Climate change affects precipitation phase, which can propagate into changes in streamflow timing and magnitude. This study examines how the spatial and temporal distribution of rainfall and snowmelt affects discharge in rain–snow transition zones. These zones experience large year-to-year variations in precipitation phase, cover a significant area of mountain catchments globally, and might extend to higher elevations under future climate change. We used observations from 11 weather stations and snow depths measured from one aerial lidar survey to force a spatially distributed snowpack model (iSnobal/Automated Water Supply Model) in a semiarid, 1.8 km
Due to increases in temperature, mountainous regions will receive less snowfall and more rainfall (Barnett et al., 2005; Stewart, 2009). This will alter the timing and amount of snowmelt, which is a significant source for water resources across the globe (Barnett et al., 2005; Marks et al., 1999; Somers and McKenzie, 2020; Viviroli et al., 2007). On the scale of the continental United States (US), a decrease in the fraction of precipitation falling as snow (snowfall fraction hereinafter) is expected to decrease stream discharge (Berghuijs et al., 2014). Earlier stream discharge peaks in response to earlier snowmelt and a decline in summer low flows across the semiarid mountainous US have been reported in both observational data records (McCabe et al., 2017; Luce and Holden, 2009; Regonda et al., 2005) and future climate projections (Naz et al., 2016; Leung et al., 2004; Milly and Dunne, 2020; Christensen et al., 2004). However, lower snowfall fractions in much of the western US have not yet led to a significant decrease in annual discharge (McCabe et al., 2017). Therefore, understanding year-round discharge responses, in particular the sensitivity of stream discharge to changes in yearly snowfall fractions, is warranted and will help us to anticipate how stream discharge might be affected by climate change.
Variations in snowfall fractions can affect the temporal distribution of
surface water inputs (SWI
Snowfall fractions may also influence the spatial distribution of SWI. In the semiarid western US, rainfall magnitudes generally increase with elevation (Johnson and Hanson, 1995). In regions with large snowfall fractions, this general elevation-driven pattern can be overlain by impacts of wind-driven redistribution of snow, which is dependent on factors such as topography, aspect, wind speed, and wind direction (Sturm, 2015; Tennant et al., 2017; Winstral and Marks, 2014; Trujillo et al., 2007). Hence, differences in the SWI distribution due to varying snow depths could be particularly substantial in areas where wind-driven redistribution of snowfall is significant. The primary controls (e.g., topography, aspect, and elevation) on snow depth and snow water equivalent (SWE) are relatively consistent from year to year, so the interannual distribution of snow is usually spatially consistent (Parr et al., 2020; Sturm, 2015; Winstral and Marks, 2002). The effects of elevation and aspect on the spatial distribution of snow depth and, thus, the potential for SWI as snowmelt, have been well studied in both high- and mid-altitude mountains (e.g., Grünewald et al., 2014; López-Moreno and Stähli, 2008; Tennant et al., 2017). Studies on snow drifting in seasonally snow-covered areas (Mott et al., 2018), prairie and arctic environments (e.g., Fang and Pomeroy, 2009; Parr et al., 2020), and in the context of avalanches (e.g., Schweizer et al., 2003) have shown that snowdrifts can strongly influence the spatial water balance. These studies have also revealed that Equator-facing slopes might only receive half as much SWI as snowmelt compared with snowdrift areas (Flerchinger and Cooley, 2000; Marshall et al., 2019). In turn, water originating from snowdrifts can locally control groundwater level fluctuations (Flerchinger et al., 1992) and contribute to streamflow into the summer season (Chauvin et al., 2011; Hartman et al., 1999; Marks et al., 2002). The relative importance of spatial snowmelt patterns is expected to increase with snowmelt magnitude, which is sensitive to snowfall fractions. Hence, quantifying spatial snowmelt patterns in areas that are not seasonally snow covered, as well as determining the importance of snowdrifts for streamflow generation in these areas, could be an important step in clarifying how stream discharge is affected by snowfall fractions.
