Investigating the development of shallow snowpacks on arable land, using comprehensive field observations and spatially distributed snow modelling

In many countries, at low altitude, snowpacks in agricultural areas are typically shallow (i.e., <1 m deep). Shallow snowpacks are much more sensitive to changes in air temperature and fluctuations between precipitation as rain or snow than deep snowpacks and have a reduced ability to dampen peak outflows compared with deep snowpacks (Wever et al. 2014). Furthermore, snowpacks less than 20 cm deep cannot insulate the soil from low air temperatures, but can amplify deep-freezing of soils due to an increase in ground albedo caused by the snow (Zhao et al. 2013). This deep-freezing of the soil profile can change the soil hydraulic properties dramatically (Stähli et al. 2001; Al-Houri et al. 2009). In southern Norway this causes severe soil erosion through impeded infiltration of snowmelt and rainwater at the end of winter (Øygarden 2003). The amount of snow at the end of winter determines how much meltwater is available for creating overland flow and soil erosion on agricultural soils (Lundberg et al. 2016). Besides damaging the nutrient rich top layer of agricultural soils, sediment transfer from fields to streams during winter and spring represents a major part of the annual loss of phosphorus and nitrogen from agricultural catchments (Su et al. 2011). Therefore, it is important to better understand the development of shallow snowpacks during winter, including the spatial distribution of snow water equivalent (SWE).

Snow is one of the most changeable elements in the hydrological cycle and poses a great range of challenges for monitoring and measuring (Doesken & Robinson 2009). Below, the main factors influencing snow processes of interest for this study are summarised. A more detailed description can be found in Gray & Male (1981) and Lundberg et al. (2016).

Catchment properties such as slope, aspect and vegetation cover have a major influence on the condition and morphology of snowpack. Slope, curvature and aspect determine exposure to wind (Lapen & Martz 1996), precipitation and solar radiation (Gray & Male 1981). Wind interacting with terrain and vegetation heavily influences the deposition and redistribution of snow, enhancing or reducing snow accumulation (Gascoin et al. 2013). More snow is usually accumulated on leeward slopes and forest edges than on windward sides, due to reduced exposure to higher wind speeds. Snow depth is often greater in depressions due to reduced wind speeds and increased turbulence. Furthermore, vegetation (e.g., crop residues) enhances the roughness of the surface during the exposed period before total snow cover, resulting in increased turbulent wind flow and causing complex snow structures and patterns. Another important factor is the air temperature at the time of snowfall, which controls the dryness, hardness and crystalline form of new snow, and thereby its erodibility by wind. For instance, heavy wet snow falling at temperatures close to 0 °C is less susceptible to wind transport and is common for low altitude areas in the temperate zone (Richard & Brun 2008), like southern Norway, where this study was conducted.

The main process responsible for differential melting of snow in a catchment is energy exchange. Important energy inputs and outputs are solar radiation, rain, incoming long-wave radiation, outgoing long-wave radiation and sensible and latent heat fluxes such as evaporation, condensation and sublimation of ice (Tarboton et al. 2000).

The complexity of snow accumulation, distribution and melt processes and the speed of change, which is typical for shallow snowpacks, make spatial investigation of snow properties challenging and costly (Gascoin et al. 2013). Therefore, the spatial resolution of snow-related data is often poor, or no data on spatial distribution are available (Tarboton et al. 2000). Distributed snowpack models, using physically based one-dimensional mass and energy balance, have been successfully used for describing spatial variability in snowpack properties (Luce & Tarboton 2001). However, most studies have focused on snowmelt modelling in mountainous, high altitude areas with usually deep snowpacks (e.g., Blöschl et al. 1991; Cline et al. 1998; Hiemstra et al. 2006; Bernhard et al. 2010; Zwaaftink et al. 2011) and forested areas (Lundberg et al. 2016). Little research has been done on modelling shallow snowpack processes with high spatial and time resolution for agricultural areas. These areas are often neglected due to the focus of most studies on water recharge (for hydropower and drinking water) by snow and ice (Lundberg et al. 2016). The effect of snowpacks on freeze-thaw and water transport in agricultural soils has received great attention (e.g., Flerchinger 2000; Parkin et al. 2013); however, most of these studies were conducted at point scale. Some studies have investigated distributed soil erosion induced by snowmelt in agricultural areas, but have either focused on a single snowmelt event (Ollesch et al. 2005, 2006; Weigert & Schmidt 2005) or could not produce satisfactory results for the whole winter period due to lack of incorporation of snowpack development in the model (Grønsten & Lundekvam 2006).

