Lower forest density enhances snow retention in regions with warmer winters: A global framework developed from plot-scale observations and modeling



[1] Many regions of the world are dependent on snow cover for frost protection and summer water supplies. These same regions are predominantly forested, with forests highly vulnerable to change. Here we combine a meta-analysis of observational studies across the globe with modeling to show that in regions with average December-January-February (DJF) temperatures greater than −1°C, forest cover reduces snow duration by 1–2 weeks compared to adjacent open areas. This occurs because the dominant effect of forest cover shifts from slowing snowmelt by shading the snow and blocking the wind to accelerating snowmelt from increasing longwave radiation. In many locations, midwinter melt removes forest snow before solar radiation is great enough for forest shading to matter, and with warming temperatures, midwinter melt is likely to become more widespread. This temperature-effect in forest-snow-climate interactions must be considered in representations of the combined ecohydrological system and can be used advantageously in forest management strategies.

1. Introduction

[2] Many independent studies have assessed the effects of trees on snow and the potential for using silviculture as a tool in water management [Kittredge, 1953; Golding and Swanson, 1978; Varhola et al., 2010a]. Forests intercept snow and emit longwave radiation but also shelter snow from wind and solar radiation compared to open sites [Varhola et al., 2010a]. Thus, decreased snow accumulation under trees is generally offset by decreased melt rates [Varhola et al., 2010a], but different processes dominate in different climatic and topographic settings [e.g., Ellis et al., 2011; Strasser et al., 2011]. The variable net effects of forest on snowpack have led to apparently conflicting conclusions regarding whether lower-density forests (from cutting, fire, or natural disturbance) result in longer or shorter-lasting snow [Kittredge, 1953; Golding and Swanson, 1978].

[3] Conventional wisdom is that as forest density decreases, interception decreases (greater accumulation) and melt rates increase (faster snow disappearance) [Varhola et al., 2010a]. Cases contrary to this, where greater melt rates occur under a denser forest, have been termed a “radiative paradox” [Ambach, 1974; Sicart et al., 2004; Lawler and Link, 2011], where increases in longwave radiation with increasing canopy cover are greater than corresponding decreases in shortwave radiation. These conditions are most likely to occur in conjunction with low atmospheric emissivities, high snow albedos, and low solar elevations [Sicart et al., 2004], and are more common on shaded, poleward facing slopes [Ellis et al., 2011]. Here we synthesize 21 plot-scale field studies to show that, despite small-scale complexities, the climate regime is a first-order control on determining how forests impact snow. Mean winter air temperatures can predict where less dense forests will contribute to enhanced snow retention.

[4] We first present the details of the site-specific studies, global climate data, and net radiation model that went into our synthesis (section 2). Next, we present our general results and conduct model sensitivity tests (section 3). Finally, we discuss our results in the context of the relevant literature with regards to site-specific observed melt rate patterns, interception, precipitation timing, and winds (section 4).

2. Methods

2.1. Meta-Analysis of Existing Studies

[5] We analyzed 21 studies with paired observations of snow under a forest canopy and in an adjacent open area to determine whether snow persisted longer in the forest or longer in the open (Table 1 and Figure 1). We focused on the difference in snow disappearance date (SDD) between a paired forest and clearing, as a metric of the net effect of the forest on snow accumulation and ablation. We took SDD as the first day with no snow, expressed as day of year. Our synthesis and analysis of previous investigations included peer-reviewed studies that reported SDD for both a forest site and a reference open site, as well as our own data (detailed in section 2.4). Forest sites ranged in canopy density, tree species, and age, and open sites ranged in the size of the open area (i.e., from a gap with the diameter of one tree height to a clear-cut), and in the presence or absence of understory or sparse overstory vegetation. Specific site characteristics are summarized in Table S1 in the supporting information.

Table 1. General Latitude, Longitude, and Climatology
No.Investigation (Reference)Longitude (DD)Latitude (DD)DJF Tempa (°C)DJF Precipa (mm)DJF Windb (m s−1)
  1. a

    From Matsuura and Willmott [2009].

  2. b

    From NASA [2007].

1Umpqua, OR [Storck, 2000]−
2Alptal, Switzerland [López-Moreno and Stähli, 2008]8.747.11.52614.1
3Alptal, Switzerland [Rutter et al., 2009]8.747.11.52614.1
4BERMS, Canada [Rutter et al., 2009]−104.753.9−16.8623.0
5Fraser, CO [Rutter et al., 2009]−105.939.9−11.3545.9
6Hitsujigaoka, Japan [Rutter et al., 2009]141.443.0−5.92496.9
7Hyytiala, Finland [Rutter et al., 2009]24.361.9−7.61093.6
8Siuntio, Finland [Koivusalo and Kokkonen, 2002]26.060.1−6.01413.1
9Mayson Lake, BC [Winkler et al., 2005]−120.451.2−8.61583.8
10Niigata, Japan [Whitaker and Sugiyama, 2005]139.638.3−0.26806.4
11Sequoia, CA [Bales et al., 2011]−119.237.1−1.03854.1
12Valles Caldera, NM [Musselman et al., 2008]−106.535.91.4803.6
13Pyrenees, Spain [López-Moreno and Latron, 2008]1432.22183.8
14UPC, BC [Thyer et al., 2004]−119.449.6−7.61763.6
15Crowsnest Pass, Alberta [Burles and Boon, 2011]−114.549.6−11.32154.0
16Fraser Lake, BC [Boon, 2009]−124.953.7−9.71263.8
17Valdai, NW Russia [Gelfan et al., 2004]33.157.6−7.51413.0
18Marmot Creek, Alberta [Ellis et al., 2011]−115.251.0−11.21064.1
19North Wasatch Mountains, Utah [LaMalfa and Ryle, 2008]−111.541.4−2.62133.8
20Turkey Lakes Watershed (TLW), central Ontario [Murray and Buttle, 2003]−84.447.1−10.83313.9
21Cedar River Watershed, near Seattle WA [Lundquist et al., this paper]−121.547.32.35443.7
Figure 1.

