21st century Wisconsin snow projections based on an operational snow model driven by statistically downscaled climate data

Authors


Abstract

Output from the Climate Model Intercomparison Project Phase 3 (CMIP3) global climate models are statistically downscaled across Wisconsin, using a method that restores the observed mean, variance, and extremes of daily temperature and precipitation. The downscaled climate data for the late 20th century, mid-21st century, and late 21st-century is used to drive the National Weather Service operational snow model, SNOW-17, to produce high-resolution (0.1° × 0.1°) projections of daily snowfall, snow depth, and snow cover for Wisconsin. These snow projections will guide wildlife scientists in climate change impact studies and the development of adaptation strategies for the state, in addition to being of value to hydrologists, agricultural scientists, and other experts. SNOW-17 simulations suggest a dramatic shortening of the Wisconsin snow season, with the greatest snowfall and snow depth reductions in spring, particularly over northern Wisconsin. Snowfall is substantially reduced in response to projected warming and only slightly offset by a projected increase in cold-season precipitation. Percent reductions in snow depth are likely to be even more impressive than in snowfall, given not only a reduced frequency that falling precipitation will be in frozen form but also an enhanced snowmelt due to rising temperatures. Copyright © 2010 Royal Meteorological Society

1. Introduction

Across most parts of the world, snow cover area has decreased since the early 20th century, particularly during spring and summer (Lemke et al., 2007). During the past 30 years, snow cover area across the Northern Hemisphere has decreased by about 10%, most distinctly in spring–summer (Serreze et al., 2000). Reductions in North American snow cover, especially during spring, have mainly occurred during the second half of the 20th century, with rapid decreases during the 1980s and early 1990s (Brown, 2000; Lemke et al., 2007). More specifically, the largest reduction in snow cover during 1967–2004 across the United States was found during March–April over the northern and western states, within the region bounded by the 0–5 °C (March–April) isotherms (Lemke et al., 2007), which includes Wisconsin. Indeed, a reduction in springtime snow cover duration is evident over Wisconsin in the study by Brown (2000), based on National Oceanic and Atmospheric Administration (NOAA) weekly snow cover data for 1966–2007.

According to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC AR4), North America should warm more than the global average during the 21st century in response to rising levels of greenhouse gases, with enhanced warming across the northern portion of the continent, particularly in winter (Christensen et al., 2007). Simultaneously, snow depth and the length of the snow season should decrease across North America, with global climate models projecting delayed autumn snowfall and earlier spring snowmelt (Christensen et al., 2007). However, these projections are complicated by the likely strengthening of the hydrologic cycle and projected increases in precipitation, particularly over the higher latitudes, which will partly offset the snow reductions associated with warming (Groisman et al., 2003; Räisänen, 2007). Kutzbach et al. (2005) and Christensen et al. (2007) identified projections of increased annual (primarily cold season) precipitation for the Midwest United States and Great Lakes Basin, based on the Climate Model Intercomparison Project Phase 3 (CMIP3) global climate models.

Given the significant climate changes that have already occurred and the continued advances in climate modelling and analysis methods, there has been a growing demand for high-resolution climate projections for climate change impact studies, requiring either statistical or dynamical downscaling. The current paper addresses such a demand for sub-county level climate projections, including snowfall and snow depth, across Wisconsin. The Wisconsin Initiative on Climate Change Impacts (WICCI: www.wicci.wisc.edu) was developed in 2007 as a collaboration between the University of Wisconsin-Madison and Wisconsin Department of Natural Resources to assess how climate will change in Wisconsin during the 21st century, identify potential impacts to the state's natural resources, and develop adaptation strategies to diminish negative impacts. Daily temperature and precipitation projections were developed through a statistical downscaling of the output from the CMIP3 global climate models, used in the IPCC AR4. For the current study, this downscaled climate data is used to drive an operational snow model to provide projections of Wisconsin snowfall and snow depth to the WICCI Wildlife Working Group, which is assessing likely impacts of climate change on Wisconsin wildlife communities.

As a motivation to downscale snow projections, wildlife researchers have found that many animal and bird species are strongly impacted by the presence and depth of snow and are therefore sensitive to changes in snow pack due to climate change. Certain wildlife in Wisconsin rely on the presence of snow cover for winter camouflage from predators, including the snowshoe hare (Lepus americanus) and ermine (Mustela erminea), and would therefore be threatened by increased mortality given a reduction in snow depth with climate change. The coat of a snowshoe hare turns white during winter due to photoperiod, helping to conceal it against the background snow cover from predators, while its large, furry back feet are ideal for moving atop snow. Likewise, an ermine's coat also turns white during winter, due to photoperiod, to help it blend into the snow, and its ability to hunt for rodents under the snow is a critical wintertime asset. For both the snowshoe hare and ermine, which rely on snow for camouflage, a shorter snow season would increase their predation rates. The timing of the change in colour of the coat of a snowshoe hare or ermine depends on photoperiod and not temperature, so these animals cannot adapt to warming by delaying the change to a white coat; besides, such an evolutionary adaptation would occur at a rate dramatically slower than projected climate change. During the period of greatest overall stress, the ruffed grouse (Bonasa umbellus) relies on snow for winter cover. During winter, it burrows into heavy snow to keep warm, reduce energy demands from heat loss, decrease the need for extended periods of foraging, and evade predators (Gullion, 1970; Thompson, 1987), while the design of its toes provides support as it walks upon the snow surface.

For the American marten (Martes americana) and Canada lynx (Lynx canadensis), the predator–prey relationship is closely tied to the presence of snow cover, with a loss of snowpack unfavourable for both animals by decreasing their competitive advantage over other carnivores, such as coyotes (Canis latrans) and bobcats (Lynx rufus) (Carroll, 2007). The small ratio of body mass to foot area in the American marten and lynx provides this competitive advantage on snow (Krohn et al., 1995; Mowat et al., 2000; Carroll, 2007). American marten are well adapted to deep snowpacks, given their ability to tunnel into the snow to track small mammals and find warmth among tree roots. Changes in snowshoe hare populations will certainly impact the Canada lynx, whose preferred prey is the snowshoe hare.

Some animals would likely benefit from a reduction in snowpack across Wisconsin, including white-tailed deer (Odocoileus virginianus). The primary wintertime stressors to white-tailed deer are air chill and snow hazard (Verme, 1968), with a positive correlation between snow depth and deer mortality and negative correlation between snow depth and deer population levels (Severinghaus, 1947; Edwards, 1956). Deer mobility is reduced by deep snow pack due to their thin legs and feet sinking into the deep snow, making them vulnerable to predators. The presence of snow contributes to heat loss in deer, reduced mobility, and the preference to seek shelter. For the American marten, Canada lynx, and white-tailed deer, the most important snow parameter is likely snow depth through its impact on mobility, while snow cover is key to the snowshoe hare and ermine for camouflage (based on research by the WICCI Wildlife Working Group); snow-water equivalent is likely of secondary importance and is therefore not a focus in this article.

