The interaction of mountain terrain with meteorological processes causes substantial temporal and
spatial variability in snow accumulation and ablation. Processes impacted by complex terrain
include large-scale orographic enhancement of snowfall, small-scale processes such as
gravitational and wind-induced transport of snow, and variability in the radiative balance such as
through terrain shadowing. In this study, a multi-scale modelling approach is proposed to simulate
the temporal and spatial evolution of high-mountain snowpacks. The multi-scale approach combines
atmospheric data from a numerical weather prediction system at the kilometre scale with process-based
downscaling techniques to drive the Canadian Hydrological Model (CHM) at spatial resolutions
allowing for explicit snow redistribution modelling. CHM permits a variable spatial resolution by
using the efficient terrain representation by unstructured triangular meshes. The model simulates
processes such as radiation shadowing and irradiance to slopes, blowing-snow transport (saltation
and suspension) and sublimation, avalanching, forest canopy interception and sublimation, and
snowpack melt. Short-term, kilometre-scale atmospheric forecasts from Environment and Climate
Change Canada's Global Environmental Multiscale Model through its High Resolution Deterministic
Prediction System (HRDPS) drive CHM and are downscaled to the unstructured mesh scale. In
particular, a new wind-downscaling strategy uses pre-computed wind fields from a mass-conserving
wind model at 50
High-mountain snowpacks are characterized by a strong spatial and temporal variability that is
associated with elevation, vegetation cover, slope steepness, orientation and wind exposure. This
variability results from processes occurring during the snow accumulation and ablation periods at a
large range of spatial scales (e.g., Pomeroy and Gray, 1995; Pomeroy et al., 1998, 2012, 2016; Clark
et al., 2011; Mott et al., 2018). Snow accumulation at the mountain range scale (1–500
The multi-scale variability of mountain snow represents a challenge for snow models used in support
of avalanche hazard forecasting (Morin et al., 2020), hydrological predictions (e.g., Warscher et
al., 2013; Brauchli et al., 2017; Freudiger et al., 2017) and climate projections (e.g., Rasouli et
al., 2014; Hanzer et al., 2018) in mountainous terrain. Several modelling strategies have been
proposed to face this challenge and to capture this multi-scale variability. At the mountain range
scale, atmospheric models at sufficient resolutions (4
Snowdrift-permitting models simulate wind-induced snow transport in the saltation and suspension layers (e.g, Pomeroy and Gray, 1995). As proposed by Mott et al. (2018), they can be divided into two main categories: (i) models solving the vertically integrated mass flux in the saltation and suspension layers (Essery et al., 1999; Durand et al., 2005; Pomeroy et al., 2007; Liston et al., 2007) and (ii) models solving the three-dimensional (3-D) advection-turbulent diffusion equation of blown snow particles in the atmosphere (Gauer, 1998; Lehning et al., 2008; Schneiderbauer and Prokop, 2011; Sauter et al., 2013; Vionnet et al., 2014). One of the main challenges for all these models is obtaining accurate driving wind fields at sufficient high resolution since they strongly impact the accuracy of simulated snow redistribution (Mott and Lehning, 2010; Musselman et al., 2015). Models of the first category need two-dimensional (2-D) driving wind fields. Liston et al. (2007), inspired by Ryan (1977), proposed the use of terrain-based parameters to adjust distributed wind fields to the local topography. These distributed wind fields can be obtained from interpolated station data (Gascoin et al., 2013; Sextone et al., 2018), hourly output from regional climate models at a convective-permitting scale (Reveillet et al., 2020) or a pre-computed wind field library using an atmospheric model (Berhnardt et al., 2010). Essery et al. (1999) used a linearized turbulence model (Walmsley et al., 1982) to build a pre-computed library of 2-D wind maps to distribute wind measurements from station data. Musselman et al. (2015) showed that this approach led to more accurate simulations of snow redistribution around an alpine crest than wind fields derived from the terrain-based parameters proposed by Liston et al. (2007). Models of the second category require a 3-D representation of the wind field and associated atmospheric turbulence. In this case, driving wind fields can be obtained from computational fluid dynamics (CFD) models (Gauer, 1998; Schneiderbauer and Prokop, 2011) or atmospheric models in large-eddy simulation (LES) mode used to generate a library of pre-computed wind fields (Lehning et al., 2008; Mott and Lehning, 2010) or fully coupled to a snowpack model (Vionnet et al., 2014). These advanced models can be used for detailed studies such as the feedbacks between blowing-snow sublimation and the atmosphere (Groot Zwaaftink et al., 2011) or the processes driving the variability of snow accumulation during a snowfall event, including preferential deposition of snowfall (Lehning et al., 2008; Mott et al., 2010; Vionnet et al., 2017).