Northern Hemisphere map of study locations and relative snow persistence under trees or in the open, overlaid on mean December-January-February temperature (°C) gridded at 0.5° spatial resolution, with example photographs of areas with more snow retention under forests or in clearings. Photos courtesy of Google Earth Imagery with the following copyrights, clockwise, from top left:


2012 Google, GeoEye;


2012 GeoEye, ZENRIN;


2012 Google, DigitalGlobe;


2012 Google.

[6] We compared the SDD between paired sites for each year using the difference in snow duration, in number of days, between the forest and open site. In many cases, these data were not directly reported, and we estimated from a published figure, as noted in the supporting information (Table S1). We classified the study locations into three groups based on their ΔSDD, where:

display math(1)

[7] Investigations with ΔSDD > 3 days were categorized as locations where snow lasts longer in the forest, investigations with ΔSDD < −3 days were categorized as locations where snow lasts longer in the open, and investigations with values in between were categorized as ties (ΔSDD ≤ 3 days and ΔSDD ≥ −3 days) because differences less than 3 days were hard to distinguish precisely from some published figures. For each category, the 25th and 75th percentile values of the distribution of earlier SDD values, representing the time of year when the first site melts out (e.g., SDD for the forest if snow lasted longer in the open site), were used in considering how time of melt relates to incoming potential solar radiation (section 3.2). In order to avoid weighting our analyses toward investigations with more sites or more years of data, we aggregated all of the sites and/or years reported by each investigation by calculating the ΔSDD for each pair of sites for each year, and then taking the median ΔSDD value for all paired sites and all years in each general location (defined as a group of sites in the same paper that would fit within one 0.5° grid cell and identified by investigation number in Table 1). The complete results from multiple years or multiple sites reported in the investigations are included in the supporting information (Table S2).

[8] For studies where the data were available, we also determined the peak snow water equivalent (SWE) measured in the open, the peak SWE measured in the clearing, and the day of year when each of these occurred. At locations with multiple years of data, we took the median value. In many cases, these data were not directly reported, and we estimated from a published figure. In some cases, we converted reported peak snow depth to SWE, using a typical density for the region. These details are noted in the supporting information (Tables S1 and S2). At over 90% of the sites, peak SWE in the forest and the clearing occurred on the same day of the year. We obtained a rough measure of average melt rate for each site by dividing peak SWE by the number of days elapsed from the date of peak SWE to the snow disappearance date (mm day−1).

2.2. Global Climate Data

[9] Because investigations were inconsistent in their reporting of local climatological values, we used large-scale mean climatology to compare the same variables across all sites (Table 1). In particular, we used mean winter (December, January, and February; DJF) air temperature (°C) and total winter precipitation (mm) based on 0.5 degree gridded monthly mean temperature and mean total precipitation for 1975–2005 [Matsuura and Willmott, 2009]. When and where possible, we compared these coarse-resolution data with measurements taken at the site. While there were some variations, most temperature differences were less than 2°C, with randomly distributed errors. The only site deviating by more than 2°C was in the Pyrenees [López-Moreno and Latron, 2008], which we point out in the results. Overall, using gridded data allowed us to include more study sites without biasing our results (see supporting information, Figure S1 and Table S3, for comparison of local versus coarse-resolution temperatures). Average winter wind speed (m s−1) values were extracted from one degree resolution global monthly values at 10 m above the surface, averaged from July 1983 to June 1993 [NASA, 2007]. These data were obtained from the NASA Langley Research Center Atmospheric Science Data Center Surface meteorological and Solar Energy (SSE) web portal supported by the NASA LaRC POWER Project (http://eosweb.larc.nasa.gov/sse/).

2.3. Google Earth Images

[10] Aerial images were obtained from Google Earth and were not used quantitatively but only for illustration. Details on these images are included in the supporting information (Table S4).

2.4. Description of Previously Unpublished Data

[11] Due to the scarcity of published studies in areas with warmer winters, we include previously unpublished results from a 2 year field campaign conducted by the authors in the Cedar River Municipal Watershed that are part of an investigation on the effect of forest structure on snowpack magnitude and duration (forests and locations discussed in Lutz et al. [2012] and Martin et al. [2013]). Snow disappearance dates for these sites were derived from temperature sensors placed 1–2 cm below the ground surface; snow presence was inferred from the temperature record by the lack of a diurnal temperature variation (following Lundquist and Lott [2008]). These temperature sensors were deployed in a gap that was approximately one tree height in diameter, and in an adjacent second growth forest plot. Additionally, an acoustic snow depth sensor was deployed in the middle of a nearby clearing (2–3 tree heights in diameter). Median SDD from the forest plot (from six to seven sensors) was compared to the median SDD from the clearing and from the center of the gap (from one sensor in each location). Snow disappeared in the clearing and the gap center within 2 days (day 136 and 138, respectively) in water year 2011.

2.5. Model Framework

[12] The energy available for snowmelt, QM, in W m−2, can be calculated as

display math(2)

where Lnet is net longwave radiation (W m−2), Snet is net shortwave radiation (W m−2), QE is the sensible heat flux (W m−2), QH is the latent heat flux (W m−2), QR is the energy transferred to the snowpack from deposited snow or rain (W m−2), QG is the ground heat flux (W m−2), and dU/dt is the change in the internal energy of the snowpack (W m−2). All fluxes are defined as positive toward the snow surface. Over 80% of the energy available for melt can be attributed to net radiation [Marks and Dozier, 1992], so here we focus our model analysis on the shortwave and longwave radiation balance, retaining only the first two terms in equation (2), which are defined in more detail below. Shortwave radiation is less under a forest than in an adjacent clearing, and longwave irradiance is higher under a forest than in an adjacent clearing, so we expect the balance between these two to be key to explaining differences between paired forest and open areas. Potential impacts from the nonradiative energy balance terms, particularly turbulent fluxes, are included in the discussion (section 4). Throughout, we describe the forest using the sky view fraction (SVF), the amount of sky visible in an upward pointing hemispherical photograph, which is a good predictor of both shortwave and longwave subcanopy regimes [Essery et al., 2008b].