Lettenmaier et al. (1999) applied a simplified statistical downscaling of three global climate models from the Second Assessment Report of the IPCC to force SNOW-17 and the Variable Infiltration Capacity (VIC) model across river basins in the North Central, Northwest, and Southeast United States for future projections of streamflow and snowmelt. Our present study is the first utilisation of the SNOW-17 to generate future snow projections in the Upper Midwest United States, in which we apply an advanced statistical downscaling to output from a large array of the state-of-the-art CMIP3 global climate models and use the downscaled climate data to force SNOW-17 in a domain centred on Wisconsin.

In this paper, the statistical downscaling methodology used to produce 0.1° × 0.1° projections of daily maximum temperature, daily minimum temperature, and daily precipitation across Wisconsin for the mid- and late 21st century is described, and this downscaled data is used to drive an operational snow model, SNOW-17, for projections of daily snowfall, snow depth, and snow cover. Previous studies have applied SNOW-17 in a lumped manner to individual watersheds (Hogue et al., 2000; Franz et al., 2008a, 2008b) or in a distributed manner by modelling a large river basin as a set of smaller sub-watersheds (Shamir et al., 2006), with simulations focused on the past observational period. In these previous studies, watershed-specific SNOW-17 parameters are used. In the current study, the Wisconsin domain for SNOW-17 has been subdivided into a 0.1° × 0.1° grid (48 × 62) and SNOW-17 is run at each grid cell using regionally identified parameters.

The statistical downscaling methodology, details of the SNOW-17 model, and calibration of the snow model are presented in Section 2. Results are shown in Section 3, and conclusions are found in Section 4.

2. Data and methods

2.1. Statistical downscaling

One key WICCI data product is a statistical downscaling of daily maximum and minimum temperature and daily precipitation amount for the domain surrounding Wisconsin (42.1–47.1°N, 93.4–86.6°W) onto a 0.1° × 0.1° grid. The downscaling is based on CMIP3 climate model output for both the 20th century (20C3M) and 21st century (A2, A1B, and B1 emission scenarios). Output from the coarse CMIP3 global climate models is converted to high spatial resolution over Wisconsin and debiased. This debiasing is vital given the significant regional biases over Wisconsin (Figure 1), with most CMIP3 global models exhibiting a wet bias in winter and warm, dry bias in summer. Using available daily climate model output from IPCC AR4, statistical downscaling of daily temperature and precipitation is performed for 1961–2000, 2046–2065 (‘mid-21st century’), and 2081–2100 (‘late 21st century’) using 15 general circulation models (GCMs) for 20C3M, 14 GCMs for A1B and B1, and 11 GCMs for A2 scenario (Table I). This statistically downscaled climate data is used to drive an operational snow model to generate daily snowfall, snow depth, and snow cover projections for Wisconsin through the 21st century. The methodology for downscaling temperature and precipitation is described below.

Figure 1.

Comparison of the mean seasonal cycle, for the late 20th century, of (a) temperature ( °C) and (b) precipitation (mm/day), averaged across Wisconsin between the observations (black—Maurer et al., 2002) for 1950–1999 and 15 CMIP3 GCMs for 1980–1999 in their 20C3M simulations (other coloured lines), prior to downscaling

Table I. List of the 15 CMIP3 GCMs used in the study, including their originating group, country, model ID, reference, and available scenarios
Originating groupCountryModel IDReference20C3MA2B1A1B
Canadian Centre for Climate Modelling & AnalysisCanadaCGCM3.1 (T47)Flato et al. (2000)XXXX
Canadian Centre for Climate Modelling & AnalysisCanadaCGCM3.1 (T63)Flato et al. (2000)X XX
Météo-France/Centre National de Recherches MétéorologiquesFranceCNRM-CM3Déqué et al. (1994)XXXX
CSIRO Atmospheric ResearchAustraliaCSIRO-Mk3.0Gordon and O'Farrell (1997)XXXX
CSIRO Atmospheric ResearchAustraliaCSIRO-Mk3.5Gordon et al. (2002)XXXX
US Dept. of Commerce/NOAA/Geophysical Fluid Dynamics LaboratoryUSAGFDL-CM2.0Delworth et al. (2006)XXXX
NASA/Goddard Institute for Space StudiesUSAGISS-AOMLucarini and Russell (2002)X XX
LASG/Institute of Atmospheric PhysicsChinaFGOALS-g1.0Yu et al. (2004)X XX
Instituto Nazionale di Geofisica e VulcanologiaItalyINGV-ECHAM4Valcke et al. (2000)XX X
Institut Pierre Simon LaplaceFranceIPSL-CM4Dufresne and Friedlingstein (2000)XXXX
Center for Climate System Research (The University of Tokyo), National Institute for Environmental Studies, and Frontier Research Center for Global Change (JAMSTEC)JapanMIROC3.2(hires)Hasumi and Emori (2004)X XX
Center for Climate System Research (The University of Tokyo), National Institute for Environmental Studies, and Frontier Research Center for Global Change (JAMSTEC)JapanMIROC3.2(medres)Hasumi and Emori (2004)XXXX
Meteorological Institute of the University of Bonn, Meteorological Research Institute of KMA, and Model and Data groupGermany/KoreaECHO-GRoeckner et al. (1996)XXXX
Max Planck Institute for MeteorologyGermanyECHAM5/MPI-OMGiorgetta et al. (2002)XXX 
Meteorological Research InstituteJapanMRI-CGCM2.3.2Kitoh et al. (1995)xXXX

The statistical downscaling consists of two stages. First, the statistical relationship between the large-scale atmospheric state and local temperature/precipitation at weather stations is determined for each calendar month from the observational record (1950–2007). Second, this established statistical relationship is applied to predict the local temperature/precipitation given a climate model's large-scale atmospheric state. The relationship between the large-scale atmospheric state and the weather stations is cross validated for all variables and seasons by alternately leaving out 3 years of data, fitting the remaining years and testing the fit on the left out data. The cross validation suggests that all statistical relationships found are robust. The downscaling approach builds upon the previous work of Wood et al. (2002, 2004) and Maurer et al. (2007).

Typically, statistical downscaling is used to relate the large-scale atmospheric state to one specific value of the temperature/precipitation at a point. With this approach, however, the downscaled variability and extremes at the point will be underestimated unless the relationship between the large-scale and the point is artificially inflated (Von Storch, 1999), which then artificially exaggerates the large-scale climate change signal. Therefore, in order to both simulate the variability and extremes and to properly account for the effect of the large-scale pattern on the weather at a point, we relate the large-scale atmospheric state to the probability density function (PDF) of temperature/precipitation at a point instead of a single value of temperature/precipitation. This approach takes into account that the large-scale atmospheric state does not completely specify the evolution of the atmosphere at small scales. Rather, the large-scale specifies the range and likelihood of particular outcomes at a point. To create a specific downscaled time series of temperature/precipitation at a point, we draw random numbers from the particular PDF for each individual day in the record. Obviously, there are an infinite number of possible time series given the large-scale atmospheric evolution, and we call these possible outcomes realisations.

