Introduction
Seasonal streamflow forecasts predict the likelihood of a difference from
normal conditions in the following months. Unlike forecasts at shorter
timescales, which aim to predict individual events, seasonal streamflow
forecasts aim at predicting long-term (i.e. weekly to seasonal) averages.
The predictability in seasonal streamflow forecasts is driven by two
components of the Earth system, the initial hydrological conditions (IHC;
i.e. of snowpack, soil moisture, streamflow and reservoir levels, etc.) and
large-scale climate patterns, such as the El Niño–Southern Oscillation
(ENSO), the North Atlantic Oscillation (NAO), the Pacific-North American
(PNA) pattern and the Indian Ocean Dipole (IOD) (Yuan et al., 2015b).
The first seasonal streamflow forecasting method, based on a regression
technique developed around 1910–1911 in the United States, harnessed the
predictability from accurate IHC of snowpacks to derive streamflow for the
following summer (Church, 1935). This statistical method recognized
antecedent hydrological conditions and land surface memory as key drivers of
streamflow generation for the following months.
Alongside the physical understanding of streamflow generation processes came
technical developments, such as the creation of the first hydrological models
and the acquisition of longer observed meteorological time series, which led
to the creation of the first operational model-based seasonal streamflow
forecasting system. This system, called extended streamflow prediction (ESP;
i.e. note that ESP nowadays stands for ensemble streamflow prediction,
although it refers to the same forecasting method), was developed by the
United States National Weather Service (NWS) in the 1970s (Twedt et al.,
1977; Day, 1985). The ESP forecasts are produced by forcing a hydrological
model, initialized with the current IHC, with the observed historical
meteorological time series available. The output is an ensemble streamflow
forecast (where each year of historical data is a streamflow trace) for the
following season(s) (Twedt et al., 1977; Day, 1985). The quality of the ESP
forecasts can be high in basins where the IHC dominate the surface
hydrological cycle for several months (the exact forecast quality depending
on the time of year and the basin's physiographic characteristics; Wood and
Lettenmaier, 2008).
In basins where the meteorological forcings drive the predictability,
however, the lack of information on the future climate is a limitation of the
ESP forecasting method and might result in unskilful ESP forecasts. This
drawback led to the investigation of the use of seasonal climate forecasts,
in place of the historical meteorological inputs, to feed hydrological models
and extend the predictability of hydrological variables on seasonal
timescales (Pagano and Garen, 2006). This investigation was made possible by
technical and scientific advances. Scientifically, seasonal climate forecasts
were improved greatly by the understanding of ocean–atmosphere–land
interactions and the identification of large-scale climate patterns as
drivers of the hydro-meteorological predictability (Goddard et al., 2001;
Troccoli, 2010). This was technically implementable with the increase in
computing resources, making it possible to run dynamical coupled
ocean–atmosphere–land general circulation models on the global scale at
high spatial and temporal resolutions (Doblas-Reyes et al., 2013). An
additional technical challenge, the coarse spatial resolution of seasonal
climate forecasts compared to the finer resolution of hydrological models,
had to be addressed. To tackle this issue, many authors have explored
different ways of downscaling climate variables for hydrological applications
(Maraun et al., 2010, and references therein).
Schematic of the EFAS-WB streamflow simulation and of the CM-SSF and
ESP seasonal streamflow hindcast generation, where P is precipitation, T
is temperature, E is evaporation and ETpot is potential evapotranspiration.
The Lisflood model diagram was taken from Burek et al. (2013).
![]()
While climate-model-based seasonal streamflow forecasting experiments are
more common outside of Europe, for example for the United States (Wood et
al., 2002, 2005; Mo and Lettenmaier, 2014), Australia (Bennett et al., 2016),
or Africa (Yuan et al., 2013), they remain limited in Europe, with a few
examples in France (Céron et al., 2010; Singla et al., 2012; Crochemore
et al., 2016), in central Europe (Demirel et al., 2015; Meißner et al.,
2017), in the United Kingdom (Bell et al., 2017; Prudhomme et al., 2017) and
at the global scale (Yuan et al., 2015a; Candogan Yossef et al., 2017). This
is because, although the quality of seasonal climate forecasts has increased
over the past decades, there remains limited skill in seasonal climate
forecasts for the extra-tropics, particularly for the variables of interest
for hydrology, notably precipitation and temperature (Arribas et al., 2010;
Doblas-Reyes et al., 2013).
