Quantifying uncertainty sources in an ensemble of hydrological
climate-impact projections
T. Bosshard,
1
M. Carambia,
2
K. Goergen,
3
S. Kotlarski,
1
P. Krahe,
2
M. Zappa,
4
and C. Sch
€
ar
1
Received 19 October 2011; revised 9 December 2012; accepted 10 December 2012.
[1] The quantification of uncertainties in projections of climate impacts on river streamflow
is highly important for climate adaptation purposes. In this study, we present a methodology
to separate uncertainties arising from the climate model (CM), the statistical postprocessing
(PP) scheme, and the hydrological model (HM). We analyzed ensemble projections of
hydrological changes in the Alpine Rhine (Eastern Switzerland) for the near-term and far-
term scenario periods 2024–2050 and 2073–2099 with respect to 1964–1990. For the latter
scenario period, the model ensemble projects a decrease of daily mean runoff in summer
(32.2%, range [45.5% to 8.1%]) and an increase in winter (þ41.8%, range [þ4.8% to
þ81.7%]). We applied an analysis of variance model combined with a subsampling
procedure to assess the importance of different uncertainty sources. The CMs generally are
the dominant source in summer and autumn, whereas, in winter and spring, the uncertainties
due to the HMs and the statistical PP gain importance and even partly dominate. In addition,
results show that the individual uncertainties from the three components are not additive.
Rather, the associated interactions among the CM, the statistical PP scheme, and the HM
account for about 5%–40% of the total ensemble uncertainty. The results indicate, in
distinction to some previous studies, that none of the investigated uncertainty sources are
negligible, and some of the uncertainty is not attributable to individual modeling chain
components but rather depends upon interactions.
Citation: Bosshard, T., M. Carambia, K. Goergen, S. Kotlarski, P. Krahe, M. Zappa, and C. Sch
€
ar (2013), Quantifying uncertainty
sources in an ensemble of hydrological climate-impact projections, Water Resour. Res., 49, doi:10.1029/2011WR011533.
1. Introduction
[2] The projected impacts of climate change on river
streamflow are associated with large uncertainties. For water
management, these projection uncertainties importantly con-
tribute to the total uncertainty, in addition to factors such as
natural variability and changes in water demand [Kundze-
wicz et al., 2008]. Although the use of climate-impact pro-
jections for water management planning purposes has been
debated [Kundzewicz and Stakhiv, 2010], most published
strategies actually make use of such information [Dessai
and Hulme,2007;Milly et al., 2008]. The quantification of
uncertainties in climate-impact projections is therefore of
particular interest [Pappenberger and Beven, 2006]. So
far, the overall uncertainty of projected climate impacts is
probably underestimated, which is due to an incomplete
sampling of the uncertainty sources [Knutti, 2008; Wilby,
2010]. Improving the knowledge about the importance of dif-
ferent uncertainty sources might thus help to design climate-
impact studies with a more complete uncertainty assessment.
[
3] Impact modeling systems that include a cascade of
different models are commonly used to assess climate
impacts (for some recent studies, see, e.g., Vicuna et al.
[2010], Campbell et al. [2011], Quintana-Segu
ıetal.
[2011], Köplin et al. [2012], and Teutschbein and Seibert
[2012]) and to provide information for water management
[e.g., Schaefli et al., 2007; Lopez et al., 2009]. Elements of
this cascade are an emission scenario, a global circulation
model (GCM), a dynamical downscaling step by means of
a regional climate model (RCM), a statistical postprocess-
ing (PP), and a hydrological model (HM). Alternatively,
the dynamical downscaling and the PP steps can be
replaced by a statistical downscaling. In the remainder of
this article, we call this cascade of emission scenarios and
models an impact modeling chain. Uncertainties in hydro-
logical climate-impact projections arise due to different
assumptions and model combinations in the whole impact
modeling chain (e.g., CM or HM structure uncertainty
[Masson and Knutti, 2011; Seiller et al., 2012]), CM or HM
parameter uncertainty [Bellprat et al., 2012; Beven, 2006],
instationarity of PP parameters [Buser et al., 2009; Boberg
and Christensen, 2012], and natural variability [see, e.g.,
Lucas-Picher et al., 2008]. For a complete analysis of
uncertainty in runoff projections, it is therefore important to
investigate the contributions of all existing sources.