For maximum and minimum temperature, we use linear regression to relate large-scale maximum/minimum temperature to local maximum/minimum temperature. In classical linear regression, the errors in the linear ‘fit’ between the predictand and the predictor are assumed to be a Gaussian distribution with variance equal to the error variance. This Gaussian distribution with mean given by the linear ‘fit’ and constant error variance is the PDF of local temperature given the large-scale. Note that linear regression only assumes that the errors from the linear fit are Gaussian. Although the temperature in certain seasons can be strongly skewed, we find that the large-scale temperature is also strongly skewed so that linear regression actually works quite well.

For precipitation, we use two separate statistical models: one for the occurrence of precipitation, to determine if a day is wet or dry, and the other for the amount of precipitation if the day is wet. Because the distributions of both precipitation occurrence and precipitation amount are highly non-normal, we use more flexible statistical models including Generalised Linear Models (GLMs) (Nelder and Wedderburn, 1972). For precipitation occurrence, we use the GLM also known as logistic regression. For precipitation amount, we find that GLMs are not general enough, so we fit the precipitation amount to the generalised gamma distribution (Stacy, 1962) with the scale parameter and one of the shape parameters dependent on the large-scale atmospheric state. Both of these precipitation models are fit using maximum likelihood. We use the 2.5° × 2.5° National Center for Environmental Prediction (NCEP) reanalysis (Kalnay et al., 1996) for the large scale and the National Weather Service's Cooperative Observer Program's (COOP) for the point observations. The COOP stations within 42.1–47.1°N, 93.4–86.6°W that report at least 10 months per year during January 1950–December 2007, and do not exhibit blatant observer biases in precipitation occurrence/weak precipitation (Daly et al., 2007), are included, consisting of a well-distributed network of 170 stations for daily maximum and minimum temperature and 164 stations for daily precipitation. Before applying the fitted relationship between the reanalysis and the COOP stations to the climate models, we quantile adjust the climate models' PDFs of the large-scale predictors to remove climate model biases (e.g. Wood et al., 2004).

For this project we have generated three realisations of gridded daily precipitation and maximum and minimum temperature. Instead of interpolating the raw atmospheric variables from the stations to the grid, the parameters of the PDFs are first interpolated to a 0.1° × 0.1° grid and then random realisations are generated using the gridded PDFs. This basically eliminates the reduction in variance and extremes that typically occurs when one interpolates the raw data.

2.2. Operational snow model

SNOW-17 is an empirically based snow accumulation and ablation model (Anderson, 1973, 2002, 2006), a component of the National Weather Service's River Forecast System (NWSRFS). This operational snow model explicitly treats most of the key physical processes that govern the behaviour of snow dynamics in a conceptual manner, requiring only commonly available data observations of temperature and precipitation as inputs. In treating a column of snow, the model addresses the following principal processes: form of precipitation, accumulation of snow, energy exchange at the snow–air interface, heat exchange at the soil–snow interface, heat storage and deficit within the snowpack, and liquid water retention and transmission of water through the snow cover.

SNOW-17 is a temperature-based model, using air temperature as the sole index for estimating energy exchange at the snow–air interface. Air temperature is a reliable indicator of snowmelt and is a readily available observation. In addition, temperature is easily extrapolated across elevational gradients using lapse rates, which is critical since elevation is the main factor explaining temperature variance across a region, in addition to land cover and land management (e.g. albedo effects), slope, and aspect. Previous studies have shown that the typical near-surface air temperature separating rain from snow is about 1.5 °C (Snow Hydrology, 1956), which can vary somewhat by region. SNOW-17 uses a single threshold temperature, specified by parameter PXTEMP, to diagnose the form of (hourly) precipitation as snow (T < PXTEMP) or rain (T > PXTEMP).

While SNOW-17 is a simplified conceptual model, Franz et al. (2008b) demonstrated that it performs consistently well and is often better than a more physically based energy balance model, the Snow–Atmosphere–Soil Transfer (SAST) model, when previously used over an Idaho watershed. Anderson (1973, 1976) likewise found that SNOW-17 produced results at least as good as energy-aerodynamic models. In general, temperature-based snowmelt models and energy balance snowmelt models exhibit similar skill levels (Ohmura, 2001; Zappa et al., 2003).

SNOW-17 requires careful, regionally specific calibration to produce quality snow simulation results. The model includes 12 key parameters, six of which are considered major parameters and can substantially affect the simulation; the allowable ranges for these parameters are specified by Anderson (2002). The following are the six major parameters:

  • 1.SCF = snow correction factor, or gage catch deficiency adjustment factor.
  • 2.MFMAX = maximum solar melt factor, assumed to occur on June 21.
  • 3.MFMIN = minimum solar melt factor, assuming to occur on December 21.
  • 4.UADJ = average wind function factor for rain-on-snow periods.
  • 5.SI = maximum water equivalent, representing the water equivalent above which 100% snow cover exists.
  • 6.ADC = areal depletion curve, which consists of eleven values and determines the extent of snow cover versus bare ground in a region.

The six minor parameters are as follows:

  • 1.MBASE = melt base temperature for snowmelt computations during non-rain periods.
  • 2.NMF = maximum negative melt factor, impacting heat gain or loss from snow pack.
  • 3.DAYGM = average daily ground melt, treated as a constant daily amount of melt which occurs at the snow–soil interface.
  • 4.PLWHC = percent liquid water-holding capacity for ripe snow.
  • 5.PXTEMP = rain-snow temperature index, used to assess the form of precipitation.
  • 6.TIPM = antecedent temperature index parameter, impacting energy exchange during non-melt periods.

In computing snow accumulation, the form of precipitation is determined by the threshold temperature, PXTEMP. The density of new snow is calculated as a function of air temperature, given that colder air masses tend to support snow with a lower liquid water equivalent. Snow cover is modelled as a single, bulk layer with a specified water-holding capacity. In determining the areal extent of snow cover, the ADC relates the areal snow extent to areal water equivalent based on the maximum water equivalent on record for a region and the water equivalent above which full snow cover exists. Snowmelt is calculated by a different set of equations for rain-on-snow and non-rain periods. For rain-on-snow events, energy and mass balance equations are applied, using several assumptions about meteorological conditions during rain events and the wind parameter UADJ. For non-rain periods, surface snowmelt is determined by temperature, using a melt factor with seasonal variations provided by a sine function using MFMAX for June 21 and MFMIN for December 21. Forest cover has a significant influence on snowmelt rates, and thus Anderson (1996) recommends initial ranges for MFMAX and MFMIN based on forest cover types. The impacts of tall forested canopies on radiation transmission likely outweighs the impacts of any other land management or land use on snow accumulation and melt. SNOW-17 computes changes in snow density, necessary for calculating snow depth, based on compaction, destructive metamorphism, and melt metamorphism (in response to the presence of liquid water).