Differences in the level of complexity of snowdrift-permitting models and associated driving wind
fields influence the spatial and temporal ranges of application of these models. Due to their
relatively low computational costs, models of the first category can be applied to simulate the snow
cover evolution over entire snow seasons at a resolution between 25 and 200
To overcome some of these limitations, Marsh et al. (2020a) developed a snowdrift-permitting scheme of intermediate complexity that solves the 3-D advection–diffusion blowing-snow transport on a variable resolution unstructured mesh. This scheme is implemented in the Canadian Hydrological Model (CHM; Marsh et al., 2020b). The landscape is discretized using a variable resolution unstructured mesh that allows an accurate representation of terrain heterogeneities with limited computation elements (Marsh et al., 2018). Marsh et al. (2020a) used the WindNinja diagnostic wind model (Forthofer et al., 2014) to build libraries of pre-computed wind fields. Wagenbrenner et al. (2016) showed that WindNinja can be used to downscale wind field from atmospheric models running at a convection-permitting scale in complex terrain.
The objective of the present study is to develop and evaluate a novel strategy for multi-scale modelling of
mountain snowpack over large regions and for entire snow seasons. Specifically, the following questions are asked. (
This work studies the evolution of the mountain snowpack around the Kananaskis Valley of the
Canadian Rockies, Alberta (Fig. 1). The study domain covers an area of 958
The digital elevation model (DEM) from the Shuttle Radar Topography Mission-SRTM (EROS Center, 2017)
at a resolution of 1 arcsec (30
Variable-resolution triangular mesh used in this study over a sub-area of the Kananaskis domain. The location of this sub-area corresponds to the red-shaded area shown in Fig. 1b. The underlining DEM was taken from the SRTM mission at 1 arcsec.
Characteristics of the mesh used in this study. The vertical error corresponds the root-mean-square error to the underlying reference topographic raster.
A dataset of tall vegetation (
Distributed snowpack simulations over the triangular mesh of the study area were performed using the
version of the Snobal scheme (Marks et al., 1999) implemented in CHM (Marsh et al. 2020b). Snobal has
been used in numerous mountainous regions across North America (e.g., Garen and Marks, 2005; Pomeroy
et al., 2016; Hedrick et al., 2018). Snobal is a physically based snowpack model that approximates
the snowpack with two layers. The surface layer was implemented here with a fixed thickness of
0.1
CHM also includes a 3-D advection–diffusion blowing-snow transport and sublimation model (Marsh et
al., 2020a): the 3-D Prairie Blowing Snow Model (PBSM-3D). This scheme uses a finite-volume method
discretization on the unstructured mesh. It deploys the parameterization of Li and Pomeroy (1997) to
determine the threshold wind speed for snow transport initiation as a function of air temperature
and snow presence. It does not depend on the properties of surface snow (e.g., density, liquid water
content) simulated by Snobal (see Sect. 4.4 for a discussion on the limitation of this approach). In
the case of blowing-snow occurrence, the steady-state saltation parameterization of Pomeroy and Gray (1990) is
used to compute the mass concentration in the saltation layer. The concentration in the saltation
layer is impacted by shear stress partitioning due to the presence of vegetation (such as shrubs)
and the upwind fetch. Upwind fetch is calculated for each triangle of the mesh using the
In steep alpine terrain, gravitational snow transport strongly affects the spatial variability of
the snowpack (e.g., Sommer et al., 2015) and the mass balance of glaciers (Mott et al., 2019) and
modifies the runoff behaviour of alpine basins (Warscher et al. 2013). For these reasons, the
SnowSlide scheme (Bernhard and Schulz, 2010) was implemented in CHM. SnowSlide is a simple
topographically driven model that simulates the effects of gravitational snow transport. SnowSlide
