In Alpine catchments, snowmelt is often a major contribution to runoff. Therefore, modeling snow processes is important when concerned with flood or drought forecasting, reservoir operation and inland waterway management. In this study, we address the question of how sensitive hydrological models are to the representation of snow cover dynamics and whether the performance of a hydrological model can be enhanced by integrating data from a dedicated external snow monitoring system. As a framework for our tests we have used the hydrological model HBV (Hydrologiska Byråns Vattenbalansavdelning) in the version HBV-light, which has been applied in many hydrological studies and is also in use for operational purposes. While HBV originally follows a temperature-index approach with time-invariant calibrated degree-day factors to represent snowmelt, in this study the HBV model was modified to use snowmelt time series from an external and spatially distributed snow model as model input. The external snow model integrates three-dimensional sequential assimilation of snow monitoring data with a snowmelt model, which is also based on the temperature-index approach but uses a time-variant degree-day factor. The following three variations of this external snow model were applied: (a) the full model with assimilation of observational snow data from a dense monitoring network, (b) the same snow model but with data assimilation switched off and (c) a downgraded version of the same snow model representing snowmelt with a time-invariant degree-day factor. Model runs were conducted for 20 catchments at different elevations within Switzerland for 15 years. Our results show that at low and mid-elevations the performance of the runoff simulations did not vary considerably with the snow model version chosen. At higher elevations, however, best performance in terms of simulated runoff was obtained when using the snowmelt time series from the snow model, which utilized data assimilation. This was especially true for snow-rich years. These findings suggest that with increasing elevation and the correspondingly increased contribution of snowmelt to runoff, the accurate estimation of snow water equivalent (SWE) and snowmelt rates has gained importance.

Snowmelt provides a dominant contribution to runoff and groundwater storages
in mountainous regions. In such areas, modeling snow processes is crucial for
resource management as well as for flood and drought forecasting. Snow
accumulates and acts as a temporary storage of water that is released as soon
as snowmelt occurs. Since erroneous simulations of snow accumulation can bias
the amount and timing of simulated snowmelt, accurately modeling both
processes is important for runoff predictions. Problems for modelers may
occur not only due to the great heterogeneity and variability of these processes, but
also due to the limited availability of necessary observational data

Characteristics of 20 Swiss catchments in this study.

Locations of snow observation stations (red stars) and 20 studied catchments (white border lines) in Switzerland.

To cover a wide range of elevations and different climatic regions, for this
study we chose 20 catchments spread over Switzerland. All of them were at
most minimally affected by human activities, such as water regulation or
abstraction. A further crucial selection criterion was the availability of
the required data. Since, especially at high elevations, the runoff regime of
many catchments in Switzerland is affected by man-made installations, the
number of possible catchments was highly limited. Catchments analyzed in this
study varied in size from 17 to 473 km

The hydrological model HBV

The external snow model framework, which we used in this study instead of the
snow routine built in the HBV model, also simulates snowmelt by a TI approach
but in addition allows for integration of observational snow data using a
data assimilation scheme. While some details on the external snow model
framework are given below, a full description of model and data assimilation
methods is available in

TI snowmelt model with data assimilation and time-varying DDF (M1): this
model is the same as that described in detail in

TI snowmelt model with time-varying DDF without data assimilation (M2): in this version, the same elaborated TI approach as in M1 was applied to simulate snow accumulation and melt at each grid cell based on the same input data grids as in M1. The DDF seasonal variations are equal to those in M1. The only difference concerns the data assimilation procedures, which were switched off in M2, such that observed SWE data were not used to update the initial estimates on snow accumulation and melt rates.

TI snowmelt model using a constant DDF without data assimilation (M3): this
version differs from M2 with respect to the DDF. Here the DDF does not show
seasonal variations but is assumed to be constant over the season. The average
DDF of 2.5 mm

Cumulative snowmelt during the snowmelt season 2007 as calculated by the snow model method M1 (full model with data assimilation, left), M2 (full model without data assimilation, middle) and M3 (simplified model, right). The sums between the three model methods differ depending on the use of observational snow data assimilation and the use of different DDFs.

