB encompasses a wide range
 Download the forcing data until the initialization date of physiographic provinces and ecoregions ranging from semiarid
 Run the dynamical hydrologic model to estimate long-term ret- in central plateaus to wet forests in the Cascade Mountains
rospective root-zone soil moisture until the initialization date (Omernik and Bailey, 1997). The mean annual precipitation ranges
(modeling period: >30 years) from about 200 mm in central plateaus to about 3550 mm in Cas-
 Develop the copula-based statistical forecasting model based on cade Mountains (Daly et al., 2002). The Columbia River is the 4th
the long-term retrospective dataset largest river in North America, originating in the Rocky Mountain
 Download the remotely sensed surface soil moisture until the and flowing to the Pacific Ocean, with a mean annual discharge
initialization date of 247 million cubic meters (Cosens and Williams, 2012). >450
 Run the DA system to assimilate remotely sensed surface soil dams have been built on this river system to provide hydroelectric-
moisture until the initialization date (assimilation period: sev- ity, flood control, irrigation, and stream regulation. The Columbia
eral months) River is dominated by snow hydrology: snow accumulating in win-
 Quantify the initial condition uncertainty in terms of a PDF ter and melting in spring. It generally shows a characteristic of low
through the ensemble members from DA flow in winter and peak flow in spring (Hamlet and Lettenmaier,
 Sample initial condition from the PDF and input into the copula- 1999). Recently, the CRB droughts have received increasing atten-
based forecasting model tion due to the low snowpack and rising temperature. For instance,
 Generate probabilistic drought forecasts the Washington Department of Agriculture (2015) calculated that
the economic loss of the 2015 state drought was more than $335
2.4. Seasonal to inter-annual drought forecasting framework million.

Based on Fig. 2, the proposed probabilistic drought forecasting

system has three components: (1) Seasonal drought forecasting. This
is the official seasonal drought forecasting product for decision-
makers and water resources managers. The system generates fore-
casts for the current season based on last season and is performed
on 1-January (winter forecasting), 1-April (spring forecasting),
1-July (summer forecasting), and 1-October (fall forecasting). (2)
Seasonal drought outlook. The seasonal outlook products offer the
seasonal drought information in an earlier time manner (up to
six-month preparation time). There are three seasonal outlook
products in the proposed system: the 1st, 2nd, and 3rd seasonal
outlook. The 1st seasonal outlook products provide drought infor-
mation for the next season given the drought status in last season.
When more information becomes available, the system updates
and generates the 2nd and 3rd seasonal drought outlook products
with four and five months of information. Due to the persistence of
soil moisture memory, the seasonal drought forecasting is
expected to provide better results than seasonal drought outlook.
(3) Inter-annual drought forecasting. The inter-annual drought fore-
casting products estimate the drought information for the next half
a year. It is performed on 1-January, 1-April, 1-July, and 1-October.
The inter-annual drought forecasting can provide long-term
drought information for decision-makers.

3. Case study

In general, droughts can be classified as meteorological drought

Fig. 7. The absolute drought extent bias between the OL/DA and synthetic truth for
(precipitation deficit), agricultural drought (soil moisture deficit), the seasonal forecasting for winter/spring/summer/fall 2015 (top-panel), seasonal
and hydrological drought (streamflow/groundwater deficit) (Van outlook for summer 2015 and inter-annual forecasting for April-September 2015
Loon, 2015). In this study, we examine the impacts of quantifying (bottom-panel).
298 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304

3.2. Data sources to the finer scale 1/8th degree NLDAS grid and to the one-hour
NLDAS temporal resolution (Xia et al., 2012). In this study, the
The meteorological forcing data, precipitation, maximum and hourly NLDAS-2 primary forcing data were first aggregated into
minimum temperature (January 1, 1979 to December 31, 2015) daily time step and then upscaled/downscaled for each HRU in
were acquired from the Phase 2 of the North American Land Data CRB.
Assimilation Systems (NLDAS-2) (Xia et al., 2014). The majority The blended microwave soil moisture climate change initiative
of NLDAS-2 atmospheric forcing data are derived from the North (CCI) products v02.2 released on February 2016 (Liu et al., 2012)
American Regional Reanalysis (NARR) which features a 32 km spa- and the Advanced Microwave Scanning Radiometer2 (AMSR2) soil
tial resolution and a three-hour temporal resolution. The NLDAS moisture products (Imaoka et al., 2010) were used in this study.
software is used to interpolate the coarse resolution NARR data The CCI soil moisture were merged from four passive and two

Fig. 8. The 1st, 2nd, and 3rd probabilistic seasonal outlook results between OL and DA for summer 2015 (according to the framework shown in Fig. 2). The inter-annual
forecasting for spring and summer 2015 given the drought status in fall 2014 and winter 2015.
 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304 299

active microwave products (Dorigo et al., 2015), including the A combination of two programs created by the USGS was used
Scanning Multichannel Microwave Radiometer (SMMR), Special in calibration, LUCA and LUMEN (Hay and Umemoto, 2006).
Sensor Microwave Imager (SSM/I), Tropical Rainfall Measure Mis- LUMEN assists in model structure during calibration so that the
sion (TRMM) Microwave Imager (TMI), Advanced Microwave Scan- model can more easily be calibrated using a top-down approach.
ning Radiometer for Earth Observing System (AMSR-E), Advanced LUCA uses the Shuffled Complex Evolution (SCE) global search
Microwave Instrument (AMI), and Advanced Scatterometer algorithm to calibrate the 31 model parameters (Duan et al.,
(ASCAT). The blended CCI data cover the period of 1978–2014. 1994). Fig. 3 presents the Kling-Gupta efficiency (KGE) (Gupta
The AMSR2 is onboard the Global Change Observation Mission1- et al., 2009) values for the 146 NRNI control points from January
Water (GCOM-W1) satellite which was launched in May 2012 by 1, 1984 to September 30, 2008. The majority of the gauges show
the Japan Aerospace Exploration Agency (JAXA). The AMSR2 soil the KGE values greater than 0.7, which indicates the good perfor-
moisture products, generated using the Land Parameter Retrieval mance of the calibrated model.