[13] We performed radiation balance calculations as a function of forest cover [Lawler and Link, 2011]. Net longwave radiation is given by,

display math(3)

where Lo is atmospheric longwave radiation, τ is the forest sky view fraction (sky visible from the forest floor), Lc is canopy-emitted longwave, and Ls is longwave emitted by the snow surface [Lawler and Link, 2011]. Incoming atmospheric longwave radiation, L0, is calculated as a function of temperature, relative humidity, and cloud cover (using the Dilley and O'Brien [1998] clear sky formulation with the Kimball et al. [1982] cloud correction, as recommended in Flerchinger et al. [2009]). The emissivity of the forest canopy and the snow surface were both approximated as 1. Forest canopy temperature was assumed equivalent to air temperature, as in Essery et al. [2008a]. While individual trunks in sparse or discontinuous canopies may heat up to much more than air temperature on sunny days [Pomeroy et al., 2009], this is in general a good approximation and has been found true for dense canopies [Sicart et al., 2004]. We assumed that the snow surface temperature tracks the surface dew point temperature (also often referred to as the frost point temperature below 0°C), as the air immediately adjacent to the snowpack is always saturated, following Andreas [1986]. Although the snow surface temperature is also a function of past snowfall events and thermal transmissivity of the underlying snowpack, the current dew point temperature is a reasonable approximation (M. S. Raleigh et al., Standard temperature and humidity approximate snow surface temperature, submitted to Water Resources Research, 2013.). Dew/frost points were calculated using the Magnus formula as specified in Alduchov and Eskridge [1996], with their coefficients. Snow surface temperature was set equal to the lower of dew point temperature or 0°C, and turbulent fluxes were neglected.

[14] Net shortwave radiation is given by,

display math(4)

where So is incoming shortwave radiation, and α is the snow albedo. Average daily clear-sky incoming shortwave radiation, So, was calculated as a function of time of year and latitude, following equations in the appendix of [Lundquist and Flint, 2006]. Cloudy-sky shortwave radiation was assumed to be 50% of clear-sky radiation for any given day, determined as an average value based on comparisons of surface clear-sky and cloudy-sky measurements [Fitzpatrick and Warren, 2005; Flerchinger et al., 2009]. Snow albedo was held constant at 0.65, chosen as an intermediate value between new (∼0.80) and old (∼0.40) snow albedos. Here shortwave transmissivity through the canopy is assumed to equal τ, the forest sky view fraction. Other albedo values and transmissivity assumptions, including a discussion of direct versus diffuse radiation, were examined in the model sensitivity tests.

3. Results

3.1. Synthesis of Existing Studies

[15] Average winter temperatures and precipitation were clearly related to whether forested or open areas retained snow longer. Locations where snow lasted longer in the open than in the forest were warmer (Figures 1 and 2a) and wetter (Figure 2b) than locations where snow lasted longer under the forest. Winter temperature and precipitation were also related (Figure 2c), such that all wetter sites were warmer, but not all warmer sites were wetter. Coarse-resolution wind speeds did not vary much across the sites, although the two sites with the greatest snow retention in the forest compared to the open (Hitsujigaoka, Japan, and Fraser, Colorado) also had higher than average wind speeds (Figure 2d).

Figure 2.

(a) Mean DJF temperature and (b) total DJF precipitation as a function of median ΔSDD (SDDforested_site – SDDopen_site). (c) DJF temperature versus DFJ precipitation and (d) mean DJF wind speed as a function of median ΔSDD.

[16] Sites subjected to mean winter temperatures less than −6°C had later snow disappearance dates under canopies, suggesting that retaining forest cover in these regions (e.g., Figure 1, Northern Finland or the east slope of the Canadian Rocky Mountains) is critical for snow retention. At the other end of the spectrum, sites with mean winter temperatures greater than −1°C had earlier snow disappearance dates under canopies, suggesting that reducing forest tree density in these regions (e.g., Figure 1, Japan or California) would enhance snow retention. Regions falling in between these extremes experienced similar snow disappearance timing between the clear and forested sites.

[17] For those sites where data were available, open sites both accumulated more snow (Figure 3a) and experienced faster spring melt rates (Figure 3b) relative to their adjacent forest sites, consistent with previous results. The ratio of peak forest SWE to peak open SWE (Figure 3c, derived from data shown in Figure 3a) generally declined with increasing winter temperatures (Figure 3c), and the ratio was also lowest for those study sites where snow lasts over 2 weeks longer in the open than under the forest (Figure 3d). The sites with snow lasting longer in the open (squares, Figure 3) exhibited both the lowest forest accumulation compared to the open (Figure 3a) and some of the fastest spring melt rates (Figure 3b), although the ratio of forest to open spring melt rates (derived from Figure 3b but not shown) did not relate directly to where snow lasts longer. Therefore, relative snow duration would seem to primarily derive from processes occurring before peak SWE, namely, interception and midwinter melt, which we define here as melt occurring prior to the date of peak SWE.

Figure 3.

For subset of studies where data were available, (a) peak SWE measured in open site versus peak SWE in forest site; (b) melt rate in open versus forest (calculated as peak SWE divided by number of days between peak SWE and snow disappearance); ratio of peak SWE in forest compared to open, based on shown in Figure 3a as a function of (c) DJF temperature; (d) snow disappearance date difference, ΔSDD (SDDforested_site – SDDopen_site); (e) average tree height; and (f) fractional forest cover. Symbols correspond to those in Figure 1.

[18] Forest characteristics were not uniformly reported across the study sites (Tables S1 and S5, in supporting information). The most frequently reported metrics were tree height (14 sites, Figure 3e) and fractional forest cover (five sites, Figure 3f). Two of the sites with snow lasting longer in the open also had very tall (∼40 m) trees (Figure 3e), but one did not. Fractional forest cover has been related to interception in other studies [Wigmosta et al., 1994; Varhola et al., 2013], and the ratio of peak forest SWE to peak open SWE was lower for sites with higher fractional coverage (Figure 3f). However, too few studies reported this metric to provide a robust comparison.