uses a snow-holding depth that decreases exponentially with increasing slope angle, limiting snow
accumulation in steep terrain. SnowSlide was initially developed for regular gridded rasters and has
been adapted here to the unstructured triangular mesh used by CHM. SnowSlide operates from the
highest triangle of the mesh to the lowest one. If the snow depth exceeds the snow-holding capacity
for a given triangle, excess snow is redistributed to the lower adjacent triangles, proportionally
to the elevation difference between the neighbouring triangles and the original one. SnowSlide uses
the total elevation (snow depth plus surface elevation) to operate. In this study, the default
formulation of the snow-holding depth proposed by Bernhardt and Schulz (2010) is used, which leads to
a maximal snow thickness (taken perpendicular to the slope) of 3.08
The impact of the presence of forest vegetation on snow interception, sublimation, snowpack accumulation and melt energetics is represented in CHM using the same canopy module as in the Cold Region Hydrological Model (CRHM; Ellis et al., 2010; Pomeroy et al. 2012). This module used leaf area index and canopy closure to compute the effect of forests on shortwave and longwave irradiance at the snow surface. Snow interception and sublimation of intercepted snow are also represented following Hedstrom and Pomeroy (1998). In this study, the canopy module was activated for the triangles covered by forest as described in Sect. 2.2.1.
Snobal and PBSM-3D require the following atmospheric forcing: air temperature, humidity, wind speed,
wind direction, liquid and solid precipitation rates, and longwave and shortwave irradiance. Due to
the scarcity of the network of meteorological stations in the region (Fig. 1), hourly atmospheric
forcings were obtained from the High-Resolution Deterministic System (HRDPS; Milbrandt et al.,
2016). HRDPS is the high-resolution NWP system running the Global Environmental Multiscale Model
(GEM) operationally over Canada at 2.5
The HRDPS atmospheric forcing at 2.5
Mountain wind fields are notoriously difficult to observe and model (Davies et al., 1995), and
obtaining high-resolution wind fields constitutes one of the greatest challenges for blowing-snow
models in mountainous terrain (e.g., Mott and Lehning, 2010; Vionnet et al., 2014; Musselman et al.,
2015; Réveillet et al., 2020). In the context of this study, hourly HRDPS near-surface wind
fields at the 2.5
The application and extension of the Barcons et al. (2018) approach for use on an unstructured mesh
and to account for direction perturbations are detailed below. First, to build the wind map library,
WindNinja was run at 50
Then, for each wind direction in the wind map library, the transfer function
To account for impacts on direction, the following approach was taken. The rasters of the wind map
library containing the horizontal
Wind speeds were then adjusted to 10
Forthofer et al. (2014) and Wagenbrenner et al. (2016, 2019) showed that the mass-conserving
version of WindNinja has difficulties simulating lee-side recirculation where flow separation
occurs. This difficulty is due to the absence of a momentum equation in the WindNinja flow
simulation (Forthofer et al., 2014). As lee-side flow strongly influences snow accumulation
(e.g., Gerber et al., 2017), an additional and optional step was added to the wind-downscaling procedure described
above. It consisted of a modification of the transfer functions
A set of CHM experiments were designed to assess the effect of the wind field downscaling and the
impact of process representation on snowpack simulations at snowdrift-permitting scales (Table 2). A
reference CHM configuration including wind downscaling accounting for recirculation and
gravitational and blowing-snow redistribution was first defined (
CHM simulations (experiments) used in this study. Rc indicates CHM simulations using wind fields from the downscaling method accounting for wind speed reduction in leeward areas. HRDPS refers to the High Resolution Deterministic Prediction System and WN to WindNinja. See text for more details.
