A methodology for statistically downscaling seasonal snow cover characteristics over the Northeastern United States


  • Lee Tryhorn,

    1. Northeast Regional Climate Center, Cornell University, Ithaca, NY, USA
    2. New York State Water Resources Institute, Cornell University, Ithaca, NY, USA
    3. Department of Earth and Atmospheric Sciences, Cornell University, Ithaca, NY, USA
    Search for more papers by this author
  • Art DeGaetano

    Corresponding author
    1. Department of Earth and Atmospheric Sciences, Cornell University, Ithaca, NY, USA
    • Northeast Regional Climate Center, Cornell University, Ithaca, NY, USA
    Search for more papers by this author

Correspondence to: A. DeGaetano, Northeast Regional Climate Center, Cornell University, Ithaca, NY 14853, USA. E–mail: atd2@cornell.edu


Snow cover amount and duration have a large role in both the natural and anthropogenic systems of the Northeast United States. Any changes in the climatology of the amount, duration, and timing of the snowpack may have long-lasting environmental and economic consequences. At present, the principal tools used to examine future climatic changes are general circulation models; however, they do not provide information at the scales required for investigating many of the impacts in which snow plays a major role. In order to address this gap, a statistical downscaling methodology for simulating climatological snow cover parameters was developed. The methodology utilizes the Statistical DownScaling Model to construct climate change scenarios at daily time scales that are subsequently evaluated based on their ability to reproduce seasonal snow cover statistics. Two different observational datasets (station and gridded) were used as predictands and empirical relationships were created between the predictands and regional scale predictors. The methodology was tested at 20 locations across the Northeast and was then applied to output from HadCM3 under the A2 and B2 emissions scenarios. The methodology performed well at capturing key properties of station-based snow cover over a range of climates and was found to perform better than the technique used in previous work in the Northeast. By the end of the century, the projections revealed characteristics that are consistent with declining snow cover, yet there are likely to be regional variations in the next several decades, especially when elevation is considered.


Snow is an important part of the terrestrial climate system (Räisänen, 2008) and any changes in the climatological amount, duration, and timing of the snowpack can have long-lasting environmental and economic consequences. Snow cover has significant effects on climate, such as regional radiative and thermal energy budgets and atmospheric circulation (Groisman et al., 1994a; Frei et al., 1999), biological systems, including species and community distributions (Walker et al., 1993, 1999), and human processes, such as winter tourism, recreation, and water supply (Hamilton et al., 2003; Scott and McBoyle, 2007). Within the northern United States rapid melt of the snowpack is a major cause of flooding (Todhunter, 2001; Graybeal and Leathers, 2006). Persistent changes in snow accumulation or melt can therefore have significant ecologic and economic consequences. In addition, snow cover is a climatically sensitive hydrologic variable and may be a useful indicator of climatic changes (Frei and Robinson, 1999; Frei et al., 1999; Brown, 2000).

Through its role in modifying surface albedo, snow cover is integrally linked with observed changes in global climate, especially for Northern Hemisphere (NH) land areas (Groisman et al., 1994b; Qu and Hall, 2006; Brown and Mote, 2009). Snow cover is generally anticipated to decrease in response to global warming, as snow cover formation and melt are closely related to a temperature threshold of 0 °C. However, the response of snow to global warming is complicated by projected increases in precipitation in certain regions (such as the Northeast) (Räisänen, 2008). Although NH snow extent has decreased during the past four decades, regional trends in snow conditions have been variable (Lemke et al., 2007; Brown and Mote, 2009). The largest decreases have typically been observed at lower elevations (Mote et al., 2005; Mote, 2006). While at some higher elevation locations there is evidence of increased snow accumulation (Mote et al., 2005; Regonda et al., 2005) in response to increasing precipitation (Zhang et al., 2007). Some studies have also identified an upward trend in lake-effect snowfall near the Great Lakes, due to warmer surface waters and decreased ice-cover (Norton and Bolsenga, 1993; Burnett et al., 2003). The snow cover response to warming is therefore likely to vary, with potential for increased accumulation in cold regions where increases in precipitation are sufficient to offset reductions in the length of the snow season (Brown and Mote, 2009).

How snow cover will continue to change is of great interest to natural resource managers, the winter tourism industry, governments, and the general public. However, climate change information is needed at a much finer spatial scale for impact and adaptation studies than what is currently provided by most general circulation models (GCMs) (whether regional or global) (Wilby et al., 2004). In addition, snow cover is sensitive to fine-scale climate forcings, which are not explicitly incorporated in GCMs. Moreover, the smoothed topography of GCMs means that orographic precipitation in general, and snowfall in particular, are among the most difficult variables to simulate in climate models. This makes the quantification of possible changes to snow cover in coming decades extremely challenging.

Dynamical and statistical downscaling techniques offer approaches for obtaining climate information at a high resolution. Statistical downscaling constitutes an empirical method to derive relationships between large-scale atmospheric variables (predictors) and observed local weather variables (predictands). These relationships are then applied to equivalent predictors from climate model output. Dynamical downscaling uses a limited-area, high-resolution model driven by boundary conditions from a GCM to derive smaller-scale information.