2.3. SNOW-17 calibration and simulations

Daily maximum and minimum temperature and daily precipitation from the WICCI gridded downscaled dataset for 1961–2000 from the 20C3M scenario is used to force the SNOW-17 model in multiple stages of calibration. The daily maximum and minimum temperature data is converted into hourly temperature data using an interpolatory cubic spline under tension, assigning the minimum temperatures to 9Z (in Zulu time, or Coordinated Universal Time) and maximum temperatures to 21Z each day (based on a climatology of hourly data for Madison, Wisconsin). Daily precipitation amounts are converted into hourly values by randomly distributing the precipitation on wet days into continuous 6-h periods (0–6Z, 6–12Z, 12–18Z, or 18–24Z), with the assumption that a typical precipitation event may last 6 h. The snow model is driven by hourly temperature and precipitation. SNOW-17 output includes daily snowfall, snow depth, snow-water equivalent in the snow pack, snow cover fraction, and rain plus snow melt. To calibrate the model parameters, mean annual totals and the mean seasonal cycles of both snowfall and snow depth from SNOW-17 (driven by climate models' output for 20C3M) are compared to observations from 14 weather stations distributed across Wisconsin. These stations are a subset of 440 U.S. stations considered homogeneous and of the highest quality by the plurality of seven expert judges in the study by Kunkel et al. (2009). After completing the model calibration at 14 stations to determine the ideal choices in parameter values, SNOW-17 is run, using those parameter values, for a series of modern and future simulations and its performance is evaluated against an extensive set of 189 stations, as described in Section 3.2.

Regarding the temperature for distinguishing snow from rain events, Anderson (2002) states that PXTEMP = 1.0 °C should be adequate. This was investigated using hourly weather observations and temperature measurements for 1980–1999 from three Wisconsin locations that span a climate gradient, Milwaukee, Rhinelander, and LaCrosse, in the NCDC Global Surface Hourly database. For these stations, it is found that an equal likelihood of snow or rain consistently occurs around 1.3–1.4 °C, so a value of PXTEMP = 1.3 °C is selected.

Given that snowmelt is sensitive to forest cover and type (Anderson, 1996, 2002; NWSRFS User's Manual), values of percent total tree (TREE), evergreen tree (EVE), and deciduous (DEC) tree cover are interpolated over the 0.1° × 0.1° grid using data from the Advanced Very High Resolution Radiometer (AVHRR) Continuous Fields Tree Cover Project (DeFries et al., 1998, 2000; Hansen et al., 2000). Similar to those recommended by Anderson (2002), the following values for MFMAX and MFMIN are applied, which result in favourable agreement between simulated and observed snow climatologies at the 14 calibration stations:

  • 1.Bare or sparsely forested (TREE < 33%): MFMAX = 1.7 mm/ °C/6 h, MFMIN = 0.4 mm/ °C/6 h.
  • 2.Evergreen forest (TREE > 33%, EVE–DEC > 20%): MFMAX = 0.5, MFMIN = 0.2.
  • 3.Deciduous forest (TREE > 33%, DEC–EVE > 20%): MFMAX = 1.1, MFMIN = 0.1.
  • 4.Mixed forest (TREE > 33%, |DEC–EVE | < 20%): MFMAX = 1.2, MFMIN = 0.15.

Snowmelt is most rapid in regions with minimal tree cover and is slowest within dense conifer forests, related to available sunlight at the surface.

The average wind function factor for rain-on-snow periods, UADJ, can be estimated by UADJ = 0.002 × U, where U is the 6-hourly wind travel in km at 1 m above the snow surface (Anderson, 1976). A dataset of recommended UADJ values for the United States (N. Mizukami, NOAA/NWS, 2009, personal communication), based on a monthly 10-m wind climatology from the North American Regional Reanalysis (Mesinger, 2006), is used to provide UADJ values for Wisconsin grid cells; values range from 0 to 0.04 mm/hPa, with forested areas experiencing lower wind speeds due to greater roughness (Anderson, 2002).

Elevation for each grid cell is retrieved from a topography dataset for the United States from GEON (geon.unavco.org); elevation is used to estimate atmospheric pressure, which affects the sensible heat flux calculations. The ADC, labelled as ‘curve B’ and identified as the most common curve in the NWSRFS User's Manual, is used for Wisconsin. Given that the downscaling approach corrects for biases in temperature and precipitation, the SCF, is set to 1, with no need for additional correction. The maximum water equivalent, SI, is assigned the value of 30 mm (typically equivalent to one foot of snow), representing the amount of snow-water equivalent needed for 100% snow cover; the estimation of 30 mm is supported by the study of Niu and Yang (2007), who developed a formula relating snow cover fraction and snow depth using satellite data over the major North American river basins. As recommended by Anderson (2002), NMF = 0.15 mm/ °C/6 h, MBASE = 0 °C, and PLWHC = 0.04. The antecedent temperature index parameter, TIPM, is specified as a function of latitude, ranging from 0.27 in southern Wisconsin to 0.12 in northern Wisconsin, to permit larger values of TIPM for regions with shallower snow covers (Anderson, 2002). Likewise, the average daily ground melt, DAYGM, is specified as a function of latitude, ranging from 0.1 mm/day in southern Wisconsin to 0 in northern Wisconsin, based on the recommendation from Anderson (2002) that DAYGM be assigned values close to zero if the soil is generally frozen under the snow. Finally, the value of the threshold density is set to 0.15 gm/cm3, a value used in earlier versions of SNOW-17 (Anderson, 2006), which produced deeper, more favourable snow depths.

SNOW-17, with parameter values identified above, and driven with 20C3M simulations, each forced with one of 15 CMIP3 global climate models, is compared against climatological snowfall and snow depths data from 1961 to 2000 for 14 Wisconsin weather stations in the Kunkel et al. (2009) database (Figure 2). Among these 14 Wisconsin sites, the mean observed annual snowfall ranges from 103.2 cm at the southernmost station, Richland Center, to 159.4 cm at the northernmost station, St. Germain, with the likelihood of precipitation to fall in frozen form increasing with latitude and the gradual rise in elevation towards the north. A comparison of annual mean snowfall between SNOW-17 and observations at these 14 sites yields a correlation of 0.77 (significance level p < 0.01), a root-mean-square-difference of 11.2 cm, and a typical percent difference (defined as the mean of the absolute value of the differences) of 7.7%. Overall, there is good agreement between annual mean snowfall totals between SNOW-17 and observations, although there is a slight tendency for the model to produce too much snowfall.

Figure 2.

(a) Comparison of the mean seasonal cycle (October–May) of (top row) snowfall and (bottom row) snow depth, given in both inches and cm, at four select Wisconsin cities between station observations (dashed—data from Kunkel et al., 2009) and SNOW-17 output for 20C3M (solid), for 1961–2000. Monthly mean (b) snowfall (cm) and (c) snow depth (cm) are compared between station observations and SNOW-17 output for all 14 Wisconsin stations from the Kunkel dataset, where each dot represents a single month during October–May from the snow climatology at a single station. The correlations for (b) and (c) are 0.96 and 0.94, respectively (N = 112)

Among the 14 Wisconsin sites, the mean observed November–April snow depth ranges from 7.1 cm at Richland Center to 27.2 cm at St. Germain. A comparison of mean November–April snow depth between SNOW-17 and observations, at these 14 sites, results in an across-station correlation of 0.86 (N = 14 stations, p < 0.01), a root-mean-square-difference of 3.0 cm, and a typical percent difference of 15.7%. Using 1 in. (2.54 cm) of snow depth as the criterion, the observed snow season across Wisconsin begins in mid-late November and ends in mid-March to mid-April. SNOW-17 tends to produce too little snow depth across northern Wisconsin, where depths are usually the greatest (Figure 2c).

