Streamflow generation and deep groundwater recharge may be vulnerable to loss of snow, making it important to quantify how snowmelt is partitioned between soil storage, deep drainage, evapotranspiration, and runoff. Based on previous findings, we hypothesize that snowmelt produces greater streamflow and deep drainage than rainfall and that this effect is greatest in dry climates. To test this hypothesis we examine how snowmelt and rainfall partitioning vary with climate and soil properties using a physically based variably saturated subsurface flow model, HYDRUS-1D. We developed model experiments using observed climate from mountain regions and artificial climate inputs that convert all precipitation to rain, and then evaluated how climate variability affects partitioning in soils with different hydraulic properties and depths. Results indicate that event-scale runoff is higher for snowmelt than for rainfall due to higher antecedent moisture and input rates in both wet and dry climates. Annual runoff also increases with snowmelt fraction, whereas deep drainage is not correlated with snowmelt fraction. Deep drainage is less affected by changes from snowmelt to rainfall because it is controlled by deep soil moisture changes over longer timescales. Soil texture modifies daily wetting and drying patterns but has limited effect on annual water budget partitioning, whereas increases in soil depth lead to lower runoff and greater deep drainage. Overall these results indicate that runoff may be substantially reduced with seasonal snowpack decline in all climates, whereas the effects of snowpack decline on deep drainage are less consistent. These mechanisms help explain recent observations of streamflow sensitivity to changing snowpack and highlight the importance of developing strategies to plan for changes in water budgets in areas most at risk for shifts from snow to rain.
Snowmelt is the dominant source of streamflow generation and groundwater recharge in many high-elevation and high-latitude locations (Regonda et al., 2005; Stewart et al., 2005; Earman et al., 2006; Clow, 2010; Jefferson, 2011; Furey et al., 2012). Soils modulate the partitioning of snowmelt into subsurface storage, deep drainage, evaporative losses, and surface runoff. Snow persistence, the fraction of time with snow cover, shows declines around the globe (Hammond et al., 2018b), and these snow losses may lead to changes in water input magnitude and timing (Harpold et al., 2015; Musselman et al., 2017; Harpold and Brooks, 2018). As areas of “at risk snow” become more apparent (Nolin and Daly, 2006), there is an urgent need for mechanistic studies that quantify the partitioning of snowmelt in the critical zone among vapor losses, surface flow, and subsurface flow and storage (Brooks et al., 2015; Meixner et al., 2016).
Changes in precipitation phase from snow to rain can modify hydrological partitioning by altering the timing and rate of inputs. Daily snowmelt rates typically do not reach the extreme intensities of rainfall (Yan et al., 2018), meaning that most areas (i.e., the Cascades) are predicted to receive more intense water inputs with more winter rainfall, whereas some other areas (i.e., Southern Rockies) will likely experience a decline in input intensity with snow loss (Harpold and Kohler, 2017). Warmer areas like the maritime western US may experience near complete loss of snowpack as snow fully transitions to rain by the end of the 21st century (Klos et al., 2014). Unlike rainfall, which is typically episodic, snow can accumulate over time and melt as a concentrated pulse of soil water input (Loik et al., 2004). This means that at 7 to 30 d scales snowmelt inputs are of greater magnitude than rainfall (Harpold and Kohler, 2017). Concentrated snowmelt can lead to a large proportion of runoff and deep drainage (Earman et al., 2006; Berghuijs et al., 2014; Li et al., 2017). With climate warming, future snowmelt rates may be reduced in many areas because earlier melt occurs when solar radiation is lower (Musselman et al., 2017). Along with warmer temperatures, increasing atmospheric humidity is leading to more frequent mid-winter melt events in humid regions yet increased snowpack sublimation and/or evaporation in dry regions (Harpold and Brooks, 2018). Given the considerable heterogeneity in climate, soils, topography, and vegetation across mountain ranges, the water budgets of different locations respond unevenly to a loss of snow.