Various combinations of techniques for downscaling snow can be found in the literature, examples include dynamical downscaling (Salathé et al., 2008), dynamical downscaling linked to a snow model, (Bavay et al., 2009), an analogue procedure used with a snow model (Martin et al., 1997), statistical downscaling (McGinnis, 1997); GCM output serving as input to a snow model (Lapp et al., 2005); physically based downscaling (Ghan and Shippert, 2006; Ghan et al., 2006), and statistical downscaling linked to a hydrological model (Hayhoe et al., 2008). The main criticism of statistical downscaling is that it cannot explicitly describe the physical processes that affect climate and therefore assumes that the derived relationships between the predictands and the predictors do not change as the climate is perturbed. While dynamical downscaling can be argued to be more physically defensible as there is the potential for processes to transition in a changing climate, the approach is much more computationally expensive than statistical downscaling and is itself subject to error due to imperfect parameterizations. As Salathé et al. (2008) note, without proper resolution of mesoscale processes and their interactions with large-scale forcings ‘… a regional climate model is unlikely to improve on results from statistical downscaling’.

In addition, statistical downscaling can also be used to downscale ‘exotic’ predictands that are not always readily available from climate models, such as wave heights, salinity, zooplankton populations, and air pollution episodes (Wilby et al., 2004). To have some level of confidence in the climate change information obtained from downscaled data, the statistical downscaling method should first reproduce the current observed characteristics of the predictand regime. Skill over the current period constitutes a necessary but not sufficient condition.

Many of the approaches in the literature downscale snow indirectly by downscaling temperature and precipitation and then feeding these data into a hydrological model or a snow accumulation and melt model. Previous studies using the Statistical DownScaling Model (SDSM) to downscale snow have favoured this approach (Dibike and Coulibaly, 2005; Choi et al., 2008; Teutschbein et al., 2011). In contrast, this study presents a methodology for directly downscaling changes in the seasonal snow cover climatology from climate model output using SDSM Version 4.2 (available for download at http://www.sdsm.org.uk) of Wilby et al. (2002). This methodology was used to analyse 20th- and 21st-century simulated changes in snow cover over the Northeast United States. Two different observational datasets (a set of station data and a gridded dataset) were used to drive the downscaling and this article examines the projected magnitude of changes in climatological snow cover.

The paper is organized as follows: data sources and methodology used in the study are described in Section 2; simulated climatological snow cover conditions in the late 20th century are presented in Section 3; future projections of snow cover over the Northeast are described in Section 4 and; discussion and concluding remarks can be found in Section 5.


The study area is the Northeastern United States (Figure 1). Two daily surface observational datasets were used for downscaling. The first was obtained from the Northeast Regional Climate Center and consisted of snow depth observations at 20 station locations (Table 1) from the US Cooperative Observer Network (COOP) with data for the period 1961–2000. The stations were selected so as to represent the diverse snow climatology of this region. SDSM only allows one station to be processed at a time, so a limited number were chosen that were representative of the diversity of the region. The second snow dataset was a 1° × 1° gridded snow depth dataset that was also produced using daily surface observations from US COOP stations (Dyer and Mote, 2006). The station data were quality controlled using criteria set forth by Robinson (1989) and used to complete a 0.25° × 0.25° grid by means of the Spheremap spatial interpolation program (modified version of Shepard's procedure; Shepard, 1968). Sixteen individual 0.25° grid cells were then used to create each 1° × 1° grid cell. Snow depth observations for 20 grid boxes encompassing the station locations were selected.

Figure 1.

Map of the Northeast with station locations indicated by numbers that correspond to those given in Table 1.

Table 1. Summary of stations analysed
Station nameLatitudeLongitudeElevation (m)
Albany, NY42.74−73.8183.82
Amherst, MA42.39−72.5443.28
Boonville, NY43.44−75.37472.44
Caribou, ME46.87−68.02190.20
Corinna, ME44.92−69.2490.53
Dobbs Ferry, NY41.01−73.8360.96
Ebensburg, PA40.47−78.73591.31
Franklinville, NY42.33−78.46484.63
Frostburg, MD39.67−78.93661.42
Galeton, PA41.74−77.65409.96
Glenmoore, PA40.10−75.75152.40
Hawley, PA41.48−75.17271.27
Hingham, MA42.23−70.9110.67
Ithaca, NY42.45−76.45292.61
Kingston, RI41.49−71.5434.75
Landisville, PA40.12−76.43109.73
Mt. Washington, NH44.27−71.301910.18
Syracuse, NY43.11−76.10125.88
Tionesta, PA41.48−79.43365.76
Wanakena, NY44.15−74.90460.25

In addition, data to drive SDSM, large-scale predictor variables derived from the National Centers for Environmental Prediction (NCEP) reanalysis dataset of 2.5° × 2.5° resolution (Kalnay et al., 1996) and 1961–2099 model output from the coupled atmosphere–ocean United Kingdom Meteorological Office Hadley Centre Climate Model version 3 (HadCM3; Pope et al., 2000) at a resolution of 3.75° × 2.5° (longitude × latitude) that utilized the Intergovernmental Panel on Climate Change (IPCC) Special Report on Emissions Scenarios (SRES) A2 and B2 scenarios, were available from the Canadian Climate Change Scenarios Network (CCCSN; http://www.ccsn.ca). HadCM3 was chosen for this analysis because it was one of two sets of climate model output provided with SDSM. It should be noted that the purpose of this study was to create a technique for downscaling long-term snow cover statistics and examine the differences in downscaled projections that rely on different observational datasets, not to compare different climate model output. The technique may perform differently when using different model output, however, that analysis is beyond the scope of this article.