3.2. Comparison of Local Versus Global Climate Data

[19] We checked local temperatures versus gridded temperatures where possible (see Figure S1 and Table S3 in supporting information), and one outlier in Figure 2a can be explained by a site-specific variation from the coarse-resolution climate data. Specifically, snow lasted 6 days longer under the forest in the Pyrenees [López-Moreno and Latron, 2008], despite a coarse-scale mean DJF temperature above 2°C (outlier in Figure 2a). The local January-February-March mean temperature plotted in the paper (their Figure 2) was about −4°C. The outlier with high precipitation (Figure 2b) and high winds (Figure 2c) is Niigata, Japan [Whitaker and Sugiyama, 2005]. They report average November to March precipitation of 1400 mm, so the 650 mm reported here for DJF is likely not an overestimate. They do not report wind speeds, but the high value shown here is likely an overestimate because the coarse-resolution grid cell for this location includes ocean areas (which have higher winds). No other anomalies were found when comparing coarse and local scale meteorological information.

3.3. Example Case Where Snow Lasted Longer in the Open: Motivation for Modeling

[20] Given the wide variety of forest characteristics across the global plot studies, it is at first surprising that relative snow duration would be related to climate characteristics. However, in reviewing plots of snow depth and SWE across all of the studies, a pattern emerged: all of the sites with snow lasting longer in the open had evidence of midwinter melt, with more winter melt occurring under the forest than in the open. We illustrate this for one site (Umpqua, Oregon, Site 1 in Table 1) for one water year (1997–1998) and then model the concept of how melt timing (winter versus spring) affects the relative role of forest cover on the snowmelt energy balance.

[21] Storck et al. [2002], working with weighing lysimeters in the Cascades of Oregon (Table 1, site 1), measured interception to be approximately 60% of total snowfall. Measured SWE in the forest accumulated at a rate about 50% of accumulation in the open (Figures 4a and 4b), where this differs from the reported interception rate because some of the intercepted snow (∼28%) sluffed off the canopy and was added to beneath-forest accumulation, although the majority (∼72%) was lost to meltwater drip [Storck et al., 2002]. To quantify the relative importance of interception versus midwinter melt on the difference in peak SWE, we summed the cumulative gains and the cumulative losses at both the forest and open sites up until the date of peak SWE (Figures 4b and 4c). We excluded the period in early February when no snow existed under the forest because there was no indication of what under-forest melt might have been had snow been present. With this in mind, there was a 172 mm difference in SWE between the open and forest on 11 March 1998 (day of peak SWE). If we discount the 11 mm lost in the open in February when there was no snow in the forest, this difference would be 183 mm. Of this difference, 131 mm (72%) were due to accumulation processes (i.e., interception), and 52 mm (28%) were due to midwinter melt processes.

Figure 4.

Time series of (a) SWE, (b) cumulative gain, (c) cumulative loss, (d) delta-SWE (daily accumulation or melt), and (d) mean daily temperature for 1997–1998 for the Umpqua Forest, Oregon site (based on data collected by Storck [2000]). Boxes identify periods of mid-winter melt when there were faster melt rates under the canopy than in the clearing. Peak SWE is marked with a vertical dashed line. Note that cumulative gains and losses were only summed when SWE was present in both the open and the forest. This neglected ∼ 11 mm of melt in the open in February at a time when no snow was observed in the forest.

[22] There were two clear, midwinter melt events, coinciding with warm winter temperatures, when melt rates under the forest exceeded those in the open (Figure 4d and 4e, marked with gray shading). Given that wind speeds at this site were low (< 2 m s −1) [Storck et al., 2002], the difference must be explained primarily by the radiation energy balance. Thus, we use modeling to explore how areas with warm winter temperatures would be likely to see more winter melt under the canopy than in the open, in contrast to the faster spring melt rates in the open (Figure 3b), which are more commonly observed and discussed.

3.4. Modeling: Framework for Understanding Midwinter Melt and the “Radiative Paradox”

[23] Incoming radiation varies as a function of day of year (Figure 5a) and temperature (Figure 5b). The amplitude of variation in average daily shortwave radiation over the year ranges from 200 to 300 W m−2, depending on latitude (Figure 5a). To a first order, the forest canopy emits longwave radiation as a function of air temperature to the fourth power [Sicart et al., 2004], while the atmosphere emits longwave radiation as a function of air temperature, humidity, and cloud cover (illustrated by the range of dots in Figure 5b). Between −12 and 6°C, longwave radiation varies by about 100 W m−2. Thus, incoming shortwave radiation is less than incoming longwave radiation throughout the winter but often exceeds incoming longwave in the summer (Figure 5).

Figure 5.

(a) Average daily clear-sky shortwave irradiance as a function of latitude and time of year, with median and interquartile ranges of snow disappearance dates, plotted at the approximate height of potential shortwave irradiance for that time-period, for locations where snow lasts longer in the open versus where snow lasts longer in the forest. (b) Incoming longwave irradiance for full forest cover and open, as a function of air temperature, with blue dots spanning a range of different humidity and cloud characteristics. Median and interquartile ranges of mean DJF temperatures, sorted as in Figure 5a, plotted near longwave irradiance at those temperatures for comparison.

[24] We hypothesize that the main difference between sites where snow lasts longer in the open versus sites where snow lasts longer in the forest is due to the shift in the dominant term in the energy balance between longwave and shortwave radiation, as dictated by time of year (shortwave) and by climatological temperatures (longwave). To illustrate the idea, we plot the median date and interquartile ranges of first snow disappearance for the two groups on the same graph as the seasonal cycle of shortwave radiation (Figure 5a). The median date of first snow disappearance at paired sites in the northern hemisphere where snow lasted longer in the forest was 25 April, when potential daily average solar radiation is about 30 W m−2 greater than the daily average on 14 April, the median date of snow disappearance at sites where snow lasted longer in the open (Figure 5a). For comparison, we plot the median and interquartile ranges of winter air temperatures for the two groups on the adjacent graph (Figure 5b), aligned with the corresponding incoming forest longwave radiation. Winter forest longwave radiation is about 25 W m−2 greater at sites where snow lasts longer in the open than at sites where snow lasts longer in the forest.

