On the causes of declining Colorado River streamflows
Mu. Xiao1, Bradley Udall2, and Dennis P. Lettenmaier1
1

Department of Geography, University of California, Los Angeles, CA, USA

2

Colorado Water Institute, Colorado State University, Fort Collins, CO, USA

Corresponding author: Dennis P. Lettenmaier (dlettenm@ucla.edu)

Key Points:
The naturalized flow of the Colorado River has decreased about 15% over the last 100
years
About half of the long-term trend is attributable to rising temperatures with the
remainder due to changes in precipitation patterns and other factors
Slightly over half of the flow reduction during the post-2000 Millenium Drought is
attributable to anomalously warm temperatures

This article has been accepted for publication and undergone full peer review but has not been
through the copyediting, typesetting, pagination and proofreading process which may lead to
differences between this version and the Version of Record. Please cite this article as doi:
10.1029/2018WR023153
© 2018 American Geophysical Union. All rights reserved.

 Abstract
The Colorado River is the primary surface water resource in the rapidly growing
U.S. Southwest. Over the period 1916-2014, the Upper Colorado River Basin naturalized
streamflow declined by 16.5%, despite the fact that annual precipitation in the UCRB over
that period increased slightly (+1.4%). In order to examine the causes of the runoff
declines, we performed a set of experiments with the Variable Infiltration Capacity (VIC)
hydrology model. Our results show that the pervasive warming has reduced snowpacks
and enhanced evapotranspiration (ET) over the last 100 years; over half (53%) of the
long-term decreasing runoff trend is associated with the general warming. Negative
winter precipitation trends have occurred in the handful of highly productive sub-basins
that account for over half of the streamflow at Lee’s Ferry. We also compared a midcentury drought with the (ongoing) post-Millennium Drought, and find that whereas the
earlier drought was caused primarily by pervasive low precipitation anomalies across
UCRB, higher temperatures have played a large role in the post-Millennium Drought. The
post-Millennium Drought has also been exacerbated by negative precipitation anomalies
in several of the most productive headwater basins. Finally, we evaluate the UCRB AprilJuly runoff forecast for 2017, which decreased dramatically as the runoff season
progressed. We find that while late winter and spring 2017 was anomalously warm, the
proximate cause of most of the forecast reduction was anomalous late winter and early
spring dryness in UCRB, which followed exceptionally large (positive) early winter
precipitation anomalies.

© 2018 American Geophysical Union. All rights reserved.

 1.0 Introduction
The Colorado River is the largest river in the Southwestern U.S. It is the source of
drinking water for many of the Colorado River Basin’s 40 million people, and provides
irrigation water to ~13,000 km2 of crops in the U.S. and Mexico (Cohen et al. 2013). It is
a lifeline for the population and agricultural economy of parts of seven U.S. states (WY,
UT, CO, NV, NM, AZ and CA) and the Mexican states of Sonora and Baja California. The
river’s naturalized streamflow (see Section 2.2 for discussion of naturalized streamflows)
at Imperial Dam (the downstream-most long-term gauging station) has averaged about
20.7 km3/yr (16.8 MAF/yr) over the last century, approximately 90% of which is
generated in the Upper Colorado River Basin (McCabe and Wolock 2007), defined as the
~289,000 km2 of drainage area upstream of the U.S. Geological Survey stream gauge at
Lees Ferry, AZ (USGS 09380000). Snowpack stored in the high elevation Rocky Mountain
headwater basins contributes about 70% of the annual streamflow (Christensen et al.
2004).
The Colorado River is heavily regulated, mostly by Glen Canyon Dam (Lake
Powell) and Hoover Dam (Lake Mead), with combined reservoir storage capacity of 67.5
km2 (54.7 maf). The importance of these reservoirs, which can store close to four times
the natural annual flow at Lees Ferry, AZ, has become especially evident during the socalled Millennium Drought, which began about 2000. This drought has coincided with
increases in water demand (Rajagopalan et al. 2009), which resulted in Lake Mead
reaching its lowest level on record in October 2016. Lakes Mead and Powell dropped
precipitously from 2000-2004 due to very low flows (71%, 74%, 41%, 71% and 64% of
average, respectively), and have not recovered due to continued high demands equal to
inflows and a lack of high flow years. Indeed, only four of the last 18 years have had above
average river discharge, limiting reservoir refill opportunities.

© 2018 American Geophysical Union. All rights reserved.

 A pronounced warming trend across the Colorado River Basin (CRB) since the
1970s (Dawadi and Ahmad 2012) has further contributed to the post-2000 imbalance
between CRB runoff and water demand. Vano et al. (2012) evaluated the temperature
sensitivity (annual average streamflow change per 1°C temperature change) and found
that the average sensitivity of annual runoff at Lees Ferry was around -5% per °C),
suggesting that warming over the last ~50 years may account for a 5-10% reduction in
annual streamflow over that period.
Several studies have investigated the effects of ongoing warming on the flow of
the Colorado River. Barnett and Pierce (2009) concluded that anthropogenic climate
change would reduce CRB runoff by 10%-30% by 2050. Reynolds et al. (2015) predicted
that minimum streamflows will decline as warming of the basin continues. Woodhouse
et al. (2016) report an increase in the frequency of warm years with low streamflow since
1988. McCabe et al. (2017) found that increases in temperature since the late 1980s have
decreased runoff generation efficiency, reducing streamflows by 7%. Udall and Overpeck
(2017) similarly found temperature-induced streamflow decreases of approximately 6%
during 2000-2014 and projected large mid-century temperature-induced declines of
20% or more should precipitation not change.
Here, we utilize a hydrological model applied for the period 1916-2014 (all data
are for water years if not specified otherwise) to evaluate the spatial and temporal
signature of the Millennium Drought in the CRB. Along with a baseline simulation forced
by gridded observations, we perform a T-detrend experiment, in which we remove the
long-term temperature trend from the model forcings, to investigate the role of the
warming on streamflow declines both over the long term and during the recent drought.
We analyze runoff in each of 20 sub-basins of the CRB, which allows us to study spatial
variations in runoff generation and anomalies. We also analyze the historical 1953-68

© 2018 American Geophysical Union. All rights reserved.

 drought in an attempt to shed light on how the hydrologic response to climate variations
has changed in recent decades, and during the Millennium Drought in particular. Finally,
we dissect the 2017 April-July streamflow forecast to understand the role of late winter
and early spring precipitation and temperature in the substantial seasonal forecast
reductions that occurred as water year 2017 progressed.
2.0 Data and approach
2.1 VIC model and forcings
The Variable Infiltration Capacity (VIC) model is a physically-based, semidistributed hydrological model which represents the land surface water and energy
budgets over a grid mesh (here 1/16th degree spatial resolution) and routes runoff
through a prescribed river network to produce streamflow estimates at specified river
nodes (Liang and Lettenmaier 1994). We applied the model at a daily time step, using
what is termed full-energy balance mode, meaning that the model iteratively solves the
surface energy budget by estimating the effective surface temperature at each time step.
Therefore, the daily average surface temperature produced by VIC is not the average of
the forcing temperatures, i.e. 0.5*(daily maximum+daily minimum). Unless stated
otherwise, the temperatures we report here are outputs from the VIC simulations.
Similar to other land surface models, the fundamental water balance equation in
VIC can be summarized as Runoff (RO) = Precipitation (P) - Evapotranspiration (ET) –
changes in Soil Moisture (∆SM) - changes in Snow Water Equivalent (∆SWE).
Groundwater is not represented in the version of VIC we used; Rosenberg et al. (2013)
found that inclusion of a parameterzation of groundwater had little effect on the model’s
streamflow simulations in the CRB. It is important to note that VIC represents snowpack
sublimation within its winter ET. Sublimation is sparsely measured but nonetheless is

© 2018 American Geophysical Union. All rights reserved.