One area where the snowfall fraction varies substantially from year to year
is the rain–snow transition zone. The rain–snow transition zone is an
elevation band in which the dominant phase of winter precipitation shifts
between snow and rain (Nayak et al., 2010), and it is often characterized by a transient snowpack in (at least) parts of the defined area. Multiple studies in the European Alps and the northwestern US have shown that snowfall fractions in the rain–snow transition zone are particularly vulnerable to increases in temperature associated with climate change (e.g., Stewart, 2009). For example, towards the end of the century the snowfall fraction in the Swiss Alps is projected to decrease by about 50 % at
In addition to the expected decrease in snowfall fractions with climate
change, annual climate variations are expected to increase almost everywhere
across the planet (Seager et al., 2012), affecting annual
runoff efficiency (Hedrick
et al., 2020) and likely also influencing stream discharge timing and
magnitude. One well-documented discharge response is that years in which
catchments receive less snow have earlier snow-driven discharge peaks (McCabe and Clark, 2005; Stewart
et al., 2005). This is relevant because earlier spring snowmelt has been
linked to an increased risk of wildfire for catchments across the western US (Westerling et al., 2006) as well as to earlier
and lower low-flows in the late-summer and fall months (Kormos et al., 2016). In some catchments and
years, portions of the stream network might also completely cease to flow,
and this drying can alter the network's ecological and biogeochemical
functioning (Datry et al., 2014). In mid-elevation
rain–snow transition zones, the annual snowpack variability is already
relatively large. For example, in the Reynolds Creek Experimental Watershed
(RCEW; Idaho, USA) the coefficient of variation (CV) of peak snow-water
equivalent (SWE) between 1964 and 2006 ranged from 0.28 to 0.37 for five
high-elevation stations (2056–2162 m) and was 0.72 for a mid-elevation
weather station at the rain–snow transition zone
(1743 m; Nayak et al., 2010). This mid-elevation
variability suggests that year-to-year differences in snowfall at the
rain–snow transition zone might already be substantial compared with nearby
catchments at higher elevations, and it allows for the investigation of
catchment hydrologic responses to snowfall variations using a relatively
short data record, especially at sites where precipitation inputs are likely
to be strongly reflected in stream discharge (i.e., with a limited memory of
past inputs). Using observations of hydro-climatically different years
(e.g., rainy vs. snowy) could reveal how discharge and stream drying at the
rain–snow transition zone has responded to past variations in water inputs,
thereby providing insight into how other small (
Thus, the overarching goal of this work is to improve our understanding of
discharge responses to year-to-year variations in precipitation phase and
magnitude. We do this at the rain–snow transition zone – a region that
experiences large year-to-year variations in snowfall fractions, covers a
significant part of the land surface, and might extend to higher elevations
due to climate change. Specifically, we address the following research
questions:
How does the spatial and temporal distribution of SWI at the rain–snow
transition zone vary between particularly wet, dry, rainy, or snowy years? How does stream discharge timing and amount respond to SWI in wet, dry,
rainy, or snowy years? Are variations in stream discharge related to variations in yearly
snowfall fractions?
Examining natural variation in snowfall fractions in the rain–snow
transition zone contrasts with other studies on snow-related processes that
focus on seasonally snow-covered catchments. While many studies of snowmelt
runoff examine seasonal responses at the landscape scale, here we focus on
hourly responses at a fine spatial resolution. This allows us to investigate
the spatial distribution of the snowpack and snowmelt as well as the phase
of precipitation and the temporal distribution of SWI. Furthermore, while
SWE is frequently used as a summarizing variable for winter precipitation
when comparing precipitation to stream discharge, SWI is more directly
related to the timing and amount of water resources; therefore SWI might be
an important variable to model in future work addressing similar questions.