[25] To summarize, locations where snow disappears earlier in the spring tend to have snow lasting longer in the open than under the forest (Figure 5a), and these same sites have warmer winter air temperatures (Figure 5b). Earlier in the year, longwave radiation exceeds shortwave radiation, and increased longwave radiation from tree canopies has the potential to cause faster melt under-canopy than in the open. Thus, changes in the timing of melt and the net radiation balance may be able to distinguish where snow lasts longer under a forest than in the open.

[26] To illustrate the interplay between air temperature, canopy cover, and time of year, we calculated the forest floor net radiative energy balance (following methods in section 2.4) assuming no clouds and 50% relative humidity for average incoming solar irradiance values typical for January (125 W m−2) and May (320 W m−2, Figure 6) in the northern hemisphere. Decreased canopy cover leads to increased net solar irradiance, with a greater difference during times of high incoming solar radiation, such as May (Figure 6a). Net longwave irradiance shifts from positive under 100% canopy cover to negative under no canopy cover (Figure 6b). Net longwave irradiance increases slowly with temperature for dew point temperatures less than 0°C and quickly for dew point temperatures above 0°C because snow longwave emissions are capped at 316 W m−2 (the value at 0°C). A 0°C dew point temperature occurs at an air temperature of about 10°C for 50% relative humidity (RH) and at an air temperature of about 2°C for 90% relative humidity. Note that for our 50% RH simulation, the difference in net longwave to the snow is greater between 10 and 15°C than between 0 and 10°C (Figure 6b). These effects combine such that total net irradiance decreases as canopy cover is removed in typical January conditions (Figure 6c) but increases as canopy cover is removed in typical May conditions (Figure 6d). The point where air temperature variations become more important than canopy cover changes depends on the relative humidity, as illustrated by the discontinuity in the contours in Figure 6c, around 10°C for this example. These general results are robust across a wide range of atmospheric conditions (humidity and cloud cover) and across a variety of shortwave attenuation and albedo assumptions, as demonstrated in the model sensitivity tests (below).

Figure 6.

(a) Net shortwave radiation and (b) net longwave radiation as a function of canopy cover, for 0.65 albedo, no clouds, and 50% relative humidity. Contours of (c) January and (d) May combined shortwave and longwave radiation as a function of both forest cover and air temperature. Arrows in c and d indicate direction of increasing energy available for melt.

3.5. Model Sensitivity Tests

[27] Here we discuss assumptions related to both longwave and shortwave irradiance in the snowmelt energy balance and present how our model varies with changes in these assumptions.

3.5.1. Cloudcover and Humidity Assumptions

[28] The relative importance of forest cover on the longwave budget also depends on the magnitude of incoming longwave irradiance from the atmosphere, which is a function of atmospheric humidity and cloud cover [Sicart et al., 2004]. For illustration, here we consider cloudy and moist conditions (100% cloud cover and 90% relative humidity) and clear and sunny conditions (no clouds and 20% relative humidity) for days with typical January and May incoming shortwave irradiance (Figure 7a).

Figure 7.

(a) Net shortwave and (b) net longwave irradiance as a function of forest cover, sky cover (clouds), temperature, and time of year.

[29] Canopy cover has a greater impact on incoming longwave when the atmosphere is clear and dry, with a lower emissivity (Figures 7b and 8d). Temperature has a stronger effect on net longwave when the dew point temperature is greater than 0°C (which occurs at around 2°C for 90% relative humidity, Figures 7b and 8a), because the outgoing longwave is capped by the 0°C snow surface temperature. January net radiation, the sum of net longwave and net shortwave, is a stronger function of temperature under moist and cloudy conditions (Figure 8b) and a stronger function of sky view fraction under sunny and dry conditions (Figure 8e). However, at both the cloudy-and-moist and the sunny-and-dry extremes, net radiation increases with greater forest cover under typical January solar irradiance values (Figure 8b and 8e). Net radiation decreases with greater forest cover under typical May solar irradiance values for cloudy and moist conditions (Figure 8c) but is relatively insensitive to forest cover for sunny and dry conditions (Figure 8f). Most locations will fall in between these two illustrated climate extremes, with relative sensitivities to forest cover and air temperature varying with atmospheric conditions.

Figure 8.

(a, d) Contours of constant net longwave radiation, and (b, c, e, f) net total radiation (including shortwave) as a function of air temperature and canopy cover, for an albedo of 0.65 and (b) January cloudy and moist conditions (top row, incoming atmospheric shortwave = 62.5 W m−2, 100% cloud cover, and 90% relative humidity) and (e) January clear and dry conditions (bottom row, incoming atmospheric shortwave = 125 W m−2, no cloud cover, and 20% relative humidity) and (c and f) May conditions with incoming atmospheric shortwave of 162.5 W m−2, and 325 W m−2, for cloudy and clear conditions, respectively. Arrows indicate direction of increasing melt rates.