 important to some aspects of our study (Andreadis et al. 2009); we describe the model’s
performance with respect to sublimation in section 4.2. The VIC model has been
successfully applied previously in a number of hydrological studies over the CRB and the
U.S. Southwest (Mote et al. 2005, 2018; Christensen et al. 2004; Christensen and
Lettenmaier 2007; Vano et al. 2012, 2014)
The VIC model simulates surface hydrological processes with parameterizations
of sub-grid vegetation, soil variability and topography, and has provided plausible
representations of CRB surface water conditions in the above-referenced studies. We
forced the model with an updated version of the Hamlet and Lettenmaier (hereafter H&L)
dataset (Hamlet and Lettenmaier 2005) at 1/16-degree resolution for the period water
years 1916-2014. We chose the H&L data set because its long-term variability is indexed
to the U.S. Historical Climatology Network (HCN; Easterling et al. 1996) stations in the
region, which have been carefully quality controlled for effects that could otherwise
result in spurious trends, such as station moves and instrument changes (e.g., the shift to
MMTS temperature sensors in the 1980s). As described in Hamlet and Lettenmaier
(2005) the H&L dataset uses HCN station data to constrain decadal variability (and hence
long-term trends), hence is in our view most appropriate for exploration of the causes of
century-scale streamflow declines over our study period 1916-2014.
2.2 Naturalized streamflows
To evaluate our model simulation results, we used naturalized streamflow data
for the Colorado River produced by the U.S. Bureau of Reclamation (USBR); see
https://www.usbr.gov/lc/region/g4000/NaturalFlow/current.html for details. The
naturalized streamflows are derived from USGS historical streamflow observations by a
process of adjustments that compensate for anthropogenic effects including consumptive

© 2018 American Geophysical Union. All rights reserved.

 uses of water, reservoir storage, trans-basin diversions, and other effects (see USBR,
1983). The naturalized streamflow data sets are produced for 29 well-distributed
tributary stations across the CRB (as well as the main stem) for the period 1906 through
2015. Others (Prairie and Callejo 2005) have noted that USBR has improved the quality
of the naturalized flow data set after 1971 and the estimates may be somewhat better
after that time.
2.3 Sub-basin analysis
We performed our analyses for the Colorado River above Imperial Dam, as well as
for the 20 sub-basins delimited by USGS WaterWatch gauges (see Figure 1) which are a
subset of the 29 naturalized streamflow points noted above. The river channel network
dataset we used is from Wu et al. (2012), based on which we determined the masks for
each of the 20 sub-basins. The Wu et al. sub-basins are similar to, but slightly different
from the more familiar 6-digit Hydrologic Unit Codes normally used in the Basin. Detailed
information about each sub-basin is reported in the SI.
It is important to note that our analysis excludes the Gila River given its distinct
hydrological and legal characteristics. The Gila River joins the Colorado River below
Imperial Dam just upstream of the U.S. border with Mexico, and in recent years has been
mostly dry at its mouth due to upstream uses by Arizona. Since 1964, the U.S. Supreme
Court has excluded it from administration under the Colorado River Compact. Although
the Gila is an important basin, its absence from this study is logical given its unique status.
2.4 Model testing and evaluation
Table 1 summarizes the long-term runoff contribution percentages from nine
major sub-basins at which naturalized streamflows are available, and for which we also
produced VIC simulations. The runoff contribution percentages from the model and

© 2018 American Geophysical Union. All rights reserved.

 naturalized flows generally are in good agreement. The Upper Basin (UCRB; defined as
the drainage area above Lees Ferry, AZ) produces more than 90% of the flow at Imperial
Dam. Therefore, we mainly focus on the UCRB here, acknowledging unusual Lower Basin
(LCRB) conditions when noteworthy.
Figure 2a shows the annual time series of naturalized streamflow (NFL) and VIC
simulations at Lees Ferry, AZ. Both the annual naturalized streamflows and VIC
simulations (r2=0.75), and their trends over the period of record (NFL: -3.3 km3/yr, VIC:
-3.4 km3/yr) are similar, suggesting that the VIC model provides a plausible
representation of natural conditions (i.e., those responding primarily to climate forcings)
and long-term hydrologic change in the basin. Hereafter, we mainly focus on VIC results
in our analysis of UCRB sub-basin long-term (1916-2014) trends and comparison
between the 1953-1968 and the Millennium drought. The annual precipitation and
average temperature (calculated by VIC as noted in section 2.1) time series plots are also
presented in Figure 2.

3.0 Results
3.1 Basin-wide trend analysis
Table 2 summarizes long-term linear (regression) trends for the UCRB for four
hydrological variables (precipitation, evapotranspiration, runoff and April 1 snow water
equivalent) from the baseline VIC simulation, and the temperature detrended (Tdetrend) simulation. We also computed trends using the non-parametric Theil-Sen slope
estimator (Theil 1950; Sen 1968) and found that they generally are in close agreement
(Table S1). Therefore, we refer to the linear trends hereafter for convenience. The Tdetrend simulation uses the same forcings as the baseline, except that annual linear
trends in the daily temperature maxima and minima are removed. We also disaggregated
© 2018 American Geophysical Union. All rights reserved.

 summer season (Apr-Sep) and winter season (Oct-Mar) for each variable (all summers
and winters mentioned hereafter are so defined).
Over the simulation period 1916-2014 the UCRB annual precipitation increased
by +1.5 km3 (1.4%) whereas winter precipitation, which is the main source for Apr 1st
snow water equivalent and streamflow in the spring and summer, had only a very small
(not statistically significant) negative trend (long-term ∆P is -0.1 km3, -0.2%). In our
baseline simulation, the long-term linear change of annual runoff (∆RO) in the UCRB is 3.4 km3 (-16.5%) and long-term change in annual evapotranspiration (∆ET) is +4.2 km3
(+4.7%). Apr 1st SWE decreased significantly (∆SWE -9.1 km3, -39.0%), which reduces
warm season streamflow from the Upper Basin, as evidenced by summer RO decreases (3.8 km3, 23.3%) even given a positive trend in summer precipitation (∆P summer is +1.6
km3). As summer RO makes up more than 3/4 of the annual RO in the UCRB, the longterm annual ∆RO is negative as noted above, although summer RO decreases are slightly
compensated by increasing winter RO (∆ROwinter +0.4 km3, 10.4%).
We performed the T-detrend simulation using the same precipitation as the
baseline simulation but with the temperature trend removed from the forcing dataset on
a grid cell by grid cell basis. In this no-warming-trend scenario, the long-term decreasing
trend in annual runoff is reduced to -1.6 km3 (-7.7%), from -3.4 km3 but not eliminated.
It suggests that 53% (-1.8 / -3.4) of the annual runoff trend is attributable to the annual
warming temperature. The increase in ET in the T-detrend simulation is smaller by 1.9
km3 (baseline: +4.2 km3, T-detrend: +2.3 km3), which explains the increase in runoff (1.8
km3) to within 0.1 km3.
The numbers in Table 2 also show that the effects of the temperature trend on
winter RO (baseline: +0.4 km3, T-detrend: +0.4 km3) and summer ET (baseline: -0.8 km3,
T-detrend: -0.6 km3) are small. Increasing temperatures cause a decrease in summer RO

© 2018 American Geophysical Union. All rights reserved.

 (baseline: -3.8 km3, T-detrend: -1.9 km3), and an increase in annual ET (baseline: +4.2
km3, T-detrend: +2.3 km3) that comes mostly in the winter (baseline: +4.9 km3, Tdetrend: +2.9 km3). On a percentage basis, both of these increasing winter trends in ET
are substantial over the 1906-2014 period: a 30% increase in the baseline ET and an 18%
increase in the T-detrend simulation ET. The summer ET changes of -1.1% and -0.8% are
comparatively small. It is worth noting that the long-term trend in UCRB winter ET is
positive in the T-detrend simulation even given no significant trend in winter
precipitation. The positive trend in winter ET is mainly caused by increased snow
sublimation. Although sublimation is strongly controlled by surface temperature, other
factors also contribute as well (see Section 4).
The remaining -1.6 km3 (-7.7%) decrease in RO in the T-detrend simulation is
curious given the increasing summer precipitation (∆Psummer +1.6 km3, 3.0%) and
negligible winter precipitation change (∆Pwinter -0.1 km3, -0.2%). In addition, although the
SWE anomaly in the T-detrend simulation is less compared with that of the baseline
simulation (baseline: -9.1 km3, T-detrend: -5.6 km3), the long-term 1906-2014 SWE trend
is still negative in the T-detrend simulation (-23.9%). Winter ∆ET in the T-detrend
simulation is only +2.9 km3 as reported in Table 2, which cannot explain all of the SWE
anomaly. One possible answer is that while the overall basin-wide precipitation changes
over time are small, precipitation declines in the most productive basins while increasing
in the less productive basins. We explore the effects of such spatial variations below.
3.2 Sub-basin conditions
Figure 1 shows that there are four sub-basins in the upper CRB (denoted by red
numbers) that produce most of the UCRB runoff: the Yampa River, Colorado River near
Cameo, Gunnison River and San Juan River (from north to south, respectively). The most
productive sub-basin is the Colorado River near Cameo (USGS 09095500) in the
© 2018 American Geophysical Union. All rights reserved.