Model (LPRM) developed by Vrije Universiteit (VU) Amsterdam
and NASA Goddard Space Flight Center (GSFC) (Owe et al., 2008),
3.4. Synthetic study
were employed in this study. These two products are selected
due to the data availability. The latest Soil Moisture Active Passive
To objectively assess the potential benefit of quantification of
(SMAP) (Entekhabi et al., 2010) L-band soil moisture products
initial condition uncertainty in drought forecasting, a synthetic
issued from March 31, 2015 will be used for future operational
study is first conducted. Following Moradkhani (2008), the syn-
drought forecasting. All the satellite soil moisture products were
thetic study includes the following steps: (1) a ‘‘truth” run of PRMS
upscaled/downscaled for each HRU in CRB. The inverse distance
with the pre-calibrated model parameters; (2) synthetic satellite
weighting (IDW) spatial interpolation technique is used in this
soil moisture observations, which are generated from the truth
study to downscale/upscale the satellite soil moisture into HRU
run by incorporating the observation errors; (3) ensemble OL run
scale.
with perturbed forcing data without DA; (4) the DA integration
Calibration of the PRMS was performed on a daily timescale uti-
that assimilates the synthetic satellite observations. The proba-
lizing a combination of unregulated U.S. Geological Survey (USGS)
bilistic initial conditions from OL and DA are then used in the cop-
streamflow data, as well as No Regulation No Irrigation (NRNI)
ula model to generate drought forecasting products. Finally, the OL
streamflow data provided by Bonneville Power Administration
and DA forecasting results are compared against the true simula-
(BPA) [https://www.bpa.gov/power/streamflow/]. The BPA-NRNI
tions to evaluate the impact of initial condition uncertainty
data cover daily streamflow from the period July 1, 1928 to
through DA. The copula probabilistic drought forecasting model
September 30, 2008, and were generated in a joint effort between
is developed based on the truth run of root-zone soil moisture. It
the U.S. Army Corps of Engineers (USACE) and the Bureau of Recla-
is noted that in the following discussions, the ‘‘DA” experiment
mation (BOR). The NRNI datasets emulate daily discharge gauges,
indicates the results with the assimilation of satellite surface soil
typically at manmade control structures, where estimated and
moisture, while the ‘‘OL” indicates the results without assimilation
measured inputs and outputs can be summed to produce daily
of satellite observations, hence the uncertainty in initial condition
streamflow data.
is not accounted for. Both DA and OL results are achieved through
hydrologic modeling and copula-based statistical modeling.
The DA was performed by assimilation of the synthetic satellite
3.3. Model calibration
surface soil moisture for the period of October 1, 2014 to Septem-
ber 30, 2015. We focused on hindcasting the drought events in
Due to the CRB’s significant extent into British Columbia, the
2015, since the CRB received historically low snowpack in this year
HRU delineation was completed in two ways. The HRUs for the
and drought emergences had been declared in Oregon (OR) and
U.S. portion were provided by the PRMS Geospatial Fabric (Viger,
Washington (WA) states in spring 2015. Considering the satellite
2014). For the area of the CRB inside British Columbia, the ESRI Arc-
data availability, we employed the CCI soil moisture products for
Map 10.3.1 was used along with a digital elevation model to pro-
DA in 2014 and AMSR2 retrievals for 2015. Contrary to the small
duce stream segment lines as well as watershed delineations.
Due to a lack of data to calibrate the Canadian portion of the
CRB, the HRUs are rather large, relative to the U.S. portion. The
two spatial shapefiles (U.S. and Canadian) were then stitched
together, ensuring no overlapping HRU areas and continuity
between stream segments between the U.S. and Canada border.
As a result, a total of 7739 HRUs and 4019 stream segments were
delineated for the CRB.
In this study, 146 NRNI data sets were used as the primary con-
trol points for calibration along with over 300 selected USGS
streamflow gauges. The NRNI data from January 1, 1979 to Decem-
ber 31, 2000 were used for calibration and January 1, 2001 to
September 30, 2008 for validation. Calibrations for USGS stream-
flow gauges varied in length due to lapsing streamflow gauge oper-
ation, however only data within the range of January 1, 1979 to
December 31, 2010 were used for calibration/validation. Monthly
averaged normal incident solar radiation atlas data [http://www.
nrel.gov/gis/data_solar.html] and monthly averaged evaporation
atlas (Farnsworth et al., 1982) were used for calibration of solar
radiation (SR) and potential evapotranspiration (PET) parameters
in the U.S. portion of the model. The SR and PET parameters for Fig. 9. Comparison of the basin-averaged daily root-zone soil moisture (m3/m3) by
the Canadian portion of the CRB were extrapolated from the U.S. the open-loop (OL) and data assimilation (DA) for spring 2013 and winter 2015
portion. across the CRB. The error bars show the 95% prediction intervals.
300 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304

ensemble size (12–20) used in the majority of previous satellite In the DA implementation, the precipitation was perturbed with
soil moisture DA studies (De Lannoy et al., 2012; Kumar et al., a lognormal distribution with a coefficient of variation of 0.25, and
2014b, 2009; Pan and Wood, 2010; Reichle et al., 2010), a large the minimum and maximum temperature were assumed to follow
ensemble size of 200 was used in this study to fully quantify the normal distributions with a coefficient of variation of 0.25. These
soil moisture posteriors. For the purpose of better visualization, values are suggested to account for errors in meteorological mea-
all PRMS simulation results were downscaled/upscaled into NLDAS surements due to spatial heterogeneity and sensor errors
1/8th degree grid cells. Results are presented only for the U.S. por- (DeChant and Moradkhani, 2015; Yan et al., 2015). The white noise
tion of CRB in this study. (standard deviation) for the CCI and AMSR2 soil moisture were

Fig. 10. The seasonal probabilistic drought forecasts for summer 2013 and spring 2015 given the drought status in spring 2013 and winter 2015, respectively. The top-panel
shows the forecasted drought areas based on MLE and the bottom-panel indicates the forecasted drought probabilities.