The need to better mitigate soil erosion and peak discharge during winter and spring from agricultural areas requires a better understanding of the temporal and spatial changes of shallow snowpacks and that spatial distributed snow models can simulate these changes, to provide continuous datasets, e.g., for soil erosion models. Therefore, the aims of this study were to: (1) quantify and visualise the temporal and spatial changes of shallow snowpacks (during three winter periods) on arable land; and (2) test a spatially distributed snow model on observed snow cover patterns. For this aim, the UEBGrid model (Sen Gupta & Tarboton 2013) was applied.

The study area, Gryteland (Figure 1), is located in the Ski municipality, approximately 30 km south of Oslo, Norway. The catchment covers an area of 0.29 km². In this area, snow cover processes have been monitored since 2008, with the main focus on arable land. A monitoring station was installed at the outlet of the sub-catchment in 2008 (Kramer & Stolte 2009). This station measures precipitation, runoff and the suspended solids content by means of automated sampling. The catchment is characterised by undulating landscape (altitude: 106–141 m a.s.l., slope 2–10%) covered by approximately 60% arable land and 40% coniferous forest. All aspects are represented except east-facing slopes.

Mean annual temperature is 5.3 °C, with minimum average daily temperatures of −4.8 °C in January/February and maximum average daily temperatures of 16.1 °C in July. Mean annual precipitation is 785 mm, with a minimum monthly amount of 35 mm in February and a maximum of 100 mm in October (Thue-Hansen & Grimenes 2015). Winter is usually relatively unstable, with alternating periods of freezing and thawing and several snowmelt events (Kværnø & Øygarden 2006). Snow data have been collected in the area since 2009 and show maximum snow depth of around 55–60 cm in the years 2009–2011 (Stolte & Kværnø 2013).

As input for modelling and for understanding the snow accumulation and melting processes, detailed measured weather data were necessary. A weather station (106 m a.s.l) was installed in the catchment at the end of 2013, providing data for the winters 2014 and 2015. Data from a station (96 m a.s.l.) 6 km from the catchment were used for winter 2013 (Thue-Hansen & Grimenes 2015).

During the three winters, SWE was sampled after weather changes were expected to result in changes in SWE. Three to five samples were collected at each measuring point, depending on the snow depth. These samples were combined to measure the mass of the snow and to calculate its density (average mass divided by average snow volume). SWE was calculated by dividing the product of average snow density and average depth of the snowpack by the density of water (1,000 kg m⁻³). The samples were taken along transects that covered all aspects and slope angles occurring in the catchment (Figure 1). To get a better understanding of the spatial development of the snow cover, snow poles (Figure 1) marked with a measuring scale were installed for winter 2015. These poles allowed snow depth to be measured from a distance by reading the value on the poles with a spyglass, keeping the disturbance of the snowpack to a minimum. Due to the strong correlation of snow depth with SWE, it has been recommended to substitute the time-consuming SWE measurements with snow depth measurements (Jonas et al. 2009).

On each measuring occasion, the SWE at each pole was estimated by multiplying the measured snow depth (at the pole) by the average snow density measured along the snow sampling transect. These spatially distributed pole measurements were then compared with the SWE maps produced by the model. In addition, aerial pictures were taken with a drone to visualise snow patterns in the catchment.

In southern Norway, the point version of the Utah Energy Balance snow model (UEB) (Tarboton & Luce 1996; You 2004; Luce & Tarboton 2010; Mahat & Tarboton 2012; Mahat et al. 2013) has been used successfully (Kramer & Stolte 2009). To extend its application in this area for spatial distributed modelling, the grid version of UEB (UEBGrid) (Sen Gupta & Tarboton 2013, 2016; Brown et al. 2014) was used in this study. It models the snow pack as a single layer to avoid over-parameterisation (Sen Gupta & Tarboton 2013). The snow pack is characterised by state variables, SWE, energy content (E_s) and the age of the snow, which is only used for albedo calculations. E_s is defined as the energy content of the snowpack plus a top layer of soil with depth (i.e., thermal active depth of the soil (D_e)). Soil temperature measurements at different depths in the catchment showed that major fluctuations usually happened between 10 and 40 cm depth (depending on the winter). Therefore, D_e, which interacts with the snow, was initially set to 10 cm (Table 1). E_s represents energy content including this soil depth. High frequency oscillating ground heat fluxes above D_e are absorbed into E_s. This procedure provides a simple approximation of the effects of frozen and unfrozen ground on the snowpack.