Snow cover estimates produced by an alternative statistical downscaling method, the Variable Infiltration Capacity (VIC) model (Liang et al., 1994, 1996; Cherkauer et al., 2002) were used for comparative purposes. These data were produced as part of the Northeast Climate Impacts Assessment (NECIA) and verification of this dataset can be found in Hayhoe et al. (2007, 2008). Monthly precipitation and temperature output from HadCM3 were statistically downscaled to daily values with a resolution of 1/8° using the Bias-Correction Spatial Disaggregation technique of Wood et al. (2002). The downscaled values were then used as input to the VIC model. This hydrological model simulates the full water and energy balance at the earth's surface by modelling processes such as canopy interception, evapotranspiration, runoff generation, infiltration, soil water drainage, and snow pack accumulation and melt. This dataset is referred to as NECIA.


Statistical downscaling model

For this study, SDSM was used to downscale projections of climatological winter snow cover characteristics over the Northeast United States. SDSM is a multivariate regression procedure that uses a stochastic weather generator to model predictands from derived regression equations while replicating the observed variability. The chaotic nature of weather cannot be fully explained by the variations in predictor variables alone, which is why the use of a stochastic weather generator is necessary (Prudhomme and Davies, 2009). SDSM therefore enables generation of multiple simulations with slightly different time series attributes, but the same overall statistical properties. Large-scale circulation pattern features, such as wind and pressure fields, are used as predictors to linearly condition local-scale weather generator parameters for the predictand series. SDSM has been applied to a host of meteorological and hydrological assessments in a range of geographical contexts (Hassan et al., 1998; Wilby et al., 1998; Hay et al., 2000; Coulibaly et al., 2005; Harpham and Wilby, 2005; Khan et al., 2006; Wetterhall et al., 2006; Tryhorn and DeGaetano, 2011). SDSM has been shown to reproduce observed climate variability well and perform better than many other statistical downscaling techniques (Diaz-Nieto and Wilby, 2005; Khan et al., 2006; Wetterhall et al., 2006).

The predictand series

Using the station snow depth observations from December to March at each location, two daily time series were created. One for days where there was an increase in snow cover (equivalent to snowfall) and one for days where there was a decrease in snow cover (equivalent to snowmelt or compaction). For example, if the snow depth on a particular day was 5.0 cm and this was a 2.5 cm increase from the day before, this was recorded as 2.5 cm in the increase time series. In the decrease series, this was recorded as zero. Within the decrease time series, days where there was no snow cover were counted as missing, since it was not possible for snow depth to decrease.

Snow increases and decreases are not entirely independent. For example, snow depth decreases may occur due to compaction, which is notably a function of the time since snowfall occurred. However, we selected this approach because we recognize that contrasting meteorological mechanisms cause snow depth to increase and decrease. A single regression model is unlikely to accurately include these different mechanisms. Accordingly, the dataset was split into increases and decreases. The entire process was then repeated using the gridded data to obtain a second set of daily snow increase and decrease time series at each location.

Downscaling the NCEP data

SDSM was calibrated at the daily time scale over the period 1961–1980. The calibration was done separately at each location for each time series. This step involved empirically linking the NCEP large-scale predictors (Table 2) from the nearest grid box with the observed snow increases and decreases. Given the obvious linkages between snow cover and temperature, temperature was considered to be a key predictor and was used in all the models. Selection of other predictors was an iterative process based on backward stepwise regression (Wilby and Wigley, 2000). Potential predictors (Table 2) were selected based on their perceived physical relationship to snow accumulation and melt, as well as to limit between-variable correlation. This screening was intended to minimize the consequences of overfitting the regression models and collinearity of the predictors. If two predictors were highly correlated with each other, only one was selected for inclusion in the model based on its partial correlation with the remaining variables. The variable with the lowest correlation with the remaining variables was selected.