 northeastern part of the UCRB. This sub-basin produces almost one quarter of the total
naturalized runoff of the UCRB. It contains not only the mainstem, but also several large
tributaries, including the Eagle, the Roaring Fork, and the Blue. A little more than 30% of
the UCRB flow is produced by the other three subbasins, and in total about 55.5% of the
total discharge of the UCRB is attributable to these four tributaries. Below, we discuss the
nature of the long-term changes in these critical sub-basins.
Figure 4 shows annual precipitation, ET and runoff changes for all sub-basins over
the 1916-2014 study period. The top row is extracted from our baseline simulation and
the bottom row is from the T-detrend simulation. We note that although some sub-basins
appear similar between baseline and original maps, the numbers are more different than
they might appear by visual inspection of the maps (Tables S3 and S4). We calculated the
changes relative to the initial value of each linear fit, shown in Table 2. Figure 4a shows a
noteworthy east-west dipole in the precipitation changes over time in the UCRB. In the
UCRB, precipitation decreases have occurred mainly in the high runoff generating
northeastern part of the basin while several sub-basins in the northwestern part of UCRB
show long-term annual precipitation increases.
Precipitation declines have also occurred in the LCRB where little runoff occurs.
These decreases in precipitation led to declines in ET and little change in sub-basin runoff
(Figures 4c and 4f), with negligible impact on total basin runoff (e.g., at Imperial Dam).
There are two sub-basins in the northeastern part of the UCRB, that have relatively
large annual precipitation decreases of -2.3 km3 (Colorado River above Cameo) and -0.7
km3 (Gunnison River) with a combined runoff decrease of -2.9 km3 (SI). These are the
same highly productive sub-basins shown in Figure 1 and are a major driver of the overall
annual runoff decline. Four basins in the northwestern part of UCRB with increasing
precipitation (the Green River downstream portion along with its San Rafael River and

© 2018 American Geophysical Union. All rights reserved.

 Duchesne River tributaries; colored in deeper blues in Figure 4) have partially offset
these long-term runoff declines by about 1.0 km3.
Figures 5 and 6 are similar to Figure 4 but for winter (Oct-Mar) and summer (AprSep), respectively. Winter runoff changes are small for both the baseline and T-detrend
simulations, as most runoff occurs during the summer season. Although the total
precipitation amounts are similar during warm and cold seasons, winter precipitation is
much more important to the UCRB’s runoff. Summer precipitation mainly contributes to
ET rather than runoff, as high summer temperatures lead to large ET, especially at lower
elevations. Winter precipitation in mountain headwater regions accumulates as
snowpack and contributes mostly to RO rather than ET, when it melts.
April 1 SWE trend plots for all the sub-basins (Figures 3a/b) show that the four
highly productive sub-basins (Yampa River, Colorado River near Cameo, Gunnison River,
and San Juan River) in the northeastern part of the basin that contribute much of the
runoff losses in the UCRB have all experienced substantial SWE decreases. Those subbasins are also snow-dominant regions as indicated by Figure 3c. Figure 5a shows that
winter precipitation has declined in all of the northeast UCRB sub-basins except for the
San Juan River, which shows a positive winter precipitation trend. Nonetheless, both SWE
(Figure 3a) and annual RO (Figure 4c) in the San Juan Basin are decreasing. The reason is
that winter ET has increased substantially: ∆Pwinter is +0.4 km3 while long-term ∆ETwinter
is +1.1 km3, with SWE decreasing by -0.7 km3, or 30.1%. Declines in SWE in the other
three basins, all of which experience declines in precipitation, are more severe and range
from -46% to -49%. The increased winter ET, along with reductions in precipitation in
these basins explain the strongly decreasing SWE and substantially explain the declines
in sub-basin runoff.

© 2018 American Geophysical Union. All rights reserved.

 As noted above, 53% (1.8 of 3.4) of the long-term runoff trend in the UCRB is
related to warming temperatures. To dissect the remaining -1.6 km3 (47%) in the Tdetrend simulation, we performed a P- and T-detrend experiment, in which we removed
both the temperature and winter precipitation trend from the original input dataset.
Importantly, under this experiment the northeast UCRB basins see increased winter
precipitation while the northwest basins see decreased winter precipitation relative to
the baseline and T-detrend simulations. Note, also, that we do not modify the summer
precipitation, which increased over the study period. Under the P&T-detrend simulation,
the UCRB’s long-term runoff losses become -0.6 km3 (1.0 km3 less than the pure Tdetrend and 2.8 km3 less than the baseline). The residual -0.6 km3 loss over the 19162014 period is attributable to increased winter ET. Section 4.2 below evaluates why
ETwinter shows a positive trend given no-P-trend and no-T-trend. The total runoff decline
of -3.4 km3 can thus be attributed to warming (-1.8 km3), insufficient P in the northeast
part of CRB (-1.0 km3) and increased winter ET (-0.6 km3).
Summer precipitation and summer ET trend spatial plots (Figure 6a-6d vs 6b-6e)
show similar patterns for both the baseline and T-detrend simulations: negative trends
have occurred over the LCRB and the eastern UCRB while some increases have occurred
in the northwestern headwaters. The spatial patterns confirm that in the summer
increases in precipitation drive increases in ET while decreases in precipitation drive
decreases in ET over both the LCRB and UCRB when surface air temperatures are
relatively high.
In the UCRB the baseline simulation Apr-Sep runoff (Figure 6c), which constitutes
almost three quarters of the CRB annual total, shows spatial patterns similar to the SWE
spatial plots in Figure 3. Taken together, the figures show where water is stored as snow
in the UCRB during winter in the cold, high-elevation headwaters regions and how SWE
© 2018 American Geophysical Union. All rights reserved.

 then contributes to runoff in the following spring and summer. Over the last century,
warming temperatures, reduced winter precipitation in the most productive mountain
sub-basins in the UCRB and slight increases in winter ET (Figure 5b) lead to reduced SWE
and consequently reduced runoff.
In the LCRB, the annual precipitation, ET and runoff plots show mostly P
decreases, ET increases and small RO changes (Figure 4). In winter, some P increases
occur in the NW portion, ET increases everywhere except in the south, and RO has little
change (Figure 5). Summer shows decreasing P, increasing ET and little RO change
(Figure 6).
3.3 Drought comparisons
In order to examine the causes of the Millennium Drought, we compare the recent
dry period from 2000-14 (D2) with the 1953-68 drought (D1). Figure 7 shows the time
series of UCRB annual streamflow volume. Long-term averages are marked as the black
horizontal baseline and the LOWESS-smoothed VIC streamflows are plotted in red. We
report basin-wide (CRB, UCRB and LCRB) annual average anomalies for four selected
variables (P, SWE, ET and RO) over the two drought periods in Table 3. Spatial anomaly
plots by sub-basin of P, SWE, ET and RO for the 1953-1968 and 2000-2014 periods are
shown in Figures 8 and 9.
Similar to the long-term trends discussed in section 3.3, comparison of the annual
anomalies of precipitation, ET and runoff during both droughts in Table 3 confirms that
the UCRB dominates total basin-wide runoff production during drought periods as in the
long term. In the Millennium Drought annual precipitation decreased more in the LCRB,
which substantially reduced ET, but not runoff. This is a very large part of the overall
basin-wide ET loss (-7.9/8.7 km3), but the LCRB ET does not make much difference to

© 2018 American Geophysical Union. All rights reserved.