1
Institute for Atmospheric and Climate Science, ETH Zurich, Z
€
urich,
Switzerland.
2
German Federal Institute of Hydrology, Koblenz, Germany.
3
D
epartement Environnement et Agro-Biotechnologies (EVA), Centre
de Recherche Publi - Gabriel Lippmann, Belvaux, Luxembourg.
4
Swiss Federal Institute for Forest, Snow and Landscape Research
WSL, Birmensdorf, Switzerland.
Corresponding author: T. Bosshard, Hydrological Research Unit, Swed-
ish Meteorological and Hydrological Institute, Folkborgsv
€
agen 1, SE-
60176 Norrköping, Sweden. (thomas.bosshard@smhi.se)
©2012. American Geophysical Union. All Rights Reserved.
0043-1397/13/2011WR011533
1
WATER RESOURCES RESEARCH, VOL. 49, 1–14, doi:10.1029/2011WR011533, 2013
[4] Numerous previous studies have investigated hydro-
logical climate-impact projections and their sensitivity to
different uncertainty sources. Here, a nonexhaustive sum-
mary of studies conducted in the Rhine basin follows.
Shabalova et al. [2003] compared two PP methods and
found that for the end of the 21st century, both methods
agree on a decrease of summer runoff and an increase of
winter runoff, but the two methods lead to different
increases of winter flood risk. Lenderink et al. [2007] also
investigated uncertainties due to PP and confirmed the
results of Shabalova et al. [2003] that the choice of the PP
rather affects the changes in runoff extremes than in the
mean runoff. Jasper et al. [2004] investigated the projected
impact on runoff of an ensemble of 17 climate scenarios
derived from seven GCMs and four emission scenarios in
two Swiss catchments. They found that changes in the sea-
sonality of runoff are robust, but the magnitude of the
changes is strongly affected by the choice of the climate
scenario. Combined uncertainties from emission scenarios,
GCMs and RCMs, were investigated by Graham et al.
[2007] who found that the choice of the GCM has a larger
impact on projected hydrological changes than the choice
of the RCM or emission scenario.
[
5] More recently, studies that systematically investigate
multiple uncertainty sources along the whole impact mod-
eling chain have been published. Wilby and Harris [2006]
assessed uncertainties from emission scenarios, GCMs, sta-
tistical downscaling, HM structure, and HM parameters.
Using a probabilistic framework, they showed that GCMs
and the downscaling step were the most important sources
of uncertainty in simulating changes of low flows in the
Thames River (UK). In a study of the impact of climate
change on hydropower production in the Mauvoisin catch-
ment (Switzerland), Schaefli et al. [2007] investigated the
importance of uncertainties due to the global mean temper-
ature projection, the regional scaling relationship, the
glacier model, HM parameters, and the hydropower man-
agement model. They found the uncertainty in the global
mean temperature and the regional scaling factors to be of
comparable magnitude and both being more important than
the other uncertainty sources. Prudhomme and Davies
[2009] assessed uncertainties due to emission scenarios,
GCMs, downscaling methods, and HMs in four mesoscale
British catchments and concluded that the driving GCM is
the dominant source of uncerta inty. They further stated that
uncertainty due to PP and the choice of the emission sce-
nario are of comparable magnitude, whereas uncertainty
due to the HMs is negligible in two out of four basins. Kay
et al. [2009] investigated the same uncertainty sources as
Wilby and Harris [2006] and also included the effect of in-
ternal variability in a case study that assessed changes of
flood frequency in two British catchments. They found
GCMs being the dominant source of uncertainty. However,
after excluding one outlier from the GCM ensemble, other
uncertainty sources such as RCMs and internal variability
became more important than GCMs. For two catchments in
Oregon (USA), Jung et al. [2011] found the natural vari-
ability and the driving GCM to be the major sources for
uncertainty with respect to flood frequency changes.