Our study location is Johnston Draw, a 1.8 km
We used hourly hydrometeorological data recorded at 11 weather stations
throughout the catchment (Fig. 1; Godsey et al., 2018a). The stations are
placed at 50 m elevation intervals on the north- and south-facing slopes, and they span a
Maps of the location of
To characterize the spatial distribution of snow depth, a 1 m resolution
snow depth product was calculated as the difference between a snow-off lidar
flight (10–18 November 2007; Shrestha and Glenn, 2016) and a
snow-on lidar flight (18 March 2009, around the time of peak accumulation),
hereafter referred to as lidar snow depth. Typical vertical accuracies for
lidar surveys are
Because only one lidar observation was available near peak snow accumulation, we also characterized snow presence throughout the season by mapping the snow-covered area (SCA) using satellite-derived surface reflectance at a 3 m resolution, which is available starting in 2016 (four-band PlanetScope scene; Planet Application Program Interface, 2017). This high-resolution imagery was critical for our analysis because snowdrifts in the rain–snow transition are relatively small in extent. No high-resolution satellite imagery was available for years that exhibited the key characteristics that we sought to study (e.g., rainy, snowy, wet, or dry; see Sect. 3.3), so we focused on a recent snow-covered period for which streamflow data and Planet imagery were available (1 November 2018 until 31 May 2019) to assess snow coverage. This targeted year was warmer than the year for which the lidar observations were available (mean annual air temperature of 8.0
We also used the surface reflectance imagery to determine the melt-out date
of the snowpack for all years in which satellite and discharge observations
were available (2016–2019). This was done by manually reviewing all
available images and visually determining when all snow had melted. Given
the high visiting frequency and limited cloudiness in early summer, we
estimate that an error of
We used the Automated Water Supply Model (AWSM; Havens et al., 2020) to obtain spatially continuous estimates of the distribution and phase of precipitation, snowpack characteristics, and surface water inputs (SWI). The two major components of the AWSM are the Spatial Modeling for Resources Framework (SMRF; Havens et al., 2017) and iSnobal (Marks et al., 1999). SMRF was used to spatially distribute precipitation and all other weather variables (air temperature, solar radiation, vapor pressure, precipitation, wind speed, and wind direction) along an elevation gradient using the hourly measurements from the weather stations. We included precipitation measurements from two stations within the basin (jd125 and jd124b) and two stations outside of the basin (jd144 and jd153; Fig. 1) to capture the elevation gradient. Precipitation at the wind-exposed site, jd124, was excluded because of precipitation undercatch issues. The interpolated vapor pressure and temperature fields were then used within SMRF to calculate the dew point and to further distinguish which fraction of precipitation fell as rain and/or snow. The output from SMRF was then used to force iSnobal, a physically based, two-layer snowpack model that accounts for precipitation advection from rain and snow (Marks et al., 1999).
The model was run at a 10 m resolution for 5 water years, namely 2005, 2009, 2010, 2011, and 2014. We selected 2009 because the snow depth lidar survey was available in this year, and we chose 2005, 2010, 2011, and 2014 because they are hydroclimatically different. The year 2005 was rainy (snowfall fraction was 63 % of the 2004–2014 mean and 23 % of 2005 total precipitation), whereas 2010 was snowy (snowfall fraction was 155 % of 2004–2014 mean and 57 % of 2010 total precipitation). The year 2014 was dry (precipitation was 86 % of the 2004–2014 mean) and 2011 was wet (precipitation was 132 % of the 2004–2014 mean) (Tables 1 and S3 and Fig. S4 in the Supplement) with snowfall fractions of 41 % and 30 % of the total precipitation for each year, respectively. The work was limited to 4 years because we aimed to focus on differences in the distribution of SWI and stream discharge for years that had different snowfall fractions and total precipitation magnitudes. Therefore, these strongly contrasting years were selected from 11 potential years of record (Godsey et al., 2018a). More stations were deployed towards the end of the 2004–2014 period, yielding additional observations to force the model with meteorological inputs and validate the model output of snow thickness; therefore, we selected later years within this period if conditions were similar. We focus on the 4 hydroclimatically distinct years in Sects. 4 and 5 of this paper, but we evaluate the model performance for all 5 years.
Precipitation, discharge, and SWI characteristics for each water
year including total precipitation (average of precipitation measured
at jd124b and jd125); the fraction of precipitation falling as snow (snowfall fraction); date and DOWY (day of water year) of the start of the snowy season (snow
In order to represent the spatial variability in snowfall and the effects of wind redistribution of snow, we use the precipitation rescaling approach proposed by Vögeli et al. (2016) that implicitly captures the spatial heterogeneity induced by these processes using distributed snow depth information (e.g., from lidar or structure from motion (SfM)). This methodology can be used to rescale the precipitation falling as snow to reproduce the observed snow distribution patterns while conserving the initial precipitation mass estimation. Given the intra- and interannual consistency of spatial patterns of snow distribution (Pflug and Lundquist, 2020; Schirmer et al., 2011; Sturm and Wagener, 2010), Trujillo et al. (2019) extended the original implementation to utilize historical snow distribution information to other years in the iSnobal model. Following these successful implementations, we used the spatial distribution of snow depth from the 2009 survey around peak snow accumulation to inform the snowfall rescaling to all years in the study period. Although using the 2009 survey to rescale snowfall in other years might have induced some uncertainty, verification of the interannual consistency in the snow distribution in this catchment by comparing the lidar snow depth and the satellite imagery indicated that this uncertainty is likely to be small.