Table 2. Candidate predictor variables
mslpMean sea level pressure
p__fSurface airflow strength
p__uSurface zonal velocity
p__vSurface meridional velocity
p__zSurface vorticity
p_thSurface wind direction
p_zhSurface divergence
5__f500 hPa airflow strength
5__u500 hPa zonal velocity
5__v500 hPa meridional velocity
5__z500 hPa vorticity
p500500 hPa geopotential
p5th500 hPa direction
p5zh500 hPa divergence
8__f850 hPa airflow strength
8__u850 hPa zonal velocity
8__v850 hPa meridional velocity
8__z850 hPa vorticity
p850850 hPa geopotential height
p8th850 hPa wind direction
p8zh850 hPa divergence
s500500 hPa specific humidity
r500500 hPa relative humidity
r850850 hPa relative humidity
shumSurface specific humidity
rhumSurface relative humidity

The predictors used for each location are shown in Table 3. Temperature was understandably a key predictor. Temperature during a winter storm plays a significant role in determining whether a location will receive rain, freezing rain, or snow and also determines how dense the snow will be. The predictors for each location vary considerably due to local differences in elevation, distance from the sea (continentality), aspect, slope, ocean currents, orographic influence, heating and cooling characteristics, and air pressure. Despite these variations there were some common predictors that emerged across both sets of data. For the increases in snow depth, zonal velocity and 850 hPa vorticity were common to many locations. For the decreases in snow depth, meridional velocity was an important predictor. Overall the station data had more predictors at each location than the gridded data and airflow strength was a common predictor for the station data, but less so for the gridded data. The predictors are consistent with the dynamic mechanisms of snowfall in the Northeast. Snow decreases are associated with warm air advection from the south, whereas snow increases are associated with storms traversing from the east and west and vorticity, which is probably a contributor to the precipitation generation process.

Table 3. Predictors used at each location for the station data (S), the gridded data (G), and both (B)
Albany, NY SSG       G   B   G   B
Amhurst, MA  BG   G  G   SB    S  B
Boonville, NY S G   S       SS      B
Caribou, ME  S GG   S     S       B
Corrinna, ME  S B    S          S  B
Dobbs Ferry, NY  S G          S       B
Ebensburg, PA S G      S    S       B
Franklinville, NY S    G       BS       B
Frostburg, MD S S   G GS    S S     B
Galeton, PA      B      S BSG S   B
Glenmoore, PA S         G S S       B
Hawley, PA  G          S B   B   B
Hingham, MA  B   S        G S S   B
Ithaca, NY SSG           B       B
Kingston, RI SB            G   B   B
Landisville, PA G           S B   S   B
Mt. Washington, NHS G   G   SS       G   B
Syracuse, NY S         S  SS       B
Tionesta, PA   S      B    B G     B
Wanakena, NY S    G G S    B   S   B
Albany, NY   B              S    B
Amhurst, MA          SSSGSGS   S  B
Boonville, NY   S   S      G     S  B
Caribou, ME  G S     GS  S        B
Corrinna, MEG S       G S S        B
Dobbs Ferry, NY            SSG        B
Ebensburg, PA   SG           S      B
Franklinville, NY   B             G     B
Frostburg, MD      S  G             B
Galeton, PA  S            G       B
Glenmoore, PA             G S       B
Hawley, PA            SS       G B
Hingham, MA SG   S       G        B
Ithaca, NY   S G    S     S      B
Kingston, RI  G G S                B
Landisville, PA                G    S B
Mt. Washington, NH     G        SB       B
Syracuse, NY   BS     S    GG      B
Tionesta, PA   S       S  G B   B  B
Wanakena, NY   S      SG G     GB GB

These predictors formed the basis of empirical relationships used to downscale the NCEP and GCM data at each location. In the case of snow increases, a fourth root transformation was applied to the predictand to account for a skewed distribution. A conditional model was used for both the increases and the decreases (as illustrated in Figure 2). Conditional models depend on an intermediate process, such as the probability of snow-day occurrence. In this case, the two-state occurrence process (i.e. snow or no snow) is first modelled as a function of the regional forcing. Then, assuming that snow increases (or separately decreases), the magnitude of increase (or decrease) is modelled conditional upon a different set of predictor weights. The final snow depth for a given day i is calculated by adding that day's increase and decrease to the amount of snow already on the ground (Figure 2). The following sections describe the methodology presented in Figure 2 in more detail and are based on Wilby et al. (1999) and Kilsby et al. (1998).

Figure 2.

Schematic diagram of the conditional models.

Snow increase (or decrease) occurrence (Oi)

Daily snow increase (or decrease) occurrence Oi was modelled by SDSM using an equation having a form similar to:

display math(1)

Equation (1) uses the four NCEP predictors that were selected for snow increase occurrence at Ithaca, NY (Table 3) as an example, with the α parameters estimated using linear least squares regression. A uniformly distributed random number r (0 ≤ r ≥ 1) was used to determine whether the snow cover increased (or decreased). For a given site and day, an increase was returned if rOi otherwise snow was assumed to decrease. Others (Fealy and Sweeney, 2007) have used logistic regression as an occurrence model. While this approach is applicable here, we have not chosen to apply SDSM as described by Wilby et al. (2002).

Snow increase or decrease amount (Si)

If snow was determined to increase (or decrease), the amount was formulated using a second regression equation based on the same predictors (Wilby and Wigley, 2000). Using the previous Ithaca example:

display math(2)

where for snow increases, Si equals the fourth root of daily snow depth increase. For snow decreases, the equation takes a similar form with Si simply the magnitude of daily snow depth reduction.

The β parameters were calculated using linear least squares regression. Once the β parameters are obtained, the residual term, ϵi, is modelled under the assumption it follows a Gaussian distribution.