[
6] In our study, we perform an ensemble of hydrologi-
cal climate-impact projections for an Alpine river catch-
ment and the two scenario periods (SCE) 2024–2050 and
2073–2099 with respect to the control period (CTL) 1964–
1990. From this ensemble, we aim to infer several sources
of uncertainty, and we use the variance as a measure for the
uncertainty. The three uncertainty sources considered are
(i) climate models (CMs) consisting of a GCM and an
RCM, (ii) PP, and (iii) HMs. The Alpine study area is a
challenging region for all three impact modeling chain ele-
ments. CMs, both GCMs and RCMs, for instance, cannot
fully resolve the complex topography ; PP methods have to
correct for potentially large biases; and HMs are chal-
lenged by complicated and spatially highly variable hydro-
logical processes such as accumulation and melt of snow.
[
7] The limited number of uncertainty sources and mod-
els included in the ensemble results in an underestimation
of the overall uncertainty associated with the hydrological
climate impacts (i.e., the uncertainty if all possible uncer-
tainty sources were fully sampled). Throughout this paper,
we call the spread in our ensemble the total ensemble
uncertainty to clearly distinguish it from the overall (true)
uncertainty.
[
8] Our two research questions are as follows: (1) how
large is the total ensemble uncertainty in the runoff projec-
tions and (2) how do different uncertainty sources contrib-
ute to the total ensemble uncertainty in the runoff
projections for both SCEs. In particular, we are interested
in how the contributions vary throughout the annual cycle
and how they affect the uncertainty in changes of different
runoff quantiles. We include fewer sources of uncertainty
than some of the previously mentioned studies, but, instead
of performing single-propagation runs, we conduct a multi-
propagation study ; that is, we vary the different CMs, PP
methods, and HMs in all possible combinations. This
approach allows for an assessment of interactions between
the uncertainty sources [Kay et al., 2009 ; Finger et al.,
2012]. We quantify the contributions of the different uncer-
tainty sources using the decomposition of the sum of squares
as described within the analysis of variance (ANOVA)
theory (see D
equ
eetal. [2007] or Yip et al. [2011] for a
detailed description). Here, we refer to this method as the
ANOVA approach. Following the idea of a multipropagation
experiment, the ANOVA approach allows to consider inter-
actions between the uncertainty sources. These interactions
represent uncertainty contributions that do not behave line-
arly. For instance, a snowmelt bias of an HM may depend
upon the temperature projection of the driving CM that could
lead to a nonlinear response in river runoff. The ANOVA is
complemented with a subsampling scheme to account for
the different sample sizes of the three uncertainty sources.
[
9] This paper is structured as follows. Section 2 briefly
describes the study area and the data. Section 3 introduces
the employed HMs, explains the PP methods, and presents
the subsampling procedure in combination with the
ANOVA. In section 4, we present the results of the hydro-
logical climate-impact projections and the ANOVA.
Section 5 summarizes our study and its main findings.
2. Study Region and Data
2.1. Study Region
[
10] The study region consists of the Alpine Rhine catch-
ment down to the gauge Diepoldsau in Eastern Switzerland
(see Figure 1). It encompasses an area of 6119 km
2
and has
BOSSHARD ET AL. : UNCERTAINTY SOURCES IN CLIMATE-IMPACT PROJECTIONS
2
a mean elevation of about 1800 m above sea level. The run-
off regime is nival, i.e., snow-dominated, but altered to some
extent by hydropower production. The hydropower main
effect is a seasonal redistribution of water from summer to
winter [Verbunt et al., 2005]. The buildup of the hydropower
capacity in the period 1945–2009 is depicted in Figure 2
(solid black line). At the end of the period, the storage
capacity amounts to about 10% of the annual runoff volume.