Hourly meteorological data collected at CRHO stations were used to evaluate the precipitation and
wind fields driving CHM (Table 3). These stations include those in Marmot Creek Research Basin (Fang
et al., 2019) and Fortress Mountain Snow Laboratory (Harder et al., 2016) (Table 3 and Fig. 1),
covering an elevation range from 1492 to 2565
Meteorological stations used for wind evaluation. TPI refers to the
topographic position index and is defined as the difference between the
elevation of the station minus the mean elevation within a 2
Airborne laser scanning (ALS) surveys were performed over the Kananaskis region on 5 October 2017
(late summer scan) and on 27 April 2018 (winter scan) using a Riegl Q-780 infrared
(1024 nm) laser scanner with a dedicated Applanix POS AV Global Navigation Satellite
System (GNSS) inertial measurement unit (IMU). The Q-780 scanner was flown at heights of
approximately 2500
The distributions of simulated and observed snow depths were compared for different 200
Wayand et al. (2018) suggested that snow persistence indices from Sentinel-2 images present a strong
potential for the evaluation of distributed snow models in mountainous areas. Hence, maps of the
snow-covered area from the Copernicus Sentinel-2 satellite mission (Drusch et al., 2012) at
20
Sentinel-2 snow cover maps at 20
The evaluation of the different wind-downscaling methods is described in Sect. 3.1. The quality of the snowpack simulations is then assessed in Sect. 3.2 using airborne lidar snow depth data and snow persistence indexes. A special emphasis is placed on the ability of the model to capture the elevation–snow depth relation as well as snow redistribution around wind-exposed ridges.
Figure 3 compares the near-surface wind field obtained from a simple bilinear interpolation of the
HRDPS wind field (Fig. 3a) with the downscaled wind field obtained with (Fig. 3c) and without
(Fig. 3b) the wind speed reduction in leeward areas. HRDPS provided a smooth wind field with
relatively higher wind speeds in the northwestern part of the region characterized by high relief
(Fig. 3a) compared to the rest of the area. HRDPS did not reflect the local terrain information due
to a horizontal resolution of 2.5
Near-surface wind field on 10 September 2017 at 18:00 UTC from
Figure 4 gives the error metrics for the wind speed (bias and RMSE) between the CHM simulations and
observations at eight automatic weather stations. The HRDPS without downscaling overestimated wind
speed (positive bias) at all stations, except the CNT station. This station is located on an exposed
crest and presents the largest TPI value among the stations used for model evaluation
(Table 3). Downscaling wind to the CHM mesh using WindNinja microscale winds (experiment
Evaluation of simulated wind speed using different downscaling methods:
The wind-downscaling method also modified the general wind direction (Figs. 3 and 5). Prevailing
winds during the study originated from the south (S; 180
Same as Fig. 4 for wind direction:
To assess the ability of CHM to simulate small-scale features of snow accumulation and transport in
alpine terrain, ALS-derived snow depths were compared with simulated snow depths for different CHM
experiments for a sub-region of approximately 77
Snow depth on 27 April 2018
The agreement between observed and simulated snow depth distributions was examined as a function of
elevation for three sub-regions (see Fig. 1) of the Kananaskis domain: Kananaskis North, Kananaskis
South and Haig (Figs. 7 and 8). For each sub-region, the median of observed snow depth increased
with elevation up to 2400
Boxplots showing the distributions of observed and simulated snow depth per 200
Wasserstein distance and RMSE between observed and simulated snow depth distribution as a function of elevation for four CHM experiments and three sub-regions. The location of these sub-regions is shown in Fig. 1.