There is a much greater variation among stations in the timing of the end of the snow season than the start of the snow season. More specifically, the observed standard deviations among the 14 sites of the start and end dates of the mean snow season (defined by at least 1 in. of snow pack) are 4.2 and 7.5 days, respectively. In late autumn–early winter, a single snow event can cover much of the state in snow, yet, in late winter–early spring, a cumulative period of warmth is necessary to melt away the present snow pack, particular in the northern Wisconsin sites. In comparing the simulated and observed snow season at the 14 sites, the root-mean-square-differences for the mean start and end dates of the snow season are 5.6 and 5.8 days, respectively; the model performs equally well at the start and end of the snow season. Cross-site correlations between the model and observations at the 14 sites for start date, end date, and duration of the snow season are 0.80, 0.89, and 0.86, respectively (N = 14 sites, p < 0.01). The observed (simulated) snow season duration ranges from 16 (19) weeks at Richland Center to 22 (22) weeks at St. Germain. Overall, simulated and observed snow depths agree fairly well, although the duration of the simulated snow pack is too long and the simulated depth is typically too shallow, particularly for the northernmost sites like St. Germain (Figure 2a,c).

Given the satisfactory agreement between observed and simulated snowfall and snow depth at the 14 calibration sites, the aforementioned parameter values are used for an array of experiments, in which SNOW-17 is driven by downscaled gridded temperature and precipitation across the Wisconsin domain for the IPCC AR4 emission scenarios of 20C3M for the late 20th century and A2, A1B, and B1 for the mid- and late 21st century. Greenhouse gas emissions are largest with the A2 scenario and smallest with the B1 scenario by the end of this century. For all scenarios, downscaled output from 11 to 15 global circulation models is used to drive SNOW-17, using three realisations of temperature and precipitation from their PDFs for each model to produce an ensemble of potential snow projections.

3. Results and discussion

3.1. Downscaled climate projections

The IPCC AR4 robustly agree that wintertime in Wisconsin will become substantially warmer and wetter during the 21st century in response to rising levels of enhanced greenhouse gases (Figure 3). By the mid-21st century, mean DJF temperature is projected to increase by 3.2 °C under the B1 scenario or 4.0 °C under the A2 scenario, based on the WICCI downscaled climate data. These scenarios diverge over time, with 4.5 °C warming by the end of the 21st century under the B1 scenario but a much more substantial 6.9 °C warming under the A2 scenario.

Figure 3.

Projections of the DJF change in Wisconsin's mean (a–d) temperature ( °C) and (e–h) precipitation (cm), based on the B1 and A2 emission scenarios and presented for both the mid- and late 21st century. Results are shown for the regridded (red dash) and downscaled (blue solid) data and are based on kernel density estimates using 14 GCMs for B1 and 11 GCMs for A2. The small tables for each plot show the 10% (90% of models expect a larger increase), mean, and 90% (10% of models expect a larger increase) projections for both the regridded and downscaled projections

These downscaled temperature changes, based on the detailed statistical downscaling method described in the previous section, are compared to temperature projections using a simple regridding of the IPCC AR4 models' output onto the same 0.1° WICCI grid (Figure 3a–d). This comparison serves to verify that the downscaling was performed reasonably and to assess what new information was acquired versus a simple interpolation. A close agreement regarding projected DJF temperature changes for Wisconsin by the late 21st century is evident between the downscaled and regridded approaches, with correlations of 0.91 (N = 14) and 0.80 (N = 11) among the 11–14 climate models' projections for the B1 and A2 scenarios, respectively. The mean warming, among the climate models, is consistently greater with the downscaling approach, mainly because the downscaled models are skewed towards greater warming. For example, for the late 21st century under the A2 scenario (Figure 3d), the 90% level of projected temperature changes (10% chance that the actual warming will be greater than this level) is + 8.2 °C from the regridding approach and + 9.9 °C from the downscaling approach. We speculate that, in the case of GCMs with too little temperature variability over Wisconsin compared to observed, the statistical downscaling method would increase the variability in order to correct towards observed, thereby making the climate more sensitive and amplifying the temperature projections. For higher emission scenarios (A2 vs B1 and late 21st century vs mid-21st century), the spread among temperature projections is greater, with increased uncertainty.

There is little difference in terms of the mean DJF precipitation change projected for Wisconsin for the mid-21st century between the A2 and B1 scenarios, with increases of 1.1 and 1.2 cm, respectively, from the WICCI downscaled data. The projected increase in DJF precipitation by the late 21st century differs more substantially between scenarios, with an increase of 1.5 cm for the B1 scenario and 2.1 cm for the A2 scenario.

Projected DJF precipitation changes for Wisconsin by the late 21st century closely agree between the downscaled and regridded approaches (Figure 3e–h), with correlations of 0.90 (N = 14) and 0.91 (N = 11) among the 11–14 climate models' projections for the B1 and A2 scenarios, respectively. Generally, the downscaled precipitation projections exhibit more agreement among the climate models, as evident by a narrower PDF.

The mean projected changes in surface air temperature and precipitation are shown spatially for the mid- and late 21st century by season in Figure 4. The projected warming is most notable by the end of the century under the A2 emission scenario, especially during winter over northern Wisconsin. An increase in cool season precipitation is robust among the models, particularly during spring over northern Wisconsin.

Figure 4.

Mean projected changes (averaged among 11–14 GCMs) in seasonal surface air temperature ( °C) and precipitation (cm/season) for the mid- and late 21st century under the B1 and A2 emission scenarios

3.2. Model performance

In the previous section, observed snow data at 14 homogeneous sites from the Kunkel et al. (2009) database were used to calibrate SNOW-17 across Wisconsin, leading to a favourable set of parameter values for the model and the creation of an extensive ensemble of modern and future simulations. The statewide performance of a calibrated SNOW-17 model is now compared to annual mean snowfall climatologies at 189 stations across Wisconsin from the NCDC State Snow Climatology and Extremes dataset (Figure 5). Comparisons are made between the model and observations both at the individual stations and on a grid, in which the observations are fit to the same 0.1° WICCI grid.

Figure 5.