 streamflow because most Lower Basin precipitation is converted to ET, drought or no
drought. Since our primary interest is on the causes of declining runoff, we again focus
on the UCRB.
Table 3 summarizes climate and hydrological differences and similarities between
the two drought periods. In particular, UCRB RO anomalies for the two drought periods
are quite similar (-2.4 vs -2.6 km3; all the numbers are D1 vs D2 in this paragraph)
whereas the SWE decrease is much greater in the Millennium Drought (-2.7 vs -4.4 km3).
Although the basin-wide annual (negative) precipitation anomaly in 1953-1968 is much
less than the Millennium Drought (-8.8 vs -11.4 km3), this order is reversed in the UCRB
(-6.1 vs -3.2 km3), where most runoff is generated. In the UCRB, the earlier 1953-68
drought has less average annual precipitation than the Millennium Drought (104.6 vs
107.5 km3), especially in winter when precipitation in the UCRB differentially contributes
to runoff production, as discussed in section 3.2. Winter precipitation in the UCRB is 51.5
km3 and 54.5 km3 for 1953-1968 and 2000-2014 respectively; whereas summer
precipitation is nearly identical (53.1 km3 and 53.0 km3; Table S6). Much higher
temperatures (+0.1 vs +1.0°C), less SWE (-2.7 vs -4.4 km3) and more winter ET (+0.4 vs
+1.8 km3) are indicative of additional key differences between the two droughts.
Table 3 combined with Figures 8a-e shows that the 1953-1968 drought mainly
resulted from a spatially widespread and consistent negative precipitation anomaly
across most of the UCRB. Temperatures were within 0.1 °C of the climatological mean.
The corresponding ET and runoff anomalies therefore mostly reflect the precipitation
reductions in each sub-basin. SWE clearly decreases uniformly in almost all parts of the
UCRB, as does runoff. Note that Figures 8a and 8c-e all have similar patterns.
Interpretation of anomalies during the Millennium Drought is more complicated
due to spatially heterogeneous conditions. Pervasive anomalously high temperatures,
© 2018 American Geophysical Union. All rights reserved.

 resulting in part from the long-term warming-trend, which emerged around the 1970s
and exacerbated by drought-specific warming, play a substantial role (Figure 9b). In
addition, D2 average ET in the UCRB (Table 3, Figure 9d), only 0.8 km3 less than the
climatological mean (despite drier conditions), combined with precipitation reductions
in the most highly productive sub-basins (Figure 9a) caused large runoff reductions in
those key basins. In the UCRB, the western sub-basins experienced positive precipitation
anomalies with commensurate increases in ET.

The northeastern sub-basins where

snow dominates and most of the UCRB runoff originates (Figure 3c), experienced
negative precipitation anomalies but without commensurate decreases in ET, which
acted to amplify the SWE reductions. Thus, Figures 9c and 9e show substantial declines
in SWE and RO from these northeastern basins along with smaller declines and even
some increases in SWE and RO from the northwestern basins. Unlike the 1950s drought,
the spatial patterns in Figures 9a and 9c-e are highly complex.
Eight basins – four from the highly productive northeast, and four from the less
productive northwest -- provide additional insights into how spatially heterogenous
precipitation, ET, and SWE combined to produce spatially variable runoff in the
Millennium drought. The four most highly productive sub-basins (marked with red
numbers in Figure 1) contributed more than 83% of the total -2.6 km3/yr RO anomaly in
the Millennium Drought; their contribution was only 34% of the -2.4 km3/yr RO anomaly
during 1953-1968 (numbers of each sub-basin are provided in the Supporting
Information). Four sub-basins on the western side of the UCRB (draining the Uinta and
Central Utah Mountains) had positive annual precipitation anomalies during 2000-2014
(leading to 0.5 km3/yr positive RO anomaly) but that positive anomaly was more than
canceled by other runoff-losing sub-basins (-2.4 km3 for the 4 highly productive
northeastern basins). Compared with 1953-68, precipitation anomalies were much more
© 2018 American Geophysical Union. All rights reserved.

 uneven in the Millennium Drought. The relatively evenly distributed positive +1°C
temperature anomalies lead to more winter ET (+0. 8 km3, 3.7% of the annual
streamflow) and reduced SWE (-4.8 km3, 23.0% of the annual streamflow), exacerbating
the precipitation reductions over the UCRB.
By combining our T-detrend and P&T-detrend simulations we can gain additional
insights into the Millennium Drought when used in a similar fashion to our long-term
trend analysis. Comparing the two simulations suggests that the temperature anomaly
was responsible for -1.4 km3 of the Millenium Drought runoff loss (total is -2.6 km3), while
the precipitation deficit caused -1.0 km3 of the remaining -1.2 km3 runoff loss. The
average runoff in the P&T-detrend results is quite close to the long-term climatology
(P&T-detrend: 18.7 km3, climatology: 18.9 km3, less than 1% difference), suggesting that
the model precipitation and temperature changes are faithfully capturing the drought
causes.
4.0 Interpretation and discussion
4.1 Long-term Trends
The Colorado River is snow-dominated, although only about 18% of the entire
basin area accumulates enough SWE to produce substantial spring and summer RO (see
Apr 1st SWE climatology >50mm as shown in Figure 3c). Basin-wide Apr 1 SWE is
approximately 20 km3, which is close to the annual runoff at Lees Ferry. Li et al. (2017)
show that for the UCRB, SWE accounts for 71% of annual runoff on average. Summer
(Apr-Sep) RO constitutes almost ¾ of the total annual RO in both the UCRB and the entire
basin. Clearly then, winter precipitation (and hence spring SWE) are closely linked to
annual runoff changes. Although the overall winter precipitation trend from 1916-2014
is not significant over the entire UCRB (-0.2%, Table 2), uneven spatial distribution

© 2018 American Geophysical Union. All rights reserved.

 causes important winter precipitation decreases in several of the snow-dominant most
runoff-productive headwater sub-basins. Warming temperatures over our nearly
hundred-year period of record in the UCRB (annual long-term ∆T is 1.8°C as in Figure 2)
induce -1.8 km3 (53%) of the annual runoff losses totaling -3.4 km3. The remaining -1.6
km3 results from negative winter precipitation anomalies, mostly in the northeastern
sub-basins of UCRB (-1.0 km3) and increasing winter ET (-0.6 km3).
4.2 Winter ET and Sublimation
We found that increasing winter ET in both the baseline (4.9 km3) and the Tdetrend (2.9 km3, Table 2) comes mainly from snow sublimation. In the T-detrend
simulation, the November to February long-term change of UCRB sublimation is 2.2 km3
(75.9% of the 2.9 km3 ∆ETwinter increase) with the remaining 0.7 km3 from increased
evaporation in March. A possible cause of these trends in individual months was our
approach using annual rather than seasonal (e.g., monthly) trend removal. Therefore, we
performed another simulation with temperature detrended on a grid cell by grid cell
basis for each month, instead of annually. This resulted in a considerable decrease in the
March ET trend, which apparently was caused primarily by the increasing annual
temperature trend. However, snow sublimation from October to February still showed
increasing trends in this monthly T-detrend simulation. We were therefore left to explain
the positive trends in snow sublimation over Oct-Feb given neither temperature nor
precipitation trends.
We considered other factors that can influence the sublimation process in VIC. We
found that the winter months had positive trends in surface aerodynamic resistance (AR),
which leads to positive trend in surface snow sublimation. The AR trend was traced to
the wind forcings in our VIC input dataset, which are based on NCEP/NCAR reanalysis,
the record for which starts in 1949. Following Livneh et al. (2013), absent wind data

© 2018 American Geophysical Union. All rights reserved.

 prior to 1948, the earlier values were set to their monthly climatological averages.
Although this approach didn’t result in a trend in wind over the 1916-2014 period, the
non-linear relationship between AR and wind speed results in larger AR values occurring
after 1948 and thus results in the long-term increasing sublimation trend. While the
resulting overall RO negative trend associated with this effect was modest (-0.6 km3), we
changed our pre-1949 wind values by randomly sampling from the later (post-1948)
record. This resulted in the long-term UCRB annual RO trend becoming essentially zero
in a new P&T detrend simulation. Livneh et al., [2013] reported that using wind
climatology had only small impacts on their long-term mean RO, but in the case of the
relatively dry CRB, the abrupt change in wind variability created artificial sublimation
that was not negligible.
4.3 Drought Comparisons
Compared to the 1953-68 drought, the causes of the Millennium Drought are more
complicated. During the 1953-1968 drought, annual precipitation anomalies were
negative across the entire CRB (Figure 8a) and temperature was close to its long-term
mean (Figure 8b). Sub-basin runoff anomalies, as well as SWE and ET anomalies, all
responded primarily to the precipitation deficits. In contrast, the upper and lower part of
CRB behaved much differently during the Millennium Drought. In the UCRB, both winter
and summer precipitation during 2000-2014 are just slightly below their climatologies
(54.4 km3/winter compared to 55.8 km3/winter long-term mean) and 53.0 km3/summer
(compared to 55.0 km3/summer long-term mean). The UCRB received approximately
normal (slightly negative anomalies) winter precipitation, which was clearly higher than
Pwinter during 1953-1968 as noted in Section 3.3, but produced less annual runoff (16.3
km3/yr vs 16.5 km3/yr).