Figure 1. Map showing the catchment of the Alpine Rhine river down to Diepoldsau. The subbasin
structures used in (left) HBV and (right) PREVAH are shown.
Figure 2. Validation of the two HMs HBV and PREVAH at the gauge Diepoldsa u. Lines show the
NSE (solid lines), 99% (dash-dotted lines), and 5% (dashed lines) runoff quantile levels (left-hand scale).
The values are calculated for moving 3 year periods. The thick lines indicate the performance of the
CTL run from 1961 to 1990 with the first 3 years being cutoff. The thin lines show the performance of
the same model configurations in a longer simulation covering the period 1954–2006 (HBV) and 1951–
2009 (PREVAH). The shaded areas depict the calibration period for each HM. The buildup of the total
hydropower reservoir vol ume is shown as a solid thick black line (right-hand scale).
BOSSHARD ET AL. : UNCERTAINTY SOURCES IN CLIMATE-IMPACT PROJECTIONS
3
2.2. Observational and CM Data
[
11] Throughout this study, we use the CT L and the two
SCEs, all three periods with an additional preceding 3 year
spin-up period, as temporal subsets of the data series.
[
12] The two HMs HBV and PREVAH (see section 3.1)
require different kinds of input data with respect to the spa-
tial resolution and the variables (see Table 1). HBV uses
subbasin-averaged daily time series of precipitation, tem-
perature, and global radiation or sunshine duration. Subba-
sin-averaged observational data have been provided by the
International Commission for the Hydrology of the Rhine
Basin (CHR ; referred to as OBS
CHR
in the remainder of
this article) [Görgen et al. 2010]. PREVAH uses daily sta-
tion data of the hydrometeorological variables precipitation
(80 stations), temperature (36 stations), relative humidity
(41 stations), sunshine duration (28 stations) , and wind
speed (43 stations) from the measurement network of
MeteoSwiss (referred to as OBS
ST
in the remainder of this
article). Both observational data sets cover the whole CTL.
[
13] For the calibration of the HMs, we used daily runoff
data of the gauges depicted in Figure 1. The data were pro-
vided by the Swiss Federal Office for the Environment (see
www.hydrodaten.admin.ch).
[
14] As climate data, we used eight transient climate
modeling chains of the ENSEMBLES project [van der Lin-
den and Mitchell, 2009], as shown in Figure 3. D
equ
eetal.
[2012] showed that, within the ENSEMBLES GCM-
RCMs, the GCM is the largest contributor to the variance
in the projections of seasonal mean temperature and precip-
itation, except for summer precipitation for which the
RCMs are the largest source of variance. Thus, it is impor-
tant to sample both the GCMs and the RCMs in an impact
modeling chain. The ensemble used in this study encom-
passes three GCMs and seven RCMs, which allows to
partly sample CM uncertainty.
[
15] All modeling chains are driven by the A1B emission
scenario [Nakicenovic and Swart, 2000]. This emission sce-
nario belongs to the A1 emission scenario family that
assumes rapid economic growth, population growth until
mid-century, a homogenization of the global wealth across
the different regions, and a rapid introduction of new
efficient technologies. The B stands for a balanced (fossil
and nonfossil) use of energy sources. The corresponding
greenhouse gas emissi ons increase until 2060 and slightly
decrease afterward, resulting in an atmospheric CO
2
con-
centration of about 700 ppm by 2100. The spatial resolution
of the RCMs is about 25 km. We used subbasin-averaged
CM time series (for HBV) and CM data interpolated to
station locations (for PREVAH).
3. Methods
[16] Figure 3 depicts the modeling chain combination
scheme employed in this study. In the following, we
describe the HMs and the PP methods and explain the
ANOVA approach in combination with a subsampling
scheme.