 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304 301

0.04 and 0.08 m3/m3, respectively, according to Kumar et al. by Madadgar and Moradkhani (2014). The copula forecasted
(2014b). Prior to DA, we scaled the CCI and AMSR2 standard devi- drought conditions using the OL and DA initial conditions were
ations by the ratio of the soil moisture time series standard devia- compared against the corresponding synthetic truth. It is noted
tion of the PRMS model as suggested by Reichle et al. (2007) and that the OL and DA initial conditions were sampled as the posterior
Liu et al. (2011). After scaling, the synthetic satellite soil moisture mean values.
observations were generated by perturbing the synthetic truths Fig. 6 presents the spatial distributions of seasonal drought
with a normal distribution with the scaled standard deviation. forecasting probabilities of OL and DA for winter/spring/summer/-
The effects of DA on soil moisture estimation. The improvement or fall 2015. The probabilistic drought conditions in each season were
degradation of DA on soil moisture predictions are assessed using forecasted using the root-zone soil moisture in the previous sea-
the normalized information contribution (NIC) (Kumar et al., son. The absolute drought extent bias (%) between the synthetic
2014b). The NIC for root-mean-square-error (RMSE) is defined as truth and MLE forecasted droughts are shown in the top panel of
follows: Fig. 7. Generally, compared with the synthetic truth, both OL and
DA seasonal forecasting products showed high probabilities
RMSEOL  RMSEDA
NIC ¼ ð17Þ (>30%) for the major drought locations in the following seasons.
RMSEOL
These results indicate the efficiency of the copula model in sea-
where RMSEOL indicates the RMSE values between OL and syn- sonal drought forecasting. Based on the drought extent bias, the
thetic truth, RMSEDA indicates the RMSE values between DA and DA estimates showed systematic improvements over the OL esti-
synthetic truth. If NIC > 0, the DA improves the OL skill; if mates in the four seasons. For instance, the drought extent bias
NIC = 0, the DA does not add any skill; if NIC < 0, the DA degrades between the MLE-OL and synthetic truth was 29.87% for winter
the OL skill; and if NIC = 1, the DA achieves the maximum skill. 2015 forecasting, and it decreased to 20.71% with MLE-DA. Simi-
Fig. 4 presents the NIC values in the surface and root-zone soil larly, drought extent bias decreased from 4.73% with MLE-OL to
moisture and their spatial distributions across the CRB. It is noted 0.95% with MLE-DA for 2015 summer.
that the NIC values for OL and DA are generated using the ensem- Fig. 8 shows the spatial distributions of three seasonal drought
ble mean estimates. From Fig. 4, the majority of the grid cells show outlook and inter-annual drought forecasting probabilities of OL
the positive NIC values, which indicates the added-value of the DA. and DA for summer 2015 and April-September (spring and sum-
Generally, the improvements in the surface soil moisture field are mer) 2015, respectively. The absolute drought extent bias (%)
consistent with the improvements in the root-zone soil moisture between the synthetic truth and MLE forecasted droughts are
field. For surface soil moisture, the daily basin-averaged RMSE shown in the bottom panel of Fig. 7. For seasonal outlook, the
(m3/m3) for the OL was 0.0213, and it decreased to 0.0114 with drought extent bias decrease when more information becomes
DA. Similarly, the daily basin-averaged root-zone soil moisture available. For OL, the drought extent bias for the 1st outlook was
RMSE value decreased from 0.0194 in the OL to 0.0116 in the DA. 33.13% for summer 2015, and it decreased to 29.67% and 24.39%
The improvements in surface field are higher than root-zone field, for the 2nd and 3rd outlook, respectively. For DA, the drought
which is consistent with the previous soil moisture synthetic stud- extent bias for the 1st, 2nd, and 3rd outlooks were 28.89%,
ies (Kumar et al., 2009, 2012, 2014a). 23.63%, and 17.13%, respectively. For summer 2015, the drought
The effects of DA on drought forecasting. Prior to investigating the extent bias for seasonal drought foresting is the lowest, which is
drought forecasting results, it is necessary to compare the initial due to the persistence of soil moisture memory. The seasonal
conditions generated by the OL and DA. This comparison can be drought forecasting provides better results than the seasonal
seen in Fig. 5. In this figure, the initial conditions for seasonal drought outlook. The inter-annual drought forecasting for April-
drought forecasting beginning on January 1 (winter forecasting), September 2015 shows the highest drought extent bias. This result
April 1 (spring forecasting), July 1 (summer forecasting), and Octo- indicates the large uncertainty associated with the long-term
ber 1, 2015 (fall forecasting) are presented. Each sub-plot contains drought forecasting.
the basin-averaged daily root-zone soil moisture for the synthetic
truth, shown as a single value; the OL and the DA, shown as a dis-
tribution of values, which represent the probability distribution of 3.5. Real case study
initial root-zone soil moisture values. From Fig. 5, it is observed
that the OL and DA display very different behavior for each season. When assimilating the real satellite soil moisture data, the sys-
For all four seasons, the DA root-zone soil moisture reduce the tematic biases between the satellite-based and model-based soil
uncertainty of OL estimations and the mean root-zone soil mois- moisture cannot be avoided (Reichle and Koster, 2004; Su et al.,
ture is closer to the synthetic truth for the DA. 2014; Yan et al., 2015; Yilmaz and Crow, 2013). Proper treatment
In this study, drought is characterized with the root-zone soil of these systematic biases is important, as the DA algorithm is
moisture percentile (Mao et al., 2015; Shukla et al., 2011; Wang designed to work with errors that are strictly random (Dee and
et al., 2011) and drought intensity is classified based on the
National Drought Mitigation Center (NDMC) United States Drought
Monitoring (USDM) (Svoboda et al., 2002). Five categories are
defined: D0 (abnormally dry, percentile 30%), D1 (moderate
drought, percentile 20%), D2 (severe drought, percentile 10%),
D3 (extreme drought, percentile 5%), and D4 (exceptional
drought, percentile 2%). Leaving out the first five years as model
spin-up period, the copula drought forecasting model was devel-
oped based on the true simulations from January 1, 1984 to
December 31, 2014. The probabilistic forecast of drought status
in the following season given the drought condition in the current
season was examined using root-zone soil moisture for each gird
cell. The dependencies between aggregated root-zone soil mois-
ture were modeled by a Gaussian copula while their marginal dis- Fig. 11. The forecasted drought extent between the OL and DA based on MLE across
tributions are modeled with lognormal distributions as suggested the CRB.
302 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304

Da Silva, 1998; Doucet and Johansen, 2011; Evensen, 1994). The forecasting skill for both summer 2013 and spring 2015. These
most common approach, the cumulative distribution function results demonstrate the added-value of DA to facilitate the state
(CDF)-matching (Reichle and Koster, 2004), was implemented here drought preparation and declaration, at least three months before
to rescale the satellite observations to the model’s climatology. The the official state drought declaration.