In Europe, the NAO is one of the strongest predictability sources of seasonal
climate forecasts; it is associated with changes in the surface westerlies
over the North Atlantic and Europe, and hence with changes in temperature and
precipitation patterns over Europe (Hurrell, 1995; Hurrell and Van Loon,
1997). It was shown to affect streamflow predictability, especially during
winter (Dettinger and Diaz, 2000; Bierkens and van Beek, 2009; Steirou et
al., 2017), in addition to the IHC and the land surface memory. It was
furthermore shown to be an indicator of flood damage and occurrence in parts
of Europe (Guimarães Nobre et al., 2017).
As the quality and usefulness of seasonal streamflow forecasts increase,
their usability for decision-making has lagged behind. Translating the
quality of a forecast into an added value for decision-making and
incorporating new forecasting products into established decision-making
chains are not easy tasks. This has been explored for many water-related
applications, such as navigation (Meißner et al., 2017), reservoir
management (Viel et al., 2016; Turner et al., 2017), drought-risk management
(Sheffield et al., 2013; Yuan et al., 2013; Crochemore et al., 2017),
irrigation (Chiew et al., 2003; Li et al., 2017), water resource management
(Schepen et al., 2016) and hydropower (Hamlet et al., 2002), but seasonal
streamflow forecasts have yet to be adopted by the flood preparedness
community.
The European Flood Awareness System (EFAS) is at the forefront of seasonal
streamflow forecasting, with one of the first operational pan-European
seasonal hydrological forecasting systems. The aim of this paper is to
bridge the current gap in pan-European climate-model-based seasonal
streamflow forecasting studies. Firstly, the setup of the newly operational
EFAS climate-based seasonal streamflow forecasting system is presented. A
Europe-wide analysis of the skill of this forecasting system compared to the
ESP forecasting approach is then presented, in order to identify whether
there is any added value in using seasonal climate forecasts instead of
historical meteorological observations for forecasting streamflow on
seasonal timescales over Europe. Subsequently, the potential usefulness of
the EFAS seasonal streamflow forecasts for decision-making is assessed.
Results
Overall skill of the CM-SSF
In the first part of the results, the skill of the CM-SSF (benchmarked with
respect to the ESP) is presented, in terms of the accuracy (MAESS),
sharpness (IQRSS), reliability (PITSS) and overall performance (CRPSS) in
the hindcast datasets. This will benchmark the added value of using Sys4
against the use of historical meteorological observations for forecasting
the streamflow on seasonal timescales over Europe.
As shown by the MAESS boxplots (Fig. 3), the CM-SSF appears on average more
accurate than the ESP for the first month of lead time only, for all seasons
excluding spring (MAM). Beyond 1 month of lead time, the CM-SSF becomes on
average as or less accurate than the ESP. There are however noticeable
differences between the different seasons. The CM-SSF shows the largest
improvements in the average accuracy compared to the ESP in winter (DJF) and
for the first month of lead time. For longer lead times (i.e. 2 to 7 months),
the accuracy of the CM-SSF is on average quite similar to that of the ESP in
autumn (SON) and winter, and on average lower in spring and summer (JJA). The
boxplots for the CRPSS look very similar to the MAESS boxplots, the main
difference being the lower average scores for 2 to 7 months of lead time in
autumn and winter (Fig. 3).
Boxplots of the MAESS, CRPSS, IQRSS and PITSS (from the top to
bottom rows) for all four seasons (SON, DJF, MAM and JJA from the left-most
to right-most columns) as a function of lead time (i.e. 1 to 7 months). The
boxplots contain the scores for all target months falling in a given season
and all 74 European regions. For all scores, values larger (smaller) than
zero indicate that the CM-SSF is more (less) skilful than the ESP
(benchmark). Where the skill is zero, the CM-SSF is as skilful as the ESP for
the hindcast period. Note that the PITSS plots have a different y-axis
scale.