© 2018 American Geophysical Union. All rights reserved.

 The situation is reversed, however, if the temperature trend is removed. In this
case the 1953-1968 drought becomes worse than the Millennium Drought. In the Tdetrend simulation, the average annual runoff for the UCRB during 1953-1968 and 20002014 was 17.2 km3/yr and 17.7 km3/yr, respectively (baseline annual runoff climatology
is 18.9 km3). Therefore, the warming temperature accounts for 54% of the annual runoff
anomaly during the Millennium Drought (-1.4 km3/yr of -2.6 km3/yr), which is very close
to its 53% contribution to the long-term decreasing runoff trend. The other half of the
runoff deficit was caused by UCRB’s negative winter precipitation anomalies in the
northeastern part of the basin where the highest runoff-generating sub-basins are. The
winter ∆P over 2000-2014 in those four highly productive sub-basins was -2.4 km3/yr,
much larger (in absolute value) than ∆Pwinter over 1953-1968, -0.9 km3/yr. Exacerbated
by above normal winter temperature in the baseline simulation, the UCRB winter ET
anomaly over 2000-2014 was 1.8 km3/yr and ∆SWE is -4.4 km3/yr (23.7% less compared
to the climatology).
These results demonstrate that warming temperature was a major driver for the
UCRB’s runoff shortage over the Millennium drought, in agreement with Udall and
Overpeck (2017). In the Lower Basin, annual precipitation had very serious negative
anomalies across the entire LCRB as shown in Figure 9a: all sub-basins exhibited
pronounced negative anomalies. While temperatures were also higher across the LCRB,
there is no need to invoke a temperature forcing to explain the drought. As noted above,
though, these LCRB precipitation anomalies have little effect on RO.
Using the Millennium Drought anomalies, we can estimate the runoffprecipitation-elasticity relationships as follows: the baseline average annual runoff for
the UCRB is 18.9 km3 and the T-detrend runoff is 17.7 km3, therefore, the 1.2 km3 runoff
decrease apparently is attributable to precipitation. Over 2000-2014 annual

© 2018 American Geophysical Union. All rights reserved.

 precipitation in the UCRB was 107.5 km3 and the climatology was 110.8 km3/yr, so ∆P/P
is -0.029. The implied elasticity is 2.12 (∆RO • RO-1 • ∆P-1 • P = -0.0616/-0.0291), which is
in good agreement with Vano et al. (2012).
4.4 Uncertainties
The results and analysis we’ve presented to this point are based on VIC
simulations forced by the extended H&L dataset. The robustness of the conclusions is
potentially dependent on both the forcings and model performance. In order to examine
the robustness of our results, we performed an exploratory uncertainty analysis of both
the model forcings and hydrological model.
First, we compared the H&L forcings to two other widely-used gridded climate
datasets: Precipitation Regressions on Independent Slope Method (PRISM; Di Luzio et al.
2008) and Livneh (Livneh et al. 2013). Over the UCRB, trends in annual precipitation of
these three datasets (H&L, PRISM and Livneh) have long-term annual trends ranging
from -6% to +2%, and for winter precipitation from -10% to +6%. As for the temperature,
on an annual basis the positive trend over UCRB range from 1.0˚C to 1.4˚C, and for winter
temperature from 1.0˚C to 1.6˚C. As noted in the Supplement (Table S7), the H&L
temperature trends generally are larger than for the other two datasets (also see section
2.1; the VIC temperature trend is not the same as the H&L trend but rather is somewhat
larger, approximately 0.4˚C, as it results from energy budget closure in the model). The
relatively large negative precipitation trend in Livneh is mostly attributable to large
annual precipitation early in the record, and in likelihood is traceable to the relatively
liberal criterion that data set uses to allow entry of stations with relatively short record
lengths.
Our choice of the H&L data set is based on its relationship with the HCN station
data (to which its decadal variability is controlled; see Hamlet and Lettenmaier, 2005).

© 2018 American Geophysical Union. All rights reserved.

 The HCN data have been carefully quality controlled, and in this sense arguably are more
appropriate for trend-related studies than are the other two data sets (or for that matter,
other data sets we might have chosen). We evaluated the H&L long-term temperature
trend over UCRB (1.4˚C increase) in comparison with the simple average over all HCN
stations in the UCRB (also 1.4˚C increase; identical to two significant figures). On this
basis, and given the criteria used in construction of the H&L data set, we believe it is most
appropriate for our purposes. We do not believe that other methods that, for instance,
might use multiple ensembles and effectively average either inputs to our outputs from
our hydrological model would be appropriate given the objectives of our analysis.
As for hydrological models, we extracted the Noah-MP and VIC results from the
UCLA Drought Monitor (Xiao et al. 2016) for model comparison (note that the forcings
for the UCLA Drought Monitor are different than H&L, but are common to the two
models). Over the entire Upper Basin and the four most productive sub-basins we
identified, the long-term trends in Noah-MP and VIC runoff are generally consistent. For
instance, for the entire UCRB, (VIC: -3.5 km3/yr; Noah-MP: -4.3 km3/yr); see also subbasin trends shown in Figure S2. Although different models would no doubt produce
somewhat different results, the fact that VIC and Noah-MP, which have essentially no
common heritage, produce similar trends gives us some confidence that our results are
reasonable model-independent.
This uncertainty analysis improves the confidence in our conclusions.
Nonetheless, more work could be done along these lines. For example, there is substantial
uncertainty in the gridded forcing data sets we used, which are sparse and especially rare
at high-elevations. More sophisticated methods could be used to represent the
uncertainty in the gridded data sets (aside from testing sensitivity to different data sets,
as we have done). Furthermore, land surface models, which simulate complex systems,

© 2018 American Geophysical Union. All rights reserved.

 contain approximations and uncertainties that produce errors that are difficult to
represent in analyses such as ours. Thus, given computational constraints, less than
complete understanding of physical processes and limited observation resolutions, stateof-the-art land surface models will inevitably produce somewhat uncertain results. We
acknowledge these uncertainties, which no doubt will motivate future work. We
nonetheless argue that our results in the larger sense transcend the effects of these
uncertainties, in particular given their robustness with respect to models and model
forcing data sets.

5.0 2017 Streamflow forecast
The Colorado Basin River Forecast Center (CBRFC) produces seasonal (Apr-Jul)
streamflow forecasts starting about January 1 with monthly updates for the CRB using its
Ensemble Streamflow Prediction (ESP) approach (Werner and Yeager 2013) based on
the Sacramento Soil Moisture Accounting model (Burnash et al. 1973).

General

characteristics of Sacramento and VIC simulations, and hence ESP forecasts, are roughly
similar (Vano et al. 2012). The CBRFC forecast utilizes historical meteorological forcings
for 1981-2010 to generate an ensemble of future streamflow series given hydrological
conditions (soil moisture and SWE) on the forecast initiation date (e.g., April 1), which
are taken from an historical model simulation. We analyzed the forecasts issued on the
first day of each month in 2017 from January through June. The official CRBFC forecast
for the UCRB 2017 Apr-Jul streamflow (natural flow at Lees Ferry) decreased
dramatically from much above normal on January 1 as the runoff season progressed.
Some media reports attributed these decreases to anomalously warm late winter and
spring conditions, and drew parallels between water year 2017 conditions, and the longterm trends analyzed above, especially in temperature.

© 2018 American Geophysical Union. All rights reserved.