Water inputs from rain or snowmelt during periods of low potential evapotranspiration and high antecedent moisture conditions are more likely to generate runoff and deep drainage (Molotch et al., 2009). Prior research has shown that near-surface soil moisture response is closely related to snow disappearance (Harpold and Molotch, 2015; Webb et al., 2015; Harpold et al., 2015), with strong links between snowmelt and soil moisture dynamics at multiple spatial and temporal scales (Loik et al. 2004; Williams et al. 2009; Blankinship et al. 2014; Kormos et al., 2014; Harpold and Molotch, 2015; Webb et al., 2015; Kampf et al., 2015). Earlier snow disappearance can lead to lower average soil moisture conditions not as conducive to streamflow generation as later snowmelt (Kampf et al., 2015; Harpold, 2016). The effects of earlier snowmelt on soil moisture dynamics may also vary with precipitation after snowmelt. Late-spring precipitation can overwrite the signal of earlier snowmelt timing on spring and summer soil moisture (Liator et al., 2008; Conner et al., 2016), whereas a lack of spring and summer precipitation can cause effects of earlier snowmelt on soil moisture to persist longer (Blankenship et al., 2014; Harpold, 2016). A transition to earlier, slower, and lesser snowmelt may increase overall evapotranspiration losses (Kim et al., 2016; Foster et al., 2015; Trujillo et al., 2012) while simultaneously decreasing the water use efficiency of conifer forests during snowmelt (Knowles et al., 2018). However, even at a well-studied location in Colorado the projected effects of shifts from snow to rain on tree water use and carbon uptake differ between modeling (Moore et al., 2008; Scott-Denton et al., 2003) and observational studies (Hu et al., 2010; Winchell et al., 2016).
Both surface runoff and deep drainage are affected by soil texture, soil depth, rooting depth (Cho and Olivera, 2009; Seyfried et al., 2005), and topography. These properties have limited variability over time spans of hydrologic analysis and can produce temporally stable spatial patterns of soil moisture, where some parts of the landscape are consistently wetter than others (Williams et al., 2009; Kaiser and McGlynn, 2018). Aspect modifies the snowpack energy balance, leading to higher sustained soil moisture content on north-facing slopes compared to south-facing slopes with the same input (in the Northern Hemisphere; Williams et al., 2009; Hinckley et al., 2014; Webb et al., 2015, 2018). Landscape evolution may lead to deeper profiles and more deeply weathered rock due to wetter conditions on north-facing slopes, making these slopes more conducive to deep drainage in some locations (Hinckley et al., 2014; Langston et al., 2015). Where soils are shallow, winter precipitation may exceed the soil storage capacity, leading to both runoff generation and deep drainage (Smith et al., 2011). Deeper soil profiles have greater storage capacity, which can sustain streamflow, even with snow loss; however, given consecutive years of low input these profiles will be depleted of storage, leading to lower flows (Markovich et al., 2016). Deeper soils can also help sustain transpiration during the late spring and summer, when shallow soils have dried (Foster et al., 2016; Jepsen et al., 2016). Streamflow can be insensitive to inputs under dry conditions, but respond rapidly once a threshold soil moisture storage value is exceeded (McNamara et al., 2005; Liu et al., 2008; Seyfried et al., 2009). McNamara et al. (2005) hypothesized that when dry-soil barriers are breached, there is sudden connection to upslope soils, leading to delivery of water to areas that were previously disconnected. In their semi-arid study area, such breaching of dry-soil barriers was only observed for periods of concentrated and sustained input from high-magnitude spring snowmelt. Together, the complex interactions of soil properties, antecedent conditions, water inputs, and evaporative demand make it challenging to determine how changes from snow to rain affect hydrologic response even in idealized settings.