The expected value was then given by:

display math(3)

where CS was an empirically derived correction ratio that allows for the bias resulting from the re-transformation of S (for snow increases) and the fact that ϵi came from a skewed distribution. The value of CS was constrained such that observed and downscaled snow depth increases (or decreases) were equal for the simulation period. In some locations, it may be appropriate to use a stochastic scaling factor φ (with a mean of 1) to increase the variance of S to agree better with observations (Hay et al., 1991). For this study, the default SDSM variance was used.

The empirical models derived from both the station and gridded data were applied to the NCEP data to create separate ensembles of 20 daily increase and decrease simulations. SDSM allows the user to select the number of ensemble members, up to a maximum of 100. Individual ensemble members are considered equally plausible local climate scenarios realized by a common set of large-scale predictors. The individually simulated increase and decrease datasets were then combined by adding up the increases and decreases over time for each ensemble member. These were then averaged to give an estimate of the snow depth at each location over the winter. Negative snow depth values were reset to zero. This process was completed using both the gridded observational data and the station data, so that two different sets of snow cover estimates were created for each location (hereby known as NCEP-grid and NCEP-station). The creation of these artificial time series (formulated from relationships between the observations and the NCEP data and representative of past climate conditions) was done with the purpose of calibrating and validating the empirical model.

Model calibration and validation

The observed data series (1961–2000) were split into two time periods (1961–1980) and (1981–2000). The first period was used for calibration of the daily models. The second set of years was used to validate the models based on their ability to replicate four features of the seasonal snow depth climatology: (1) annual maximum snow depth; (2) date of annual maximum snow depth; (3) number of days without snow on the ground; and (4) starting and ending dates of persistent snow cover. In these comparisons, the snow season was defined as November to May and persistent described a 7-d period with 2.54 cm or more of snow cover on each day.

Despite the daily time step of the downscaling approach, these seasonal parameters were the desired output given the intended applications of the data and to facilitate comparison with the NECIA data. In the NECIA, snowfall data, downscaled at a daily resolution, were also used to specify features of the longer-term seasonal snow depth climatology. These parameters were identified as being necessary to assess climate change vulnerabilities related to water resources, ecosystem health, agriculture, and tourism.

Within the calibration period, the R2 value for the Ithaca model used as illustration in Equations (1) and (2) was 0.311. This was representative of the R2 values associated with the initial daily model at other stations and grids, which generally varied from 0.10 to 0.40. For spatially conservative variables such as temperature, explained variances in excess of 70% are typically reported in approaches using SDSM. However, heterogeneous variables, such as daily snowfall or precipitation, are typically associated with smaller explained variances similar to those obtained here (Wilby et al., 2002). The CCCSN suggests that R2 values on the order of 10–15% (and in some cases below 10%) are typical and should not discourage users from working with these variables.

Given the expected low level of explained variance and the ultimate application of accurately describing relevant features of the long-term snow climatology a detailed evaluation of daily model performance in the calibration period was not conducted. Rather, validation focused on the ensemble mean of the metrics enumerated above during the independent 1981–2000 period.

Downscaling HadCM3

The same statistical models were applied to output from HadCM3 (i.e. the NCEP predictors were replaced with analogous data from HadCM3). Each downscaled dataset consisted of an ensemble of 20 daily simulations of snow depth increases and decreases, which were used to compute ensemble means of the seasonal metrics listed in the previous subsection. These snow cover datasets are hereby known as HadCM3-grid and HadCM3-station.

Historical snow cover estimates

Validation period (1981–2000)

Overall, the downscaling technique captured the main features of seasonal snow cover cycle. However, depending upon whether the station or gridded dataset was used to specify the predictands, different estimates of daily average snow cover were obtained (Figure 3). This not unexpected and is due to the inherent differences in the two observational datasets. The grid boxes represent region-wide averages that include a range of elevations, while the station dataset is representative of a point measurement at a particular location. This does not mean that either one of the datasets is better than the other for downscaling, as their potential application may be very different. For example, some impact models require point measurements (e.g. ski area viability) while others (e.g. total watershed snowpack) require gridded data.

Figure 3.

Comparison of downscaled daily average snow depth for the validation period 1981–2000 using the observed station (black) and gridded (grey) datasets as predictands for (a) Wanakena, NY, (b) Albany, NY, and (c) Ithaca, NY. Solid lines are the observed data and dotted lines are the NCEP downscaled values.

Figure 3 illustrates the differences between the gridded and station datasets (both observed and downscaled) during the validation period using three representative sites. At the 20 station locations, daily average observed snow cover during December to February (DJF) based on the gridded data is, on average, 72% of (3.6 cm less than) that based on the observations. However, there is considerable station-to-station variation depending upon the gradient of snow cover in the 1° grid. For instance at Wanakena (Figure 3(a)), station-based snow depth is on average three times higher than that for the grid, while at Albany the station-based snow depth is less than half of that for the grid (Figure 3(b)). In general, the observed gridded snow cover is less than the station data at the snowiest locations and greater at less snowy sites. This feature is also present in the downscaled NCEP datasets.

Regardless of whether the station- or grid-based data are considered, Figure 3 also indicates that the agreement between the observed and downscaled average (1981–2000) snow depth on individual days is modest, at best. However, when temporally aggregated agreement between the observed and downscaled values increases. For instance, across the stations, daily average downscaled December to February snow cover is 85% of the observed value.