 We evaluated the causes of the changes in the 2017 forecasts using the same ESP
approach as used by CRBFC, but using the VIC rather than the Sacramento model. Because
the ESP method requires near real-time records and meteorological forcings, we used the
UCLA/UW Drought Monitor data set (see Xiao et al. 2016) to perform the retrospective
ensemble forecasts.
Figure 10 shows the predicted naturalized streamflow at Lees Ferry for each
forecast initialized on the first day of each month. The red line shows the official forecasts
produced by CRBFC and the blue line is the average of the ensemble predictions
generated using the UCLA/UW drought monitor dataset. The green line shows the
streamflow predictions that would have been made with a perfect precipitation forecast
(they come from a VIC simulation with observed 2017 precipitation) but with
temperature ensembles taken from observations for 1981-2010.
We performed this experiment to separate the effect of precipitation and
temperature on the ESP results. In interpreting the forecasts, it is important to note that
the forecast period is the same (April-July) for all forecasts, even though for post-April 1
forecasts, part of the forecast period has already occurred, and some of the water literally
has already gone “under the bridge”. It is clear that both the red and blue curves exhibit
peaks around February-March with forecasts declining later. The CBRFC forecasts are
higher than those made with VIC, which most likely is attributable to a different
hydrologic model and different model forcing data sets; however, both sets of forecasts
have the same general patterns. Also, both sets of forecasts are still above climatology
for the last forecast (June 1), due to anomalously high SWE early in the forecast period.
From the green line we can infer that the differences between the perfect precipitation
forecasts initialized at each time and climatology are considerably smaller than the
differences between either of the ESP forecast sets and climatology.

© 2018 American Geophysical Union. All rights reserved.

 Given perfect precipitation forecasts, the forecasts vary from 100% to 110% of the
mean, which are close to the true value (observed flow relative to climatology) of 105.3%.
Anomalously warm temperatures in February and March 2017 (plots are not shown
here) caused some error in the forecasts: the streamflow forecasts initialized on February
1st and March 1st are both higher than observed because the climatology is cooler, but the
differences are modest. In general, warm temperatures lead to less runoff and vice versa
but this appears not to be the primary explanation for the rapid decrease in the two ESP
ensemble means through the winter and early spring.
Figure 11 shows the monthly time series plots of precipitation, SWE, runoff and
soil moisture change (P-∆SWE-ET-RO) for the UCRB for both 2017 and climatology from
1981-2010. The precipitation plot (Figure 11a) shows that the UCRB received
anomalously high precipitation in January and February (with the highest anomaly in
January), but the precipitation later in the forecast period was less than climatology. The
direct effect is that in February 2017 there was a large positive SWE anomaly (Fig. 11(b)),
but the anomaly decreased thereafter. This explains why the ESP forecast peak was in
February.
The RO time series plot in Figure 11(c) is more complicated: RO production was
anomalously high in March, April and June, but lower than climatology in May. The
question of interest is, where did the snowpack that accumulated in Jan and Feb go? From
Figure 11(b), about 5 km3 of SWE melted in February and March. However, precipitation
anomalies were in the range (negative) 2-4 km3 for each month from Mar 2017 on. The
fact that 2017 ET during the forecast period was close to climatology (plot not shown)
suggests that enhanced early season snowmelt supplied water to the soil column, but
reduced subsequent precipitation hindered runoff production.

© 2018 American Geophysical Union. All rights reserved.

 Figure 11(d) shows the modeled water balance for the soil column (P-ET-RO∆SWE). Figure 11(d) shows that ∆SMs in Mar and Apr are larger than climatology, but not
by much. Furthermore, runoff generation (Figure 11c) is above climatology during that
period. However, as the precipitation deficit persisted into late spring and summer, SM
began to decrease substantially. The RO actually produced was less than the early
forecasts (initialized in February and March) because the ESP ensemble mean effectively
corresponds to normal precipitation, which is higher than actually occurred from late
winter on in 2017. In summary, the sharp reduction in forecasts through late winter and
spring appears to be primarily related to negative anomalies in late winter precipitation,
with anomalously warm late winter temperatures having a secondary effect.
6.0 Summary and Conclusions
Both long-term (~100 years) trends in streamflow and comparisons of two major
drought periods (1953-68 and Millennium) point to ongoing changes in the relative
control of precipitation and temperature on the river’s runoff. Udall and Overpeck (2017)
have argued that a transition is occurring, which is especially evidenced by the different
responses of the 1953-68 and ongoing Millennium drought to precipitation and
temperature anomalies. We find that, while there is strong evidence for such a transition,
the situation is complicated by spatial variations across the sub-basins that contribute
most to both long-term trends and drought variations in the basin, as well as to seasonal
differences in temperature and precipitation trends and anomalies. Specifically, we
conclude that:
Over the UCRB (which produces about 90% of the entire basin’s runoff), the longterm 1916-2014 decreasing trend of annual runoff is -3.4 km3 (or 16.5% over the
entire record). The increasing trend in annual temperature averaged over the
basin over the same period has been 1.8 °C. When the annual temperature trend
© 2018 American Geophysical Union. All rights reserved.

 is removed, the negative trend in annual runoff becomes -1.6 km3, which suggests
that warming caused a little over half (1.8 km3 or 53%) of the annual runoff trend.
Four snow-dominated, sub-basins in the northeast part of the basin that in
combination account for over half of the UCRB runoff have experienced modest
declines in winter precipitation, which account for a substantial part of the UCRB
runoff trend (-1.0 km3) that is not attributable to warming. The remainder of the
runoff loss (-0.6 km3) is mostly associated with increased winter ET (mainly snow
sublimation).
Compared to the 1953-1968 drought, which was caused by a basin-wide
precipitation deficit, the Millennium Drought reflects a strong influence of warmer
temperatures. The UCRB experienced low streamflow (2.6 km3/yr below average,
slightly more severe than the 2.4 km3/yr negative anomaly for 1953-1968) during
the Millennium Drought years (2000-2014 in our analysis). The four sub-basins in
the northeastern part of the UCRB with the largest negative long-term trends are
also the major contributors to Millennium Drought runoff anomalies. The
decrease of runoff for the Colorado River near Cameo was especially prominent it alone accounts for over half of the 2000-2014 runoff anomalies. Although subbasins with positive runoff anomalies on the south side of Uinta Mountains such
as the Duchesne and San Rafael Rivers counteract some of the deficit, UCRB
Millennium Drought runoff was well below normal due primarily to deficits in the
northeastern sub-basins.
During the Millennium Drought years, the UCRB’s precipitation was close but
slightly below the long-term climatology (annual: 107.5 vs 110.8 km3/yr; winter:
54.5 vs 55.8 km3/yr). However, Millennium Drought annual precipitation was
higher than the average for 1953-1968 (104.6 km3/yr). Winter precipitation

© 2018 American Geophysical Union. All rights reserved.

 during the Millennium Drought was also higher than in the 1953-68 drought; only
summer precipitation was slightly lower. However, the highly productive subbasins in the northeastern portion of the UCRB had comparatively large winter
precipitation deficits during 2000-2014, which resulted in 1.0 km3/yr of the UCRB
streamflow total reductions (2.6 km3/yr) that were not attributable to warming.
Warming temperatures caused 1.4 km3/yr runoff losses.
By reforecasting the 2017 Apr-Jul natural streamflow at Lees Ferry using the same
ESP approach used by CBRFC, we reproduce similar reductions in forecasted
runoff to the CBREFC forecasts through the forecast season in what started as a
large positive forecast anomaly in Apr-Jul runoff forecast on Jan 1. The Apr-Jul
forecast peaked around March 2017 due to abundant SWE in the UCRB induced
by high early winter precipitation. Anomalously high snowmelt increased runoff
in March and April. However, precipitation from March on continued below
normal, and the forecast trended downward in the later months, eventually
ending with only modestly above normal Apr-Jul runoffs. Anomalously warm
temperatures from late winter on in 2017 aggravated the situation but appear not
to be the major cause of the forecast declines, which rather was relatively dry
conditions from mid-winter on.
Given the importance of the Colorado River Basin to the rapidly growing U.S. Southwest,
others likely will address the causes of the both the long-term and recent changes in CRB
runoff, and the future implications of these findings as the 21st century continues to warm
. As we note in section 4.4, our results and conclusions are tightly linked with the forcing
dataset, and the model(s) we used.

The gridded forcings (for precipitation and

temperature, as well as other variables derived from them) propagate through the
hydrologic modeling and in turn our diagnosis of runoff changes. We opted to use the
© 2018 American Geophysical Union. All rights reserved.

 Hamlet and Lettenmaier (2005) model forcing data set because it is closely linked to the
U.S. Hydroclimatic Network (HCN; Easterling et al., 1996), which is based on a set of
stations with relatively complete long-term records that have been corrected for station
moves and instrument changes.