Including blowing-snow redistribution strongly affected model results. As expected, it increased the
spatial variability of simulated snow depth within each elevation band compared to experiments
NoWndTr NoAv and NoWndTr Av (Fig. 7). When the wind speed reduction in leeward
areas was not simulated (experiment WndTr Av NoRc), CHM underestimated the median snow
depth (as well as the first and third quartiles) above 2500
The observed and simulated snow depth distributions were compared for the upper slopes of the domain
(defined in Sect. 2.3.2), particularly exposed to wind-induced snow transport (Fig. 9). The CHM
simulation without lateral redistribution of snow, NoWndTr NoAv, presented a systematic
overestimation of snow depth for all slope orientations (Fig. 9a–c) and yielded the worst
Wasserstein distance metric among all simulations (Fig. 9d–f). Including gravitational
redistribution reduced the positive bias and the Wasserstein distance. This reduction was not found
for some slope orientations, however (W, SW and S orientations for Kananaskis North, Fig. 9a; SW
and S orientations for Kananaskis South, Fig. 9b). The moderate values of the slope angle generally
found for these orientations were not sufficient to trigger gravitational snow redistribution in
SnowSlide. For example, the percentage of slope values larger than 40
Bias
Figure 10 shows the maps of observed snow persistence indexes as well as the indexes derived from
two CHM simulations. Observed SP (Fig. 10a) presented similar patterns compared to the observed
distribution of snow depth in late April (Fig. 6a), showing that snow persistence patterns are
primarily controlled by the patterns of peak snow accumulation (Wayand et al., 2018). Avalanche
deposits identified in Fig. 6a corresponded to maximal SP values, whereas low SP values were found
near ridge lines, exposed to wind. Overall, the Pearson correlation coefficient between observed
snow depth and SP reached 0.69 (
Maps of snow persistence index (SP)
Figure 11 shows how accurately the different CHM experiments were able to reproduce the observed SP
distributions as a function of elevation. The simulation-derived and observed SP distributions are
shown in Fig. S2. Model performances for snow persistence were generally in
agreement with those for snow depth presented in Fig. 8. Experiments without blowing-snow
redistribution (NoWndTr NoAv and NoWndTr Av) overestimated snow persistence at all
elevations with a positive bias increasing with elevation for experiment NoWndTr
NoAv. Including blowing-snow redistribution in experiments WndTr Av NoRc and WndTr Av Rc significantly decreased snow persistence, mainly above 2300
Bias and Wasserstein distance between observed and simulated snow persistence index as a function of elevation for four CHM experiments and three sub-regions. The location of these sub-regions is shown in Fig. 1.
Same as Fig. 9 but for the snow persistence index (SP).
This study presents a new high-resolution modelling strategy for mountain snowpack, combining atmospheric forcing from a NWP system at convection-permitting scale with the multi-scale, snowdrift-permitting model CHM. Several CHM configurations were tested to highlight how omitting physical processes influenced the performances of snowpack simulations at snowdrift-permitting
scales (50
Results of blowing-snow redistribution simulations in CHM were sensitive to the quality of the
driving wind field, in particular the impact of recirculation areas, at the mountain range scale
(
The evaluation of the wind-downscaling methods versus point measurements did not show systematic improvements compared to the original HRDPS wind field, consistent with studies of high-resolution wind modelling in complex terrain (e.g., Horvath et al., 2012; Vionnet et al., 2015). Model results in Sect. 3.1 highlight the challenge of evaluating wind simulations at locations near peaks or ridges due to approximation in the location of the stations as previously mentioned in Fiddes and Gruber (2014) and Winstral et al. (2017). On the other hand, differences between the wind-downscaling methods were clearly identified and quantified when evaluating the snow simulations using distributed data. ALS snow depth and snow persistence indexes derived from Sentinel-2 allowed for targeted model evaluation in areas of interest such as the upper slopes exposed to wind-induced snow transport. These results confirm the large potential of ALS snow depth data for detailed model evaluation (e.g., Hanzer et al., 2016; Hedrick et al., 2018). In addition, they show that snow persistence indexes derived from freely available Sentinel-2 images (Wayand et al., 2018) can generally support more similar conclusions than those derived from ALS snow depth. This highlights the fact that these indexes can be used to evaluate large-scale snowpack simulations at snowpack-permitting scales in regions that are not covered by lidar. As illustrated by Wayand et al. (2018), the snow persistence index is influenced by variability in both snow accumulation and ablation, so that this index can only be used to evaluate snow redistribution models if variable insolation effects are also simulated. This is the case in the simulations presented in this paper (Sect. 2.2.5).