In terms of the four seasonal snow cover climatology statistics of interest, there is close correspondence between the observed at downscaled datasets across the set of stations, both in terms of mean error (e) and correlation (r) (Figures 4 and 5). However, there is better correspondence between the downscaled data and the station values than the gridded observations (Figure 4). On average for the downscaled station data, seasonal maximum depth is underestimated (−3.5 ± 8.1 cm; 0.94), the date of the maximum snow depth is later than observed (3.4 ± 7.9 d; 0.78), the starting date of persistent snow cover is well specified (3.1 ± 5.8 d; 0.96) and the end of the persistent snow cover is later than in the observations (3.4 ± 10.1 d; 0.95). Here the corresponding values of e and r are given parenthetically. The number of days without snow cover (not shown) is also similar (2.1 ± 8.6 d; 0.99).

Figure 4.

Scatter plots showing observed station versus downscaled NCEP data left-hand panels and observed gridded versus downscaled NCEP data right-hand panels for maximum snow depth (a and b), date of maximum snow depth (c and d), starting date of persistent snow cover (e and f), and the ending date of persistent snow cover (g and h). The 1 : 1 line is shown in each panel for reference.

Figure 5.

Comparison of observed and downscaled station-based snow cover parameters at each station during the 1981–2000 validation period. Each group of four alternating grey and white columns shows from left to right the observed, downscaled NCEP, downscaled HadCM3 (average of A2 and B2 scenario, ensembles), and NECIA values. The lines at the top of each column represent the 1 standard deviation interval. In panel (a), values of maximum snow depth (top) and date of maximum snow depth (bottom) are shown. In panel (b), the dates of the start (top) and end (bottom) of the persistent snow period are shown.

For the gridded dataset, the tendency for the downscaled data to underestimate observed maximum snow depth is not as strong (2.5 ± 9.0 cm; 0.80); however there is more variability between the observed and downscaled NCEP data (Figure 4(b)). There is little agreement between the observed and downscaled date of maximum snow cover (7.1 ± 11.1 d; 0.48) other than a clustering of values around February first (Figure 4(d)). Likewise, the relationship between the downscaled and observed starting (−0.3 ± 7.7 d; 0.71) and ending (−5.1 ± 13.0 d; 0.67) dates of persistent snow cover is not as strong for the gridded values when compared with the station data (Figure 4). The difference between observed and downscaled days without snow cover is larger for the gridded data (−5.4 ± 17.3 d; 0.69) than with the station data (not shown).

Twentieth century GCM snow cover estimates

Hadley Centre Climate Model version 3

Next the ability of the downscaled simulations from HadCM3 to estimate the climatological snow cover metrics during the 1981–2000 validation period was examined for the station-based datasets (Figure 5). Given the weaker correspondence between the 1° × 1° gridded observations and the downscaled NCEP data, only the downscaled station results are presented. In addition, metrics are shown based on the downscaled NECIA data. Although the NECIA dataset represents grid boxes, rather than station points, its relatively fine 0.125° × 0.125° resolution is more representative of the station data than the 1° × 1° gridded values, particularly in areas with complex terrain.

In terms of snow depth (Figure 5(a)), the downscaled GCM data replicate the observations well at all stations, having similar means and interannual variation. Like the NCEP simulation, downscaled maximum snow depth is underestimated with a mean error of −3.0 ± 8.9 cm. There is no relationship between the size (or sign) of the error and the observed maximum snow cover. The correlation between the downscaled GCM depth and the observations (r = 0.62) is weaker than that based on the NCEP data.

The date of downscaled maximum snow depth is on average earlier than the observed date (−3.3 ± 10.8 d). There is a tendency for the estimates to be earlier at stations that experience maximum snow depth earlier in the year (Figure 5(a)). Correlation between the downscaled GCM dates and the observations is 0.58 which is weaker than that using the NCEP data. The downscaled GCM estimates match the observations well in terms of the starting (e = 2.2 ± 5.2 d; r = 0.96) and ending (e = 2.9 ± 9.2 d; r = 0.95) dates of persistent snow cover, values that are very similar to the NCEP-based estimates (Figure 5(b)) as well as days without snow cover (e = 3.9 ± 7.6 d; r = 0.98).

Northeast Climate Impacts Assessment Comparison

The NECIA maximum snow depth, on average, underestimates the observed values by −10.7 ± 27.1 cm. This error is affected by a large overestimation of snow depth at Mt. Washington (Figure 5(a)). Without Mt. Washington, the underestimation of maximum snow depth is greater, averaging −16.3 ± 10.7 cm. Excluding Mt. Washington, the correlation between the NECIA estimates and observations increases from 0.38 to 0.93. The NECIA data produces earlier estimates of the dates of maximum snow depth (e = 6.1 ± 8.8 d; r = 0.77); start of persistent snow cover (e = 3.9 ± 11.4 d; r = 0.76); and end of the snow cover season (e = 9.1 ± 11.6 d; r = 0.89) than are indicated by the observations. On average, the NECIA data indicate fewer days without snow cover (e = 20.1 ± 10.7 d, r = 0.96) than occur based on the observations.