The model is driven by input of air temperature, precipitation, wind speed, humidity and solar radiation at time steps sufficient to resolve the diurnal cycle (six hours or less). Sen Gupta & Tarboton (2013) reorganised the input–output workflow of UEB to use network Common Data Form map files (netCDF) to allow the model to be run over a grid with input varying in space and time, resulting in the UEBGrid model used in this study. The model runs separately at each grid cell and the melt outputs are accumulated for predefined sub-watersheds.

Measured weather data (precipitation, wind speed, solar radiation, air temperature, humidity) at hourly time intervals were used here as input for the model (Figures 2 and 3). For the different years, the monthly Bristow–Campbell B parameter was given as an input (Table 1) for the model to calculate atmospheric transmissivity. Required maps for watershed delineation (Figure 1), slope and aspect were derived in ArcMap 10.2.2 from a digital elevation model (DEM), created from light detection and ranging data (LIDAR) (Starkloff & Stolte 2014). The calculations for the agricultural area were done with a raster resolution of 5 m × 5 m. The forest was not further considered in this study, because in all three winters the snow depth was negligible.

UEBGrid allows definition of grid cells for which detailed output can be produced. In this study, grid cells containing the location of a snow pole or SWE measuring point were selected for output, resulting in a total of 41 points. Six points were chosen for calibration and are discussed in detail in this paper. These six points represent the major characteristics of the catchment, with a north facing slope (‘North 1’), a south facing slope (‘South 1’), a northwest facing plain (‘Point 12’), two depressions (‘Station 1’ and ‘Station 2’) and a point influenced by the forest (‘Station 3’) (Figure 1).

The spatially distributed snow observations of 2015 allowed evaluation of the performance of the model for all locations in the study area. Therefore, the best calibration results for 2015 were spatially and temporally compared with the measured SWE measurements. To statistically evaluate the spatial performance of UEBGrid for 2015, the Nash–Sutcliffe model efficiency coefficient (NSE) was calculated for each measuring occasion.

For calibrating the model, the user guide of UEBGrid (Tarboton & Luce 1996) suggests adjusting the parameters surface aerodynamic roughness height (z₀), surface head conductance (K_s), K_sat and new snow visible band reflectance (a_vo). A previous study carried out with the point version of UEB for the same area as used in this study (Kramer & Stolte 2009) showed that the parameters K_s and a_vo had no significant effect on the modelling results and suggested that correction of the threshold temperatures T_r and T_s might be necessary. In a later study, Tarboton et al. (2000) suggested adjusting the drift factor (df) to better account for snow transport by wind. Additionally, due to the importance of the thermal interaction of shallow snowpacks with the underlying soils (Norum et al. 1976), the parameter D_e was chosen as a calibration parameter. Considering this, the parameters K_sat, z₀, df, T_r, T_s, and D_e were used for calibrating the model. The initial input parameters used for this study are presented in Table 1.

The df parameter is used in the model to assign different cells a multiplication factor (Equation (2)) to account for spatial differences in SWE due to snow drift by wind.

Measured air temperature, precipitation, and wind speed and direction are presented in Figures 2 and 3, respectively. Solar radiation and relative humidity are not presented, because no significant difference was observed between the years, which means that differences in snow cover properties were not caused by these factors.

As can be seen in Figure 2, the periods with measurable snow cover had different durations in different years.

Wind speed and wind direction are important factors for the distribution and shape of snowpacks, especially during snowfall and immediately after deposition. Already wind speed of about 3 m s⁻¹ can redistribute light and dry snow crystals (Gray & Male 1981). In Figure 3, measured wind speed together with measured precipitation for the snow measuring periods in the three winters are presented. As can be seen, the highest (hourly average; 8 m s⁻¹) measured wind speed occurred in winter 2015, possibly resulting in the redistribution and densification of deposited snow. For comparison, an average typical blizzard occurring in the USA has an average wind speed of 10 m s⁻¹. Furthermore, during the main snowfall episodes in 2015, the wind speeds exceeded 4 m s⁻¹, reaching 5–6 m s⁻¹. In comparison, during winter 2013 and 2014, the wind speeds reached only 3.5 m s⁻¹ during snowfall.