CDF-matching approach can correct all the moments of the distri- It is also noted that with increasing ensemble size, the compu-
bution regardless of its shape. Leaving out the first five years as tational cost will increase. Therefore, the DA in this study is per-
model spin-up period, the CCI soil moisture products were rescaled formed on a high-performance computing infrastructure. All the
from 1984 to 2014. The AMSR2 soil moisture products were DA simulations are performed on the Linux Hydra Cluster located
rescaled from 2012 to 2015. For real data study, the perturbation at Portland State University (PSU) with 24 nodes, 384 processors.
errors were the same as the synthetic study except for the model The ensemble size of 200 was selected based on the DA perfor-
error. The model error is normally distributed with a coefficient mance accuracy and computational demand. Since we use the
of variation of 0.15 (Yan et al., 2015). Since the satellite soil mois- PMCMC approach in this study, it requires larger ensemble size
ture data quality control plays an important role in the final DA than the EnKF. However, similar to the approach developed for
performance (Champagne et al., 2011; Yin et al., 2014), before the EnKF (Yin et al., 2015), the optimal ensemble size in the
the rescaling procedure, both the CCI and AMSR2 soil moisture PMCMC can be estimated.
dataset are screened by the data provider for larger water bodies
(percent of land in pixel <95%), frozen soils (surface temperature 4. Conclusions
<273 K), and dense vegetation (vegetation optical depth >0.8).
Since no ‘‘true” drought data exist for real case study, the state In this study, we proposed a probabilistic drought forecasting
drought declarations are used as the references to assess the system, with a combination of dynamical and statistical compo-
drought forecasting skill (Shukla et al., 2011). Two case studies nents. The dynamical hydrologic modeling is coupled with the
are presented here to indicate the added-value of DA for improving copula-based statistical forecasting. Moreover, the ensemble data
drought forecasting skill. (1) In spring 2013, drought declarations assimilation technique is used to improve state initialization in
were issued for nine counties in the southern Idaho (ID). Three the copula-based probabilistic forecasting framework by allowing
months later, a total of 19 counties in ID issued drought emer- for uncertainty in the initial condition. This probabilistic drought
gence. (2) In winter 2015, the PNW received historically low snow- forecasting system was implemented in the Columbia River Basin,
pack conditions. In June 2015, WA Governor declared the statewide Pacific Northwest. We examine the impacts of assimilating remo-
drought and OR Governor declared drought emergencies for 19 out tely sensed surface soil moisture for quantifying the initial condi-
of 36 Oregon counties (about 80% of the state’s landmass). As a tion uncertainty and their subsequent contributions toward an
result, the DA was performed by assimilation of the real satellite improved forecasting of agricultural droughts. Results from both
surface soil moisture for six-month period until the forecast initial- synthetic and real case studies suggest that the proposed drought
ization date. forecasting system significantly improves the seasonal agricultural
The drought conditions of the PNW in summer 2013 and spring drought forecasting skills and can facilitate the state drought
2015 can be predicted based on the drought conditions in spring preparation and declaration. Similar to what has been explained
2013 and winter 2015, with the copula-based seasonal drought here, other satellite or in-situ data, e.g. precipitation and total
forecasting system. The DA can further improve the drought fore- water storage, can also be assimilated in the same way to improve
casting skill as demonstrated in the synthetic study. Similar to the forecasting of meteorological and hydrological droughts.
the synthetic study, prior to investigating the seasonal drought
forecasting results, it is necessary to compare the initial conditions Acknowledgements
of OL and DA. In Fig. 9, the initial conditions for seasonal drought
forecasting beginning on July 1, 2013 (summer forecasting) and Partial financial support for this project was provided by the
April 1, 2015 (spring forecasting) are presented. Each sub-plot con- National Oceanic and Atmospheric Administration (NOAA) Model-
tains the basin-averaged daily root-zone soil moisture for the OL, ing, Analysis, Predictions, and Projections (MAPP) (Grant No.
shown as a single value; and the DA, shown as a distribution of val- NA140AR4310234) and National Science Foundation (NSF) Cyber-
ues, which represent the probability distribution of initial root- Innovation for Sustainability Science and Engineering (CyberSEES)
zone soil moisture values. For both seasons, the DA root-zone soil (Grant No. CCF-1539605).
moisture show lower values, which is more consistent with the
real drought situations. References
Fig. 10 presents the spatial distributions of seasonal drought
forecasting probabilities and MLE forecasted droughts for summer Ahmadalipour, A., Moradkhani, H., Yan, H., Zarekarizi, M., 2017. Remote Sensing of
2013 and spring 2015, respectively. The MLE forecasted drought Drought: Vegetation, Soil Moisture, and Data Assimilation. In: Lakshmi, V. (Ed.),
Remote Sensing of Hydrological Extremes. Springer International Publishing
extent (%) over the U.S. portion of the CRB between the OL and Switzerland, pp. 121–149. http://dx.doi.org/10.1007/978-3-319-43744-6_7.
DA is shown in Fig. 11. Given the initial conditions in spring Beechie, T., Imaki, H., 2014. Predicting natural channel patterns based on landscape
2013, the basin-averaged forecasted drought probabilities for sum- and geomorphic controls in the Columbia River basin, USA. Water Resour. Res.
50, 39–57. http://dx.doi.org/10.1002/2013WR013629.
mer 2013 were 51.80% and 54.64% for OL and DA, respectively. The Champagne, C., McNairn, H., Berg, A.A., 2011. Monitoring agricultural soil moisture
forecasting drought probabilities for spring 2015 were 32.86% and extremes in Canada using passive microwave remote sensing. Remote Sens.
49.58% for OL and DA, respectively. For both cases, the DA suggests Environ. 115, 2434–2444. http://dx.doi.org/10.1016/j.rse.2011.04.030.
Chen, Y.D., Zhang, Q., Xiao, M., Singh, V.P., Zhang, S., 2015. Probabilistic forecasting
a higher probability of drought in summer 2013 and spring 2015, of seasonal droughts in the Pearl River basin. China. Stoch. Environ. Res. Risk
which is more consistent with the state declaration. The MLE fore- Assess. 1–10. http://dx.doi.org/10.1007/s00477-015-1174-6.
casted drought extents increased from 64.67% and 25.85% in the OL Cosens, B.A., Williams, M.K., 2012. Resilience and water governance: adaptive
governance in the Columbia River Basin. Ecol. Soc. 17. http://dx.doi.org/
to 70.57% and 52.83% in the DA for summer 2013 and spring 2015,
10.5751/ES-04986-170403.
respectively. Especially for the spring 2015, the OL forecasts under- Daly, C., Gibson, W.P., Taylor, G.H., Johnson, G.L., Pasteris, P., 2002. A knowledge-
estimated the severe drought conditions for WA and OR. In terms based approach to the statistical mapping of climate. Clim. Res. 22, 99–113.
of both forecasted probabilities and MLE forecasted drought http://dx.doi.org/10.3354/cr022099.