Two types of metrics were used when using ALS snow depth data for model evaluation: RMSE and Wasserstein distance. RMSE corresponds to a traditional “point-to-point” verification metric. Such a metric may favour homogenous snow cover simulations. Indeed, a snow cover simulation including avalanching may present a degree of realism, but errors in the exact location of the avalanche deposits may increase RMSE compared to a simulation without avalanching due to the double-penalty problem (e.g., Nurmi, 2003). This issue is often encountered when evaluating the ability of high-resolution atmospheric models to simulate localized events such as convective precipitation (e.g., Clark et al., 2016). The Wasserstein distance (Rüschendorf, 1985) was used in this study as a complementary metric to evaluate the agreement between observed and simulated distributions for specific areas (elevation bands or specific slope orientations). This metric may lead to a perfect match even if the observations and the simulations are not co-located, however. This highlights the need to consider several verification metrics with identified strengths and limitations. In the future, more advanced verification methods such as the neighbourhood method developed in the atmospheric community (Ebert et al., 2013) could be considered.
This study used a wind-downscaling method inspired by Barcons et al. (2018) and developed for large
areas. Part of the uncertainty associated with this method comes from the value of the radius of
influence used to compute the transfer functions (Sect. 2.2.4). A value of 1
All the atmospheric driving data for CHM were obtained from the HRDPS, the Canadian NWP system using
GEM at 2.5
The model evaluation for the upper slopes exposed to wind showed that CHM simulations including
blowing snow tend to overestimate snow redistribution across slopes subject to wind erosion and
deposition. These results were obtained for a CHM mesh with a typical area of
Gravitational snow redistribution is simulated in CHM with the SnowSlide scheme (Bernard and Schulz, 2010). Model results showed that CHM can reproduce the formation of snow accumulations due to avalanching that visibly correspond with the observations. However, the increase in RMSE for snow depth at low elevation for all simulations including avalanching suggests that CHM does not effectively capture the true location of these deposits. SnowSlide relies on a maximum holding capacity of snow that only depends on the slope angle and does not consider the small-scale terrain roughness, limiting the ability of the scheme to reproduce snow accumulation for steep faces (Sommer et al., 2015). In addition, the exact location of avalanche deposits is influenced by avalanche dynamics (Pudasaini and Hutter, 2007), which are not reproduced in SnowSlide. CHM also does not represent snowfall enhancement due to interactions between the flow field and the local cloud formation as well as the preferential deposition of snowfall resulting from pure particle flow interaction (Lehning et al., 2008; Vionnet et al., 2017; Mott et al., 2018). Gerber et al. (2019) suggested that, when combined, these two effects can increase snow accumulation on the leeward side of mountain ridges by 26 %–28 %. In the current version of CHM, wind-induced snow transport is the only process responsible for additional snow deposition on leeward slopes. The parameterization of Dadic et al. (2010) could be tested in CHM but would require an estimation of the vertical wind speed that could be provided by WindNinja. A study is in progress in the Canadian Rockies to better assess the impact of terrain–flow–precipitation interactions on snow accumulation in the region. Finally, uncertainties associated with the Snobal snowpack scheme were not quantified in this study. In particular, errors in simulated snow density can affect the comparison between observed and simulated snow depth (Raleigh and Small, 2017; Lv and Pomeroy, 2020), despite the use of an improved snow density algorithm for Snobal (Hedrick et al., 2018). Inaccurate estimations of the ground heat flux may also affect the simulation of the snow cover duration (Slater et al., 2017). Pritchard et al. (2020) showed how multi-physics ensemble snow modelling can be applied to assess uncertainties on distributed snowpack simulations and a similar framework could be applied to CHM, including uncertainties in PBSM-3D and SnowSlide.