(a) Mean snowfall based on SNOW-17 simulations for years 1981–2000 of 20C3M, shown in inches and cm. (b) Mean snowfall from 189 stations (for the entire period of record for each station) using the NCDC State Snow Climatology and Extremes dataset, fit to the same WICCI grid. (c) Standard deviation of annual mean snowfall based on SNOW-17 simulations for 20C3M related to the late 20th century, shown in inches and cm. (d) Standard deviation of observed annual snowfall from 148 stations using the NCDC DS3220 monthly summary dataset, fit to the WICCI grid

Spatially, the simulated and observed annual snowfall maps agree reasonably well, with a spatial correlation of 0.88 (p < 0.01) for 1756 Wisconsin land grid cells (observations at 189 stations are fit to the 0.1° WICCI grid using tension spline interpolation). The Wisconsin mean snowfall is 128.5 cm in the model and 121.3 cm in the observations, with 6% too much snowfall simulated statewide. The largest simulated biases are about 20 cm too much snowfall along the southern edge of the state and 20 cm too little snowfall along the northern edge. The correlation between annual mean snowfall in the SNOW-17 simulations of 20C3M and the observations at the 189 stations is 0.80 (p < 0.01) (Figure 6). One notable bias is the simulation of too much snowfall at the sites with the least observed snowfall (in southern Wisconsin) and too little snowfall at the sites with the most observed snowfall (in northern Wisconsin); this may be related to the use of the same PXTEMP value statewide. Overall, both the calibration analysis at 14 sites and the performance evaluation at 189 sites yield favourable comparisons of the SNOW-17 simulations with observations.

Figure 6.

Scatter plot of the annual mean snowfall, in both inches and cm, for 189 Wisconsin stations, comparing the station observations from the NCDC State Snow Climatology and Extremes dataset to output from SNOW-17 for 20C3M; results are shown for the late 20th century. The curved line represents a curve fit for the data

The standard deviation of annual snowfall in the SNOW-17 simulations of 20C3M agrees well spatially with the observed standard deviation (using NCDC DS3220 dataset), with a spatial correlation r = 0.79 (N = 1756 Wisconsin grid cells, p < 0.01). However, the simulated standard deviation is about 17% too large statewide, most notably across southern Wisconsin.

3.3. Multiple realisations

In this study, for each GCM and emission scenario, three realisations of downscaled climate projections are considered from the entire PDF of potential local responses to the large-scale circulation pattern; the realisations differ by a stochastic component, given that the large-scale atmospheric pattern does not fully explain the temperature and precipitation at a site. By considering an array of climate models and multiple realisations, a careful assessment of the range of potential climate change impacts in Wisconsin may be developed.

To illustrate the importance of considering multiple realisations from the downscaled data, two case studies of Wisconsin storms are presented from the 20C3M simulation of the high-resolution version of Model for Interdisciplinary Research on Climate (MIROC3.2-hires: Hasumi and Emori, 2004) global circulation model (model is arbitrarily chosen); case studies in which the realisations differ substantially are purposely selected. Recall that each realisation of the downscaled climate data consists of a ‘best guess,’ that is uniquely determined by the large-scale pattern, and a random component that varies between realisations but is consistent with the downscaled probability distribution.

In both case studies, the large-scale weather pattern consisted of a surface cyclone over Arkansas and surface anticyclone over either the Southeast United States or just off the Atlantic coast of the southeastern states, favouring southerly winds and moisture transport from the Gulf of Mexico into Wisconsin. The first case study occurs on 10 January 1961 (this date is relevant to the simulation, not the observed record) in the MIROC3.2-hires 20C3M simulation (Figure 7). This case study illustrates the fact that the same large-scale atmospheric pattern can produce differing precipitation patterns across Wisconsin in the WICCI downscaled dataset. In all three realisations, substantial precipitation falls across southern Wisconsin in response to moisture advection from the Arkansas cyclone. However, the peak precipitation is located in southwest Wisconsin in the first realisation (Figure 7d), south-central Wisconsin in the second (Figure 7e), and more disperse in the third (Figure 7f). Given the widespread cold conditions across the state on that date, the SNOW-17 simulations show that all the precipitation falls as snow, as evident by the close spatial match between the precipitation maps and snowfall maps in Figure 7.

Figure 7.

(a–c) Daily snowfall (cm) and (d–f) precipitation (cm) for three individual realisations (rows) of a storm case study on 10 January 1961 as simulated by MIROC-hires in a 20C3M simulation. Note the irregular contour intervals in the colour bar

The second case study occurs on 13 January 1970 in the MIROC3.2-hires 20C3M simulation (Figure 8). This case study demonstrates that temperature differences among the realisations can lead to substantially different snowfall simulations given the same large-scale circulation pattern. In the first and second realisations, nearly all precipitation falls as snow according to the SNOW-17 simulation since most of the state is below the cutoff temperature of 1.3 °C. In the third realisation, the downscaled precipitation pattern includes a vast area of precipitation across southern Wisconsin and two localised areas in the northwest and northeast state (Figure 8f). Given that this realisation includes mild temperatures in excess of 1.3 °C across the region of vast precipitation in southern Wisconsin (Figure 8i), it nearly all falls as rain, with just the local precipitation areas in the northwest and northeast state falling as snow (Figure 8c).

Figure 8.

(a–c) Daily snowfall (cm), (d–f) precipitation (cm), and (g–i) temperature ( °C) for three individual realisations (rows) of a storm case study on 13 January 1970 as simulated by MIROC-hires in a 20C3M simulation. Note the irregular contour intervals in the colour bar

3.4. State-averaged snow projections

Projections for Wisconsin-average snowfall, snow depth, and snow cover are now described, based on SNOW-17 simulations forced by three realisations of downscaled WICCI temperature and precipitation data for 11–14 global climate models for scenarios A2, A1B, and B1 (Figure 9). A comparison of the A2, A1B, and B1 snow simulations with those from 20C3M shows reductions in Wisconsin snowfall of 34.0 cm (−28%), 36.1 cm (−30%), and 25.4 cm (−21%), respectively, by the mid-21st century and 57.9 cm (−47%), 53.6 cm (−44%), and 37.4 cm (−31%), respectively, by the late 21st century. The greatest snowfall reductions are projected at the flanks of the snowfall season in November and March–April, representing a shortening of the season; likewise, the largest projected decreases in Wisconsin snow depth and cover occur in February–March.

Figure 9.

Projected (October–May) Wisconsin-averaged (a) snowfall, (b) snow depth, and (c) snow cover, in both inches and cm, for ‘modern’ (late 20th century—black), mid-21st century for B1 (green) and A2 (orange) scenarios, and late 21st century for B1 (blue) and A2 (red) scenarios, as simulated by SNOW-17. For example, in (a), the red curve represents snowfall averaged spatially across Wisconsin and averaged among 11 GCMs and 3 realisations for each GCM

According to the SNOW-17 simulations for 20C3M, Wisconsin, on average, maintains a snow pack of at least 2.54 cm (1 in.) for 140 days, from mid-November to mid-April, peaking in mid-February around 19 cm (Figure 9b). In other words, the mean climatology of a location in Wisconsin includes a 140-day period with at least 1 in. of snow on the ground. The average duration of a snow pack of at least 1 in. is projected to reduce by 16, 24, or 28 days by the mid-21st century according to scenarios B1, A1B, and A2, respectively, and by 28, 44, and 50 days by the late 21st century for these scenarios. In most cases, the snow season shortens more in the spring than in the autumn. Similarly, according to the SNOW-17 simulations for 20C3M, Wisconsin maintains a snow cover of at least 50% from late November to late March, representing 119 days (Figure 9c). In other words, for a period of 119 days, at least half of the state is covered by snow on average. The average duration of a snow cover of at least 50% is projected to diminish by 22, 38, or 34 days by the mid-21st century according to scenarios B1, A1B, and A2 and by 37, 58, and 68 days by the late 21st century for these scenarios. Based on the A2 scenario, by the late 21st century, an average Wisconsin snow cover of at least 50% might be limited to late December through mid-February, totaling only 50 days. Snow cover may be a better measure of the Wisconsin snow season than snow depth given that the latter is largely weighted by the heavy snowpack of northern Wisconsin.