Figure 6 offers a different perspective for comparing the SDSM-downscaled HadCM3 and NECIA snow cover metrics with the observed values. Overall, the SDSM methodology better replicated the observed data. For maximum snow depth (Figure 6(a)), the tendency for underestimation by the NECIA data is widespread. With the exception of Mt. Washington, the NECIA values are consistently less than the observed data and nearly always smaller than the SDSM-downscaled HadCM3 data. Arguably, a component of this underestimation may be attributable to spatial smoothing across the NECIA grids. Likewise for the dates of maximum snow depth and the end of the persistent snow periods, the NECIA values are consistently earlier than both the observations and SDSM-downscaled values (Figure 6(b) and (d)). For the date of the start of the snow season (Figure 6(c)), underestimation (too early) by the NECIA values can be attributed to a subset of five stations (Ithaca, Kingston, Landisville, Hawley, and Amherst). These sites represent a range of geographic locations and elevations but typically have the latest observed snow season start dates in the station network.

Figure 6.

Scatter plots showing observed station versus downscaled HadCM3 (open circles) and NECIA (solid circle) values at each station during the 1981–2000 validation period. Comparisons are shown for (a) maximum snow depth, (b) date of maximum snow depth and persistent snow cover (c) starting, and (d) ending dates. The 1 : 1 line is shown in each panel for reference.

Comparison with alternative approaches

Despite providing a more accurate representation of the seasonal snow climatology than the NECIA data, it could be argued that the SDSM approach is inefficient in that a daily time step is used to arrive at long-term (in this case 20 years) seasonal averages. This is also the case with the NECIA data, which uses a physical hydrological model to obtain daily snow depths estimates that are summarized to obtain climatological averages. Nonetheless, the true value of either approach should be weighed against less cumbersome approaches that do not require the generation of seasonal snow climatology parameters from daily data.

A search of the literature did not identify any approaches in which the snow climatology was described via an empirical relationship with more widely available variables such as average temperature or precipitation. Therefore, to provide a measure of the performance of such an approach, a statistical model relating winter average temperature to seasonal maximum snow depth was developed. At best, this model explained 76% of the variance in maximum snow depth across the 20 station network. This level of explained variance could only be achieved if Boonville and Mt. Washington were excluded. Including these stations, whose unique snow climatologies are influenced by extreme elevation and optimal Lake-effect snow mechanisms, reduced the explained variance to 35%. These explained variances are similar to station-specific values relating annual snow depth to seasonal temperature and precipitation at a set of Swiss stations (Rebetez, 1996). Even across a relatively small region such as the Northeastern United States, local influences affect the mechanisms and weather patterns associated with snow accumulation and ablation. To some degree this necessitates station- (or at least subregion-) specific downscaling models (Table 3).

Snow depth can also be obtained from GCM and regional climate model output. Such models typically compute snow water equivalent, which is transformed to snow depth using a parameterization for typical snow density. Roesch (2006) shows that snow depth estimates obtained from coupled climate models typically suffer from delayed snow melt in the spring while replicating the early-winter pattern of snow accumulation well. These features are reflected in the SDSM estimates. They attribute this to excessive spring snowfall. Using regional climate models, Duffy et al. (2006) conclude that all models significantly underestimated snow water equivalent in the western United States, and attributed this in part to deficiencies in the land surface models used. Similar to the method used in the NECIA, Brun et al. (2012) simulate daily snow depth using a physical snowpack model. On the basis of reanalysis data, the onset and end of continuous snow cover are, according to the authors, ‘well reproduced’ with simulated average biases of −5.1 and +7.2 d. These biases are similar to, albeit larger than, those obtained in the Northeast using SDSM and reflect the tendency for most methods to delay the ablation of snow cover.

Future GCM snow cover estimates

Future projections of seasonal snow depth characteristics across the Northeast were examined for the following time periods: the 2020s (2010–2039), the 2050s (2040–2069), and the 2080s (2070–2099) using HadCM3-station. These time periods were compared with the full climatological period 1971–2000. By the end of the century, the regional changes were predominately in the same direction and showed patterns that are consistent with declining future snow (Tables 4 and 5). However, in the earlier time periods there is potential for snow depth increases at some stations highlighting regional variations in the magnitude of the changes.

Table 4. Differences in the number days without snow on the ground (days), annual maximum snow cover (%) and date of maximum snow depth (days) in the 2080s, relative to 1971–2000
Station nameCool season days without snowAnnual maximum snow cover (% change)Date of maximum snow depth
A2 2080sB2 2080sA2 2080sB2 2080sA2 2080sB2 2080s
Dobbs Ferry+16+15−22.1−20.8−4−1
Mt. Washington−1+2+7.9−1.8−4−4
Table 5. Differences in the starting and ending dates of persistent snow cover (days) in the 2080s, relative to 1971–2000
Station namePersistent snow starting datePersistent snow ending date
A2 2080sB2 2080sA2 2080sB2 2080s
Dobbs Ferry43−14−13

The snow cover estimates from HadCM3-station project decreases in annual maximum snow depth of on average 21.6% by the end of the century under the A2 emissions scenario and an average increase of 23 d without snow (Table 4). The largest decreases in maximum snow depth are near 40% and tend to occur at sites with the highest maximum snow depths. Mt. Washington is a notable exception at which maximum snow depth is projected to increase under the A2 scenario. Many stations experience increases of 30 or more days without winter snow cover.