One of the model outputs from iSnobal is surface water input (SWI), which represents snowmelt from the bottom of the snowpack, rain on snow-free ground, or rain percolating through the snowpack. Rainfall is directly counted as SWI when it falls over snow-free ground, and it is included in the energy and water balances when it falls onto the snowpack. To calculate snowmelt, iSnobal solves each component of the energy balance equation for each model time step using the best available estimations of forcing inputs. Melt occurs in a pixel when the accumulated input energy is greater than the energy deficit (i.e., cold content) of the snowpack. If the accumulated energy input is smaller than the energy deficit, the sum of current hour melt and previous hour liquid water content will be carried over into the next hour. If that hour's input energy conditions are negative, the liquid mass is refrozen into the column. Sublimation and evaporation of liquid water from the snow surface and condensation of liquid water onto the snow surface is computed as a model output term, although these quantities were not considered here. Canopy interception must be handled a priori when developing the model forcing input, and it was also not considered here. Although not accounting for the latter introduces some uncertainty, we expect this to be small with the shrub and grass vegetation types in Johnston Draw catchment. Lastly, iSnobal is limited to snow processes only, which means that SWI “exits” the modeling domain. In reality, SWI travels to the stream as surface or subsurface runoff, could be stored in the soil until it evaporates or is transpired, or could recharge deeper groundwater storages. The route that SWI takes depends on the overall catchment wetness as well as the local energy balance (e.g., incoming radiation) and vegetation activity. In this paper, we computed SWI for each pixel and time step and assumed that all SWI generated in simulated snow-free pixels was rain and that all SWI generated in simulated snow-covered pixels was snowmelt.
Model results were evaluated in two ways: first, the simulated snow depths
were compared to lidar snow depths covering the entire basin on 18 March 2009; second, the temporal variation of the simulated snow depths were
compared to snow depths measured at each of the weather stations for all
simulated years. The latter comparison was done using model results from a
30 m
The phase and magnitude of precipitation and the magnitude and temporal distribution of SWI were compared to annual discharge and the stream dry-out date. The stream dry-out date is the day when the stream first ceased to flow at the catchment outlet. For comparisons across seasons, we defined winter as December, January, and February; spring as March, April, and May; summer as June, July, and August; and fall as September, October, and November. To compare SWI with the dry-out date, we also calculated how much SWI occurred during the water year before the stream dried. No delays were considered when comparing SWI to discharge (e.g., discharge as a fraction of SWI in January results from dividing discharge in January by SWI in January). Discharge metrics were also compared to the flashiness of SWI inputs, which was calculated as the sum of the difference in total SWI from day to day, divided by the sum of SWI (also known as the Richards–Baker flashiness index; Baker et al., 2004). Further metrics included the fraction of time that more than half of the catchment was snow-covered as well as the melt rate between the date on which there was 40 % snow coverage in the catchment and the date on which the catchment was snow-free. A threshold of 40 % snow coverage was chosen because this resulted in an approximately linear melt rate for all years.