3.1. Hydrological Models
[
17] In our study, we use HMs that have already been set
up in the catchment. Both the HBV and the PREVAH
model are semidistributed conceptual rainfall-runo ff mod-
els. Both models use the hydrological response unit (HRU)
approach to cluster the spatial units according to their
hydrological characteristics. Table 1 summarizes the char-
acteristics of the two HMs in the Alpine Rhine catchment,
Table 1. List of Key Characteristics of the Two Employed HMs
HBV PREVAH
Model type Conceptual, semidistributed Conceptual, semidistributed
Meteorological input data Precipitation (P), temperature (T), and
global radiation (R) or sunshine duration (S)
Precipitation (P), temperature (T), sunshine
duration (S), cloud cover (C), relative
humidity (H), wind speed (V)
Number of subbasins 2 6
Spatial resolution of the underlying
digital elevation model
1000 m 1000 m 500 m 500 m
Internal time step 1 day 1 h
Land use classes 4 29
HRU definition Elevation, land use Elevation, land use, aspect, soil type
Snow/glacier melt modeling approach Degree-day factor Degree-day factor with aspect and slope correction
Evapotranspiration parameterization Penman-Wendling Penman-Monteith
Calibration period 1970–1984 1985–1990
References Eberle et al. [2005], Lindström et al.
[1997], Görgen et al. [2010]
Gurtz et al. [1999], Zappa and Gurtz [2003],
Viviroli et al. [2009]
Figure 3. Modeling chain combination scheme. The three
analyzed modeling chain elements are depicted from left to
right. The naming of the CM chains provided by the
ENSEMBLES project is based on the following pattern:
institution that did run the RCM, employed GCM and
RCM [van der Linden and Mitchell, 2009].
BOSSHARD ET AL. : UNCERTAINTY SOURCES IN CLIMATE-IMPACT PROJECTIONS
4
and Figure 1 shows the subbasin structures of the two
HMs. For the HBV model, we use the HBV134 setup of
the German Federal Institute of Hydrology. The HBV134
requires subbasin-averaged hydrometeorological time
series as input data and applies a lapse rate correction to
disaggregate basin-averaged temperature to zones of dif-
ferent elevations. For the PREVAH model, we use the
setup by Verbunt et al. [2006], which we have recalibrated
in the period 1985–1990 to use sunshine duration instead
of global radiation as an input variable from which short-
wave radiation is derived according to Schulla [1997].
PREVAH requires hydrometeorological station data as an
input. The station data are interpolated using inverse
distance weighting. For temperature, a height-dependent
regression is applied (detrended inverse distance weighting).
[
18] Both HMs correct the observed precipitation to
account for undercatch and interpolation errors [Sevruk and
Nevenic, 1998]. In the HBV model, a linear precipitation
correction factor is applied to precipitation, and no distinc-
tion is made between rain and snow. In the PREVAH model,
two correction factors for rain and snow are used. The pre-
cipitation correction factors are estimated by calibration
using simulations driven by observed data (OBS
CHR
or
OBS
ST
). The water-balance-corrected precipitation is then
used as a reference in the PP. This provides a consistent way
to process the CM data without artificially modifying the
precipitation change signal by a nonlinear precipitation cor-
rection (i.e., different corrections for snow and rain). Con-
sistent with the above, all figures in the remainder of this
article show water-balance-corrected precipitation.
[
19] For further details about the HMs, we refer to the
references listed in Table 1.
3.2. Statistical PP
[
20] We use a bias-correction (BC) and delta-change
(DC) approach for the PP step in the impact mo deling
chain. The parameters of the PP methods were estimated
based on the full 30 year periods 1961–1990 (CTL), 2021–
2050 (SCE1), and 2070–2099 (SCE2), including a 3 year
spin-up period.