Graphical explanation of how to calculate

Replacing a TI model with another TI model, and not with an energy-balance or
snowpack-physics model, may appear unusual at first glance. However, if
concerned with conceptual hydrological modeling at a daily timescale, the TI
model framework used here constituted an ideal testing environment. To
provide daily snowmelt rates, the dynamic data assimilation framework within
M1 represents current state-of-the art methodology in operational snow
hydrological monitoring. Since it accounts for measured snow depletion rates
at hundreds of monitoring sites, it provides the best possible input to the
hydrological model. Even with data assimilation switched off (M2), if
validated against snow lysimeter data at daily time steps, the performance is
almost on par with the output of top-notch energy-balance models

Timing of snowmelt onset and of runoff events due to snowmelt affects the
availability of water resources and influences flooding and droughts

Both efficiency metrics were calculated for (a) each catchment and (b) each of the two calibration experiments. The performance statistics are discussed separately for each of the three groups of catchments depending on mean elevation.

To look for differences between the three snow model methods, individual catchments and years were selected. Representing a catchment at high elevations, results for the Dischma catchment (EZG 2327, gauge Davos Kriegsmatte) with a mean elevation of 2349 m a.s.l. are shown in Fig. 4. The yellow background displays the catchment-specific snowmelt season during which the bulk of the snowmelt typically occurs. The blue and gray lines at top of the graph indicate the snowmelt input to the hydrological model from M1 excluding and including rain, respectively, in this example for the record-high snow year 1999. The observed runoff is shown by the black curve, while the different colored curves indicate the simulations with M1, M2 and M3. The curves as well as the performance metrics achieved by the differential split-sample experiment demonstrate that for this catchment, the M1 model as input to the hydrological framework provided the best runoff simulations, even though the differences are small. Note however, that in this example M1 particularly outperforms the other models in the month of July, which is outside the standard evaluation period.

Observed and modeled runoff for the Dischma catchment for 1999, as well as water input from snowmelt and rain modeled with method M1. The upper benchmark model BM in red.

Results of the leave-one-out approach.

First, we used the leave-one-out approach to calibrate the hydrological
model. The leave-one-out approach represents a typical scenario in
operational conceptual runoff modeling, i.e., to use as much data as possible
for calibration and to apply the resulting parameter values to the current
season. Results grouped according to mean catchment height are presented
below (Fig. 5). Using this calibration procedure for catchments with mean
elevation below 1000 m a.s.l., the hydrological model showed good results
independent of which snow model was used as input to the hydrological model
framework. Even without using a snow model at all (i.e., the lower
benchmark), the runoff model resulted in lower but still positive performance
values, indicating that the choice of snow model within a conceptual runoff
modeling framework is of less importance when dealing with catchments at
lower elevations. Similarly for catchments with mean elevation between 1000
and 2000 m a.s.l. the differences between the three model runs were small.
While

Results of the differential split-sample approach.

For the differential split-sample approach, snow-rich years were used to
validate the runoff models. As expected, the analysis using the differential
split-sample approach revealed similar performance patterns compared to the
leave-one-out approach, but with increased differences between model runs
(Fig. 6). As seen before, at low and mid-elevation classes the differences
between the three model versions as well as between calibration and
validation were comparably small. The median values of efficiencies for each
model version ranged between 0.7 and 0.8 (

Results of the leave-one-out approach for catchments with mean
elevation above 2000 m a.s.l. Median (solid lines) and interquartile
(25th–75th percentile; shading) range of

The validation of the differential split-sample experiment showed that the
three external snow models provided the best runoff simulations for snow-rich
years, specifically for catchments with a mean elevation of above
2000 m a.s.l. In a further analysis, we ordered the single validation
years individually by catchment for the leave-one-out approach from snow poor
to snow rich based on peak SWE. This procedure allowed testing of whether
there was a trend in the runoff performance metrics associated with the snow
amount of single years. Such a trend was indeed evident, as seen in Fig. 7.
Independent of the snow model used, the best results were achieved when
validating the model performance during snow-rich years regarding both

Based on daily runoff data measured over a period of 15 years at 20
catchments in Switzerland, we evaluated the sensitivity of a conceptual
hydrological modeling framework to snowmelt input from snow models of
different complexity. The most complex snow model integrated
three-dimensional sequential assimilation of snow monitoring data with a
snowmelt model based on the temperature-index approach. In contrast, the
simplest snow model represented snowmelt with a constant degree-day factor,
and did not include any data assimilation. The snow models were combined with
the HBV model in the version HBV-light

This study was partly funded by the Federal Office of the Environment (FOEN). We thank MeteoSwiss for access to the meteorological data and FOEN for providing river-runoff observations used in this study. Thanks to Manfred Stähli and Massimiliano Zappa for helpful discussions and to Nathalie Chardon for reviewing the English of this article.Edited by: R. Woods Reviewed by: J. Parajka, O. Rössler, and one anonymous referee