3.5.2. Shortwave Transmissivity and Albedo Assumptions

[30] Net solar radiation is sensitive to the incoming solar irradiance (S0), to the transmission of solar radiation through the canopy (τ) and to the albedo of the snowpack (α), which evolves through time and also depends on leaf litter (see equation (3)). We made the simplest assumption that solar transmissivity equals the sky view fraction ( math formula, one option in Sicart et al. [2004]). This implies that, on average, all solar irradiance is blocked according to the fraction of sky blocked by trees overhead. Sicart et al. [2004] also considered the second option that

display math(5)

where a=0.45 and b=0.29, as found empirically by Pomeroy et al. [2002], where LAI math formula is the one-sided leaf area index. Using Beer's law (described in Nijssen and Lettenmaier [1999]):

display math(6)

where k is the average projected foliage area in the direction of the sun, depending on the sun angle and specific canopy characteristics. Alila and Beckers [2001] suggested that mean values of k could vary from 0.1 to 0.5 in British Columbia, and Sicart et al. [2004] suggested k≈0.8 in Fraser, Colorado and Wolf Creek, Canada. For a dense canopy (k≈0.8), this formulation is very similar to the simpler approximation of math formula (Figure 9a). However, for a sparse, or more transmissive, canopy (k<0.3), much more solar radiation reaches the forest floor even with 100% canopy cover. Therefore, our simple assumption provided the greatest shortwave attenuation in the forest relative to the clearing and the maximum expected effect of forest shading. More specific shortwave transmission models have been developed for cases where precise canopy structures are known [e.g., Ni et al., 1997; Musselman et al., 2012], but given our interest in a generic forest defined by sky view fraction, we do not consider these here.

Figure 9.

(a) Illustration of average transmissivity of shortwave radiation through a canopy as a function of canopy sky view fraction, both for a linear assumption and for Beer's Law with different k values (attenuation due to foliage), assuming an empirical relationship between SVF and LAI. (b) Illustration of combined effects of transmissivity and albedo on net shortwave radiation, for cases of spatially-constant albedo of 0.65 (dashed lines) and of albedo decreasing linearly from 0.65 under no canopy to 0.45 under full canopy (solid lines).

[31] With the Beer's law form of canopy attenuation (k=0.3), net shortwave irradiance only drops off for sky view fractions less than 0.5 (Figure 9a and 10a). January net radiation still increases with increasing canopy cover (Figure 10c). However, May net radiation has a maximum at about 0.45 SVF (Figure 10d). This is a direct result of the functional form of transmissivity (Figure 9a), which has an inflection point at about 0.45 SVF. This sensitivity test highlights the importance of considering site specific canopy architecture in locations where shortwave irradiance is important to melt rates. For example, canopy gaps [Golding and Swanson, 1978; Lawler and Link, 2011] and strips [Kittredge, 1953; Anderson and Hoover, 1976] have been proposed as optimal shapes for minimizing the rate of shortwave irradiance increase with larger sky view fractions.

Figure 10.

Contour plots as in Figure 4 (no clouds and 50% RH), but for math formulaand k=0.3. (a) Net shortwave a function of time of year, (b) net longwave as a function of temperature, and (c) January and (d) May net total radiation (W m−2) contoured as a function of SVF and temperature. Albedo is held fixed at 0.65. Arrows indicate direction of increasing net radiation.

[32] Changes in snow albedo are also a critical component of the shortwave energy balance. Albedo changes from near 0.85 after fresh snowfall to ≈0.4 for older, dirtier snow [United States Army Corps of Engineers (USACE), 1956], changing net shortwave irradiance by a factor of 2. In most models, albedo changes with time, increasing when new snow falls and decreasing over the days following, but is constant across vegetation types (e.g., DHSVM) [Wigmosta et al., 1994]. Several studies have identified a more rapid decrease in albedo with time under a forest canopy due to the accumulation of leaf litter [Hardy et al., 2000; Storck, 2000; Melloh et al., 2002; Burles and Boon, 2011], with some studies showing forest floor albedo values dropping as low as 0.2 (a value more typical for vegetation than snow). Roughly, at any given point in time, pine litter lowered the albedo about 20% under the forest compared to the open [Hardy et al., 2000; Melloh et al., 2002].

[33] To illustrate the effects of changing snow albedo with canopy cover, we modeled albedo as a linear function of SVF, such that α=0.65 in the clearing and decreased to 0.45 under complete forest cover, and examined the combined effects of this albedo change with changes in solar transmittance (Figure 9b). Increased leaf litter and the associated lower albedo under denser canopies offset the effects of increased radiation attenuation. The effects of albedo changes are greatest in cases where more solar radiation is transmitted through the canopy at lower sky view fractions (low k values) because more of that radiation is also reflected back, resulting in negligible net solar radiation differences between the open and closed canopy cases for the case of k=0.1. Again, the base case of math formula with spatially constant albedo results in the greatest shortwave radiation increases with increases in sky view fraction (Figure 9b). The general patterns for net irradiance under forest cover (Figure S2, in supporting information) are similar to those shown in Figure 8 for constant albedo, suggesting that forest impacts on albedo are likely of secondary importance compared to forest impacts on incoming shortwave attenuation in terms of which SVF would be most or least optimal for snow retention.

[34] In addition to the above considerations, direct and diffuse radiation, which we considered together, have different behavior with respect to canopy transmission [Reifsnyder et al., 1971] and albedo [Warren, 1982; Gardner and Sharp, 2010]. In particular, Stähli et al. [2009] showed that forest transmissivity is higher on cloudy days than on sunny days because a larger fraction of diffuse radiation passes through the canopy. Thus, our formulation underestimates diffuse radiation under the canopy on cloudy days, and the illustration of forest versus open net shortwave cloudy-day radiation in Figure 7a is likely less variable (i.e., closer to a flat line) than shown. Because so few measurements exist of both direct and diffuse radiation transmission through canopies, we do not explicitly model these effects.

4. Discussion

[35] Our results suggest that mean winter temperatures indicate where the net effect of forest-cover shifts from increasing snow duration to decreasing snow duration relative to an adjacent open area (Figures 1 and 2a). These differences generally set up before peak SWE is reached (Figure 3a) and thus are primarily due to interception and/or midwinter melt. Where mean winter temperatures are greater than −1°C, melt occurs at a time of year when solar radiation is relatively smaller, and thus net radiation under a forest is greater than in a clearing. These general results hold across the northern hemisphere and can be indexed with coarse-resolution mean winter temperatures.