This study presents a new multi-scale modelling strategy for mountain snowpacks over large regions. It
combines (i) atmospheric forcing from the Canadian GEM NWP system at a convective-permitting scale
(Milbrandt et al., 2016), (ii) a meteorological downscaling module including a wind-downscaling
strategy relying on the diagnostic wind model WindNinja (Forthofer et al., 2014) and (iii) the
multi-scale snowdrift-permitting model CHM (Marsh et al., 2020a, b). This system was used to
simulate the snowpack evolution for an entire snow season over a domain of 958
The main conclusions of this study are as follows.
Pre-computed wind fields at 50 Snowpack simulation without lateral snow redistribution (blowing-snow and gravitational snow
redistribution) cannot capture the spatial variability of snow cover in alpine terrain and
overestimates snow depth and snow cover duration at high elevations. Including gravitational
redistribution improved model results and reduced snow depth at high elevations. Snow depth and
snow cover duration were still overestimated around ridge lines exposed to winds. Snowpack simulation including blowing-snow and gravitational snow redistribution provided the
best estimates of the shape of the elevation–snow depth relation across the Kananaskis region and
reproduced the decrease in mean snow depth found at high elevation. These results were obtained
for a CHM experiment driven by a wind field including the wind speed reduction in leeward
areas. Removing this reduction led to a systematic underestimation of snow depth around ridges,
partially due to an underestimation of snow deposition on leeward slopes. These results highlight
that wind fields without lee-side slowdown are not sufficient to simulate snow redistribution in
mountainous terrain. Snowpack simulations including blowing-snow and gravitational snow redistribution
overestimated snow redistribution from windward to leeward slopes and subsequent avalanching. This
is potentially due to the absence of subgrid topographic effects in the driving wind field and in
the snow transport equations in CHM. High-resolution snow persistence indexes derived from Sentinel-2 present a strong potential
for the detailed evaluation of distributed snowpack models, in particular in regions where
airborne lidar snow depth data are not available. These indices can be used for model evaluation
targeting specific areas (e.g., ridge lines exposed to intense wind-induced snow redistribution,
avalanche deposition areas).
The results of this study demonstrate that CHM at a snowdrift-permitting scale constitutes a promising tool for large-scale modelling of mountain snowpack. Future work will combine (i) improvements in the physical parameterizations in CHM and in the driving wind fields, (ii) large-scale simulations across the western Canadian Cordillera, and (iii) improvements of CHM simulations with assimilation of high-resolution observations such as ALS snow depth or Sentinel-2 snow cover.
The open-source CHM model code (Marsh et al., 2020b,
CRHO meteorological and snow observations are available through the web portal
The supplement related to this article is available online at:
VV, CM, BM and JP designed the study and the modelling strategy. VV and CM developed the wind-downscaling module in CHM. VV was responsible for the analysis of the results and the preparation of the manuscript. BM, JS and KM processed and provided the airborne lidar snow depth data. SG processed and provided the Sentinel2 snow cover images. NW provided a pre- and post-processing toolkit for CHM. All authors contributed to the preparation of the manuscript.
The authors declare that they have no conflict of interest.
Martyn Clark (USask) and Barbara Casati (ECCC) are thanked for their helpful scientific discussions. Special thanks are due to Logan Fang (USask) and Greg Galloway (USask) for providing quality-controlled snow and meteorological data. Simon Gascoin acknowledges the Centre National d'Etudes Spatiales (CNES) for granting him access to its high-performance computer. We thank Rebecca Mott and Tobias Sauter for their careful reading and useful comments, which improved the manuscript.
This research has been supported by the Canada First Research Excellence Fund (Global Water Futures), the Natural Sciences and Engineering Research Council of Canada (Discovery Grants), Alberta Innovates (Water Innovation Program), Canada Research Chairs (CRC in Water Resources and Climate Change, CRC in Glacier Change), and the Canadian Foundation for Innovation and Tula Foundation.
This paper was edited by Jürg Schweizer and reviewed by Rebecca Mott and Tobias Sauter.