![]()
The boxplots of the IQRSS show that the CM-SSF predictions are on average as
sharp as those of the ESP for the first month of lead time (slightly sharper
in autumn; Fig. 3). For 2 to 7 months of lead time, in autumn and winter, the
CM-SSF predictions are on average sharper than those of the ESP, whereas in
spring and summer, the CM-SSF predictions are on average slightly less sharp
than the ESP predictions.
Plots of the percentage of the ESP (a) and the CM-SSF
(b) hindcasts falling into each reliability category (reliable – in
terms of both spread and bias, too large, too narrow, over-predicting and
under-predicting) for all four seasons (SON, DJF, MAM and JJA from the
left-most to right-most bars in each reliability category). The results are
shown as bar charts for the first month of lead time and as circles for the
seventh month of lead time. These lead times were selected for display to
highlight the evolution of reliability between the first and last months of
the hindcast. The percentages were calculated from hindcasts for all target
months falling in a given season and all 74 European regions.
![]()
As shown by the boxplots of the PITSS (Fig. 3), the CM-SSF predictions are
less reliable than the ESP predictions for all seasons and months of lead time. For the first month of
lead time and all seasons, 10–20 % of the ESP hindcasts and less than
5 % of the CM-SSF hindcasts are reliable (Fig. 4). About 40–60 % of
the ESP hindcasts are not reliable for the first month of lead time and all
seasons due to the ensemble spread. Approximately half of these hindcasts are
too large, while the other half (slightly more in autumn and winter) are too
narrow. Furthermore, 50–80 % of the ESP hindcasts
under-predict the simulated streamflow for the first month of
lead time and all seasons. The percentage of reliable (unreliable) ESP
hindcasts increases (decreases) with lead time, as the effect of the IHC
fades away. About 70–90 % of the CM-SSF hindcasts are too narrow for the
first month of lead time and all seasons. With increasing lead time, the
percentage of CM-SSF hindcasts that are too narrow (large) decreases
(increases), especially in spring. Approximately 40–50 % of the CM-SSF hindcasts over-predict the simulated
streamflow in spring and summer for the first month of lead time (and
increasingly over-predict with longer lead times). In autumn and winter,
about 70 % of the CM-SSF hindcasts under-predict the simulated streamflow
for the first month of lead time (and increasingly under-predict with longer
lead times).
For all verification scores, the boxplots for autumn and winter are slightly
smaller than for spring and summer, hinting at a smaller variability in the
verification scores amongst regions and target months in autumn and winter
than in spring and summer. Furthermore, the presence of the boxplots above
the zero line (i.e. no skill line) for all lead times suggests that the
CM-SSF is more skilful than the ESP for some regions and target months,
beyond the first month of lead time.
Potential usefulness of the CM-SSF
In the second part of the results, the potential usefulness of the CM-SSF
compared to the ESP is described for decision-making. Here, potential
usefulness is defined as the ability of the forecasting systems to predict
lower or higher streamflows than normal, as measured with the ROC score.
Generally, either of the two forecasting systems (CM-SSF or ESP) is capable
of predicting skilfully whether the streamflow will be anomalously low or
high in the coming months (Fig. 5). However, for a few seasons and regions,
none of the two forecasting systems is skilful at predicting lower and/or
higher streamflows than normal. This is especially noticeable in winter.
Maps of the best system (as measured with the ROC score) for all
four seasons (SON, DJF, MAM and JJA) and the lower and upper simulated
streamflow seasonal terciles (left-most and right-most columns, respectively)
in each region from (a) to (h). The pie charts display the
best system for each lead time (i.e. 1 to 7 months), as shown in the example
pie chart on the bottom right of this figure. There are three possible cases:
(1) neither the ESP nor the CM-SSF is skilful (red colours), (2) the ESP is
skilful and better than the CM-SSF (yellow colours), and (3) the CM-SSF is
skilful and better than the ESP (blue colours).