These measurements confirmed the observations in the field that wind strongly shaped the snowpack in winter 2015, which had to be considered in the modelling.

Measured snow depth and SWE for the three winter periods are presented in Figure 4, which shows results from the six locations used for model calibration and validation.

The smallest snow depths and SWEs measured at Station 2 in 2013 were set to zero because running meltwater on snow and ice made it impossible to measure during these dates (Figure 4). The measurements at Station 1 after the second measuring occasion in winter 2015 were not representative because of a thick (10–20 cm) ice layer on top of the soil, that was established after snowmelt followed by refreezing. Any SWE or snow depth measurements carried out at this location afterwards represent measurements on top of this ice layer until the end of winter.

The development of snow depth and SWE differed between the three winters (Figure 4). In winter 2013, there was an early decrease in snow depth and SWE due to a melting event at the end of January (Figure 2), followed by a sudden increase in snow depth and SWE, and a constant snow depth and slow increase of SWE over time until the spring melt. In total, 10 measuring occasions were possible in 2013.

The short winter of 2014 resulted in only four measuring occasions (note: the unusually high first snow depth value at Station 1 might be a measuring mistake). While the snow depth decreased already after the second measuring occasion, SWE continued to increase until all snow melted rapidly after the fourth measuring occasion. An increase in SWE followed by rapid snowmelt is typical behaviour of snowpack before each melting episode (Gray & Male 1981). As the snow depth decreases, the snow density increases due to retention of water (meltwater, often supplied by rain) in the snowpack, until the snowpack reaches a point where it is saturated with water (generally 3–5% water by weight) and the internal temperature of the pack reaches 0 °C (Gray & Male 1981). When the snowpack has reached this point, i.e., when it is ‘ripe’, then all additional water is quickly released and the snowpack collapses. In 2014 no major precipitation events (Figure 2) occurred before all the snow was gone. The air temperature (Figure 2) stayed below 0 °C for almost the whole period with snow, until melting started (measuring occasions 3 and 4).

During winter 2015, snow cover was established for a longer period than in 2014, resulting in seven measuring occasions. Similar to winter 2013, winter 2015 was more complex than winter 2014 with an extensive melting event before sampling was started, followed by an increase in snow depth and SWE and an additional snowmelt event on 24–26 January. Meltwater that had accumulated in the depression froze during the following cold period beginning on 26 January, which resulted in thick ice layers in the depressions (Figure 5).

Figure 5 shows an image of the arable part of the study area patched together from aerial pictures taken at the end of the melting period in 2015. In large parts of the area, all the snow had melted except in the south.

The aerial images (Figure 5) visualised spatial patterns of the snow cover. From these patterns, it was possible to obtain information about the processes determining the distribution of snow at catchment scale. Examples are: (1) ice development due to refreezing of accumulated meltwater in depressions, which were the main pathways for surface runoff; (2) the closed snow cover in the southern parts of the catchment, on arable land, showing the shadowing effect of the forest against solar radiation and deposition of snow along the forest edge due to reduced wind speeds.

The modelling results for winter 2013 are presented in Figure 6. Changing the parameters df, lh, K_sat, T_r and T_s did not improve the modelling results. The best possible fit was achieved with a slight change of z₀ from 0.01 to 0.006 m and the adjustment of D_e to 0.3 m (Figure 6, black line).

However, it should be noted that similar fits were achieved by setting the parameters lh, T_s and T_r to values of 0.8, 3 and 3.8 °C, respectively. Snowfall at temperatures around 3 °C has been observed in the study area; however, setting T_s to 3 °C alone did not change the modelling results, and the highest measured lh in soaked snow cannot greatly exceed 0.15 (Kinar & Pomeroy 2015).

That the adjustment of D_e was necessary to achieve better modelling results for 2013 indicated that the interaction of the snowpack with the underlying soil was a major process occurring in winter 2013. Observations carried out during winter 2013 supported this assumption. Winter 2013 was characterised by a major melting event during the measuring period (between 30 January and 6 February) (Figure 2). During this event an (approximately 10 cm thick) ice layer, which had developed earlier in January on the soils, prevented infiltration of meltwater (image 3 in Figure 7), resulting in high water contents in the bottom layer of the snowpack. During snow sampling on 30 January, it was observed that a slight disturbance of the snow resulted in a total collapse of the lower 5 cm snow into water (Figure 7). D_e directly influences T_a of the snowpack in UEBGrid, as can be seen from Equations (4) and (6). An increasing of D_e results in a decrease of T_a and therefore in a reduction of the liquid content of the snowpack and a slowdown of melting.