[36] The coherence of the signal across disparate studies and very different forest types suggests that climate effects transcend site-specific and forest structural differences. Simple modeling demonstrates that midwinter melt under forests in warmer regions is a plausible explanation for the patterns observed. While seldom highlighted in paper discussions, many other authors have reported faster melt rates under forests compared to under open sky during winter or early spring melt events (e.g., Storck et al. [2002]: December and January melt in Oregon; Bales et al. [2011]: January melt in California; López-Moreno and Stähli [2008]: multiple locations near Alptal, Switzerland; and López-Moreno and Latron [2008]: several periods in April in Spain).

[37] However, for any particular location, site-specific effects should also be considered, as illustrated by the wide range of results presented for the Sierra Nevada, California by Kittredge [1953]. Mountains span a wide range of elevations and microclimates within a short horizontal distance. Thus, snow at lower elevations may last longer in clearings while snow at higher elevations lasts longer under the forest, and the precise elevation where this shift occurs likely varies between years and between different forest structures. For example, two different studies near Alptal, Switzerland (Table 1, studies 2 and 3) reached different conclusions about snow duration (longer in the open in López-Moreno and Stähli [2008], but about the same in Rutter et al. [2009]). López-Moreno and Stähli [2008] concluded that maximum differences between peak SWE in forested and open areas near Alptal occurred during warm and wet years (and the minimum difference during cold and dry years), similar to the climatological patterns shown here (Figure 2).

[38] Additionally, the importance of solar radiation changes not only with time of year but also with location on the landscape. Faster melt rates occur under the forest than in the open on north facing slopes [Ellis et al., 2011], and frequently no differences in melt rates are observed between openings and forests on south facing slopes [Bales et al., 2011]. The combined interactions of slope, aspect, and forest cover on melt were nicely illustrated in an idealized modeling study of patches of forests and clearings around a conical mountain by Strasser et al. [2011]. As shown in the modeling section, temperatures and time of melt are indices of the net radiation balance. Thus, this framework can readily be extended to include site-specific characteristics, such as slope, aspect, and cold air pooling.

[39] Our work here has focused on the importance of the radiative energy balance in controlling where snow is retained longer. Other factors that play important roles, and must be considered on a more site-specific basis, include canopy interception, timing of precipitation, and turbulent mass and energy transfer (wind). These processes relate critically to regional and site-specific climatology, as well as to specific canopy structure. In the two sections below, we discuss how these factors relate to the results presented here.

4.1. Canopy Interception

[40] Canopy interception affects snow accumulation and thus, snow duration. A smaller fraction of total snowfall accumulating under trees will contribute to snow there disappearing earlier. Reported interception rates vary from 28 to 83% [Martin et al., 2013, their Table 1]. While some study sets suggest that interception rates are greater in warm, maritime climates (60% in the Pacific Northwest, [Storck et al., 2002] than in cold, continental climates; 30–45% in the boreal forest [Hedstrom and Pomeroy, 1998]); numbers reported in other studies do not fit this pattern (e.g., 60% in black spruce in the boreal forest, [Pomeroy and Schmidt, 1993]; 37% for snow in sitka spruce in central Scotland, [Johnson, 1990]).

[41] Total annual plot-scale interception is likely due more to forest structure and density (e.g., Figure 3f) than to climate characteristics, but existing data are too few and disparate to draw conclusions with certainty. Increasing sky view fractions decrease interception and increase snow accumulation in most settings [Kittredge, 1953; Varhola et al., 2010a].

[42] Because interception is a fractional multiple of precipitation, up to a maximum canopy storage capacity, the absolute difference between snow accumulated under the forest canopy and in the opening is greater for places that receive more total snow (Figure 3a). Snow in warmer climates is associated with more total precipitation (Figure 2c). However, because these sites also received much of that winter precipitation in the form of rain, the warm-winter sites did not have the highest snow accumulations (Figure 3a) but did have the lowest forest:open ratios for peak accumulation (Figure 3d).

[43] One secondary effect of climate on interception could be the role of rain-on-snow and of midwinter melt in regularly removing intercepted snow from the canopy, thus increasing the canopy capacity for total interception over a season. Anecdotal evidence from some colder locations suggests that snow intercepted in the first winter snowfall remains in the forest canopy throughout the season, thus reducing the forest's capability to intercept more snow in subsequent events. Further investigation into canopy snow storage across locations and climates is essential to test this hypothesis.

4.2. Timing of Precipitation

[44] The study locations where snow lasted longer in the clearing not only had more winter precipitation (Figure 2b), but many of them also had a clear Mediterranean Climate (i.e., three out of four investigations were in the maritime mountains of the Western United States), with less spring and summer precipitation. This illustrates another reason why snow disappears earlier there. New spring snowfall would maintain snowcover both under the canopy and in the open regardless of winter melt. However, where precipitation falls primarily in the winter, spring snowfall is much less likely to occur and have this effect. The watershed-specific societal and ecological importance of enhancing snow water retention also depends on the timing of precipitation. Retaining winter snow is more important in regions where rain is not likely to fall later to make up the difference.

4.3. Wind and Turbulent Transfer

[45] Wind speeds, though highly variable in complex terrain, are invariably less under forest canopies than in an adjacent open area (e.g., 15–20% less, based on a Pacific Northwest Douglas-fir forest) [Chen et al., 1993]. Both wind redistribution and turbulent fluxes are dependent on wind speed. For example, measured sublimation rates were 3–10 times higher in open areas compared to a forested location in Idaho [Reba et al., 2012]. Where wind deposition and redistribution play a large role in snow accumulation [Vajda et al., 2006], or where turbulent fluxes are the primary contributor to snowmelt [Marks et al., 1998; Koivusalo and Kokkonen, 2002], snow lasts longer under the forest than in a clearing. In general, the mean DJF wind speeds at our sites were not related to the differences in snow disappearance dates between the forest and the open (Figure 2d), although the two locations with the longest snow retention in the forest relative to the clearing had higher-than-average wind speeds (Figure 2d). In complex terrain, local wind variations are likely as large as the global variations shown and may have very important local effects. The importance of canopy sheltering from wind also depends on the size of the open area used for comparison, which ranged from large clear-cuts to small gaps in the forest in our studies. Published descriptions of openings (Table S1) were not sufficient to assess these effects in a quantitative manner.