![]()
For most seasons and regions, the ESP is more skilful than the CM-SSF at
predicting lower and higher streamflows than normal. However, in winter for
most regions and during other seasons for several regions, the CM-SSF appears
more skilful than the ESP. Regions where the CM-SSF best predicts lower and
higher streamflows than normal at most lead times are summarized in Table 1
for all four seasons and the lower and upper terciles of the simulated
streamflow.
Regions where the CM-SSF is more skilful than the ESP at predicting
anomalously low (lower tercile; first column) or high (upper tercile; second
column) streamflows for all four seasons (SON, DJF, MAM and JJA from the top
to bottom rows). This is a summary of the information displayed in Fig. 5.
Lower tercile
Upper tercile
SON
– Few regions in Fennoscandia – Po River basin (northern Italy) – Elbe River basin (south of Denmark) – Upstream of the Rhine River basin – Upstream of the Danube River basin – Duero River basin (Iberian Peninsula)
– Few regions in Fennoscandia – Iceland – Parts of the Danube River basin – Segura River basin (Iberian Peninsula)
DJF
Many regions except – in most of Fennoscandia north of the Baltic Sea, – parts of central Europe.
Same as lower tercile.
MAM
– Few regions on the Iberian Peninsula – Few regions in the western part of central Europe
Same as lower tercile.
JJA
– Few regions in the United Kingdom (UK) – Ireland – North-western edge of the Iberian Peninsula – Regions in Fennoscandia around the Baltic Sea – Regions south of the North Sea
– Northern part of the UK – Ireland – North-western edge of the Iberian Peninsula – Regions in Fennoscandia around the Baltic Sea – Around the Elbe River basin – Upstream of the Danube River basin – Along the Adriatic Sea in Italy
Discussion
Does seasonal climate information improve the predictability of seasonal
streamflow forecasts over Europe?
On average over Europe and across all seasons, the CM-SSF is skilful (in
terms of hindcast accuracy, sharpness and overall performance, using the ESP
as a benchmark) for the first month of lead time only. This means that, on
average, Sys4 improves the predictability over historical meteorological
information for pan-European seasonal streamflow forecasting for the first
month of lead time only. At longer lead times, historical meteorological
information becomes as good as or better than Sys4 for seasonal streamflow
forecasting over Europe. Crochemore et al. (2016) and Meißner et
al. (2017) similarly found positive skill in the seasonal streamflow forecast
(Sys4 forced hydrological model compared to an ESP) for the first month of
lead time, after which the skill faded away for basins in France and central
Europe, respectively. Additionally, on average over Europe and across all
seasons, the CM-SSF is less reliable than the ESP for all lead times. This is
due to a combination of too narrow and biased CM-SSF hindcasts, where the
bias depends on the season that is being forecasted. As mentioned in the
methods section of this paper, the ESP is not a “naive” benchmark, which
might partially explain the limited predictability gained from Sys4.
The predictability varies per season and the CM-SSF predictions are on
average sharper than and as accurate as the ESP predictions in autumn and
winter beyond the first month of lead time (and increasingly sharper with
longer lead times). The CM-SSF however tends to systematically under-predict
the autumn and winter simulated streamflow (and increasingly under-predicts
with longer lead times). In spring and summer, the CM-SSF predictions are on
average less sharp and less accurate than the ESP predictions, and they tend
to systematically over-predict the simulated streamflow (and increasingly
over-predict with longer lead times).
The added predictability gained from Sys4 was shown to lead to skilful CM-SSF
predictions of lower and higher streamflows than normal for specific seasons
and regions. The CM-SSF is more skilful at predicting anomalously low and
high streamflows than the ESP in certain seasons and regions, and noticeably
in winter in almost 40 % of the European regions, mostly clustered in
rainfall-dominated areas of western and central Europe. Several authors have
discussed the higher winter predictability over (parts of) Europe, with
examples in basins in France (Crochemore et al., 2016), central Europe
(Steirou et al., 2017), the UK (Bell et al., 2017) and the Iberian Peninsula
(Lorenzo-Lacruz et al., 2011). Bierkens and van Beek (2009) additionally
showed that there was a higher winter predictability in Scandinavia, the
Iberian Peninsula and around the Black Sea. Our results are mostly consistent
with these findings, except for Scandinavia, where the ESP is more skilful
than the CM-SSF in winter. Bierkens and van Beek (2009) produced the seasonal
streamflow forecast analysed in their paper by forcing a hydrological model
with resampled years of historical meteorological information based on their
winter NAO index. However, Sys4 has difficulties in forecasting the NAO over
Europe (Kim et al., 2012), which could have led to these inconsistent results
with the ones presented by Bierkens and van Beek (2009).