Day, G.N., 1985. Extended streamflow forecasting using NWSRFS. J. Water Resour.
extents, the DA is much more consistent with the state declaration. Plan. Manage. 111, 157–170. http://dx.doi.org/10.1061/(ASCE)0733-9496
In summary, compared with the OL, the DA improves the drought (1985) 111:2(157).
 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304 303

De Lannoy, G.J.M., Reichle, R.H., Arsenault, K.R., Houser, P.R., Kumar, S., Verhoest, N. Kumar, S.V., Reichle, R.H., Koster, R.D., Crow, W.T., Peters-Lidard, C.D., 2009. Role of
E.C., Pauwels, V.R.N., 2012. Multiscale assimilation of advanced microwave subsurface physics in the assimilation of surface soil moisture observations. J.
scanning radiometer–EOS snow water equivalent and moderate resolution Hydrometeorol. 10, 1534–1547. http://dx.doi.org/10.1175/2009JHM1134.1.
imaging spectroradiometer snow cover fraction observations in northern Kumar, S.V., Reichle, R.H., Harrison, K.W., Peters-Lidard, C.D., Yatheendradas, S.,
Colorado. Water Resour. Res. 48, W01522. http://dx.doi.org/10.1029/ Santanello, J.A., 2012. A comparison of methods for a priori bias correction in
2011WR010588. soil moisture data assimilation. Water Resour. Res. 48. http://dx.doi.org/
DeChant, C.M., Moradkhani, H., 2011. Improving the characterization of initial 10.1029/2010WR010261.
condition for ensemble streamflow prediction using data assimilation. Hydrol. Kumar, S.V., Harrison, K.W., Peters-Lidard, C.D., Santanello, J.A., Kirschbaum, D.,
Earth Syst. Sci. 15, 3399–3410. http://dx.doi.org/10.5194/hess-15-3399-2011. 2014a. Assessing the impact of L-band observations on drought and flood risk
DeChant, C.M., Moradkhani, H., 2012. Examining the effectiveness and robustness of estimation: a decision-theoretic approach in an OSSE environment. J.
sequential data assimilation methods for quantification of uncertainty in Hydrometeorol. 15, 2140–2156. http://dx.doi.org/10.1175/JHM-D-13-0204.1.
hydrologic forecasting. Water Resour. Res. 48, W04518. http://dx.doi.org/ Kumar, S.V., Peters-Lidard, C.D., Mocko, D., Reichle, R., Liu, Y., Arsenault, K.R., Xia, Y.,
10.1029/2011WR011011. Ek, M., Riggs, G., Livneh, B., Cosh, M., 2014b. Assimilation of remotely sensed soil
DeChant, C.M., Moradkhani, H., 2014. Toward a reliable prediction of seasonal moisture and snow depth retrievals for drought estimation. J. Hydrometeorol.
forecast uncertainty: addressing model and initial condition uncertainty with 15, 2446–2469. http://dx.doi.org/10.1175/JHM-D-13-0132.1.
ensemble data assimilation and sequential Bayesian combination. J. Hydrol. Lavers, D., Luo, L., Wood, E.F., 2009. A multiple model assessment of seasonal
519, 2967–2977. http://dx.doi.org/10.1016/j.jhydrol.2014.05.045. climate forecast skill for applications. Geophys. Res. Lett. 36, L23711. http://dx.
DeChant, C.M., Moradkhani, H., 2015. Analyzing the sensitivity of drought recovery doi.org/10.1029/2009GL041365.
forecasts to land surface initial conditions. J. Hydrol. 526, 89–100. http://dx.doi. Leavesley, G.H., Lichty, R.W., Thoutman, B.M., Saindon, L.G., 1983. Precipitation-
org/10.1016/j.jhydrol.2014.10.021. Runoff Modeling System: User’s Manual. US Geological Survey, Colorado CO.
Dee, D.P., Da Silva, A.M., 1998. Data assimilation in the presence of forecast bias. Q. J. Li, H., Luo, L., Wood, E.F., Schaake, J., 2009. The role of initial conditions and forcing
R. Meteorol. Soc. 124, 269–295. http://dx.doi.org/10.1002/qj.49712454512. uncertainties in seasonal hydrologic forecasting. J. Geophys. Res. 114, D04114.
Dong, J., Steele-Dunne, S.C., Judge, J., van de Giesen, N., 2015. A particle batch http://dx.doi.org/10.1029/2008JD010969.
smoother for soil moisture estimation using soil temperature observations. Adv. Liu, Q., Reichle, R.H., Bindlish, R., Cosh, M.H., Crow, W.T., de Jeu, R., De Lannoy, G.J.M.,
Water Resour. 83, 111–122. http://dx.doi.org/10.1016/j.advwatres.2015.05.017. Huffman, G.J., Jackson, T.J., 2011. The contributions of precipitation and soil
Dorigo, W.A., Gruber, A., De Jeu, R.A.M., Wagner, W., Stacke, T., Loew, A., Albergel, C., moisture observations to the skill of soil moisture estimates in a land data
Brocca, L., Chung, D., Parinussa, R.M., Kidd, R., 2015. Evaluation of the ESA CCI assimilation system. J. Hydrometeorol. 12, 750–765. http://dx.doi.org/10.1175/
soil moisture product using ground-based observations. Remote Sens. Environ. JHM-D-10-05000.1.
162, 380–395. http://dx.doi.org/10.1016/j.rse.2014.07.023. Liu, Y.Y., Dorigo, W.A., Parinussa, R.M., De Jeu, R.A.M., Wagner, W., McCabe, M.F.,
Doucet, A., Johansen, A., 2011. A tutorial on particle filtering and smoothing: fifteen Evans, J.P., Van Dijk, A.I.J.M., 2012. Trend-preserving blending of passive and
years later. In: The Oxford Handbook of Nonlinear Filtering. Oxford University active microwave soil moisture retrievals. Remote Sens. Environ. 123, 280–297.
