JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 107, NO. D20, 4429, doi:10.

1029/2001JD000659, 2002

Long-range experimental hydrologic forecasting for the eastern

United States
Andrew W. Wood and Edwin P. Maurer
Department of Civil and Environmental Engineering, University of Washington, Seattle, Washington, USA

Arun Kumar
Climate Prediction Center, NOAA National Center for Environmental Prediction, Camp Springs, Maryland, USA

Dennis P. Lettenmaier
Department of Civil and Environmental Engineering, University of Washington, Seattle, Washington, USA

Received 22 March 2001; revised 26 November 2001; accepted 18 January 2002; published 23 October 2002.
[1] We explore a strategy for long-range hydrologic forecasting that uses ensemble
climate model forecasts as input to a macroscale hydrologic model to produce runoff and
streamflow forecasts at spatial and temporal scales appropriate for water management.
Monthly ensemble climate model forecasts produced by the National Centers for
Environmental Prediction/Climate Prediction Center global spectral model (GSM) are bias
corrected, downscaled to 1/8° horizontal resolution, and disaggregated to a daily time
step for input to the Variable Infiltration Capacity hydrologic model. Bias correction is
effected by evaluating the GSM ensemble forecast variables as percentiles relative to the
GSM model climatology and then extracting the percentiles’ associated variable values
instead from the observed climatology. The monthly meteorological forecasts are then
interpolated to the finer hydrologic model scale, at which a daily signal that preserves the
forecast anomaly is imposed through resampling of the historic record. With the resulting
monthly runoff and streamflow forecasts for the East Coast and Ohio River basin, we
evaluate the bias correction and resampling approaches during the southeastern United
States drought from May to August 2000 and also for the El Niño conditions of December
1997 to February 1998. For the summer 2000 study period, persistence in anomalous
initial hydrologic states predominates in determining the hydrologic forecasts. In contrast,
the El Niño-condition hydrologic forecasts derive direction both from the climate model
forecast signal and the antecedent land surface state. From a qualitative standpoint the
hydrologic forecasting strategy appears successful in translating climate forecast signals to
hydrologic variables of interest for water management. INDEX TERMS: 1860 Hydrology:
Runoff and streamflow; 1863 Hydrology: Snow and ice (1827); 1836 Hydrology: Hydrologic budget (1655);
1833 Hydrology: Hydroclimatology; KEYWORDS: climate downscaling, hydrologic forecast, seasonal forecast,
streamflow forecast, eastern United States
Citation: Wood, A. W., E. P. Maurer, A. Kumar, and D. Lettenmaier, Long-range experimental hydrologic forecasting for the eastern
United States, J. Geophys. Res., 107(D20), 4429, doi:10.1029/2001JD000659, 2002.

1. Introduction ocean-atmosphere general circulation models (OAGCMs)

[Barnston et al., 1999; Livezey, 1990; Kumar et al., 1996;
[2] Great strides have been made over the last decade in Livezey et al., 1996] or statistical methods such as canonical
understanding climate teleconnections, as manifested by correlation analysis [e.g., Barnston and He, 1996]. A recent
ocean-atmosphere phenomena resulting in large part from trend has been to use ensemble forecasting approaches in
thermal inertia of the oceans [see, e.g., Livezey et al., 1997; which a global land-atmosphere-ocean model (initialized
Shukla, 1998; Koster et al., 1999], such as El Niño-South- with atmospheric, land surface, and ocean conditions at
ern Oscillation (ENSO), the Pacific Decadal Oscillation forecast time) is run into the future for forecast horizons of
(PDO), and the North Atlantic Oscillation. Exploitation of months to years, using prescribed sea surface temperatures
the understanding of these phenomena has resulted in (SSTs) derived using one of a variety of forecast methods.
demonstrable improvements in long-lead (months to years) Although the atmosphere is essentially chaotic, the pre-
climate forecasting. These forecasts are based on coupled scribed SSTs effectively constrain the evolution of model
forecasts. By perturbing the initial atmospheric conditions
Copyright 2002 by the American Geophysical Union. and repeating the simulation a number of times, an ensem-
0148-0227/02/2001JD000659 ble of forecasts is constructed which represents the range of