In spring, the CM-SSF is more skilful than the ESP at predicting lower and
higher streamflows than normal beyond 1 month of lead time in approximately
15 % of the European regions, and mostly in regions of western Europe.
This could be due to a persistence of the skill from the previous winter
through the land surface memory (i.e. groundwater-driven streamflow or
snowmelt-driven streamflow), as highlighted by Bierkens and van Beek (2009)
for Europe, Singla et al. (2012) for parts of France, Lorenzo-Lacruz et
al. (2011) for the Iberian Peninsula and Meißner et al. (2017) for the
Rhine. Moreover, it could be that most of the gained predictability occurs in
March, a transition month between the more predictable winter (as mentioned
above) and spring, as discussed by Steirou et al. (2017). The ESP is overall
more skilful than the CM-SSF at predicting the spring streamflow in
snow-dominated regions (e.g. most of Fennoscandia and parts of central and
eastern Europe). This hints at the importance of the IHC (i.e. of snowpack)
and the land surface memory for forecasting the spring streamflow in
snow-dominated regions in Europe.
The added predictability from Sys4 for forecasting lower and higher
streamflows than normal is limited in summer and autumn for most regions. The
CM-SSF is more skilful at predicting anomalously low and high streamflows
than the ESP in about 10–20 % of the European regions during those
seasons. Other studies have found similar patterns for (parts of) Europe;
these include less skill in summer than in winter overall for basins in
France (Crochemore et al., 2016), less skill for the low flow season (July to
October) for basins in central Europe (Meißner et al., 2017), negative
correlations in summer and autumn seasonal streamflow forecasts in central
Europe as the influence of the winter NAO fades away (Steirou et al., 2017),
and less skill overall in summer than in winter in Europe (Bierkens and van
Beek, 2009). The lower CM-SSF skill for predicting lower and higher
streamflows than normal in summer could additionally be due to the convective
storms in summer over Europe, which are hard to predict, and to the fact that
it is the dry season in most of Europe, where rivers are groundwater fed.
Therefore, in this season, the quality of the IHC controls the streamflow
predictability.
While the CM-SSF is most skilful (in terms of hindcast accuracy, sharpness
and overall performance, using the ESP as a benchmark) in autumn and winter
and most potentially useful in winter, this does not appear to correlate with
high performance in the Sys4 precipitation and temperature hindcasts (as seen
on the maps of correlation for Sys4 precipitation and temperature for all
four seasons (SON, DJF, MAM and JJA) and with 2 months of lead time (as
identified in this paper); available at
https://meteoswiss.shinyapps.io/skill_metrics/, Forecast skill metrics,
2017). Over Europe, the Sys4 precipitation and temperature hindcasts are the
most skilful in summer and the least skilful in autumn and winter. Moreover,
the regions of high CM-SSF skill for predicting lower and upper streamflows
than normal do not clearly correspond to regions of high performance in the
Sys4 precipitation and temperature hindcasts. These differences could be
partially induced by the different benchmark used to evaluate the skill of
the CM-SSF (i.e. the ESP) compared to the one used to look at the performance
of the Sys4 precipitation and temperature hindcasts (i.e. ERA-Interim).
However, these results clearly indicate that looking at the performance of
the Sys4 precipitation and temperature hindcasts only does not give a good
indication of the skill and potential usefulness of the seasonal streamflow
hindcasts over Europe, and that marginal performance in seasonal climate
forecasts can translate through to more predictable seasonal streamflow
forecasts, and vice versa. The added predictability in the CM-SSF could be
due to the combined predictability in the precipitation and temperature
hindcasts, as well as a lag in the predictability from the land surface
memory.