Nonetheless, the stations included in HCN are

predominantly at low elevations, and various avenues (e.g., assimilation of available
surface and/or satellite observations into a coupled land/atmosphere model) could be
pursued to better represent the role of high elevation climatic changes, which may well
not have occurred in concert with changes at lower elevations. We also note in section
4.4 (and explore, via limited experiments with a second model, Noah-MP) the possible
sensitivity of our results to the form of the LSM, but much more could be done in this
respect. Finally, we note that all of our experiments are off-line, hence we partition CRB
runoff changes into those associated with warming temperatures and other factors
(mostly precipitation changes); however, these multivariate changes may well be linked
in ways that we have not explored. For instance, the modest changes in precipitation that
we examined may be coupled with temperature changes, and/or changes in the
atmospheric radiative balance, and such linkages certainly are worth exploring.

Acknowledgements: The research on which this paper was based was supported in part
by the National Oceanic and Atmospheric Administration’s Modeling, Analysis,
Prediction, and Projections Program (MAPP) under award no. NA14OAR4310293 to the
University

of

California,

Los

Angeles.

https://doi.org/10.6084/m9.figshare.6995012.v1;

Model
contact

simulation

archive

muxiao@ucla.edu

is
for

questions about the data.

© 2018 American Geophysical Union. All rights reserved.

 References
Andreadis, K. M., P. Storck, and D. P. Lettenmaier, 2009: Modeling snow accumulation
and ablation processes in forested environments. Water Resour. Res., 45, 1–13,
doi:10.1029/2008WR007042.
Barnett, T. P., and D. W. Pierce, 2009: Sustainable water deliveries from the Colorado
River in a changing climate. Proc. Natl. Acad. Sci. U. S. A., 106, 7334–7338,
doi:10.1073/pnas.0812762106.
Burnash, R. J. C., R. L. Ferral, and R. A. McGuire, 1973: A generalized streamflow
simulation system – Conceptual modelling for digital computers. 204 pp.
Christensen, N. S., and D. P. Lettenmaier, 2007: A multimodel ensemble approach to
assessment of climate change impacts on the hydrology and water resources of the
Colorado River Basin. Hydrol. Earth Syst. Sci. Discuss., 11, 1417–1434.
Christensen, N. S., A. W. Wood, N. Voisin, D. P. Lettenmaier, and R. N. Palmer, 2004: The
effects of climate change on the hydrology and water resources of the Colorado
River basin. Clim. Change, 62, 337–363,
doi:10.1023/B:CLIM.0000013684.13621.1f.
Cohen, M., J. Christian-Smith, and J. Berggren, 2013: Water to Supply the Land: Irrigated
Agriculture in the Colorado River Basin.
http://water.usgs.gov/watercensus/colorado.html.
Dawadi, S., and S. Ahmad, 2012: Changing climatic conditions in the Colorado River
Basin: Implications for water resources management. J. Hydrol., 430–431, 127–
141, doi:10.1016/j.jhydrol.2012.02.010.
http://dx.doi.org/10.1016/j.jhydrol.2012.02.010.
Easterling, D. R., T. R. Karl, E. H. Mason, P. Y. Hughes, and D. P. Bowman, 1996: United
States Historical Climatology Network (US HCN), Monthly Temperature and
Precipitation Data. 280 pp. http://cdiac.esd.ornl.gov/ndps/ushcn/abstract_r3.html.
Hamlet, A. F., and D. P. Lettenmaier, 2005: Production of Temporally Consistent Gridded
Precipitation and Temperature Fields for the Continental United States. J.
Hydrometeorol., 6, 330–336, doi:10.1175/JHM420.1.
Li, D., M. L. Wrzesien, M. Durand, J. Adam, and D. P. Lettenmaier, 2017: How much runoff
originates as snow in the western United States, and how will that change in the
future? Geophys. Res. Lett., 44, 6163–6172, doi:10.1002/2017GL073551.
http://doi.wiley.com/10.1002/2017GL073551.
Liang, X., and D. Lettenmaier, 1994: a simple hydrologically based model of land surface
water and energy fluxes for general circulation models. J. Geophys. …, 99.
http://onlinelibrary.wiley.com/doi/10.1029/94JD00483/full (Accessed April 16,
2014).
Livneh, B., E. A. Rosenberg, C. Lin, B. Nijssen, V. Mishra, K. M. Andreadis, E. P. Maurer,
and D. P. Lettenmaier, 2013: A long-term hydrologically based dataset of land
surface fluxes and states for the conterminous United States: Update and
extensions. J. Clim., 26, 9384–9392, doi:10.1175/JCLI-D-12-00508.1.
http://journals.ametsoc.org/doi/abs/10.1175/JCLI-D-12-00508.1 (Accessed
October 22, 2014).
Di Luzio, M., G. L. Johnson, C. Daly, J. K. Eischeid, and J. G. Arnold, 2008: Constructing
retrospective gridded daily precipitation and temperature datasets for the
conterminous United States. J. Appl. Meteorol. Climatol., 47, 475–497,
doi:10.1175/2007JAMC1356.1.
McCabe, G. J., and D. M. Wolock, 2007: Warming may create substantial water supply
shortages in the Colorado River basin. Geophys. Res. Lett., 34, 1–5,
© 2018 American Geophysical Union. All rights reserved.

 doi:10.1029/2007GL031764.
——, ——, G. T. Pederson, C. A. Woodhouse, and S. McAfee, 2017: Evidence that recent
warming is reducing upper Colorado river flows. Earth Interact., 21,
doi:10.1175/EI-D-17-0007.1.
Mote, P. W., A. F. Hamlet, M. P. Clark, and D. P. Lettenmaier, 2005: Declining Mountain
Snowpack in Western North America*. Bull. Am. Meteorol. Soc., 86, 39–49,
doi:10.1175/BAMS-86-1-39.
——, S. Li, D. P. Lettenmaier, M. Xiao, and R. Engel, 2018: Dramatic declines in snowpack
in the western US. npj Clim. Atmos. Sci., 1, 2, doi:10.1038/s41612-018-0012-1.
http://www.nature.com/articles/s41612-018-0012-1.
Prairie, J., and R. Callejo, 2005: Natural Flow and Salt Computation Methods, Calendar
Years 1971-1995. Bur. Reclam., 1–112.
http://digitalcommons.usu.edu/govdocs/135/%5Cnpapers2://publication/uuid/
6183E843-86FF-4150-9535-C0407CB5477D.
Rajagopalan, B., K. Nowak, J. Prairie, M. Hoerling, B. Harding, J. Barsugli, A. Ray, and B.
Udall, 2009: Water supply risk on the Colorado River: Can management mitigate?
Water Resour. Res., 45, 1–7, doi:10.1029/2008WR007652.
Reynolds, L. V., P. B. Shafroth, and N. LeRoy Poff, 2015: Modeled intermittency risk for
small streams in the Upper Colorado River Basin under climate change. J. Hydrol.,
523, 768–780, doi:10.1016/j.jhydrol.2015.02.025.
http://dx.doi.org/10.1016/j.jhydrol.2015.02.025.
Sen, P. K., 1968: Estimates of the Regression Coefficient Based on Kendall ’ s Tau Pranab
Kumar Sen. J. Am. Stat. Assoc., 63, 1379–1389.
http://www.jstor.org/stable/10.2307/2285891.
Theil, H., 1950: A rank-invariant method of linear and polynomial regression analysis I.
Ned. Akad. Wetenchappen, 53, 386–392, doi:10.1007/978-94-011-2546-8.
http://link.springer.com/chapter/10.1007/978-94-011-2546-8_20.
U.S. Department of Interior, Bureau of Reclamation, 1983: DRAFT Colorado River
Simulation System Hydrology Data Base.
Udall, B., and J. Overpeck, 2017: The twenty-first century Colorado River hot drought
and implications for the future. Water Resour. Res., 1–15,
doi:10.1002/2016WR019638. http://doi.wiley.com/10.1002/2016WR019638.
Vano, J. A., T. Das, and D. P. Lettenmaier, 2012: Hydrologic Sensitivities of Colorado
River Runoff to Changes in Precipitation and Temperature*. J. Hydrometeorol., 13,
932–949, doi:10.1175/JHM-D-11-069.1.
——, and Coauthors, 2014: Understanding Uncertainties in Future Colorado River
Streamflow. Bull. Am. Meteorol. Soc., 95, 59–78, doi:10.1175/BAMS-D-12-00228.1.
http://journals.ametsoc.org/doi/abs/10.1175/BAMS-D-12-00228.1.
Werner, K., and K. Yeager, 2013: Challenges in Forecasting the 2011 Runoff Season in
the Colorado Basin. J. Hydrometeorol., 14, 1364–1371, doi:10.1175/JHM-D-12055.1. http://journals.ametsoc.org/doi/abs/10.1175/JHM-D-12-055.1.
Woodhouse, C. A., G. T. Pederson, K. Morino, S. A. Mcafee, and G. J. Mccabe, 2016:
Increasing influence of air temperature on upper Colorado River streamflow. 1–8,
doi:10.1002/2015GL067613.Received.
http://onlinelibrary.wiley.com/doi/10.1002/2015GL067613/full.
Wu, H., J. S. Kimball, H. Li, M. Huang, L. R. Leung, and R. F. Adler, 2012: A new global river
network database for macroscale hydrologic modeling. Water Resour. Res., 48, n/an/a, doi:10.1029/2012WR012313.
http://doi.wiley.com/10.1029/2012WR012313 (Accessed October 17, 2014).