Press, pp. 656–704. http://dx.doi.org/10.1016/j.rse.2012.03.014.
Duan, Q., Sorooshian, S., Gupta, V.K., 1994. Optimal use of the SCE-UA global Luo, L., Wood, E.F., 2007. Monitoring and predicting the 2007 U.S. drought. Geophys.
optimization method for calibrating watershed models. J. Hydrol. 158, 265–284. Res. Lett. 34, L22702. http://dx.doi.org/10.1029/2007GL031673.
http://dx.doi.org/10.1016/0022-1694(94)90057-4. Madadgar, S., Moradkhani, H., 2013. A Bayesian framework for probabilistic
Entekhabi, D., Njoku, E.G., O’Neill, P.E., Kellogg, K.H., Crow, W.T., Edelstein, W.N., seasonal drought forecasting. J. Hydrometeorol. 14, 1685–1705. http://dx.doi.
Entin, J.K., Goodman, S.D., Jackson, T.J., Johnson, J., Kimball, J., Piepmeier, J.R., org/10.1175/JHM-D-13-010.1.
Koster, R.D., Martin, N., McDonald, K.C., Moghaddam, M., Moran, S., Reichle, R., Madadgar, S., Moradkhani, H., 2014. Spatio-temporal drought forecasting within
Shi, J.C., Spencer, M.W., Thurman, S.W., Tsang, L., Van Zyl, J., Neill, P.E.O., Kellogg, Bayesian networks. J. Hydrol. 512, 134–146. http://dx.doi.org/10.1016/j.
K.H., Crow, W.T., Edelstein, W.N., Entin, J.K., Goodman, S.D., Jackson, T.J., jhydrol.2014.02.039.
Johnson, J., O’Neill, P.E., Kellogg, K.H., Crow, W.T., Edelstein, W.N., Entin, J.K., Madadgar, S., Moradkhani, H., 2016. Copula function and drought. In: Handbook of
Goodman, S.D., Jackson, T.J., Johnson, J., Kimball, J., Piepmeier, J.R., Koster, R.D., drought and water scarcity. Principles of Drought and Water Scarcity, vol. 1.
Martin, N., McDonald, K.C., Moghaddam, M., Moran, S., Reichle, R., Shi, J.C., Francis and Taylor.
Spencer, M.W., Thurman, S.W., Tsang, L., Van Zyl, J., 2010. The soil moisture Mao, Y., Nijssen, B., Lettenmaier, D.P., 2015. Is climate change implicated in the
active passive (SMAP) mission. Proc. IEEE 98, 704–716. http://dx.doi.org/ 2013–2014 California drought? a hydrologic perspective. Geophys. Res. Lett. 42,
10.1109/JPROC.2010.2043918. 2805–2813. http://dx.doi.org/10.1002/2015GL063456.
Evensen, G., 1994. Sequential data assimilation with a nonlinear quasi-geostrophic Markstrom, S.L., Niswonger, R.G., Regan, R.S., Prudic, D.E., Barlow, P.M., 2008.
model using Monte Carlo methods to forecast error statistics. J. Geophys. Res. GSFLOW-Coupled Ground-Water and Surface-Water FLOW Model Based on the
99, 10143–10162. http://dx.doi.org/10.1029/94JC00572. Integration of the Precipitation-Runoff Modeling System (PRMS) and the
Farnsworth, R.K., Thompson, E.S., Peck, E.L., 1982. Evaporation Atlas for the Modular Ground-Water Flow Model (MODFLOW-2005). US Geological Survey.
Contiguous 48 United States. NOAA Technical Report NWS 33. Markstrom, S.L., Regan, R.S., Hay, L.E., Viger, R.J., Webb, R.M., Payn, R.A., LaFontaine,
Genest, C., Rémillard, B., 2008. Validity of the parametric bootstrap for goodness-of- J.H., 2015. PRMS-IV, The Precipitation-Runoff Modeling System, Version 4. US
fit testing in semiparametric models. Ann. l’Institut Henri Poincaré Probab. Stat. Geological Survey.
44, 1096–1127. http://dx.doi.org/10.1214/07-AIHP148. Milly, P.C.D., Betancourt, J., Falkenmark, M., Hirsch, R.M., Kundzewicz, Z.W.,
Gupta, H.V., Kling, H., Yilmaz, K.K., Martinez, G.F., 2009. Decomposition of the mean Lettenmaier, D.P., Stouffer, R.J., 2008. Stationarity is dead: whither water
squared error and NSE performance criteria: implications for improving management? Science (80-) 319, 573–574. http://dx.doi.org/
hydrological modelling. J. Hydrol. 377, 80–91. http://dx.doi.org/10.1016/j. 10.1126/science.1151915.
jhydrol.2009.08.003. Mishra, A.K., Desai, V.R., 2005. Drought forecasting using stochastic models. Stoch.
Hamlet, A.F., Lettenmaier, D.P., 1999. Effects of climate change on hydrology and Environ. Res. Risk Assess. 19, 326–339. http://dx.doi.org/10.1007/s00477-005-
water resources in the Columbia River basin. J. Am. Water Resour. Assoc. 35, 0238-4.
1597–1623. http://dx.doi.org/10.1111/j.1752-1688.1999.tb04240.x. Mishra, A.K., Singh, V.P., 2010. A review of drought concepts. J. Hydrol. 391, 202–
Hay, L.E., Umemoto, M., 2006. Multiple-Objective Stepwise Calibration Using Luca. 216. http://dx.doi.org/10.1016/j.jhydrol.2010.07.012.
US Geological Survey. Mishra, A.K., Singh, V.P., 2011. Drought modeling – a review. J. Hydrol. 403, 157–
Hayes, M., Svoboda, M., Le Comte, D., Redmond, K.T., Pasteris, P., 2005. Drought 175. http://dx.doi.org/10.1016/j.jhydrol.2011.03.049.
monitoring: new tools for the 21st century. In: Wihite, D.A. (Ed.), Drought and Moradkhani, H., 2008. Hydrologic remote sensing and land surface data
Water Crises: Science, Technology, and Management Issues. CRC Press, pp. 53– assimilation. Sensors 8, 2986–3004. http://dx.doi.org/10.3390/s8052986.