A slight overprediction for the North 1 and Station 1 locations can be seen, but not exceeding 2 cm difference between measured and modelled SWE. For these locations, the initial z₀ value of 0.01 m resulted in a better fit (Figure 6). The model predicted a longer period of snow cover for Stations 1 and 3 and Point 12 compared with the measured values (Figure 6, arrow markings). However, these measurements require some explanation. As can be seen from Figure 7, there was still snow present on 3 April 2013.

There was no snow on the south facing slopes on 3 April (image 1 in Figure 7), where the model predicted the timing of total exposure correctly. The north facing slopes were still covered with snow (image 1), which the model predicted correctly. At Point 12 (image 4), Station 3 (image 2) and Station 1 (not shown), the snow had disappeared in areas where the stubble had broken the snow surface (therefore zero measurements at these locations), but not in the wheel tracks, which had no stubble (image 4 in Figure 7). In comparison, neighbouring areas (not shown) without any plant cover or covered with only poorly developed winter wheat were still completely covered with snow on 9 April. The model was unable to account for the change in albedo and increased heat conductance due to stubble. For bare soil, the model would probably have correctly predicted the timing of total disappearance of snow at all locations. The last measurement points in Figure 6 indicate when all snow had disappeared at all locations (3 April at South 1 and 15 April at the other locations), which fits well with the timing predicted by the model.

The modelling results for 2014 are presented in Figure 8. Calibration was done by adjusting z₀ visually from 0.01 to 0.006. Increasing D_e from 0.1 to 0.3 m, as used for the simulation of winter 2013, did not improve the modelling results.

However, it should be noted that the limited number of measurements in 2014 made detailed evaluation of the model results difficult, but it can be seen that the modelled SWEs had the same magnitude as the measured values and the model simulated the measured increases and decreases in SWE (Figure 8). Winter 2014 was the least complex of the three observed winters, which was also mirrored by the simple calibration necessary.

For winter 2015, no satisfactory fit was achieved by adjusting D_e and z₀ to values used for winter 2013 and 2014 (Figure 9, dotted line). This indicated that other processes, which did not occur or had only insignificant effects in 2014 and 2013, needed to be taken into account for the 2015 simulations. That wind played a greater role in winter 2015, compared with the other winters, was confirmed by the wind measurements (Figure 3). In addition, during the conducted field measurements it was observed that the following interacting processes occurred: (1) different snow patterns due to the redistribution of snow by wind, depending on the terrain (slope, curvature and aspect); (2) the forest acting as an obstacle to wind (reducing wind speeds, causing snow to be deposited along the forest edge); (3) the shading effect of the forest against solar radiation. More snow could accumulate in the southern part of the catchment due to process 2, because the forest edge protected the area against the major wind direction of southwest to southeast in 2015 (Figure 3), contrary to 2014 and 2013, where major wind directions were north to southeast (Figure 3). These interacting processes cannot be directly simulated with UEBGrid or, indeed, with most available snow models (Essery et al. 2013). However, in UEBGrid the drift factor (df) can be used to account for the effect of processes 1 and 2 (Tarboton et al. 2000; Luce & Tarboton 2001). Accordingly, the df (Figure 8) was calibrated for the six locations (Figure 8). As can be seen, depending on the location, different values for df were necessary to fit the modelled to the measured SWE.

With these adjustments, a satisfying fit was achieved. However, the complete disappearance of snow was two days late for the South 1 location, while for North 1, Station 3 and Point 12 this was one day too early. A slight overprediction occurred at all locations, especially after the fifth measuring occasion, with a maximum difference of approximately 0.03 m for the seventh measuring occasion at the North 1 location.

That the highest df value (1.5) had to be applied to the areas close to the forest edge suggested that process 2 played a major role in winter 2015. Also, the south facing slopes required a value of 1.5, suggesting increased deposition of snow on these slopes during winter 2015. Based on this calibration, a map with different values for df, depending on the location, was created (Figure 10), enabling an evaluation of the model performance for the whole area. The highest df value (1.5) was given to the areas close to the forest edge (similar to Station 3) and south facing slopes (similar to South 1). Areas with similar characteristics to Point 12 and Station 2 were given a value of 1.4, while north facing slopes with similar characteristics to the points North 1 and Station 1 were given a value of 1.2.