[
21] The two PP methods differ distinctively regarding
the treatment of changes in the variability. In the BC, the
time series used to drive the HM are based on CM data,
which are corrected toward the climatological mean of the
observations in the CTL. Thus, the variability in the CM
data determines the variability in the bias-corrected forcing
data, e.g., the succession of wet and dry days. The DC
approach scales observed time series according to a climate
change signal estimated from CM data. The scaled obser-
vational time series are used to force the HMs in the SCE.
It is therefore the variability in the observed time series that
determines the variability in the HM’s forcing data. Also,
the two PP methods differ in the number of variables they
include. The BC approach corrects all required variables
from the output of the CMs, whereas the DC method is
applied to temperature and precipitation only, and unscaled
observed time series are used for the other variables.
3.2.1. Bias Correction
[
22] FortheBCoftheCMdata,weusealinearscalingas
employed by Lenderink et al. [2007] and Görgen et al. [2010].
[
23]LetX be a meteorological input variable for the HM.
As the observational reference, we use subbasin-averaged
time series and call it X
obs
avg
. For OBS
ST
used by PREVAH,
we first spatially interpolate the station data and derive sub-
basin-averaged time series. Next, for each CM j,wecalcu-
lateasubbasinmeantimeseriesasareaweightedgridcell
averages and denote it with X
j
avg
. We estimate the correction
parameters a
X
(m) between the subbasin mean time series
X
obs
avg
and X
j
avg
for each month m in the annual cycle. We use
the following linear correction models for the various mete-
orological variables (see Table 1 for an explanation of the
abbreviations):
P
j
avg
¼ a
P
P
j
avg
; a
P
mðÞ¼
P
obs
avg
mðÞ
P
j
avg
mðÞ
(1)
T
j
avg
¼ T
j
avg
þ a
T
; a
T
mðÞ¼T
obs
avg
mðÞT
j
avg
mðÞ (2)
S
j
avg
¼ min S
0
; a
S
S
j
avg
hi
; a
S
mðÞ¼
S
obs
avg
mðÞ
S
j
avg
mðÞ
(3)
R
j
avg
¼ a
R
R
j
avg
; a
R
m
ðÞ
¼
R
obs
avg
mðÞ
R
j
avg
mðÞ
(4)
H
j
avg
¼ max 0; 100 a
H
100 H
j
avg
hi
; a
H
mðÞ¼
100 H
obs
avg
mðÞ
100
H
j
avg
mðÞ
(5)
V
j
avg
¼ a
V
V
j
avg
; a
V
mðÞ¼
V
obs
avg
mðÞ
V
j
avg
mðÞ
; (6)
where the overbar denotes climatological monthly means in
the CTL, and the superscript * stands for the bias-corrected
subbasin-averaged daily time series. The equations in the
left column show the linear BC model, whereas the equa-
tions in the right column explain how the correction param-
eter is derived. Note that, although not explicitly indicated
in the left column, the correction factor a
X
varies according
to the month. As indicated in Table 1, each HM uses only a
selection of the meteorological variables listed above.
[
24] Some CMs do not provide sunshine duration.
Thus, we use cloud fraction (C) as a proxy and derive S
according to
S
j
avg
¼ S
0
1 C
j
avg
(7)
with S
0
representing the maximum possible sunshine
duration.
[
25] After the BC on the basin level, a further spatial disag-
gregation step is necessary for the PREVAH model to account
for the finer spatial input structure. In the CTL, we derive for
every month m and every meteorological subarea i a disaggre-
gation relation based on the observations according to
r
i
mðÞ¼X
obs
i
mðÞX
obs
avg
mðÞ for temperature ; and (8)
r
i
mðÞ¼
X
obs
i
mðÞ
X
obs
avg
mðÞ
for the other variables ; (9)
BOSSHARD ET AL. : UNCERTAINTY SOURCES IN CLIMATE-IMPACT PROJECTIONS
5