Note the dual peak in Wisconsin snowfall evident both in the SNOW-17 simulations in Figure 9a and also in the observations (Figure 2), particularly true of the northern state; this dual peak consists of maxima in snowfall during January and March and a local minimum during February. Compared to January, temperatures also primarily remain below freezing in February but precipitation is less abundant. By March, precipitation tends to be substantially higher, given the larger moisture holding capacity of warmer air and greater climatological ascent in the atmospheric column; events in March often either occur as heavy snowstorms or rain/mixed precipitation events due to higher temperatures. Reflecting this seasonal pattern, Changnon et al. (2006) found that the greatest number of Wisconsin snowstorms was observed in March. By the late 21st century, especially under the A2 scenario, the secondary snowfall peak in March becomes much less distinct (Figure 9a), as March precipitation events become much more likely to fall in the form of rain.

Based on observations from Kunkel et al. (2009) and SNOW-17 output for the late 20th century and late 21st century (A2 scenario) at the 14 stations, the typical distribution of daily snowfall amounts, for events with daily snowfall of at least 1 cm, in Wisconsin is shown in Figure 10. Observations reveal that Wisconsin typically receives 22.1 days per year with at least 1 cm of snowfall, including 9.9 days with just 1–3 cm of snowfall (the frequency is first computed at each of 14 stations and then averaged). The SNOW-17 simulations using 20C3M downscaled climate data produce 23.7 days per year with at least 1 cm of snowfall, close to observed. The SNOW-17 simulations had too many days with snowfall amounts of 1–5 cm but too few with amounts of 5–11 cm. Under the A2 scenario, by the end of the 21st century, the frequency of snowfall days, of at least 1 cm, is projected to decrease by 45.1% to 13.0 days per year across Wisconsin. There is a tendency for larger percent decreases in the frequency of days with heavier snowfall amounts. For instance, the frequency of 1–3 cm snowfall days in Wisconsin is projected to decrease by 41% by the end of the century (A2 scenario), while the frequency of 15–17 cm snowfall days is projected to decrease even more, by 54%.

Figure 10.

Distribution of daily snowfall amounts, based on averaging among 14 stations (Kunkel et al., 2009) scattered across Wisconsin, based on late 20th century station observations (green) and simulations by SNOW-17 for the late 20th century (blue) and late 21st century (red—A2 scenario). Values in (a) are the number of snowfall days per year, while (b) shows the log of the frequency. The difference between the simulated frequency of snowfall days in the late 21st century and late 20th century is shown in black. Daily snowfall is shown in 2 cm bins, such that ‘10 cm’ represents 9–11 cm

Changnon (1969) and Changnon and Changnon (1978) found that Midwest snowstorms that produced at least 15.2 cm (6 in.) of snowfall resulted in major economic and human impacts. It appears, from the present study, that the frequency of such potent snowstorms should diminish across Wisconsin during the 21st century. According to Figure 10, Wisconsin typically receives 15–27 cm of snowfall about 9.3 days per decade in the observations and 9.6 days per decade in the SNOW-17 simulations for 20C3M, suggesting an accurate model result. However, based on the A2 scenario, the frequency of such devastating snowstorms should decrease to 4.8 days per decade, representing a 50% reduction in their occurrence.

Projected changes in Wisconsin-averaged snowfall by the end of the 21st century, from the SNOW-17 simulations based on the A2 scenario, range from − 30.5 cm for the Commonwealth Scientific and Industrial Research Organization Mark version 3.5 (CSIRO-Mk3.5: Gordon et al., 2002) to − 86.4 cm for MIROC3.2-medres (Hasumi and Emori, 2004) (Figure 11). Of the 11 GCMs, MIROC3.2-medres features the largest projected November–April warming, + 9.3 °C, and only a small increase in November–April precipitation, + 3.5 cm. Alternatively, a more modest snowfall reduction of − 30.5 cm is projected for Wisconsin according to SNOW-17 simulations for CSIRO-Mk3.5, given a weaker November–April warming of + 5.3 °C and a substantial increase in November–April precipitation by + 6.7 cm. Nonetheless, all SNOW-17 simulations forced by downscaled IPCC model output indicate a projected reduction in Wisconsin snowfall. The correlation between projected changes in November–April temperature and annual snowfall for three realisations of 11 GCMs under the A2 scenario is − 0.90 (N = 11 models × 3 realisations = 33) (Figure 11a), which distinctly relates larger warming with larger snowfall reductions. Likewise, the correlation between projected changes in November–April precipitation and annual snowfall is + 0.65 (Figure 11b), suggesting that projected increases in cold-season precipitation will favour smaller reductions in snowfall.

Figure 11.

(a) Scatter plot of the projected changes in Wisconsin-averaged November–April temperature ( °C) and annual snowfall (cm), based on comparing the late 21st century under the A2 scenario with the late 20th century under the 20C3M scenario. Each individual dot represents one of three realisations for a CMIP3 GCM. (b) Same as (a), except comparing projected changes in November–April precipitation (cm) and annual snowfall

The benefit of considering multiple realisations of climate projections is clear in Figure 11b. The range of November–April temperature trends for Wisconsin among three realisations for a single GCM is typically 0.06 °C (2081–2100 vs 1961–2000 under the A2 scenario), suggesting that each realisation agrees closely with the others in terms of temperature trends. However, this is not true of precipitation trends among realisations. For a single GCM, the typical range of November–April precipitation trends, among three realisations, is 1.13 cm. In the most extreme example of differences among realisations, ECHAM5 simulates a Wisconsin-average increase in November–April precipitation of 7.53 cm in one realisation and 9.82 cm in another realisation. Given the stochastic nature of precipitation, there is a significant benefit to downscaling climate data as a PDF and considering multiple realisations in assessing potential future precipitation trends. Snowfall projections for Wisconsin are sensitive to variations in precipitation trends among realisations of the climate projections. For instance, when three realisations of the downscaled projections of temperature and precipitation from CNRM-CM3, under the A2 scenario, are used to force SNOW-17, the projected reduction in annual snowfall by the late 21st century ranges from − 49.43 cm to − 60.52 cm, largely in response to significant variations in precipitation trends among the realisations.

3.5. Spatial snow projections

By the mid-21st century, Wisconsin's mean snowfall is projected to decrease by 25.4 cm (−21%) or 34.0 cm (−29%), based on SNOW-17 simulations for scenarios B1 and A2, respectively. By the late 21st century, snowfall is simulated to decrease by 37.3 cm (−31%) or 57.9 cm (−49%) for these scenarios. The snowfall projections substantially diverge between the two scenarios by the end of the century. In general, the absolute reduction in snowfall is greatest across northern Wisconsin, while the percent reduction is largest across southern Wisconsin where annual mean snowfall is lowest (Figure 12).