The lidar snow depth ranged from 0 to 5.3 m on the date of acquisition (18
March 2009), which was near peak snow cover (median of 0.4 m and CV of 0.91; Fig. 2a). The south-facing slopes had little to no snow cover (mean of 0.3 m), whereas the north-facing slopes were covered with 0.7 m of snow on average. For the years studied here, during the approximate duration of the snowy season between 15 November and 15 April, the average snow depth for all
north-facing stations was more than 5 times that of the average snow depth at south-facing stations (0.20 vs. 0.04 m, respectively), and the snowpack lasted almost 90 d longer on average (132 vs. 43 d, respectively). Weather stations on north-facing slopes and at higher elevations generally had deeper snowpacks and were covered with snow for longer than sites on the south-facing slopes or at lower elevations (Godsey et al., 2018a). The snowpack distribution was also affected by wind-driven redistribution of snow. For instance, snow depths at jdt2 (north facing) and jdt3b (south facing) were consistently lower than at the weather stations directly below them in elevation (jdt1 and jdt2b, respectively). Large snowdrifts formed in some western parts of the watershed, up to a maximum depth of 5.3 m (90th percentile of all snow depths was 1.2 m; Fig. 2a). Wind-driven redistribution of the snow in Johnston Draw is facilitated by a relatively consistent southwesterly wind direction (average of 225
Simulated snow depths on the day of the lidar survey agreed well with the
lidar snow depth (
The median NSE for the hourly simulated snow depths compared to observations
at the weather stations ranged from 0.22 (wet 2011) to 0.86 (snowy 2010) for
all modeled years and weather stations, with the RMSE ranging from 0.008 to
0.097 m (Table 2; see Fig. S7 in the Supplement for time series of all simulated and observed snow depths). The RMSE was equal to or lower than 0.1 m for all years, with the year in which the NSE performance was lowest (wet 2011) having an RMSE of 0.046 m. The temporal variation of the snowpack at each of the weather stations was well captured by the model; the median NSE for the normalized snowpack depths (NSE
The Nash–Sutcliffe efficiency (NSE; Nash and Sutcliffe,
1970) and root-mean-square error (RMSE) for simulated and observed snow
depths at each weather station as well as the NSE for normalized
(
The spatial pattern of SWI was similar for all years, with the highest SWI
occurring in the snowdrifts (maximum SWI (SWI
Maps showing the yearly sum of surface water inputs (SWI, mm) for
Weekly sums of SWI ranged from 0 to
Weekly sums of surface water inputs (SWI, summation of rainfall and
snowmelt, green polygons, mm), rainfall (gray polygons, mm), and specific
discharge (black line graph, mm) for
Streamflow was least responsive to SWI at the beginning of each water year (Fig. 5). For instance, in 2005 and 2010, 174 and 108 mm of SWI occurred before 1 February (31 % and 20 % of annual SWI), whereas discharge amounted to only 7 % and 1 % of its yearly total during that same period. Similarly, 82 mm of SWI in October 2011 resulted in less than 1 mm discharge, whereas 180 mm of SWI in November–January led to 62 mm of discharge. SWI generally resulted in most discharge when SWI rates were high, such as during a 3 d rain-on-snow event in February 2014 (SWI of 75 mm and discharge of 29 mm) or during spring snowmelt in April 2011 (SWI of 108 mm and discharge of 102 mm). Such individual precipitation events had a strong influence on the annual runoff efficiency. For instance, 2014 had a slightly higher runoff efficiency (0.16) than 2005 (0.11) and 2009 (0.14), mostly due to the high runoff generation during one rain-on-snow event (29 mm, 36 % of yearly discharge).
Cumulative surface water inputs (SWI, dashed lines, mm) and discharge (colored polygons, mm) for each of the water years (dark green – rainy 2005, light blue – snowy 2010, dark blue – wet 2011, and light green – dry 2014). Circles indicate the day on which the stream ceased to flow at the catchment outlet (dry-out date; please note that the stream did not cease to flow in 2011), and diamonds indicate the day on which all snow had melted from the catchment (melt-out date).
Annual discharge was highest in 2011 (307 mm, 43 % of SWI) and lowest in
2005 (62 mm, 11 % of SWI). Despite similar SWI in 2005 and 2010
(SWI
Scatterplots of
Comparison of annual discharge and the stream dry-out date to metrics describing the phase and magnitude of precipitation, the temporal distribution of SWI, and key characteristics of the snowpack highlighted the importance of the magnitude and timing of SWI (Fig. 7). Significant relationships with annual discharge were found for annual precipitation (Fig. 6a) and the sums of precipitation and snowfall in spring (Figs. 7 and S9 in the Supplement). The dry-out date of the stream was significantly correlated with annual precipitation, the sum of winter and spring precipitation and spring snowfall, spring precipitation as a fraction of SWI, the melt-out date of the snowpack, and the sum of SWI before the dry-out date (Figs. 7 and S9 in the Supplement). No significant correlation was found between the annual, winter, and spring snowfall fraction and annual discharge and the stream dry-out date (Fig. 7).