5. Conclusions

[46] Although forest structure and interception, as well as fine-scale site differences, such as slope, aspect, elevation, and wind exposure, will continue to influence snow retention at hectare scales [Ellis et al., 2011; Strasser et al., 2011], our results demonstrate, to a first order, that where mean DJF temperatures exceed −1°C, forests with lower total canopy cover are likely to enhance snow retention by minimizing mid-winter and early spring melt. Globally, the applicable temperature zone coincides with areas where snow water storage is important for human water supplies [Barnett et al., 2005], where snow is most at-risk to global warming [Nolin and Daly, 2006], and where forests are already extensively managed [Hassan et al., 2005]. Therefore, we must carefully consider the framework presented here both for future research and for its implications for joint forest and water management under changing climatic conditions.

5.1. Framework for Future Research

[47] Given the disparate nature of the forest plots and site-specific studies comprising our meta-analysis, it is remarkable that a global pattern emerges. The concept of climate and melt timing allows us to compare these studies and to make predictions for unstudied locations. The energy balance model also lets us downscale these predicted differences using local temperature gradients (for mean winter temperature) and local topographic shading (for potential solar radiation). The SDD metric is relatively easy to measure (e.g., from low-cost temperature sensors [Lundquist and Lott, 2008] or repeat photography), which allows for widespread evaluation of such predictions.

[48] The studies included here did not describe their forests with enough detail or consistency to evaluate forest structural effects on interception or melt. However, recent research developments suggest that this limitation will be overcome in the near future. Varhola et al. [2013] describe the four main forest metrics important to hydrology as leaf area index (LAI), forest cover (FC), sky-view factor (SVF), and forest height (H). Techniques now exist to derive all of these from airborne Light Detection and Ranging (LiDAR) measurements [Varhola et al., 2010b, 2012] and/or from satellite remote sensing (e.g., Landsat [Varhola and Coops, 2013; Varhola et al., 2013]; MISR [Chopping et al., 2009]; and MODIS [Chopping et al., 2011]). Modeling work has also recently been completed to use site-specific forest structure information to calculate the forest floor energy budget [Musselman et al., 2012, 2013; Seyednasrollah et al., 2013]. For the first time, this allows forest structure to be explicitly accounted for at landscape scales.

[49] Recent work has also demonstrated a technique to directly and nondestructively measure forest interception through stem compression [Friesen et al., 2008; Martin et al., 2013] or potentially, vibration frequency [Stewart et al., 2012]. These developments will allow subsequent studies to explicitly address forest structure and interception processes with much more detail than was possible here.

5.2. Application to Climate Change

[50] The context of mean winter temperatures allows us to extend these recommendations both to other locations on the globe and to future time periods of projected warming. Washington, Oregon, and California along the west coast of the US stand out as places where snow often lasts longer in clearings. Many forested areas of Europe (e.g., the Pyrenees and Alps), as well as the foothills of the Himalayas and parts of Japan, are in similar temperature regimes or are likely to be in a similar temperature regime to the US west coast following moderate warming (Figure 1). These results help us to bring together a large number of disparate studies into an overarching framework and provide guidelines for how changes in temperatures and changes in vegetation must be considered together.

[51] Climate models project warmer winter temperatures leading to earlier snow disappearance from the landscape [Stewart et al., 2004], which can have deleterious effects on human and natural systems, including a loss of soil insulation [Brown and DeGaetano, 2011], declines in surface albedo [Sturm, 2005], loss of late-season soil moisture for ecosystem health [Tague et al., 2009], and declines in late-season streamflow [Elsner et al., 2010]. In locations where winter temperatures warm from a regime where no winter melt occurred to a regime where winter melt does occur, the results presented here suggest that snow under forest cover would be more sensitive to climate change than snow in unforested areas. Model results suggest this is the case for the Tuolumne River Watershed in the Sierra Nevada, California [Cristea et al., 2013]. Specifically, streamflow timing was more sensitive to climate change and advanced earlier in the year under simulations with 100% forest cover than under simulations with no forest cover.

5.3. Application to Management

[52] The results presented here (both modeling and synthesis of other studies) suggest that if management goals are to retain snow on the landscape, forest cover should be strategically preserved in some locations (e.g., those with cold winter temperatures) and reduced, through thinning or gap-cutting, from others (e.g., those with warm winter temperatures). These results corroborate many forest management recommendations (e.g., cut gaps to enhance snow-fed water supplies) that were made in warmer maritime regions in the 1950s [e.g., Kittredge, 1953; Anderson et al., 1958; Anderson and Hoover, 1976].

[53] Even where no desire exists to directly manipulate the forest, these results provide key insight into how water managers must consider joint climate-vegetation change. For example, past and present management practices of fire suppression, timber harvesting, and silviculturally motivated replanting have led to very dense forests in many regions [Rautiainen et al., 2011]. However, recent observations and projected trends indicate increasing forest stressors (e.g., drought, insect infestations, and fire) and rapid forest die-backs [e.g., Kurz et al., 2008; Allen, 2009] Depending on climatic and topographic location, these changes could have quite contrasting impacts on snow and hydrology.

[54] Some management actions, such as strategically introducing canopy gaps to fire-suppressed forests, could not only optimize snow retention, but also increase fire resilience and landscape heterogeneity [Larson and Churchill, 2012]. The framework presented here provides a starting template for expected forest-snow-climate interactions that can guide regional experiments to determine the best paths forward for joint forest-water management.


[55] We would like to thank those study authors who answered questions and/or shared data with us, notably, Pascal Storck, Konstantinos Andreadis, Andrés Varhola, Ignacio López-Moreno, and Harri Koivusalo. We would also like to thank the Associate Editor and three anonymous reviewers, whose comments greatly helped improve the manuscript. This work was supported by the National Science Foundation under grant CBET-0931780.