Conceptual illustration of study hypotheses indicating the
importance of concentrated snowmelt input
The goal of this study is to improve our understanding of how changes in
precipitation phase from snow to rain affect hydrological partitioning in a
one-dimensional (1-D) representation of the critical zone. Partitioning of
precipitation input,
We hypothesize that reducing the fraction of precipitation falling as snow leads to lower runoff and deep drainage because it reduces the concentration of input in time (Fig. 1). Concentrated input during melt of a seasonal snowpack often saturates soils, causing saturation excess runoff and deep drainage below the root zone (Hunsaker et al., 2012; Kampf et al., 2015; Webb et al., 2015; Barnhart et al., 2016). Diffuse input over time reduces the likelihood of saturation because it allows more water redistribution and evapotranspiration between inputs. We also hypothesize that snowmelt is critical for runoff generation and deep drainage in dry climates and deep soils, where snowmelt is the dominant cause of soil saturation (McNamara et al., 2005; Tague and Peng, 2013), whereas the partitioning of rain and snowmelt may be more similar in wet climates and shallow soils, which are more frequently saturated by either rain or snowmelt inputs (Loik et al., 2004) (Fig. 1).
To evaluate soil moisture response to rainfall and snowmelt over a wide range of climate and soil conditions, we used HYDRUS-1D (Šimůnek et al., 1998), a physically based finite-element numerical model for simulating one-dimensional water movement in variably saturated, multi-layer, porous media.
We utilized daily input data from five United States Department of Agriculture Natural Resources Conservation Service (NRCS) snow telemetry (SNOTEL) sites in each of three regions that span a wide range of climate and snow conditions: the Cascades, Sierra Nevada, and Uinta mountains, for a total of 15 sites. Daily rather than hourly data were chosen to limit the effects of missing and incorrect values on the analyses. The three regions were chosen to represent dominant climate types in the western US, and within each region, sites were selected to span a snow persistence (SP) gradient, where SP is the mean annual fraction of time that an area is snow covered between 1 January and 3 July (Moore et al., 2015) (Fig. 2a, Table 1).
SNOTEL station properties including the start and end of data
records, site elevation, and mean annual climatic characteristics:
precipitation (
With each climate scenario we simulated vertical profiles of volumetric
water content (VWC), which were depth-integrated to compute soil moisture
storage (
Daily precipitation (
We chose three SNOTEL sites (432 Currant Creek, 698 Pole Creek R.S., 979 Van Wyck) to represent soil profile characteristics. While 365 of the 747 SNOTEL sites in the western US have soil moisture sensors, only a fraction of these sites have detailed soil profile data. The sites with soil profile data have information obtained from soil samples taken in the soil pits and processed in the NRCS Soil Survey Laboratory in Lincoln, NE, for texture, water retention properties, and hydraulic conductivity. We obtained detailed soil profile data in the form of pedon primary characterization files from the NRCS, and selected profiles (Fig. 2b, Table 2) that represent the range of soil textures and hydraulic conductivity values present at SNOTEL locations. Each had detailed NRCS pedon primary characterizations to depths greater than 100 cm and
Soil profile properties derived from NRCS pedon reports and the ROSETTA (Ros.) neural network. Columns are SNOTEL site, soil profile horizon, depth range of horizon, rock percent of sample volume, organic carbon percent of sample volume, sand percent of sample weight, silt percent of sample weight, clay percent of sample weight, Db
In HYDRUS-1D, we simulated water flow and root water uptake for a vertical
domain 10 m deep (Fig. 2b). The domain was discretized into 500 nodes with
higher node density near the surface (
We created soil layers with depths and textures taken from the NRCS soil
pedon measurements. From this information we applied the neural network
capability of HYDRUS-1D, which draws from the USDA ROSETTA pedotransfer
function model (Schaap et al., 2001), to determine soil hydraulic
parameters. Using the NRCS pedon primary characterizations we input percent
sand, silt and clay, bulk density, wilting point, and field capacity. The
neural network model then estimates soil hydraulic parameters based on these
inputs. Below the depth of the soil pedon measurements, we configured the
simulations to have a deep “bedrock” or regolith layer with lower
saturated hydraulic conductivity (
We applied climate scenarios from each of the 15 SNOTEL sites selected
(Table 1) to each of the soil profiles to examine how climate and soil type
affect partitioning. We then conducted additional experiments to modify
inputs using just the loam profile. First, to examine whether snowmelt and
rainfall are partitioned differently, we changed all precipitation to rain
and compared the simulation output to those with the original climate data.