Heat plot showing the Pearson correlation coefficients (
Snow drifting and aspect-driven differences in snow dynamics caused a strong
variability in the spatial pattern of the snowpack (Fig. 2a) and SWI (Fig. 3). We found that the spatial pattern in simulated SWI was similar across
all years, with snowdrifts receiving up to 7 times more SWI than the
catchment average (SWI
Snowdrifts delivered 4.2 % (2005) to 7.2 % (2010) of the basin-total
annual SWI on just
We found that the majority of SWI occurred in winter and spring and that
catchment-average SWI was more uniform in time in snowy 2010 than in the
other years (CV of daily SWI in 2010 was 1.7, whereas other years ranged from 2.14 to 2.65). The steadier water inputs in the snowmelt period might explain why annual discharge in snowy 2010 was double that of rainy 2005 despite similar total SWI. More stable water inputs from snowmelt rather than flashy water inputs from rain could have led to wetter soils and higher soil conductivity rates, allowing more water to pass through the subsurface towards the stream or towards deeper storage (Hammond et al., 2019). Previous work in the nearby Dry Creek Experimental Watershed (Idaho) showed that water stored in the soil dries out approximately 10 d after snowmelt (McNamara et al., 2005). For the years on record here, streamflow was sustained for a minimum of 59 d after the melt-out date (Table 1), even though SWI is generally low after June each year (Fig. 4). This underscores that it is likely that deeper flow paths contribute to the stream in early summer. This is also consistent with stream discharge being nearly unresponsive to SWI during the dry catchment conditions in the beginning of each water year (Fig. 5). During fall, subsurface water storage across the catchment is low; thus, any SWI during this period likely results in recharge or evaporation rather than stream discharge (Seyfried et al., 2021). Air temperature also
has a small effect on the runoff efficiency, particularly in the summer
season. The runoff efficiency, calculated as summer discharge divided by
summer precipitation for the 2004–2014 record, was significantly correlated
with summer air temperatures (
In contrast with our hypothesis and what has been suggested in the literature
(e.g., based on the comparison of 420 catchments in the continental US using
the Budyko framework; Berghuijs et al., 2014), neither annual discharge nor
the stream dry-out date were correlated with snowfall fraction (Figs. 6,
7). Instead, annual discharge and the stream dry-out date were more
correlated with total precipitation and the snowpack melt-out date. This
highlights the importance of the temporal distribution of SWI, which is not
captured in an annual snowfall fraction. The temporal distribution of SWI
might be less important for predicting stream discharge and cessation in
more humid catchments (in which precipitation is more evenly distributed over
the year and/or in which more precipitation events occur) or in larger
catchments, such as those considered in Berghuijs et al. (2014; range of
catchment areas between 67 and 10 329 km
Bilish et al. (2020) similarly found that streamflow was not correlated with the snowfall fraction for a small catchment with an ephemeral snowpack in the Australian Alps. They attributed this to the frequent occurrence of mid-winter snowmelt: the snowpack melted out several times each year, independent of the annual snowfall fraction, and, thus, did not store a significant amount of water. Field observations at Dry Creek, a nearby semiarid catchment that includes a rain-dominated and a snow-dominated area, also suggested that the snowfall fraction was not related to annual discharge for a small sub-catchment at the rain–snow transition zone (treeline sub-catchment, 0.015 km
Although the model adequately reproduced the spatial snowpack patterns and
dynamics (Fig. 3 and Table 2), temporal variations in the snow depths (i.e.,
melt and accumulation) recorded at the weather station locations were
simulated better than the absolute snow depths. To investigate why simulations of snow depths were poor for some stations and years, we
calculated the average and precipitation-weighted average wind directions,
wind speeds, and snow densities for all events during which the snowfall
fraction was higher than 0.2 (i.e., 20 %; see Table S12 and
Fig. S13 in the Supplement) from the station data. Although wind speed and directions were generally consistent (Fig. S13 in the Supplement), in 2011, the combination of higher snow densities (stronger cohesion of snow particles; 122 kg m
In addition to the uncertainty in the spatial redistribution of snow
depending on wind speeds, wind direction, and snow densities, we suggest
three additional reasons for the differences between simulated and observed
snow depths. First, the varying performance at jd125 might be related to
inaccuracies in calculating the phase of precipitation, which would most
strongly affect lower elevations at which the phase shifts more often from
rain to snow. Any uncertainty in the magnitude or phase of precipitation
would decrease model efficiency because precipitation was interpolated based
on elevation, after which the proportion of precipitation falling as snow
was redistributed based on the lidar snow depths (see Sect. 3.3). Second,
the simulated snow depths reflect all processes occurring in each 10 m grid
cell (our model resolution), whereas the ultrasonic snow depth measurements
represent processes at
Discrepancies between simulated and observed snow depths are challenging to solve, especially for areas with an ephemeral snow cover (Kormos et al., 2014) or with complex vegetation patterns, such as the sagebrush in Johnston Draw. Shallow snow covers are more sensitive to small variations in energy fluxes than deeper seasonal snow covers (Pomeroy et al., 2003; Williams et al., 2009). As a result, small errors in the spatial extrapolation of the forcing data or in the forcing data itself (e.g., uncertainty in the observed relative humidity or temperature) can introduce uncertainties in the model results (Kormos et al., 2014). For instance, the transition from snow-covered to snow-free areas results in a large change in albedo, which influences solar radiative fluxes. The snowpack at the rain–snow transition zone can melt out several times per year, even within a single day, and melt-out dates are variable across the catchment. Therefore, a small error in the simulated melt-out date for each cell can result in a larger error in the basin-average or yearly results. Perhaps these challenges are also a reason for the limited number of studies that have simulated warm snowpacks (Kormos et al., 2014; Kelleners et al., 2010), despite multiple regional studies highlighting that the rain–snow transition zone is expanding and that regional climates are changing rapidly (Klos et al., 2014; Nolin and Daly, 2006). Challenges linked to snow ephemerality likely also affected our results, but the agreement between the observed and simulated snow depths indicates that at least the general patterns of accumulation and melt in space and over time were represented by the simulations, at a scale that was small enough to characterize the snowdrifts.