ACL 6-1
ACL 6-2 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING

global atmospheric conditions that may occur over the preclude direct use in hydrologic modeling of OAGCM
forecast period. output fields such as surface precipitation and temperature
[3] Much of the research in this area has focused on [Leung et al., 1999; Chen et al., 1996; Roads et al., 1999].
atmospheric simulation outputs having the large subcon- [6] Hydrologists have tended to surmount this difficulty
tinental region (e.g., the southwestern United States) as the using simple strategies such as OAGCM-conditioned com-
minimum scale, in part because of the long-recognized positing, in which OAGCM output is used to guide the
difficulty of atmospheric models in reproducing observed construction or weighting of an ensemble of historically
climate at smaller scales (e.g., <107 km2) [IPCC, 1996, observed meteorological time series, which subsequently is
section 6]. As a result, connecting climate forecasts to the used as input to a hydrologic model [e.g., Georgakakos et
scales and features of the hydrosphere in which human al., 1998; Leung et al., 1999]. In theory, at least, the
society is often most interested (regional and smaller, for probabilistic assessment of differences between streamflow
land surface variables) has been problematic. In addition to ensembles resulting from these hydrologic simulation and
the scale problem, the land surface variables of greatest streamflow ensembles based on observed climatology
interest to society, such as surface precipitation and runoff, (inputs without conditioning) may then support recommen-
are generally predicted less reliably than features of large- dations for operational decisions by water resources system
scale circulation [see, e.g., Risbey and Stone, 1996]. In managers. A variation on this approach is to derive the
parallel with global scale predictions, however, a number of conditioning signal from a typecast of the present year
downscaling methods have evolved, including dynamical related to SST-classified climate modes. Hamlet and Let-
approaches that use finer resolution (mesoscale) atmos- tenmaier [1998, 1999], for instance, demonstrated a sim-
pheric models [e.g., Cocke and LaRow, 2000; Giorgi and plified method of long-range forecasting (up to a 1-year
Mearns, 1991], statistical approaches [Wilby and Wigley, lead) for the Columbia River basin. The method utilized
1997; Wilby et al., 1998], and climate-analog approaches resampling of previous observed hydrometeorological data
[IPCC, 1996; Leung et al., 1999; Georgakakos et al., 1998]. for years with apparently analogous climatic characteristics,
Recent comparisons of dynamical and statistical methods determined by ENSO- and PDO-based compositing. The
are given by Murphy [1999] and Kidson and Thompson shortcoming of this method is that it requires partitioning
[1998]. of the historic record into climate categories, which can
[4] At much smaller scales, hydrologists have long been result in statistical problems when the number of years in a
concerned with understanding and reproducing the dynam- given class is small. Furthermore, there is an implicit
ics of the land surface water and energy balance. Hydrologic assumption that the classification method is stable over
study has mostly focused on the local scale of catchments or time.
basins (on the order of 102 – 103 km2) at which water [7] One-way climate model-hydrology model linkages
management is effected. Much applied hydrologic predic- were explored by Kim et al. [2000], who applied a meso-
tion work, for instance, efforts to abstract the physics of scale regional climate model over northern California for
runoff generation and groundwater behavior, has been downscaling one member of an OAGCM-based 3-month
intended to benefit water resources end uses, such as forecast ensemble during the 1997 – 1998 El Niño event.
irrigation, water supply, hydropower generation, fisheries Using the spatially distributed TOPMODEL [Beven and
management, and navigation. An intersection of the inter- Kirkby, 1979] for hydrologic simulation, they found that the
ests of hydrologists and climate modelers has occurred over largest hydrologic forecast errors resulted from general
the last decade as the difference in spatial scales has circulation model (GCM) errors in precipitation prediction.
narrowed. While general circulation models now often For a smaller catchment in Colorado, Wilby et al. [2000]
operate at spatial resolutions of 1 – 3°, macroscale hydro- compared statistical and dynamical downscaling methods,
logic models (e.g., those of Liang et al. [1994], Leavesley using a regional climate model, for translating National
and Stannard [1995], and Beven and Kirkby [1979]) have Centers for Environmental Prediction (NCEP)/National
increased in scale and geographical coverage so that mod- Center for Atmospheric Research (NCAR) reanalysis [Kal-
eling of continental scale river basins (e.g., the Columbia, nay et al., 1996] output into local precipitation and temper-
the Mississippi) is now possible. Furthermore, the land ature forcing time series for a hydrologic model. Their
surface parameterizations in coupled land-atmosphere- results underscored the need for bias correcting climate
ocean models increasingly resemble or borrow from macro- model outputs and confirmed the view that while statistical
scale hydrology model representations, and vice versa and dynamical approaches yield similar downscaling skill,
[Koster et al., 2000; Ducharne et al., 2000]. In consequence statistical techniques are less computationally demanding.
an operational linkage of hydrologic and climate forecasting [8] This paper describes an exploratory hydrologic fore-
models is now being pursued at a number of research cast system that uses monthly ensemble climate forecasts of
centers. monthly total precipitation (Ptot) and monthly average
[5] Although the motivation for development of macro- temperature (Tavg) for 6-month lead times produced by the
scale hydrologic models has been in part to improve NCEP global spectral model (GSM), an OAGCM. We test a
representation of the land surface in coupled land-atmos- relatively simple approach for linking global ensemble
phere-ocean models, they can also be implemented using forecasts from coupled ocean-land atmosphere models with
one-way forcing from OAGCMs, such as ensemble climate macroscale hydrologic models, with the intent of improving
forecasts. While conceptually simpler than operation in a hydrologic prediction capabilities for soil moisture, runoff,
fully coupled mode, the one-way linkage is still hampered and streamflow.
by the need to address regional biases in OAGCM climate [9] The region chosen for the study was the eastern
simulation outputs. These biases are substantial enough to United States, defined as the area east of the Mississippi
 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING ACL 6-3

Figure 1. Experimental long-lead hydrologic forecasting approach.

River drainage plus the Ohio River basin but excluding the among other variables, are archived. In the forecast runs,
Laurentian Great Lakes basin, for a forecast period from predicted SSTs over the tropical Pacific domain are speci-
May to October 2000. This period was selected because a fied on the basis of the NCEP OAGCM [Ji et al., 1998]. At
severe drought was anticipated for the southeastern United the time of our research, GSM forecasts were run at T42
States as a result of much below normal soil moisture and horizontal resolution (2.8125° latitude/longitude).
streamflow during late winter and early spring 2000 [12] GSM is actually run at time steps on the order of an
(reflected as early as December 1999 in federal agency hour or less, and therefore in principal the temporal dis-
outlooks such as the National Drought Mitigation Center’s aggregation is not necessary. Use of monthly ensembles,
weekly Drought Monitor and the Climate Prediction Cen- however, greatly reduces the data distribution and handling
ter’s U.S. Drought Monitor). During this period, SST overhead. Because our downscaling approach (section
anomalies in the tropical Pacific were returning to near 2.3.2) imposes plausible daily temporal structure onto the
normal from a prior ENSO cold phase (La Niña) episode. monthly GSM forecast products as part of the same process
Lingering effects of the cold phase, which in the south- that spatially disaggregates the GSM products to 1/8° spatial
eastern United States have been correlated with dryness, had resolution, the use of the monthly GSM output is not only
the potential to compound the existing soil moisture deficits. adequate for purposes of our hydrologic forecast objectives
We also evaluated the method for a study period (beginning but also streamlines the process considerably.
in November 1997) during which SST anomalies reflected a
strong El Niño event. 2.2. VIC Macroscale Hydrology Model
[13] The VIC model [Liang et al., 1994, 1996, 1999] is
a semidistributed, grid-based hydrological model that
2. Approach parameterizes the dominant hydrometeorological processes
[10] Our forecast approach uses GSM’s surface forecast taking place at the land surface-atmosphere interface. A
fields (Ptot and Tavg) to create daily forcing ensembles for mosaic representation of land surface cover and parameter-
the Variable Infiltration Capacity (VIC) macroscale physical izations for infiltration and the spatial variability of pre-
hydrology model. Hydrologic model forecasts are produced cipitation account for subgrid scale heterogeneities in key
by first initializing VIC model states with a spin-up period hydrological processes. The model uses three soil layers
based on observed meteorology prior to the start of forecast and one vegetation layer, with energy and moisture fluxes
and then driving the model with ensemble forecast meteor- exchanged between the layers. The model has been
ology through the end of the forecast period. This basic applied to such large continental rivers as the Columbia
framework is illustrated in Figure 1. [Nijssen et al., 1997], the Arkansas-Red [Abdulla et al.,
1996], and the Mississippi [Maurer et al., 1999; Cherka-
2.1. Global Spectral Model Ensemble Generation uer and Lettenmaier, 1999], and, as part of the Land Data
[11] Each month, NCEP’s Climate Modeling Branch Assimilation System (LDAS) project [Mitchell et al.,
generates a 20-member ensemble of 6-month lead climate 2000], to the continental United States [Wood et al.,
forecasts, simulated with GSM. The forecasts are accom- 1998]. A more complete description of model processes
panied by a 210-member ensemble of climate hindcasts is given by Liang et al. [1994, 1996]. Runoff generated
(also 6 months long, matching the calendar months of the within a grid cell is routed to the stream gauge locations
forecasts) representing the period 1979 – 1999 (21 years). using methods described by Lohmann et al. [1998a,
The 20 forecast ensemble members are produced by using 1998b]. The VIC model uses vegetation and soil parameters
20 different atmospheric initializations with predicted SSTs produced for use by LDAS and described by Maurer et al.
in the tropical Pacific Ocean as of the date of the forecasts. [2001].
The hindcast ensemble generation process is similar, except 2.2.1. Forcings
that the ensemble members are produced by using 10 [14] VIC model forcings are used both in driving the
different atmospheric initializations with observed SSTs hydrologic model during a 1-year spin-up period, and, via
for each of the 21 years in the 1979 – 1999 hindcast. The resampling, in assembling the daily forecast sequences.
different atmospheric initializations are in each case drawn Because the meteorological variables most widely avail-
from a sequence of atmospheric analysis fields at the able in long-term data archives are daily precipitation and
beginning of the forecast initialization month (the month daily temperature minimum and maximum, we estimate
prior to the 6-month forecast period), spaced 12 hours apart. most of the other forcing variables required by the VIC
This process is repeated every month, and Ptot and Tavg, model (e.g., downward solar and longwave radiation,
ACL 6-4 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING

discrepancy somewhat. Currently, we use the LDAS prod-

uct without adjustment.
2.2.2. Eastern U.S. Application
[15] The VIC model was implemented at 1/8° latitude/
longitude resolution (150 km2 cell area) over the domain
shown in Figure 2. The domain includes the Ohio River
basin, which drains the easternmost portion of the Mississippi
River basin, and an east coast region which includes 24
coastal drainage basins, 17 of which flow east-southeast to
the Atlantic Ocean and 7 of which drain southward to the
Gulf of Mexico. Within the model domain, runoff in smaller
subbasins was routed to produce streamflow estimates at U.S.
Geological Survey (USGS) river-gauging station locations.
[16] The Ohio River basin and east coast models simulate
areas of 600,000 and 1.1 million km2, respectively. Sub-
basins calibrated for streamflow forecasting included the
Ohio River, the Delaware River, the Potomac River, the
Apalachicola-Chattahoochee-Flint (ACF) River system and
the Alabama-Coosa-Tallapoosa (ACT) River system. Prin-
cipal streamflow routing nodes were the Apalachicola River
at Sumatra, Fla. (USGS station 12359170), the Alabama
River at Claiborne Lock and Dam near Monroeville, Ala.
(USGS station 02428400), the Potomac River near Wash-
ington, D.C., Little Falls (USGS station 01646500), the
Delaware River at Trenton, N.J. (USGS station 01463500),
and the Ohio River at Metropolis, Ill. (USGS station
Figure 2. Hydrologic forecasting model domain, includ- 03611500).
ing the Ohio River basin (light gray) and the east coast [17] The model-forcing data spanned the period 1950-
drainages (dark gray). Runoff in numbered basins was current, where the current date evolved during the experi-
routed to produce streamflow: (1) Ohio; (2) Alabama- ment. Model calibration was accomplished by varying
Coosa-Tallapoosa (ACT); (3) Apalachicola-Chattahoochee- parameters related to infiltration and subsurface drainage,
Flint (ACF); (4) Potomac; and (5) Delaware River. with the aim of reproducing monthly streamflow volumes
while preserving the general features of the daily response
(e.g., daily average flow peaks and recessions). Sample
calibration results are shown in Figure 3.
humidity) from this minimum set of variables using
methods described by Maurer et al. [1999]. Wind speed 2.3. Hydrologic Forecasting Approach
data are taken from the NCEP/NCAR reanalysis [Kalnay [18] To translate long-range climate predictions to the
et al., 1996], which are available to within a month of real realm of hydrology, regional biases and the temporal and
time. The observational data are typically taken from spatial scale mismatch between models must be resolved.
National Climatic Data Center (NCDC) Cooperative This section describes our bias-correction and downscaling
Observer (Co-op) Stations, which are available on the approach and its demonstration as a proof of concept
Web within 2 – 4 months of real time. For real-time data exercise.
(bridging the gap between the end of the available Co-op 2.3.1. Unbiasing of Climate Model Ensembles
data and the forecast date), we used data available from [19] The premise of the bias correction step is that despite
the LDAS project. The LDAS precipitation data are so- biases in GSM-simulated climate, the GSM forecasts may
called Stage IV observations, a combination of radar and have a useful signal if interpreted relative to the GSM
station data produced by NCEP. Temperature data are from climatology rather than the observed climatology. The
the Eta Data Assimilation System (EDAS) and are essen- GSM climatology is defined by the monthly distributions
tially an analysis product from the NCEP Eta weather (for months 1 – 6 in the forecast period, separately) of
forecast model run over the continental United States. simulated GSM Ptot and Tavg taken from the GSM hindcast
LDAS also produces a real-time wind data set (another simulations (i.e., the 210 simulated values for each of the 6
EDAS product). All LDAS data are produced on a geo- forecast period months, for each variable). The monthly
graphic 1/8° grid, for the area from latitude 25°N–53°N, observed climatology spans the same time period as the
longitude 67°W –125°W. Typical monthly biases in this GSM output (1979 – 1999) and was created from Co-op
product revealed by our preliminary verification were a station daily observations averaged to a monthly timestep
spatially averaged (over the study domain) bias of
5% in and to the GSM grid resolution; hence the observed
monthly precipitation totals,
1.5°C in average maximum monthly distributions for Ptot and Tavg are defined by only
temperature, and +1.5°C in average minimum temperature. 21 values per variable. Bias correction is achieved by
LDAS wind speeds appear to be significantly higher than replacing GSM forecast values for Tavg and Ptot with values
those produced by the NCEP/NCAR reanalysis, although having the same percentiles (nonexceedence probabilities)
recent EDAS modifications appear to have reduced the with respect to the observed climatology that the original
 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING ACL 6-5

Figure 3. Calibration results for four streamflow forecasting basin gauging locations.

GSM values had with respect to the GSM climatology, for a Tavg in months February – June, and the entire process is
given month. The forecasts are subsequently expressed as repeated for each of the ensemble forecast members. (4)
anomalies (temperature shift and precipitation percentage) Finally, the bias-corrected forecasts are expressed as addi-
with respect to the observed monthly means for the 21-year tive (for Tavg) and multiplicative (for Ptot) anomalies.
climatology period. Bias correction is performed at the [21] In the precipitation and temperature bias correction
GSM scale, and each GSM cell (23 cells spanned the study scheme, when either the GSM output or the associated
region) is treated individually, defining its own set of percentile falls above or below the range of empirical Weibull
monthly distributions. percentiles (equal to 1/(N + 1) and N/N + 1, where N is the
[20] For example, bias correcting a monthly Tavg forecast number of members from which the probability distribution
for January – June requires the following steps: (1) The is estimated), theoretical probability distributions are fit to the
January GSM Tavg is assigned a nonexceedence probability data to extend the empirical distributions. This becomes
(or percentile) within the 210-value GSM climatology necessary because the historical climatology is defined by
distribution for January Tavg. (2) A January Tavg having the 21 years of historical observations, whereas the model
the same nonexceedence probability in the observed clima- ensembles consist of a larger 210-member data set. For low
tology is then calculated. (3) Steps 1 and 2 are repeated for precipitation, an Extreme Value Type III (Weibull) function
ACL 6-6 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING

Figure 4. April 2000 global spectral model (GSM) climatology for monthly total precipitation and
average temperature, compared with observations averaged over the corresponding geographic area. The
data are for the GSM computational cell centered on latitude 37.97°N, longitude 87.19°W, in the Ohio
River basin.

was used, with a minimum lower bound of zero; whereas for member e (e = 1 – 20). TVICmean(m) is the observed 1979–
extreme high precipitation an Extreme Value Type I (Gum- 1999 mean Tavg for month m, and TANOMfcst(m, e) is the
bel) distribution was employed. For temperature, a normal additive Tavg forecast anomaly for month m and ensemble
distribution was used for both minimum and maximum. member e. Likewise, PVICfcst(m, e) is the forecast monthly
[22] The need for bias correction is demonstrated by an Ptot for a given VIC cell in month m of a forecast ensemble
example that compares GSM’s precipitation and temper- member e, PVICmean(m) is the observed 1979 – 1999 mean
ature climatology for one typical cell (centered on latitude Ptot for month m, and PANOMfcst(m, e) is the multiplicative
37.97°N, longitude 87.19°W, in the Ohio River basin) with Ptot forecast anomaly for month m and ensemble member e.
observed values (Figure 4). As alluded to in the introduc- The addition of temperature anomalies will hereafter be
tion, biases of the magnitude shown (e.g., up to 15°C for referred to as shifting, and the multiplication by precipita-
Tavg in August) are occasionally found in climate model tion anomalies will be referred to as scaling.
simulations of surface variables, particularly if one exam- [24] The final step in preparing the forecasts for input to
ines the output for individual cells rather than for a region or the VIC model is to replace the monthly mean sequences
continent. The lack of agreement with observations stems in by daily sequences. For each month (e.g., January) in each
part from poor climate model resolution of subgrid scale forecast ensemble, one year from the climatology period is
land surface related heterogeneities, such as orography or randomly selected (e.g., 1988). For each VIC cell, the
soil wetness. Such biases preclude the use of the climate observed daily values of precipitation for the selected year
model output as a direct input to the hydrology model. and month (e.g., 1988, January) are scaled so that the
2.3.2. Downscaling of Long-Range Ensemble Forecasts monthly total precipitation is equal to the forecast Ptot for
[23] Following bias correction, the monthly GSM scale the ensemble member and month. The resulting values of
forecast anomalies are translated to the spatial and temporal daily precipitation become the daily sequence for that
scale of VIC model inputs. The Tavg and Ptot anomalies are month of the particular forecast ensemble member. Daily
spatially interpolated to the 1/8° VIC cell centers and applied Tmin and Tmax from the same selected year (e.g., 1988) are
to the monthly observed 1979 – 1999 1/8° cell means shifted equally so that their average, (Tmin + Tmax)/2,
(derived from Co-op station observations as described in reproduces the monthly forecast Tavg for the ensemble
section 2.2.1), to create monthly forecast sequences at the member and month, and the resulting values of Tmin and
VIC model scale, in the following manner: Tmax become the daily sequence for that month of the
particular forecast ensemble member. Daily wind speed is
TVICfcst ðm; eÞ ¼ TVICmean ðmÞ þ TANOMfcst ðm; eÞ taken without adjustment from the VIC daily values for the
selected year and month, forming the fourth daily forcing
used by the VIC model. The same year is used to select
the daily data for a given month of an ensemble forecast
PVICfcst ðm; eÞ ¼ PVICmean ðmÞ  PANOMfcst ðm; eÞ: member in every cell of a study area (the Ohio River basin
and east coast). Using the same year-month combination
Here TVICfcst(m, e) is the forecast monthly Tavg for a given for resampling over the large-scale hydrologic units helps
VIC cell in month m (m = 1– 6) of a forecast ensemble to preserve a degree of spatial synchronization in the
 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING ACL 6-7

Figure 5. Climatology period (1979 – 1999) streamflow distribution simulated from daily variable
infiltration capacity (VIC) 1/8° observations, compared with a parallel simulation from monthly GSM-
scale (2.8125°) spatially averaged observations, after the downscaling and disaggregation procedure.

weather components driving hydrologic response. The [27] In addition to the forecast ensemble, we also gen-
random sampling of a climatology year for selection of daily erated a hydrologic ensemble hindcast by applying the
sequences is repeated for each month in each forecast procedures described in sections 2.3.1 – 2.3.2 to the ensem-
ensemble member. ble members of the GSM hindcast in place of the forecast.
[25] We performed a test of this method using observed This hydrologic ensemble hindcast yields a hydrologic
total monthly precipitation and average temperature times- climatology for 1979– 1999 derived by the methods used
eries for 1979 – 1999, aggregated to the GSM scale, as raw to derive the hydrologic ensemble forecast. Rather than
forcings over the Ohio River drainage area. These large- comparing the forecast ensemble results directly with the
scale forcings were processed (using the interpolation and empirial probability distribution of observed streamflow or
temporal disaggregation steps) into daily VIC scale forc- of model-simulated fields (e.g, grid cell runoff or soil
ings, with which we simulated streamflow. Figure 5 shows moisture) based on observations, we compared the hydro-
that the method is able to reproduce the mean and variance logic ensemble forecasts with the hydrologic ensemble
of the basin streamflow climatology without introducing hindcast. In this experiment, we wanted to ensure than
substantial method-related bias. any incidental forecast error associated with the approach
2.3.3. Producing Hydrologic Forecasts would also arise in the climatology distributions with which
[26] Near the tenth day of each month the GSM ensemble the forecasts were compared. Upon completion of each
forecasts become available, and we process the monthly forecast or hindcast run, monthly total precipitation, evap-
GSM output into format suitable for input to VIC. By the oration and runoff (surface plus base flow), and monthly
twentieth day of the month, during the period April – Sep- average soil moisture and temperature were archived, and
tember 2000, the hydrology model state was initialized the daily streamflow routing was performed for the selected
through that current date using a 1-year spin-up simulation, subbasins (shown in Figure 2).
the forcings for which were the gridded observational data 2.3.4. Retrospective ENSO Event Forecast Simulation
described in section 2.2.1. Once the current hydrology model [28] As an additional test of the method, we performed a
moisture states were obtained, the 20 forecast ensemble retrospective comparison of the 10-member hindcast ensem-
members were run to produce an ensemble of six month ble associated with the November 1997 SSTs ( just prior to
long hydrologic forecasts, beginning the following month. the strong 1997 –1998 El Niño), which was extracted from
ACL 6-8 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING

Figure 6. Observed climatology for April through September 2000, defined as the monthly gridded
observations of total precipitation and average temperature, and associated simulated analyses of average
soil moisture and total runoff. These are shown as percentiles of the variables’ observed and simulated
(21 year) climatological distributions, respectively.