Second, to examine the effects of input concentration, the temporal
clustering of input through time, we artificially produced intermittent
input (four 5-day periods of low magnitude) and concentrated input (one
20-day period of high magnitude) of the same annual total for one wet (559) and one dry (375) site using the loam profile (1056) for all years of data. Third, to examine how soil depth affects partitioning, we altered the
depth of rooting zones to 0.5, 1.5, and 2 times their original depth. For 0.5 depth scenarios, we replaced soil layers deeper than 0.5 times the original depth with the bedrock/regolith layer. For
Using the simulation results, we examined how rain and snowmelt were
partitioned into soil storage (
To analyze hydrologic partitioning at event timescale we defined rainfall
events as days with precipitation while SWE equaled zero and snowmelt events
for days with declining SWE and no simultaneous precipitation. To focus on
differences between rainfall and snowmelt, only events with entirely
rainfall or entirely snowmelt input were considered in this analysis; mixed
events were excluded, though mixed input accounts for an average of 47 %
of annual input across all sites and years (Table S6). Events could last as
long as the conditions were continuously satisfied, and only those followed
by at least 5 days of no input were used in analysis. Total depths of
each variable were computed for each defined event time period. Input rain
and snowmelt were summed over the event time period, and response variables
(
At the annual scale, soil water input and partitioning components (rain,
snowmelt,
To characterize climate conditions at the mean annual scale, each site was
classified as dry (precipitation deficit, PET
Using both the event and annual results, we examined (1) whether partitioning of rainfall input differed from that of snowmelt input, and (2) how partitioning was affected by climate, soil texture, and soil depth. For question 1, we tested for differences in event partitioning between input type (rain or snowmelt) and differences in annual partitioning between historical and all rain scenarios using ANOVA. For question 2, we tested for differences in annual partitioning between climate (wet, dry) and soil depth groupings, also using ANOVA. Additionally, for question 2 we tested the pairwise difference in linear regression slopes using the regression with interaction test in JMP (SAS-based statistical software) to determine whether the rate of change between explanatory and response variables differed by climate or soil depth grouping. By comparing the slopes of regressions run on standardized data, it is possible to assess the influence of independent variables on dependent variables in different groupings. In this study, we use this test to assess the influence of snowmelt fraction of input and input concentration index on runoff and deep drainage response for all, wet, and dry groupings as well as soil texture groupings.
Simulations for each of the 15 climate scenarios exhibit substantial
variability at the annual scale in precipitation (
Our first research question asks whether snowmelt and rainfall are
partitioned differently. At the event scale, input rates are significantly
greater on average for snowmelt than for rainfall in each of the three
regions and for the full dataset (ANOVA
Boxplots of event input rate (cm d
At the annual scale, input at all sites is a mixture of rain and snowmelt.
To examine the importance of snow to partitioning, we used snowmelt
fraction, defined as the fraction of snowmelt to total precipitation, and
ICI. Snowmelt fraction and snow persistence are
generally positively correlated with ICI at dry sites in the Uinta and
Sierra, but this correlation declines with wetter sites in the Cascades
(Fig. S7).
Ratio of runoff (
We then compared the hypothetical scenarios where we treated all
precipitation as rain to snow-dominated historical scenarios. All rain leads
to significantly lower
Boxplots of
Mean values of unitless response variables
Another effect of snow loss can be a decrease in input concentration.