Figure 12.

Projected changes in mean snowfall, shown as (a–d) absolute differences (cm) and (e–h) percent differences. Differences are based on comparing the late 20th century simulations from 20C3M with both mid- and late 21st century simulations for scenarios B1 and A2

The largest reduction in snow depth is projected for late winter, representing a shortening of the snow season. To illustrate this point, Figure 13 displays the simulated changes in the mean snow depth for mid-December, early February, and mid-March by the mid- and late 21st century, representing the early, mid-, and late Wisconsin snow season. By the mid-21st century, the mean snow depth on March 15 is projected to decrease by either 6.9 cm (−48%) or 9.4 cm (−68%), for scenarios B1 and A2, respectively. By the late 21st century, the projected reduction in snow depth for these scenarios is 10.2 cm (−68%) or 12.4 cm (−85%). Similar to changes in snowfall in Figure 12, the simulated reduction in mean snow depth for mid-March is greatest in northern Wisconsin, while the percent reduction is larger in southern Wisconsin. The projected percent reduction in snow depth, throughout the entire winter season, is substantially greater than the percent reduction in snowfall, given that the reduction in snow depth reflects both diminished snowfall and accelerated snowmelt.

Figure 13.

Projected changes in mean December 15th snow depth (based on December 10–20th), shown as (a–d) absolute differences (cm) and (e–h) percent differences. Differences are based on comparing the late 20th century simulations from 20C3M with both mid- and late 21st century simulations for scenarios B1 and A2. Similar plots are shown in (i–p) for February 1st (based on January 27–February 6th) and in (q–x) for March 15th (based on March 10–20th). The three time periods represent early, mid-, and late in the Wisconsin snow season

4. Conclusions

The observational record offers strong evidence that warming has resulted in diminishing snow cover, from across the globe down to the state of Wisconsin, while the CMIP3 global climate models simulate substantial snow losses throughout this century. This is of particular concern to wildlife scientists across Wisconsin, given numerous studies that have linked the behaviour, survival, and competition of birds and animals to the presence and depth of snowpack. In response to the demand among WICCI wildlife researchers for high-resolution, sub-county level projections of daily snowfall, snow depth, and snow cover, the SNOW-17 snow model is forced with downscaled climate data from the CMIP3 models to generate the necessary downscaled snow projections.

The downscaling method, by which 0.1° × 0.1° projections of daily maximum and minimum temperature and precipitation are developed for Wisconsin from 15 global climate models, corrects not only the mean, but also the variance and extremes. The method is based on PDFs, thus leading to multiple realisations of the local weather in response to the large-scale atmospheric pattern. Based on the downscaled data, by the end of the 21st century, Wisconsin will warm by an average (among 11–14 GCMs) of 4.5 or 6.9 °C and receive 1.5 or 2.1 cm more precipitation, according to the B1 and A2 emission scenarios. After carefully calibrating the parameters of SNOW-17 against observations, the snow model is forced by this downscaled climate data for the late 20th century, mid-21st century, and late 21st century. This study represents the first statewide application of the stand-alone SNOW-17 model for simulating future snow patterns in response to climate change. SNOW-17 generally produces a reasonable distribution of mean snowfall across Wisconsin, although it underestimates the higher observed snowfall amounts across the northern state and under-simulates in southern Wisconsin. For each GCM and each scenario, SNOW-17 is forced by three unique realisations of the WICCI downscaled data, which allows for multiple evolutions of the precipitation and derived snowfall patterns for a given large-scale atmospheric circulation pattern; consideration of multiple GCMs, scenarios, and realisations allows for a better estimation of the range of possible climate change and supports probabilistic risk assessments. For a single GCM, different realisations of the statistically downscaled climate data can be characterised by moderately different precipitation trends, due to the stochastic nature of precipitation, significantly affecting snowfall projections from SNOW-17; this supports the PDF-based downscaling approach and the need to analyse multiple realisations of climate projections.

The length of the Wisconsin snow season is projected to dramatically decline, with the largest reductions in snowfall likely in springtime. By the end of the 21st century, mean snowfall is projected to decline by 31–49%, based on the B1 and A2 scenarios, with the greatest percent reduction in frequency for heavier snowstorms. By the end of this century, mean snowfall is projected to diminish, even among SNOW-17 simulations driven by CMIP3 models with projected increases in winter precipitation. Reductions in snow depth will likely be even more pronounced, as more precipitation falls in the form of rain and snowmelt is accelerated. In particular, mid-March snow depth, averaged across Wisconsin, is projected to decrease by 68% or 85%, based on B1 and A2 scenarios, by the end of the 21st century. Spatially, the largest reductions in snowfall are likely in northern Wisconsin, while the greatest percent reductions are likely in southern Wisconsin. While white-tail deer may benefit from a loss of snowfall, these 21st century snow projections for Wisconsin suggest a serious threat to the state's wildlife, including the snowshoe hare, ermine, ruffed grouse, American marten, and Canada lynx, thus supporting the need for wildlife impact studies and the development of adaptation strategies.

Clearly the impact of reduced snowfall in Wisconsin will extend beyond wildlife into many aspects of the state, including its hydrology and agriculture. Reduced snow cover and earlier snowmelt across northern and central Wisconsin will likely cause the soils to warm up and dry out earlier in spring, allowing farmers to drive their machinery onto the soil sooner to begin corn and soybean planting. However, alfalfa, wheat, and some other crops may become vulnerable to winterkill without the protection of the snow pack (Leep et al., 2001).

Several limitations are identified in this study. Regarding the downscaling methodology, the lack of consideration of lake ice and wind likely impaired the ability to accurately represent lake-effect snowfall across Douglas and Bayfield Counties along Lake Superior and Door County along Lake Michigan. However, unlike Michigan, the snowfall patterns across Wisconsin are not dominated by lake-effect snow, given the state's position on the windward side of Lake Michigan. Several challenges are noted when applying SNOW-17 on a statewide grid, including the calibration of grid-scale output against localised station observations and the difficulty in assigning appropriate parameter values spatially across the state (e.g. PXTEMP). Other limitations are model-specific, such as the inability of a temperature-based model to accurately represent heavy snowmelt during high wind events, given wind is not an input variable.

Acknowledgements

This study was funded by the Wisconsin Focus on Energy and the University of Wisconsin-Madison Nelson Institute. The authors are grateful to Naoki Mizukami (NOAA) for supplying parameter datasets for SNOW-17, Mark Raleigh (University of Washington) for assistance with the use of the model, Ken Kunkel for providing snow data, and members of the WICCI Wildlife Working Group, namely Michael Meyer, Keith McCaffery, James Woodford, and Karl Martin, for insight into animal–snow relationships. CMIP3 data was obtained through PCMDI. Comments from two anonymous reviewers were very beneficial. CCR Contribution #993.

Ancillary

Advertisement