the 210-member climatology ensemble for November. The prescribed SSTs and forecast SSTs would not have differed
purpose of this analysis was to evaluate forecast perform- as greatly as in other periods. The hydrologic forecast
ance made from a month exhibiting strong SST anomalies in ensemble based on the 10-member El Niño ensemble
the tropical Pacific Ocean, that is, conditions favorable for members was compared with ensembles yielded from use
skillful climate forecasting. The NINO3 index, which meas- of the entire 210-member November hindcast.
ures the deviation from normal of the sea surface temper-
ature in the eastern Pacific, and which is high during an El
Niño event, reached its highest value in decades in winter 3. Results
1997. We treated the 10-member ensemble for 1997 as a [29] We evaluated the results of the experiment using two
surrogate for an actual forecast made at that time, even types of output: (1) spatially distributed variables such as
though the SSTs used in the hindcast from which it was surface forcings, hydrologic model runoff, and soil moisture
drawn were prescribed according to observations, rather and (2) streamflow at selected locations. These outputs were
than projected. In that particular month, however, forecast generated for the six monthly forecast dates beginning in
skill for tropical Pacific SSTs was relatively high; thus the April 2000. We report here a representative sample of the
 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING ACL 6-9

Figure 7a. April 2000 GSM forecast ensemble medians for May, July, and September monthly total
precipitation and average temperature and GSM forecast-based (VIC simulated) ensemble medians of
average soil moisture and total runoff, shown as percentiles of the 21-year GSM hindcast climatology
distribution for each respective variable.

forecast results for three starting dates, 20 April, 20 June, moisture and runoff. Data and model deficiencies notwith-
and 20 August, and observed conditions for May, July, and standing, these are treated as surrogate observations for the
September. For spatial forecast results the forecast ensemble summer 2000 period. The broad features of the results for
medians are plotted as a percentile of GSM climatology simulated soil moisture and runoff are consistent with the
ensembles derived from GSM hindcasts. These percentiles signals in precipitation and temperature, modulated by the
are verified against the observation-based, retrospective simulated antecedent soil moisture conditions (characterized
forecast fields, shown as percentiles of the observation- by general deficits in the southeast throughout the study
based, retrospective climatology. For streamflow forecast period). Soil moisture and runoff percentiles were quite
results, GSM forecast and hindcast (climatology) distribu- similar, which reflects the VIC model tendency for precip-
tions (discussed in section 2.3.3) are shown, in addition to itation inputs to elevate runoff and base flow in concert with
observations that have become available since the forecasts soil moisture, especially considered from a monthly stand-
were made. point. Against these we contrast Figures 7a and 7b, which
show GSM-based forecasts of the same fields.
3.1. Spatial Analyses [31] In the observational analyses, extremely low May
[30] Figure 6 shows observation-based gridded precipita- precipitation coupled with high temperatures deepened
tion and temperature fields and corresponding simulated soil drought conditions throughout the Ohio River valley and
ACL 6 - 10 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING

Figure 7b. (a) June and (b) August 2000 GSM forecast ensemble medians for July and September
monthly total precipitation and average temperature and GSM forecast-based (VIC simulated) ensemble
medians of average soil moisture and total runoff, shown as percentiles of the 21-year GSM hindcast
climatology distribution for each respective variable.

southeast, while the northeastern United States experienced stricken region, which initially included Florida, shifted west
slightly higher than average precipitation (with respect to the toward Alabama and Louisiana.
1979 – 1999 climatology) (Figure 6). By July, temperatures [32] The April ensemble forecasts (spanning the period
along the east coast and the northern Ohio River valley were May– October) showed above-normal precipitation in the
cooler, and precipitation had risen to above normal in many southeastern United States in May and on the east coast
locations, while the relative dryness and heat persisted from (excepting Florida) in July, but then it showed slightly
the Gulf Coast to the southern Ohio River valley. September below-normal precipitation everywhere except Florida in
brought high temperatures everywhere except Florida and September (Figure 7a). The median forecast was for temper-
Georgia, and the region of low precipitation shifted north, atures slightly above normal everywhere except Florida in
while the drought eased along the Gulf Coast. In response to May, then for temperatures more strongly above normal west
these forcings, anomalously low soil moisture and runoff, of the Atlantic states in July and September. Normal to cool
which had been general over the entire domain south of New conditions were forecast in the Atlantic states during the
England, recovered gradually along the east coast and the study period. Consequently, initially dry soil moisture and
northern Ohio River valley. The center of the drought- below-normal runoff were predicted to recover (at least in
 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING ACL 6 - 11

Figure 8. April, June, and August 2000 monthly average streamflow forecast and hindcast
(climatology) ensembles compared with observed values, for the Potomac River near Washington,
D.C., Little Falls (location shown in the upper right).

the median forecast) in Florida and the mid-Atlantic states by (upper right) and the forecast to climatology ensemble
July but to linger in parts of Alabama and Louisiana. In comparisons for the same three forecast start dates detailed
southern New England and the Ohio River valley, however, in Figure 7.
the median forecast called for a continuation of dryness and [36] Figure 8, for the Potomac River near Washington,
heat and hence low soil moisture and runoff. D.C., Little Falls, shows forecast streamflow distributions
[33] The June forecasts (Figure 7b, top, shows July and for the April forecast which are similar to the hindcast
September) anticipated, in the median, above-normal pre- climatology distributions but a bit lower in May and
cipitation in the southeastern United States in July and on the September, when it can be seen from Figure 7a that the
East Coast and Ohio River valley (except in New England) entire watershed is drier than normal. The observed stream-
in September and cooler temperatures everywhere in both flows in May were lower than either distribution suggested,
months. Even so, the forecasts indicated that the dry soil but the forecast correctly indicates the direction of the
moisture and below normal runoff would fail to recover fully anomaly. In June and August the forecasts and climatology
in the southern and Gulf states, while in southern New distributions have similar medians except for July stream-
England and the Ohio River valley, they would to transition flow in the June forecast, when the forecast is erroneously
to above-normal levels. higher, reflecting the above-normal precipitation forecast.
[34] The median August forecasts (Figure 7b, bottom, for The forecasts in June and August agree fairly well with
September) called for dry and hot conditions in Florida, the observations, which were not far from normal for the June
Gulf states, and the Ohio River valley, with above-normal to September period.
precipitation and cooler to normal temperatures from the [37] In Figure 9, showing the April forecast for the
mid-Atlantic states to New England. These circumstances Alabama River at Claiborne Lock and Dam, the forecast
would serve to aggravate the low August soil moisture ensemble distributions are slightly lower than the climatol-
centered on Alabama. ogy ensembles in the first 2 months of the forecast period
but thereafter are similar. The forecasts give a slight
3.2. Streamflow Evaluation indication that streamflows will be low early in summer
[35] Predicted streamflows reflected the condition of soil but give no indication of the severity of the streamflow
moisture and runoff, although the deviation from normal anomaly that is observed. This result is consistent with the
was generally small relative to the variability exhibited by above-normal simulated antecedent soil moisture for April
both the climatological and forecast ensemble distributions. (Figure 6) in half of the watershed and the forecast of
Figures 8 and 9 show the location of the streamflow sites above-normal precipitation in July (Figure 7a). Subsequent
ACL 6 - 12 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING

Figure 9. April, June, and August 2000 monthly average streamflow forecast and hindcast
(climatology) ensembles compared with observed values, for the Alabama River at Claiborne Lock
and Dam, Ala. (location shown in the upper right).

forecasts in June and August, however, are significantly temperatures were mostly above normal throughout the
lower than the climatology ensemble distributions, respond- domain. The net result for soil moisture and runoff for fall
ing to the dry initial conditions for those forecasts (Figure 6) and winter 1997 – 1998 was wetter than average in the
and near- or below-normal precipitation thereafter (Figure southeastern United States and along the east coast, with
7b). The June and August forecasts distinctly anticipated the the largest soil moisture anomaly in the southeastern United
severe declines in summer streamflow that were observed. States in December 1997 to March 1998 and in the north-
eastern United States in March and April 1998.
3.3. El Niño Period Forecast Results [39] The spatial (perfect SST) forecasts made in Novem-
[38] Consistent with the expectation of an anomalously ber 1997 appear to correctly indicate the direction of soil
wet winter season in the southeast associated with the El moisture and runoff anomalies; that is, the median forecast
Niño climate phase [Changnon, 1999; Barnston et al., was for wetter than normal conditions. Exceptions were the
1999], the gridded observed precipitation and temperature northeastern United States and the Ohio River basin, where
and associated hindcast hydrologic simulations of soil the median forecasts were wetter than average as early as
moisture and runoff for November 1997 (not shown here) December 1997 and drier than average in April and May
showed wetter than normal (again, with respect to the 1998, whereas the retrospective analysis showed the oppo-
1979 – 1999 climatological period) precipitation in the site condition. Also, the forecast percentiles tended toward
southeastern United States and the Atlantic drainages, with the median relative to the hindcast percentiles, perhaps as a
drier than average conditions in the Ohio River basin, a result of the finer resolution afforded the probability scale
pattern echoed in the relative soil moistures and runoff. At by the use of 210 ensemble members for the climatology
the same time, temperatures throughout the study domain compared with the 21 ensemble members used to provide
were relatively and uniformly cool. In December, the south- statistical context for the forecasts. A comparison of months
eastern states were again very wet while the rest of the 1– 3 (December 1997 to February 1998) of the hindcast and
domain received about normal precipitation, and the entire the forecast is shown in Figure 10.
region was still relatively cool, but less so than in the [40] A sample of streamflow results for the El Niño
previous month. In January through March, the entire period forecast is given in Figure 11 for the Potomac River
domain received above-average precipitation, but by April near Washington, D.C., Little Falls. The Potomac River
drier weather appeared to move north from Florida, even- watershed’s slightly below-normal precipitation in Decem-
tually extending to the entire southeast. During this time, ber and near-normal soil moisture (Figure 10) led to a
with the exception of a cool March in the southeast, forecast ensemble with a similar median to the climatology
 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING ACL 6 - 13

Figure 10. (a) December 1997 – February 1998 gridded observed monthly total precipitation and
average temperature and associated analyses of average soil moisture and total runoff, shown as
percentiles of the observed climatology; (b) November 1997 GSM-derived forecasts for the same period,
shown as percentiles of the GSM-based hindcast climatology.

ensemble in December. Both ensembles were slightly higher ogy forecasts without undertaking to determine the skill of
than the observed streamflow. The forecasts’ above-normal the climate model and resulting hydrologic forecasts quan-
precipitation in January and February produced above- titatively (hence the term ‘‘skill’’ is here used loosely to
normal forecast distributions for streamflow in those indicate general consistency of forecasts with observed
months, although the forecasts, as in the Alabama River values). We did not focus on climate model forecast skill
example shown in Figure 9, failed to anticipate the magni- because our primary purpose was to develop a framework
tude of the anomaly actually observed, which continued within which ensemble climate forecasts could be used for
throughout the forecast period. hydrological purposes. A secondary objective was to deter-
mine whether the climate model forecast signal or hydro-
logic (soil moisture) persistence would dominate in
4. Discussion situations where the climate forecast anomalies were sig-
[41] We evaluated the spatial forecasts from a qualitative nificant. We conclude from this exercise that the down-
standpoint only, broadly assessing the consistency of anom- scaling procedure successfully transfers the climate forecast
aly direction in the climate forecasts and resulting hydrol- signals to the hydrologic variables. It is especially encour-
ACL 6 - 14 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING

in the antecedent hydrologic model state. The initial soil

moisture signature, characterized by a decreasing gradient
in soil moisture and runoff percentile from south to north
(which resulted from the antecedent conditions) was largely
erased by the normal to wetter than average precipitation
forecasts coupled with normal to cooler than average
temperature forecasts (after the December warmth in the
northeast). By February (forecast month 3) the wet anomaly
over the entire region was consistent with, although weaker
than, the anomaly revealed in the hindcast analysis.
[45] One issue that bears mention is the high variance in
forecast ensemble and climatology ensemble spatial fields
(for a given point) and hence in the streamflow ensembles
(Figures 8, 9, and 11). Given a high variance, a large shift in
Figure 11. El Niño condition streamflow forecast for the the mean of the forecast distribution is required to produce a
Potomac River near Washington, D.C., Little Falls, using a statistical difference in the forecast outcome; hence the
10-member GSM hindcast ensemble for November 1997 as discrimination of the forecast system (in the sense described
a forecast surrogate, compared with the GSM hindcast by Wilks [1995]) is weaker than it would be for a narrower
ensembles for November 1979 – 1999 (a climatology) and distribution of ensembles. A wide forecast distribution
with observed streamflows. prohibits water managers from making decisions that effec-
tively rule out one end or the other (or, in some cases, both)
of the climatological distribution of expected hydrologic
aging that the hydrologic model, which performs a non- conditions for the forecast period. Future work in refining
linear transformation of temperature, precipitation, and methods to bias-correct and downscale climate forecasts for
other inputs to streamflow, was able to retrieve the observed use in hydrologic prediction must therefore take care to
streamflow climatology from the downscaled, GSM-scale minimize the addition of method-related uncertainty.
observed precipitation and temperature climatology. [46] Further exploration of the approach should also
[42] One feature of this approach that may require further include a broader range of climate and land surface con-
evaluation is the spatial interpolation of bias-corrected GSM- ditions than were examined here. Where snowpack plays a
scale output anomalies, rather than their associated proba- major role in the seasonal cycle, for example, contributions of
bility values, directly to the VIC 1/8° resolution grid cell the hydrologic and climate components of the forecasts are
centers. Interpolation of anomaly quantiles (in probability expected to be significantly different at different times of the
space) to the VIC resolution would yield a joint probability year. An effort to determine a priori where and when long-
anomaly for the thousands of VIC cells in each basin far range forecasts are likely to have skill, purely on the basis of
exceeding the original GSM anomaly quantile, so the latter land surface and climate considerations, would be useful.
approach was chosen as a method of bridging the scale gap [47] Our experience in this study suggests that future
between GSM output and VIC input. It also appears that the work to apply the climate-hydrology model forecasting
downscaling procedure may increase the variability of the approach in real time and to assess the outputs quantita-
forecasts somewhat by producing high outliers in the precip- tively should also concentrate on improvements in two
itation fields (as a result of the rescaling of sampled daily areas. First, the accurate estimation of initial hydrologic
patterns to match monthly anomalies), and refinements will conditions for the forecasts is critical for capturing the
be pursued in future applications to resolve this problem. influence of land surface anomalies that persist into the
[43] For the summer 2000 study period and region the forecast period; hence an improvement in real-time access
climate model forecasts had a mixed performance, as to meteorological data for hydrologic simulation of initial
estimated spatially from the precipitation, temperature, run- conditions would increase the accuracy of the forecasts.
off, and soil moisture ensemble medians and from the Second, quantitative evaluation of the climate forecasts
streamflow comparisons. For example, in the April and would benefit from the existence of retrospective meteoro-
June month 1 median forecasts (Figure 7), the southeast is logical forecast data sets generated with the identical
slightly wetter than the climatology median, while our methods (to the extent possible) used in producing the
retrospective analysis shows that the region continued to current forecasts, for a climatology period of several deca-
be dry. In each case, however, the persistence of antecedent des. The retrospective perfect SST-based surrogate forecast
low soil moisture from the hydrology model maintained low of the type explored here hints only at an upper bound on
soil moisture and runoff several months into the forecast forecast performance, rather than an estimate of perform-
period, so that the forecast results still agreed reasonably ance consistent with the forecasts produced in real time.
well with the retrospective analysis for lead times of several [48] Nonetheless, from an end user standpoint, the fore-
months. Here the weak and at times incorrect direction of casting approach appears to have potential utility for con-
the climate model forecasts was balanced by a degree of ditioning water resources related outlooks, particularly
skill derived from persistence in the hydrologic states. when there is either a strong anomaly signal in the climate
[44] For the El Niño-condition forecasts of November forecasts or highly anomalous antecedent conditions in the
1997 (Figure 10), in contrast to the summer 2000 results, hydrologic model state. A quantitative exploration into the
the hydrologic forecast results appeared to be determined by suitability of the spatial and streamflow forecasts for partic-
the climate model forecast signal, as well as by persistence ular water resources applications appears to be warranted.
 WOOD ET AL.: SEASONAL HYDROLOGIC FORECASTING ACL 6 - 15