Experimental scenarios with constant
Soil stores water that may later be partitioned into
When these same relationships are separated by soil texture rather than
wet/dry climate (Fig. 7b, Table S5), the response patterns are similar
between soil types except for the sandy loam profile, which displays higher
To assess the influence of soil profile depths on partitioning, we altered
the loam soil profile to be
Figure 8b displays daily time series of surface runoff, deep saturation,
deep drainage, and cumulative deep drainage during an example period for the
four different soil root zone depth scenarios. The shallowest rooting zone
of
The initial hypotheses for this study were that runoff and deep drainage would be greater from snowmelt than rainfall and that snowmelt is more important to generating runoff and deep drainage in deep soils and dry climates than in shallow soils and wet climates. Our results indicate that snowmelt is an efficient runoff generator, though not necessarily an efficient generator of deep drainage. Deep drainage is less affected by input type because it is controlled by deep soil moisture patterns over longer timescales. Soil texture modifies daily wetting and drying patterns but has limited overall effects on annual partitioning, whereas increases in soil depth decrease runoff and increase deep drainage. Overall these results indicate that runoff may be substantially reduced with seasonal snowpack decline in all climates, whereas the effects of snowpack decline on deep drainage are less consistent. We expand on these key findings in the paragraphs below and suggest that areas in dry watersheds with storage similar to peak SWE may be most likely to experience reductions in deep drainage with continued slow loss.
Multiple lines of evidence confirm snowmelt as a more efficient runoff generator on average than rainfall. At event scale runoff efficiency was elevated for snowmelt because of the 22 % greater input rate and 17 % wetter soils than rainfall. This is consistent with previous studies showing that snowpack development and subsequent melt tend to occur when soils are at elevated moisture contents due to lower ET (Liu et al., 2008; Williams et al., 2009; Bales et al., 2011). The effects of snowmelt vs. rainfall are weaker at annual timescales (Fig. 5, Table S3) because these longer time periods include a combination of snow, mixed, and rainfall inputs in contrast to the event analysis in which we analyzed only events that were exclusively snowmelt- or rainfall-dominated. Forcing all input into the extreme case of all rain produces 67 % declines in runoff efficiency (dry: 0.13 vs. 0.04; wet: 0.46 vs. 0.29) (Table 3, Fig. 6), likely because the input becomes less concentrated in time for the all-rain scenario, allowing more ET. We also hypothesized that the effects of changing snowpacks would be greatest in dry climates, where soil saturation is less frequent. However, simulations suggest that both wet and dry climates are as likely to show reduced surface runoff with declining snow water inputs.
The effects of snow loss on
Soil texture and depth generally did not change partitioning at the annual
timescale as much as the varying climate scenarios (Fig. 6), with the
exception of changes in the shallowest soils (
Focusing on the influence of soil texture, simulations indicate that shorter
durations of deep drainage for the coarser sandy loam compared to the finer
texture soils are offset by higher rates of flux during deep drainage in the
coarser profile (Fig. 8a). Similarly, lower likelihood of surface saturation in the sandy loam soil compared to other soils is offset by greater likelihood of infiltration excess runoff. Therefore, in this 1-D approach, soil depth exerts a stronger control on annual total input partitioning to
Given the complex nature of soil water movement in heterogeneous mountain
topography, this study makes several assumptions and simplifications. The
simulations do not include the intricacies of vegetation water use, assuming
a static leaf area index (LAI) with root uptake controlled only by PET and
soil moisture, and we assume free drainage from the bottom boundary of the
modeled domain. Changing static LAI has a substantial effect on soil
moisture dynamics (Chen et al., 2014), though model performance to match
simulated and observed soil moisture does not necessarily improve with the
assimilation of dynamic LAI values (Pauwels et al., 2007). Incorporating
site-specific constant LAI from field measurements or remotely sensed data
may have improved model performance, especially during spring green-up and
fall senescence, and is recommended for future site-specific studies. The
water balance in hydrologic models can be highly sensitive to the method
chosen to represent root uptake and plant water use (Gerten et al., 2004),
and hydrologic models generally poorly capture or replicate the interactions
between soil, vegetation, and atmospheric properties that combine to control
plant water use (Gómez-Plaza et al., 2001; Gerten et al., 2004; Zeng et
al., 2005). In addition, we did not allow for frozen soils in our
simulations, but this can be a strong influence on soil input partitioning
in places where snow depth was
Additionally, simulations are generally wetter than measured water contents;
therefore, the representation of partitioning shown here displays relative
response between climates and soil profiles rather than absolute
quantification of these partitioned components. The profile depths we
simulated represent the minimum likely soil depth, as the collection of the
pedon reports was limited by the depth of refusal for sample collection.