In most regions and for most seasons, at least one of the two forecasting
systems (CM-SSF or ESP) is able to predict lower or higher streamflows than
normal. However, in winter, the number of regions and lead times for which
none of the forecasting systems are skilful increases. This could be because
in winter, many regions experience weather-driven high streamflows and the
performance of Sys4 is limited at this time of year (as mentioned above). In
those regions, the seasonal streamflow forecasts could be improved either by
improving the IHC, through for example data assimilation, or by improving
the seasonal climate forecasts.
Overall, the ESP appears very skilful at forecasting lower or higher
streamflows than normal, showing the importance of IHC and the land surface
memory for seasonal streamflow forecasting (Wood and Lettenmaier, 2008;
Bierkens and van Beek, 2009; Yuan et al., 2015b).
What is the potential usefulness and usability of the EFAS seasonal
streamflow forecasts for flood preparedness?
What appears like little added skill does not necessarily mean no skill for
the forecast users and can in fact be a large added value for decision-making
(Viel et al., 2016). The ability of a seasonal streamflow forecasting system
to predict the right category of an event months ahead is valuable for many
water-related applications (e.g. navigation, reservoir management,
drought-risk management, irrigation, water resource management, hydropower
and flood preparedness). From the results presented in this paper, it appears
that either of the two forecasting systems (CM-SSF or ESP) is capable of
predicting lower or higher streamflows than normal months in advance, thanks
to the predictability gained from the IHC, the land surface memory and the
seasonal climate hindcast in some regions and for certain seasons.
However, as highlighted by White et al. (2017), there is currently a gap
between usefulness and usability of seasonal information. What is a useful
scientific finding does not automatically translate into usable information
which will fit into any user's decision-making chain (Soares and Dessai,
2016). While several authors have already investigated the usability of
seasonal streamflow forecasts for applications such as navigation
(Meißner et al., 2017), reservoir management (Viel et al., 2016; Turner
et al., 2017), drought-risk management (Sheffield et al., 2013; Yuan et al.,
2013; Crochemore et al., 2017), irrigation (Chiew et al., 2003; Li et al.,
2017), water resource management (Schepen et al., 2016) and hydropower
(Hamlet et al., 2002), its application to flood preparedness is still left
mostly unexplored. One exception being Neumann et al. (in review),
who look at the use of the CM-SSF to predict the 2013/14 Thames basin floods.
This is partially due to the complex nature of flood
generating mechanisms, still poorly studied on seasonal timescales beyond
snowmelt-driven spring floods, as well as the fact that seasonal forecasts
reflect the likelihood of abnormal seasonal streamflow totals, but without
much skilful information on the exact timing, location and severity of the
impact of individual flood events within that season. Coughlan de Perez et
al. (2017) looked at the usefulness of seasonal rainfall forecasts for flood
preparedness in Africa and highlighted the complexities behind using these
forecasts as a proxy for floodiness (for a discussion on floodiness, see
Stephens et al., 2015). Furthermore, decision-makers in the navigation,
reservoir management, drought-risk management, irrigation, water resource
management and hydropower sectors are familiar with working on long
timescales (i.e. several weeks to months ahead). In contrast, the flood
preparedness community is currently mostly used to working on timescales of
hours to a couple of days.
The Red Cross Red Crescent Climate Centre has recently designed a new
approach that harnesses the usefulness of seasonal climate information for
decision-making for disaster management. This approach, called
“Ready-Set-Go!”, is made up of three stages. The “Ready” stage is based
on seasonal forecasts, where they are used as monitoring information to drive
contingency planning (e.g. volunteer training). The “Set” stage is
triggered by sub-seasonal forecasts, used as early-warning information to
alert volunteers. Finally, the “Go!” stage is based on short-range
forecasts and consists in the evacuation of people and the distribution of
aid (White et al., 2017). Using a similar approach, seasonal streamflow
forecasts could complement existing forecasts at shorter timescales and
provide monitoring and early-warning information for flood preparedness. Such
an approach however requires the use of consistent forecasts from short to
seasonal timescales. In this context, moving to seamless forecasting is
becoming vital (Wetterhall and Di Giuseppe, in review).