© 2018 American Geophysical Union. All rights reserved.

 Xiao, M., B. Nijssen, and D. P. Lettenmaier, 2016: Drought in the Pacific Northwest, 19202013. J. Hydrometeorol., JHM-D-15-0142.1, doi:10.1175/JHM-D-15-0142.1.
http://journals.ametsoc.org/doi/10.1175/JHM-D-15-0142.1.

© 2018 American Geophysical Union. All rights reserved.

 Table 1: Naturalized (NFL) and VIC runoff contribution percentages for selected USGS
gauges. Values are computed relative to the annual streamflow climatology at the
Imperial Dam, AZ-CA. The percentages are relative to long-term averages for water year
1971-2014.
Station name

NFL

VIC

COLORADO RIVER NEAR CAMEO (09095500)

22.8%

21.8%

GUNNISON RIVER NEAR GRAND JUNCTION
(09152500)

14.6%

10.3%

SAN JUAN RIVER NEAR BLUFF (09379500)

12.4%

10.9%

GREEN RIVER NEAR GREENDALE
(09234500)

12.2%

10.7%

WHITE RIVER NEAR WATSON (09306500)

3.5%

3.5%

DUCHESNE RIVER NEAR RANDLETT
(09302000)

4.8%

6.5%

YAMPA RIVER AT DEERLODGE PARK
(09260050)

8.0%

9.2%

COLORADO RIVER AT LEES FERRY LEEFY
(09380000)

91.8%

91.0%

COLORADO RIVER ABOVE IMPERIAL DAM
(09429490)

100%

100%

© 2018 American Geophysical Union. All rights reserved.

 Table 2: UCRB annual and seasonal changes in water balance variables over water-years
1916-2014 in km3/yr (km3 for SWE) and percentages relative to the starting value of the
fit. P is precipitation, T is temperature in Celsius, ET is evapotranspiration, RO is total
runoff and SWE is Apr 1 snow water equivalent. Dashed “D” denotes results from Tdetrend simulation. Winter period is Oct-Mar and summer period is Apr-Sep.
P

T

ET

ET-D

RO

RO-D

SWE

SWE-D

Annual

1.5(1.4%)

1.8

4.2(4.7%)

2.3(2.6%)

-3.4(-16.5%)

-1.6(-7.7%)

-9.1(-39.0%)

-5.6(-23.9%)

Winter

-0.1(-0.2%)

1.9

4.9(30.5%)

2.9(18.0%)

0.4(10.4%)

0.4(9.0%)

Na

Na

Summer

1.6(3.0%)

1.7

-0.8(-1.1%)

-0.6(-0.8%)

-3.8(-23.3%)

-1.9(-11.9%)

Na

Na

© 2018 American Geophysical Union. All rights reserved.

 Table 3: Annual average anomalies during the mid-century drought D1 (1953-1968) and Millennium Drought D2 (2000-2014) for CRB,
UCRB and LCRB. Long-term climatologies are also provided. Results are relative to the 1916-2014 baseline simulation (Table 2); units are
km3 (except temperature is Celsius). The climatologies are extracted from the baseline simulation. (Table S6 includes the summer and
winter anomalies for UCRB).

P
anomaly

PClimatol
ogy

SWE
anomaly

SWEClimato
logy

CRBD1

-8.8

163.8

-2.9

19.3

CRBD2

-11.4

UCRBD1

-6.1

UCRBD2

-3.2

LCRBD1
LCRBD2

-2.7
-8.2

-4.8
110.8

-2.7

-0.2
-0.4

-6.0

ETClimatolo
gy

RO
anomaly

ROClimatolo
gy

T
anomaly

TClimat
ology

143.0

-2.7

20.7

0.0

8.5

-8.7
18.7

-4.4
53.0

ET
anomaly

-3.7

-2.8
91.9

-0.8
0.6

-2.3
-7.9

-2.4

1.0
18.9

-2.6
51.1

-0.3
-0.2

0.1

5.6

1.0
1.8

-0.2

13.0

1.0

© 2018 American Geophysical Union. All rights reserved.

 Figure 1: Percent of total CRB runoff (at Imperial Dam) originating from 20 sub-basins,
calculated based on long-term average from VIC simulation for water years 1971-2014.
The sub-basins shown in Figure 1 were extracted from a published dataset by Wu et al.
(2012).

© 2018 American Geophysical Union. All rights reserved.

 Figure 2: Annual time series and linear regression trend plots for Colorado River Basin
above Lees Ferry; (a) annual (naturalized) runoff, (b) annual precipitation and (c) annual
average surface temperature calculated by VIC. Changes are calculated relative to the
starting value of the fit. Note that precipitation (b) is from an extended version of the
Hamlet and Lettenmaier (2005) data set at 1/16th degree spatial resolution while
temperature (c) is calculated from VIC and is approximately 0.4˚C warmer than the
Hamlet and Lettenmaier input temperature.

© 2018 American Geophysical Union. All rights reserved.

 Figure 3: Spatial plots of Apr 1 SWE trends for (a) baseline simulation; (b) T-detrend
simulation over each sub-basin. The changes over 1916-2014 are calculated relative to
starting value of the linear regressions; (c) Long-term average Apr 1st SWE.

© 2018 American Geophysical Union. All rights reserved.

 Figure 4: Spatial changes of (a) annual precipitation from gridded observations; (b) ET;
and (c) runoff from baseline VIC simulation over 1916-2014 for CRB above Imperial Dam.
Changes are calculated relative to the starting value of linear fits. Panels (d)- (f) are the
same as (a)-(c) but variables are extracted from the T-detrend simulation. Panels (a) and
(d) are identical.

© 2018 American Geophysical Union. All rights reserved.

 Figure 5: Same as Figure 4 but for winter (Oct-Mar).

© 2018 American Geophysical Union. All rights reserved.

 Figure 6: Same as Figure 4 but for summer (Apr-Sep).

© 2018 American Geophysical Union. All rights reserved.

 Figure 7: Time series of VIC simulations of annual runoff (top), winter runoff (middle)
and summer runoff (bottom) at Lees Ferry (UCRB). The black horizontal lines are the
long-term means, and red lines result from LOWESS filtering of VIC results.

© 2018 American Geophysical Union. All rights reserved.

 Figure 8: Average annual anomaly plots for each sub-basin during the drought period
1953-1968. The variables in each panel are: (a) precipitation; (b) temperature; (c) SWE;
(d) ET; and (e) runoff (panels (c)-(e) are from VIC simulations.

© 2018 American Geophysical Union. All rights reserved.

 Figure 9: Same as Figure 8 but for 2000-2014 Millennium Drought.

© 2018 American Geophysical Union. All rights reserved.

 Figure 10: Apr-Jul 2017 streamflow forecasts at Lees Ferry initialized on the first day of
each month expressed as percentages relative to 1981-2010 climatology. Red line
represents the official forecasts published by CBRFC; blue line represents equivalent VIC
reforecasts; green line is forecast with perfect precipitation forecast and temperature
climatology. The horizontal dashed line is from a forecast with perfect precipitation and
temperature.

© 2018 American Geophysical Union. All rights reserved.

 Figure 11: Monthly time series plots of (a) precipitation, (b) SWE, (c) runoff, and (d) soil
moisture change. The blue line is the 2017 forecast; red line is historical climatology.

© 2018 American Geophysical Union. All rights reserved.