Regardless of the challenges that come with studying an intermittent snow cover, the relationship between the snowpack melt-out date and stream dry-out date poses interesting opportunities to inform hydrological models or evaluate model results with independent observations. Measurements of SCA can be obtained through satellite imagery and are, thus, easier and cheaper to obtain than SWE or snow depth measurements (e.g., Elder et al., 1991). Satellite observations can be particularly helpful to investigate remote areas that exceed a feasible modeling domain, and they can be used to inform or evaluate models. Given the restrictions for satellite imagery imposed by clouds and the visit frequency, particularly for areas with an ephemeral snow cover that might melt out in a single day, a combination of satellite imagery and snowpack modeling seems a promising way to leverage these observations while ensuring the fine temporal resolution that might be needed to study stream cessation.
As a result of climate change, the rain–snow transition zone will receive more rain and less snow, which may influence the spatial and temporal distribution of surface water inputs (SWI, summation of rainfall and snowmelt). The goal of this work was to quantify the spatial and temporal distribution of SWI at the rain–snow transition zone as well as to assess the sensitivity of annual stream discharge and stream cessation to the temporal distribution of SWI and to the annual snowfall fraction. To this end, we used a spatially distributed snowpack model to simulate SWI during 5 years, of which 4 had contrasting climatological conditions. We found that the spatial pattern of SWI was similar between years and that snow drifting and aspect-controlled processes caused large differences in SWI across the watershed. Snowdrifts received up to 6 times more SWI than other sites, and the difference between SWI from the snowdrifts and catchment average SWI was highest for the year with the highest snowfall fraction. This highlights that the snowfall fraction affects the spatial variability in SWI, with more rain leading to less variability. The majority of SWI occurred in winter or spring, which was also the time that the percentage of SWI becoming streamflow was highest (up to 94 % in April 2011). Over the 2004–2014 data record, annual discharge was insensitive to snowfall fraction and depended more on total and spring precipitation. The stream dry-out date was also sensitive to total and spring precipitation. In addition, stream cessation was positively correlated with the last day on which there was snow present anywhere in the catchment, which indicates that the persistence of snowdrifts in small parts of the catchment is critical for sustaining streamflow. This study highlights the heterogeneity of SWI at the rain–snow transition zone, its impact on stream discharge, and, thus, the need to spatially and temporally represent SWI in headwater-scale studies that simulate streamflow.
The hydrometeorological and discharge data used in this paper are available at
The supplement related to this article is available online at:
LK and SEG developed the concept for the study. LK, SH, ET, AH, and KH performed and/or contributed to the simulations. LK prepared the first draft of the manuscript. All co-authors provided recommendations for the data analysis, participated in discussions about the results, and edited the paper.
The contact author has declared that neither they nor their co-authors have any competing interests.
Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
USDA is an equal opportunity provider and employer.
This research has been supported by the Swiss National Science Foundation (grant no. P2ZHP2_191376) and the US National Science Foundation (grant no. EAR-1653998). The Reynolds Creek Critical Zone Observatory Cooperative Agreement (grant no. EAR-1331872) provided support for processing the snow depth data.
This paper was edited by Markus Hrachowitz and reviewed by two anonymous referees.