Soares and Dessai (2016) also identified the accessibility to the
information, enhanced by collaborations and ongoing relationships between
users and producers, as a key enabler of the usability of seasonal
information. International projects, such as the Horizon 2020 IMPREX
(IMproving PRedictions and management of hydrological EXtremes) project (van
den Hurk et al., 2016), alongside promoting scientific progress on
hydrological extremes forecasting from short to seasonal timescales over
Europe, gather together forecasters and decision-makers and can effectively
demonstrate the added value of the integration of seasonal information in
decision-making chains. The Hydrologic Ensemble Prediction EXperiment
(HEPEX) is another international initiative that brings together researchers
and practitioners in the field of ensemble prediction for water-related
applications. It is an ideal environment for collaboration and fosters
communication and outreach on topics such as the usefulness and usability of
seasonal information for decision-making.
Aspects for future work
In this paper, terciles of the simulated streamflow are used. However, and
because the application of the EFAS seasonal streamflow forecasts is of
particular relevance for flood preparedness, the evaluation of the hindcasts
for lower and higher streamflow extremes (for example the 5th and 95th
percentiles, respectively) would be more relevant and might give very
different results. This was not done in this paper as the time period covered
by the seasonal streamflow hindcasts (i.e. approximately 27 years) was not
long enough for statistically reliable results for lower and higher
streamflow extremes. The limited hindcast length is a common problem in
seasonal predictability studies. Increasing the hindcast length back in time
could lead to more stable Sys4 hindcasts and hence to more stable and
potentially skilful seasonal streamflow hindcasts (Shi et al., 2015).
Furthermore, in this paper, the hindcasts were analysed against simulated
streamflow, used as a proxy for observed streamflow. This is necessary
because it enables an analysis of the quality of the hindcasts over the
entire computation domain, rather than at non-evenly spaced stations over
the same domain (Alfieri et al., 2014). Further work could however include
carrying out a similar analysis for selected river stations in Europe, in
order to account for model errors in the hindcast evaluation.
The calculation of the verification scores (excluding the ROC) was made by
randomly selecting 15 ensemble members from the 51 ensemble members of the
CM-SSF hindcasts, for starting dates for which the ensemble varies between 15
and 51 members (i.e. hindcasts made on 1 January, March, April, June, July,
September, October and December; this is due to the split between 15 and 51
ensemble members in the Sys4 hindcasts, as described in Sect. 2.1.2 of this
paper). In order to investigate the potential impact of this evaluation
strategy on the results presented in this paper, the CRPSS was calculated for
15 and 51 ensemble members of the CM-SSF hindcasts for starting dates for
which 51 ensemble members are available for the full hindcast period (i.e.
hindcasts made on 1 February, May, August and November). This is displayed in
Fig. 6 for all hindcast starting dates, lead times (i.e. 1 to 7 months) and
regions combined. Overall, it is apparent that the impact of this evaluation
strategy on the results presented in this paper should be minimal, as all
points align themselves approximately with the 1-to-1 diagonal.
CRPSS calculated for the CM-SSF against the ESP (benchmark) for
hindcasts made on 1 February, May, August and November, all lead times (i.e.
1 to 7 months) and all 74 European regions. The x-axis (y-axis) contains
the CRPSS calculated from 15 (all 51) ensemble members of the CM-SSF.
![]()
The next version of the ECMWF seasonal climate forecast, SEAS5, was released
in November 2017. Future work could include forcing the Lisflood model with
SEAS5 and comparing the obtained seasonal streamflow hindcasts to the CM-SSF
presented in this paper. This should indicate whether developments to the
seasonal climate forecast translate through to better pan-European seasonal
streamflow forecasts, which is of particular interest for regions and seasons
when neither the ESP nor the CM-SSF is currently skilful.