69. Moradkhani, H., Hsu, K.-L., Gupta, H., Sorooshian, S., 2005. Uncertainty assessment
Howitt, R., Medellín-Azuara, J., MacEwan, D., Lund, J., Sumner, D., 2014. Economic of hydrologic model states and parameters: sequential data assimilation using
Analysis of the 2014 Drought for California Agriculture. University of California, the particle filter. Water Resour. Res. 41, W05012. http://dx.doi.org/10.1029/
Davis. 2004WR003604.
Imaoka, K., Kachi, M., Kasahara, M., Ito, N., Nakagawa, K., Oki, T., 2010. Instrument Moradkhani, H., Dechant, C.M., Sorooshian, S., 2012. Evolution of ensemble data
performance and calibration of AMSR-E and AMSR2. Int. Arch. Photogramm. assimilation for uncertainty quantification using the particle filter-Markov
Remote Sens. Spat. Inf. Sci. ISPRS Arch. 38, 13–16. chain Monte Carlo method. Water Resour. Res. 48, W12520. http://dx.doi.org/
Joe, H., 1997. Multivariate models and dependence concepts. Analysis. Export Date 10.1029/2012WR012144.
21 May 2013. NCDC, 2012. Billion-Dollar Weather/Climate Events. NOAA [WWW Document].
Keyantash, J., Dracup, J.A., 2002. The quantification of drought: an evaluation of Omernik, J.M., Bailey, R.G., 1997. Distinguishing between watersheds and
drought indices. Bull. Am. Meteorol. Soc. 83, 1167–1180. http://dx.doi.org/ ecoregions. J. Am. Water Resour. Assoc. 33, 935–949. http://dx.doi.org/
10.1175/1520-0477(2002) 083<1191:TQODAE>2.3.CO;2. 10.1111/j.1752-1688.1997.tb04115.x.
Koster, R.D., Mahanama, S.P.P., Yamada, T.J., Balsamo, G., Berg, A.A., Boisserie, M., Owe, M., de Jeu, R., Holmes, T., 2008. Multisensor historical climatology of satellite-
Dirmeyer, P.A., Doblas-Reyes, F.J., Drewitt, G., Gordon, C.T., Guo, Z., Jeong, J.-H., derived global land surface moisture. J. Geophys. Res. Earth Surf. 113. http://dx.
Lawrence, D.M., Lee, W.-S., Li, Z., Luo, L., Malyshev, S., Merryfield, W.J., doi.org/10.1029/2007JF000769.
Seneviratne, S.I., Stanelle, T., van den Hurk, B.J.J.M., Vitart, F., Wood, E.F., 2010. Pan, M., Wood, E.F., 2010. Impact of accuracy, spatial availability, and revisit time of
Contribution of land surface initialization to subseasonal forecast skill: first satellite-derived surface soil moisture in a multiscale ensemble data
results from a multi-model experiment. Geophys. Res. Lett. 37, L02402. http:// assimilation system. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 3, 49–56.
dx.doi.org/10.1029/2009GL041677. http://dx.doi.org/10.1109/JSTARS.2010.2040585.
304 H. Yan et al. / Journal of Hydrology 548 (2017) 291–304

Plaza, D.A., De Keyser, R., De Lannoy, G.J.M., Giustarini, L., Matgen, P., Pauwels, V.R. Wood, A.W., Lettenmaier, D.P., 2008. An ensemble approach for attribution of
N., 2012. The importance of parameter resampling for soil moisture data hydrologic prediction uncertainty. Geophys. Res. Lett. 35, L14401. http://dx.doi.
assimilation into hydrologic models using the particle filter. Hydrol. Earth Syst. org/10.1029/2008GL034648.
Sci. 16, 375–390. http://dx.doi.org/10.5194/hess-16-375-2012s. Wood, A.W., Hopson, T., Newman, A., Brekke, L., Arnold, J., Clark, M., 2016.
Reichle, R.H., Koster, R.D., 2004. Bias reduction in short records of satellite soil Quantifying streamflow forecast skill elasticity to initial condition and climate
moisture. Geophys. Res. Lett. 31, L19501. http://dx.doi.org/10.1029/ prediction skill. J. Hydrometeorol. 17, 651–668. http://dx.doi.org/10.1175/JHM-
2004GL020938. D-14-0213.1.
Reichle, R.H., Koster, R.D., Liu, P., Mahanama, S.P.P., Njoku, E.G., Owe, M., 2007. Xia, Y., Mitchell, K., Ek, M., Sheffield, J., Cosgrove, B., Wood, E., Luo, L., Alonge, C., Wei,
Comparison and assimilation of global soil moisture retrievals from the H., Meng, J., Livneh, B., Lettenmaier, D., Koren, V., Duan, Q., Mo, K., Fan, Y.,
advanced microwave scanning radiometer for the earth observing system Mocko, D., 2012. Continental-scale water and energy flux analysis and
(AMSR-E) and the scanning multichannel microwave radiometer (SMMR). J. validation for the North American land data assimilation system project
Geophys. Res. 112, D09108. http://dx.doi.org/10.1029/2006JD008033. phase 2 (NLDAS-2): 1. Intercomparison and application of model products. J.
Reichle, R.H., Kumar, S.V., Mahanama, S.P.P., Koster, R.D., Liu, Q., 2010. Assimilation Geophys. Res. Atmos. 117, D03109. http://dx.doi.org/10.1029/2011JD016048.
of satellite-derived skin temperature observations into land surface models. J. Xia, Y., Sheffield, J., Ek, M.B., Dong, J., Chaney, N., Wei, H., Meng, J., Wood, E.F., 2014.
Hydrometeorol. 11, 1103–1122. http://dx.doi.org/10.1175/2010JHM1262.1. Evaluation of multi-model simulated soil moisture in NLDAS-2. J. Hydrol. 512,
Ross, T., Lott, N., 2003. A Climatology of 1980–2003 Extreme Weather and Climate 107–125. http://dx.doi.org/10.1016/j.jhydrol.2014.02.027.
Events. National Oceanic and Atmospheric Administration. Yan, H., 2016. From Drought Monitoring to Forecasting: A Combined Dynamical-
Sheffield, J., Wood, E.F., Chaney, N., Guan, K., Sadri, S., Yuan, X., Olang, L., Amani, A., Statistical Modeling Framework (Ph.D. Dissertation). Portland State University.