[49] Acknowledgments. This publication was supported in part by a Liang, X., D. P. Lettenmaier, E. F. Wood, and S. J. Burges, A simple
grant to the Joint Institute for the Study of the Atmosphere and Ocean hydrologically based model of land surface water and energy fluxes for
(JISAO) at the University of Washington, under NOAA Cooperative general circulation models, J. Geophys. Res., 99(D7), 14,415 – 14,428,
Agreement NA17RJ1232, Contribution 911, and as part of the GEWEX 1994.
Continental-Scale International Project (GCIP). Liang, X., E. F. Wood, and D. P. Lettenmaier, Surface soil moisture para-
meterization of the VIC-2L model: Evaluation and modifications, Global
Planet. Change, 13, 195 – 206, 1996.
References Liang, X., E. F. Wood, and D. P. Lettenmaier, Modeling ground heat flux in
Abdulla, F. A., D. P. Lettenmaier, E. F. Wood, and J. A. Smith, Application land surface parameterization schemes, J. Geophys. Res., 104(D8),
of a macroscale hydrologic model to estimate the water balance of the 9581 – 9600, 1999.
Arkansas-Red River Basin, J. Geophys. Res., 101(D3), 7449 – 7460, 1996. Livezey, R., Variability of skill of long-range forecasts and implications for
Barnston, A. G., and Y. He, Skill of canonical correlation analysis forecasts their use and value, Bull. Am. Meteorol. Soc., 71, 300 – 309, 1990.
of 3-month mean surface climate in Hawaii and Alaska, J. Clim., 9, Livezey, R. E., M. Masutani, and M. Ji, SST-forced seasonal simulation and
2579 – 2605, 1996. prediction skill for versions of the NCEP/MRF model, Bull. Am. Meteor-
Barnston, A. G., A. Leetmaa, V. Kousky, R. Livezey, E. A. O’Lenic, H. Van ol. Soc., 77, 507 – 517, 1996.
den Dool, A. J. Wagner, and D. A. Unger, NCEP forecasts of the El Nino Livezey, R. E., M. Masutani, A. Leetmaa, H.-L. Rui, M. Ji, and A. Kumar,
of 1997 – 98 and its U.S. impacts, Bull. Am. Meteorol. Soc., 80(9), 1999. Teleconnective response of the Pacific/North American region atmo-
Beven, K. J., and M. Kirkby, A physically based, variable contributing area sphere to large central equatorial Pacific SST anomalies, J. Clim., 10,
model of basin hydrology, Hydrol. Proc., 6, 279 – 298, 1979. 1787 – 1820, 1997.
Changnon, S., Impacts of 1997 – 98 El Nino-generated weather in the Uni- Lohmann, D., E. Raschke, B. Nijssen, and D. P. Lettenmaier, Regional
ted States, Bull. Am. Meteorol. Soc., 80(9), 1999. scale hydrology, II, Application of the VIC-2L model to the Weser River,
Chen, M., R. E. Dickinson, X. Zeng, and A. N. Hahmann, Comparison of Germany, J. Sci. Hydrol., 43, 143 – 158, 1998a.
precipitation observed over the continental United States to that simulated Lohmann, D., et al., The project for intercomparison of land-surface para-
by a climate model, J. Clim., 9, 2233 – 2249, 1996. meterization schemes (PILPS) phase-2c Red-Arkansas River Basin ex-
Cherkauer, K., and D. P. Lettenmaier, Hydrological effects of frozen soils in periment, 3, Spatial and temporal analysis of water balance fluxes,
the upper Mississippi River basin, J. Geophys. Res., 104(D16), 19,599 – J. Global Planet. Change, 19, 161 – 179, 1998b.
19,610, 1999. Maurer, E. P., D. P. Lettenmaier, and J. O. Roads, Water balance of the
Cocke, S., and T. E. LaRow, Seasonal predictions using a regional spectral Mississippi River basin from a macroscale hydrologic model and NCEP/
model embedded within a coupled ocean-atmosphere model, Mon. NCAR reanalysis, Eos Trans. AGU, 80(46) Fall Meet. Suppl., F409 –
Weather Rev., 128, 689 – 708, 2000. F410, 1999.
Ducharne, A., R. D. Koster, M. J. Suarez, M. Stieglitz, and P. Kumar, A Maurer, E. P., G. M. O’Donnell, D. P. Lettenmaier, and J. O. Roads,
catchment-based approach to modeling land surface processes in a gen- Evaluation of NCEP/NCAR reanalysis water and energy budgets using
eral circulation model, 2, Parameter estimation and model demonstration, macroscale hydrologic simulations as a benchmark, in Observations and
J. Geophys. Res., 105(D20), 24,823 – 24,838, 2000. Modeling of the Land Surface Hydrological Processes, Water Sci. Appl.,
Georgakakos, K. P., A. P. Georgakakos, and N. E. Graham, Assessment of vol. 3, edited by V. Lakshmi, J. Albertson, and J. Schaake, 137 – 158,
benefits of climate forecasts of reservoir management in the GCIP region, AGU, Washington, D. C., 2001.
GEWEX News, 8(3), 1998. Mitchell, K., et al., Recent GCIP-sponsored advancements in coupled land-
Giorgi, F., and L. O. Mearns, Approaches to the simulation of regional surface modeling and data assimilation in the NCEP Eta mesoscale mod-
climate change: A review, Rev. Geophys., 29, 191 – 216, 1991. el, paper P1.22 presented at 15th Conf. on Hydrol., Am. Meteorol. Soc.,
Hamlet, A. F., and D. P. Lettenmaier, Effects of climate change on hydrol- Long Beach, Calif., 2000.
ogy and water resources in the Columbia River basin, J. Am. Water Murphy, J., An evaluation of statistical and dynamical techniques for down-
Resour. Assoc., 35(6), 1597 – 1623, 1998. scaling local climate, J. Clim., 12(8), 2256 – 2284, 1999.
Hamlet, A. F., and D. P. Lettenmaier, Columbia River streamflow forecast- Nijssen, B., D. P. Lettenmaier, X. Liang, S. W. Wetzel, and E. F. Wood,
ing based on ENSO and PDO climate signals, J. Water Resour. Plann. Streamflow simulation for continental-scale river basins, Water Resour.
Manage., 125(6), 333 – 341, 1999. Res., 33, 711 – 724, 1997.
IPCC (Intergovernmental Panel on Climate Change), Climate Change Risbey, J. S., and P. H. Stone, A case study of the adequacy of GCM
1995: The Science of Climate Change: Contribution of Working Group simulations for input to regional climate change assessments, J. Clim.,
I to the Second Assessment Report of the Intergovernmental Panel on 9, 1441 – 1467, 1996.
Climate Change, edited by J. T. Houghton et al., 572 pp., Cambridge Roads, J. O., S. C. Chen, M. Kanamitsu, and H. Juang, Surface water
Univ. Press, New York, 1996. characteristics in NCEP global spectral model and reanalysis, J. Geophys.
Ji, M., D. W. Behringer, and A. Leetmaa, An improved coupled model for Res., 104(D16), 19,307 – 19,327, 1999.
ENSO prediction and implications for ocean initialization, II, The Shukla, J., Predictability in the midst of chaos: A scientific basis for climate
coupled model, Mon. Weather Rev., 126, 1022 – 1034, 1998. forecasting, Science, 282, 728 – 731, 1998.
Kalnay, E., et al., The NCEP/NCAR 40-year reanalysis project, Bull. Am. Wilby, R. L., and T. M. L. Wigley, Downscaling general circulation model
Meteorol. Soc., 77, 437 – 471, 1996. output: A review of methods and limitations, Prog. Phys. Geogr., 21,
Kidson, J. W., and C. S. Thompson, A comparison of statistical and 530 – 548, 1997.
model-based downscaling techniques for estimating local climate varia- Wilby, R. L., T. M. L. Wigley, D. Conway, P. D. Jones, B. C. Hweitson, J.
tions, J. Clim., 11, 735 – 753, 1998. Main, and D. S. Wilks, Statistical downscaling of general circulation
Kim, J., N. L. Miller, J. D. Farrara, and S.-Y. Hong, A seasonal precipitation model output: A comparison of methods, Water Resour. Res., 34,
and stream flow hindcast and prediction study in the western United 2995 – 3008, 1998.
States during the 1997/98 winter season using a dynamic downscaling Wilby, R. L., L. E. Hay, W. J. Gutowski Jr., R. W. Arritt, E. S. Takle, Z. Pan,
system, J. Hydrometeorol., 1(4), 311 – 329, 2000. G. H. Leavesley, and M. P. Clark, Hydrological responses to dynamically
Koster, R. D., M. J. Suarez, and M. Heiser, Variance and predictability of and statistically downscaled climate model output, Geophys. Res. Lett.,
precipitation at seasonal-to-interannual timescales, J. Hydrometeorol., 27(8), 1199 – 1202, 2000.
1(1), 26 – 46, 1999. Wilks, D. S., Statistical Methods in the Atmospheric Sciences, Academic,
Koster, R. D., M. J. Suarez, A. Ducharne, M. Stieglitz, and P. Kumar, A San Diego, Calif., 1995.
catchment-based approach to modeling land surface processes in a gen- Wood, E. F., X. Liang, D. Lohmann, and D. P. Lettenmaier, The project for
eral circulation model, 1, Model structure, J. Geophys. Res., 105(D20), intercomparison of land-surface parameterization schemes (PILPS) phase-
24,809 – 22,822, 2000. 2(c) Red-Arkansas River experiment, 1, Experiment description and sum-
Kumar, A., M. Hoerling, M. Ji, A. Leetmaa, and P. Sardeshmukh, Assessing mary intercomparisons, J. Global Planet. Change, 19, 115 – 135, 1998.
GCM’s suitability for making seasonal predictions, J. Clim., 9(1), 115 –
129, 1996.






















Leavesley, G. H., and L. G. Stannard, The precipitation-runoff modeling A. Kumar, Climate Prediction Center, NOAA National Center for
system—PRMS, in Computer Models of Watershed Hydrology, edited by Environmental Protection, Camp Springs, MD 20748, USA. (arun.kumar@
V. P. Singh, pp. 281 – 310, Water Resour. Publ., Highlands Ranch, Colo., noaa.gov)
1995. D. P. Lettenmaier, E. P. Maurer, and A. W. Wood, Department of Civil
Leung, L. R., A. F. Hamlet, D. P. Lettenmaier, and A. Kumar, Simulations and Environmental Engineering, Box 352700, University of Washington,
of the ENSO hydroclimate signals in the Pacific Northwest Columbia Seattle, WA 98195, USA. ([email protected]; edm@hydro.
River Basin, Bull. Am. Meteorol. Soc., 80(11), 2313 – 2329, 1999. washington.edu; [email protected])