The operational EFAS medium-range streamflow forecasts are currently
post-processed as a means to improve their reliability (Smith et al., 2016,
and references therein). Results from this paper have shown that the CM-SSF
is mostly unreliable (with regards to the EFAS-WB) and could hence benefit
from post-processing of the seasonal climate forecast. However,
post-processing techniques used for the EFAS medium-range streamflow
forecasts might not be suitable for the CM-SSF, as the seasonal climate
forecast used for the latter should be post-processed in terms of its
seasonal anomalies rather than for errors in the timing, volume and magnitude
of specific events. This is currently being considered for operational
implementation within EFAS and is an active area of discussion within the
EFAS user community.
For the analysis presented in this paper, the CM-SSF was benchmarked against
the ESP. Several other techniques exist for seasonal streamflow forecasting,
such as statistical methods using predictors ranging from climate indices to
antecedent observed precipitation and crop production metrics, to mention a
few (e.g. Mendoza et al., 2017; Slater et al., 2017). Further analysis could
include benchmarking the CM-SSF against one or multiple statistical methods,
to assess the relative benefits of various seasonal streamflow forecasting
techniques.
In this paper, the ability of both systems (CM-SSF and ESP) to forecast lower
and higher streamflows than normal was explored, with several hypotheses made
to link the streamflow predictability to regions' hydro-climatic processes.
This includes the higher potential usefulness of the ESP in forecasting the
spring streamflow in snow-dominated regions and the summer streamflow in
regions where rivers are groundwater fed. In these regions and for these
seasons, the IHC and the land surface memory drive the predictability. The
CM-SSF provides an added potential usefulness in winter in the
rainfall-dominated regions of central and western Europe, where the skill
appears to persist through to spring due to the land surface memory (i.e.
groundwater-driven streamflow and snowmelt-drive streamflow). While further
exploration of these hypotheses is outside of the scope of this paper, future
work is required to disentangle the links between the added predictability
from Sys4 and the basins' hydro-climatic characteristics, for example,
understanding the predictability in snow-dominated basins, arid regions and
temperate groundwater-fed basins.
In this context, additional work to further disentangle and quantify the
contribution of both predictability sources (seasonal climate forecasts
versus IHC) to seasonal streamflow forecasting quality over Europe could be
carried out by using the EPB (end point blending) method (Arnal et al.,
2017).
Conclusions
In this paper, the newly operational EFAS seasonal streamflow forecasting
system (producing the CM-SSF forecasts by forcing the Lisflood model with the
ECMWF System 4 seasonal climate forecasts (Sys4)) was presented and
benchmarked against the ESP forecasting approach (ESP forecasts produced by
forcing the Lisflood model with historical meteorological observations) for
the hindcast period 1990 to 2017. On average, Sys4 improves the
predictability over historical meteorological information for pan-European
seasonal streamflow forecasting for the first month of lead time only (in
terms of hindcast accuracy, sharpness and overall performance). However, the
predictability varies per season and the CM-SSF is more skilful on average at
predicting autumn and winter streamflows than spring and summer streamflows.
Additionally, parts of Europe exhibit a longer predictability, up to 7 months
of lead time, for certain months within a season. In terms of hindcast
reliability, the CM-SSF is on average less skilful than the ESP for all lead
times, due to a combination of too narrow and biased CM-SSF hindcasts, where
the bias depends on the season that is being forecasted.
Subsequently, the potential usefulness of the two forecasting systems (CM-SSF
and ESP) was assessed by analysing their skill in predicting lower and higher
streamflows than normal. Overall, at least one of the two forecasting systems
is capable of predicting those events months in advance. The ESP appears the
most skilful on average, showing the importance of IHC and the land surface
memory for seasonal streamflow forecasting. Nevertheless, for certain regions
and seasons the CM-SSF is the most skilful at predicting anomalously low or
high streamflows beyond 1 month of lead time, noticeably in winter for almost
40 % of the European regions. This potential usefulness could be
harnessed by using seasonal streamflow forecasts as complementary information
to existing forecasts at shorter timescales, to provide monitoring and
early-warning information for flood preparedness.
Overall, patterns in skill in the CM-SSF are however not mirrored in the Sys4
precipitation and temperature hindcasts. This suggests that using seasonal
climate forecast performance as a proxy for seasonal streamflow forecasting
skill is not adequate and that more work is needed to understand the link
between meteorological and hydrological variables on seasonal timescales over
Europe.