Ali, A., Demuth, S., Ogallo, L., 2014. A drought monitoring and forecasting Yan, H., Edwards, F.G., 2013. Effects of land use change on hydrologic response at a
system for sub-sahara African water resources and food security. Bull. Am. watershed scale. Arkansas. J. Hydrol. Eng. 18, 1779–1785. http://dx.doi.org/
Meteorol. Soc. 95, 861–882. http://dx.doi.org/10.1175/BAMS-D-12-00124.1. 10.1061/(ASCE)HE.1943-5584.0000743.
Shukla, S., Lettenmaier, D.P., 2011. Seasonal hydrologic prediction in the United Yan, H., Moradkhani, H., 2016. Combined assimilation of streamflow and satellite
States: understanding the role of initial hydrologic conditions and seasonal soil moisture with the particle filter and geostatistical modeling. Adv. Water
climate forecast skill. Hydrol. Earth Syst. Sci. 15, 3529–3538. http://dx.doi.org/ Resour. 94, 364–378. http://dx.doi.org/10.1016/j.advwatres.2016.06.002.
10.5194/hess-15-3529-2011. Yan, H., DeChant, C.M., Moradkhani, H., 2015. Improving soil moisture profile
Shukla, S., Steinemann, A.C., Lettenmaier, D.P., 2011. Drought monitoring for prediction with the particle filter-Markov Chain Monte Carlo method. IEEE
Washington State: indicators and applications. J. Hydrometeorol. 12, 66–83. Trans. Geosci. Remote Sens. 53, 6134–6147. http://dx.doi.org/10.1109/
http://dx.doi.org/10.1175/2010JHM1307.1. TGRS.2015.2432067.
Sklar, A., 1959. Fonctions de Répartition à n Dimensions et Leurs Marges, vol. 8. Yilmaz, M.T., Crow, W.T., 2013. The optimality of potential rescaling approaches in
Publ. l’Institut Stat. l’Université Paris, pp. 229–231. land data assimilation. J. Hydrometeorol. 14, 650–660. http://dx.doi.org/
Stainforth, D.A., Aina, T., Christensen, C., Collins, M., Faull, N., Frame, D.J., 10.1175/JHM-D-12-052.1.
Kettleborough, J.A., Knight, S., Martin, A., Murphy, J.M., Piani, C., Sexton, D., Yin, J., Zhan, X., Zheng, Y., Liu, J., Hain, C.R., Fang, L., 2014. Impact of quality control
Smith, L.A., Spicer, R.A., Thorpe, A.J., Allen, M.R., 2005. Uncertainty in predictions of satellite soil moisture data on their assimilation into land surface model.
of the climate response to rising levels of greenhouse gases. Nature 433, 403– Geophys. Res. Lett. 41, 7159–7166. http://dx.doi.org/10.1002/2014GL060659.
406. http://dx.doi.org/10.1038/nature03301. Yin, J., Zhan, X., Zheng, Y., Hain, C.R., Liu, J., Fang, L., 2015. Optimal ensemble size of
Steinemann, A.C., 2006. Using climate forecasts for drought management. J. Appl. ensemble Kalman filter in sequential soil moisture data assimilation. Geophys.
Meteorol. Climatol. 45, 1353–1361. http://dx.doi.org/10.1175/JAM2401.1. Res. Lett. 42, 6710–6715. http://dx.doi.org/10.1002/2015GL063366.
Su, C.H., Ryu, D., Crow, W.T., Western, A.W., 2014. Beyond triple collocation: Yossef, N.C., Winsemius, H., Weerts, A., Van Beek, R., Bierkens, M.F.P., 2013. Skill of a
applications to soil moisture monitoring. J. Geophys. Res. Atmos. 119, 6419– global seasonal streamflow forecasting system, relative roles of initial
6439. http://dx.doi.org/10.1002/2013JD021043. conditions and meteorological forcing. Water Resour. Res. 49, 4687–4699.
Svoboda, M., LeComte, D., Hayes, M., Heim, R., Gleason, K., Angel, J., Rippey, B., http://dx.doi.org/10.1002/wrcr.20350.
Tinker, R., Palecki, M., Stooksbury, D., Miskus, D., Stephens, S., 2002. The drought Yuan, X., 2016. An experimental seasonal hydrological forecasting system over the
monitor. Bull. Am. Meteorol. Soc. 83, 1181–1190. http://dx.doi.org/10.1175/ Yellow River basin – Part 2: The added value from climate forecast models.
1520-0477(2002) 083<1181:TDM>2.3.CO;2. Hydrol. Earth Syst. Sci. 20, 2453–2466. http://dx.doi.org/10.5194/hess-20-
Van Loon, A.F., 2015. Hydrological drought explained. Wiley Interdiscip. Rev. Water 2453-2016.
2, 359–392. http://dx.doi.org/10.1002/wat2.1085. Yuan, X., Wood, E.F., Luo, L., Pan, M., 2011. A first look at climate forecast system
Viger, R.J., 2014. Preliminary Spatial Parameters for PRMS Based on the Geospatial version 2 (CFSv2) for hydrological seasonal prediction. Geophys. Res. Lett. 38,
Fabric, NLCD2001 and SSURGO. US Geol Surv. 10.5066/F7WM1BF7. L13402. http://dx.doi.org/10.1029/2011GL047792.
Wang, A., Lettenmaier, D.P., Sheffield, J., 2011. Soil moisture drought in China, Yuan, X., Wood, E.F., Roundy, J.K., Pan, M., 2013. CFSv2-Based seasonal
1950–2006. J. Clim. 24, 3257–3271. http://dx.doi.org/10.1175/2011JCLI3733.1. hydroclimatic forecasts over the conterminous United States. J. Clim. 26,
Washington Department of Agriculture, 2015. Interim Report: 2015 Drought and 4828–4847. http://dx.doi.org/10.1175/JCLI-D-12-00683.1.
Agriculture. Yuan, X., Ma, F., Wang, L., Zheng, Z., Ma, Z., Ye, A., Peng, S., 2016. An experimental
WCRP, 2010. Drought Predictability and Prediction in a Changing Climate: seasonal hydrological forecasting system over the Yellow River basin – Part 1:
Assessing Current Capabilities. World Climate Research Programme, Understanding the role of initial hydrological conditions. Hydrol. Earth Syst. Sci.
Barcelona, Spain. 20, 2437–2451. http://dx.doi.org/10.5194/hess-20-2437-2016.