Shallow soil profiles can also lead to a wet bias in simulations, and this
can artificially elevate saturation excess flow, leading to our observations
of greater
Sub-daily dynamics in snowmelt and ET are not captured by our use of a daily time step. We chose to model soil water response to rainfall and snowmelt at the daily time step due to better data quality, but processes affecting partitioning of these inputs take place at sub-daily scales. Comparisons of results from simulations using daily vs. hourly input demonstrate similar timing of response but greater cumulative surface runoff from hourly simulations and greater cumulative deep drainage from daily simulations (Table S2). The short hourly time period allows for higher intensity input, which causes infiltration excess overland flow, whereas daily input is of lower intensity, allowing for greater deep percolation. Additionally, SNOTEL sites do not have measured values of PET, so we relied on a modeled 4 km gridded product (Abatzoglou, 2013), which may better represent some sites than others. It was beyond the scope of this study to perform a sensitivity analysis of PET data source.
Hydrologic response in hillslopes and catchments is not fully captured in
the 1-D modeling approach. Water partitioned into
The simulations used here only allow for matrix flow, excluding macropore
flow, for a simplified representation of soil water movement. Preferential
flow though the profile can enhance deep drainage relative to surface
runoff, which is another potential reason why soil moisture simulations were
biased wet; 60 %–80 % of deep drainage has been shown to occur as
preferential rather than interstitial flow (Wood et al., 1997; Jaynes et
al., 2001; Sukhija et al., 2003), although the amount of preferential flow
varies substantially between climates and soils. The magnitudes of fluxes in
our simulations are consistent with observation studies, however, lending
more confidence to the simplified modeling approach. Simulated annual
Future work could examine the potential sensitivity of the results to these limiting assumptions, In particular, researchers could examine the extent to which adding spatially and temporally varying vegetation processes and/or preferential flow pathways would change water balance partitioning. Simulations could expand to two dimensions to examine how downslope affects partitioning from ridgelines to valley bottoms or to three dimensions to examine effects of flow convergence and exfiltration in hillslope hollows. Because of the complexity of subsurface properties, this work would also benefit from more information about the hydraulic properties of the deep subsurface below the measured soil pedons. This could be linked with model analyses examining how both subsurface properties and boundary conditions affect the simulations.
This study helps to explain the mechanisms that lead to greater runoff from snowmelt. At event scale snowmelt generates more runoff because it tends to have a greater input rate and occurs on wetter soils than rainfall. Seasonal snowmelt elevates runoff in both wet and dry climates. Deep drainage can also decline with loss of snow, but it has a weaker response because soil storage buffers the impacts of snow loss. Soil properties can mediate the effects of snowmelt on rainfall changes, with soil depth having a greater effect than texture on input partitioning, particularly where soil water storage is less than mean annual precipitation. Soils that are shallower than observed soil depths generate the greatest runoff and deep drainage, indicating that shallow soils may show the largest changes in partitioning as input transitions from snowmelt to rainfall. Increasing soil depth above observed depths gradually reduces surface runoff while increasing deep drainage. Soil texture modifies short-term timing of soil moisture and runoff generation, but these effects are not large enough to alter the annual response of different soil types to changes in snow. The 1-D simulations provide basic hypotheses for hydrologic partitioning under changing snowmelt that should be further explored with 2-D or 3-D hydrological models and direct observations. Although more work is necessary to translate these findings to watershed-scale streamflow response, the findings highlight the importance of precipitation-phase shifts for runoff generation and groundwater recharge.
Underlying model output data along with description on the CUASHI hydroshare repository are available at:
The supplement related to this article is available online at:
JH, AH, and SK designed the experiments and JH and SW carried them out. JH and SW performed the simulations. JH conducted statistical analyses on model outputs. JH prepared the manuscript with contributions from all the co-authors.
The authors declare that they have no conflict of interest.
This research has been supported by the National Science Foundation, Division of Earth Sciences (grant no. 1446870).
This paper was edited by Nunzio Romano and reviewed by Bettina Schaefli and one anonymous referee.