Figure 11 shows the modelled SWE maps for the seven measuring occasions in 2015. In addition, for each date the modelled SWE is compared with the measured SWE at the snow pole locations.

To better quantify the performance of the model, for each measuring occasion, the calculated NSE is presented in the graphs of Figure 11. Except for 27 January, the model performed well, with NSE ranging between 0.55 and 0.73. The NSE for 27 January was 0.06, indicating that the model was only slightly better than the average of the measured data. As can be seen from Figure 11, the model predicted lower SWE at many locations compared with the measurements. The measurements taken on 27 January were characterised by a decrease in snow depth, but an increase in SWE compared with 23 January (Figure 4). This suggests that the snow pack was actually maintaining more water from precipitation (snow and rain at 1.5 °C), which had occurred on 26 January (Figure 2), than UEBGrid anticipated. The parameter lh can be used in the model to account for an increased water storage of the snowpack. However, the given value for lh is constant for the whole modelling period, which would have affected all results.

Based on the modelling experiences gathered in this study, four suggestions on how the UEBGrid model could be improved were found. (1) To improve the prediction of different snow patterns due to the combination of wind speed, wind direction and terrain, the df could be linked to slope angle, aspect and input of measured wind directions. (2) To account for accumulation of meltwater in snowpacks at the bottom of slopes and in depressions, a flow routing routine could be implemented in the model. This would also help in the context of soil erosion prediction due to snowmelt. (3) To allow for ice layers within snowpacks and on soil surfaces, theoretically, using a multiple layered snow pack and (4) including the possibility of a soil profile below the snow in the model (as in the case of some non-distributed snow models) could improve the simulation results. However, suggestions 3 and 4 could result in a difficult to calibrate, overparameterised and measurement-intensive model, reducing its applicability. Therefore, when improving the model, a good balance between model complexity and input data requirements, which UEBGrid provides, has to be maintained (Avanzi et al. 2016).

Overall, the UEBGrid model proved capable of simulating the development and spatial distribution of SWE in the shallow snowpacks in the three winter periods. However, the different calibration approaches that were necessary for winter 2013 and 2015 showed the delicacy of modelling the change in shallow snowpacks. Due to the rapid changes occurring in such snowpacks, as shown by the observations and measurements, a certain simulation error has to be expected. Especially, the refreezing of meltwater and prevention of runoff and infiltration of meltwater by frozen soils and ice layers can make simulations of shallow snowpacks challenging. Simulation results therefore have to be treated with care, and correct results can probably only be acquired in combination with detailed field observations and measurements. Many additional tools for improving the observation and measurements of rapidly changing snowpacks already exist (e.g., Egli et al. 2009; Hopkinson et al. 2012; Garvelmann et al. 2013; Kinar & Pomeroy 2015; De Michele et al. 2016), which should be more widely utilized.

The three winter periods studied differed in length, amount of snow, and processes shaping the snow cover. These interacting processes, caused by different weather conditions and state of the soil, resulted in higher (2013, 2015) or lower (2014) complexity of the modelling. By considering weather data and quantitative (measurements) and qualitative (aerial and ground images) field observations, it was possible to correctly validate the model results. Especially, the importance of detailed photographic image and video documentation, as a substitute for quantitative measurements, for process understanding cannot be stressed enough.

UEBGrid proved capable of simulating the development and spatial distribution of SWE in the study area for 2013, 2014 and 2015. The acquired calibration values, especially for df, should be tested in coming winters and in other areas with similar characteristics, to evaluate how consistent they are. UEBGrid captured the essential physics of accumulation and melt processes of shallow and patchy snow covers. The measurements and calibrations underlined how sensitive such snowpacks are to wind, small changes in the composition of precipitation and soil temperature conditions, requiring a greater focus on automated, differentiating precipitation and spatially distributed wind and SWE measurements. Also, more research into processes related to shallow snowpacks is necessary, with the prospect of decreasing snow depths in areas where snowpacks now are usually deep, due to global warming (Barnett et al. 2005).

However, the fact that it was possible to satisfactorily simulate temporal and spatial changes of shallow snowpacks made us confident that we are a step closer to the overall goal: to better quantify and finally reduce soil erosion during winter and spring.

The research was funded by the SIS-Catchy project of NIBIO, Norwegian Institute of Bioeconomy Research.