Under the B2 emissions scenario the decline in maximum snow depth and increase in snow-free days are smaller, averaging 18.4% and 17 d, respectively. The date of maximum snow cover is projected to be on average between 2 and 3 d earlier in the season under both emissions scenarios, with many stations experiencing peak snow depth conditions slightly (2–3 d) later in the year.

The progression of changes in the average annual cycle of snow cover under the A2 emissions scenario is illustrated for two relatively snowy locations in Figure 7. These stations have fairly high elevations, but are located in the extreme south and general northern parts of the study area. The decrease in maximum snow depth is greatest between the 2050s and 2080s simulation, with the smallest decline between the 2020s and 2050s simulation. Changes at the end of the snow season are also more dramatic than those occurring earlier in the season. In both cases there is little difference in average daily snow depth from November into early January, except in the 2080s period. The change in snow depth later in season is more pronounced between the 30 year periods, particularly at Boonville (Figure 7(b)).

Figure 7.

Daily average snow depth during 1970–2000 (black solid); 2010–2039 (black dotted); 2040–2069 (grey dotted) and 2070–2099 (grey solid) using the HadCM3-station A2 simulation at (a) Frostburg, MD and (b) Boonville, NY.

Time series of projected change in annual maximum snow depth and the number of days without winter snow cover under both the A2 and B2 emissions scenarios are shown for two representative stations in Figures 8 and 9. Both have present-day maximum snow depths near the median of the stations in the study network. In the case of annual maximum snow depth at Galeston, the difference between the two scenarios is subtle (Figure 8) during most of the century. It is not until after 2060, that the maximum snow depth for the A2 scenario becomes consistently less than that for the B2 scenario. The 20-year running mean depths for the A2 scenario decline steadily through most of the 21st century, whereas those for the B2 scenario vary from decade to decade.

Figure 8.

Time series of HadCM3 A2 (thin black) and B2 (thin grey) projections of annual maximum snow depth at Galeton, PA. The downscaled NCEP values for the 1961–2000 period are also given (dotted). The heavy lines show a 20-year running mean of the annual values.

Figure 9.

Time series of HadCM3 A2 (thin black) and B2 (thin gray) projections of the number of days without snow on the ground at Franklinville, PA. The downscaled NCEP values for the 1961–2000 period are also given (dotted). The heavy lines show a 20-year running mean of the annual values.

At Franklinville, the increase in days without snow cover follows a similar pattern (Figure 9), with the 20-year running mean of snow-free days under the A2 scenario showing a consistent increase from 2010 through the end of the century. At both locations, the changes based on the NCEP-station data during the 1961–2000 period are greater than those indicated by the HadCM3-station data under either scenario. Although the sample size during the period is small, especially in terms of a 20-year running mean, this gives reason to consider the magnitude of change in the projections to be conservative.

By the end of the century, the winter snow season is projected to shorten across the Northeast, with snow appearing later in the winter and disappearing earlier in the spring (Table 5). According to HadCM3-station, the start dates of the persistent snow season are on average projected to become 9 d later and the end dates are projected to become 10–13 d earlier, depending on emissions scenario. However, there are some regional variations, with some high snow areas showing little change or small increases in the snow season (e.g. Mt. Washington). This is likely due to the increase in temperature being offset by an increase in precipitation and is similar to the findings of Martin et al. (1997) in the French Alps.

To explore what was driving these counterintuitive changes more closely, the number of days with an increase or decrease of snow cover was examined separately. In HadCM3-station under both emissions scenarios there were slightly fewer increase days in the 2080s relative to 1971–2000 across the station network. However the change was very small. Most of the change in snow cover was driven by a large increase in the number of decrease days (20–16 d). Mt. Washington provides an interesting example (Figure 10). Decrease days showed an increasing trend in the downscaled HadCM3 station data similar to the other stations (Figure 10(b)). Surprisingly, the increase days also increased (Figure 10(a)). This indicates that, particularly under higher emissions scenarios, in places where relatively cold temperatures will remain common, some of the decrease in snow cover associated with increasing temperatures may be offset by increases in snowfall. However, increases in the potential for snow cover decreases suggests that snow depth, when present will be less. This supports the conclusions of work done in New Hampshire by Campbell et al. (2010) who found that the number of snow cover events (the number of times that the snowpack forms and dissipates) is expected to increase by 2–3 events per year.

Figure 10.

Time series of HadCM3 A2 (thin black) and B2 (thin gray) projections of the number of days experiencing an increase (a) or decrease (b) in snow depth at Mt. Washington, NH. The downscaled NCEP values for the 1961–2000 period are also given (dotted). The heavy lines show a 20-year running mean of the annual values.

In addition, there is evidence for greater relative snow losses at higher elevations (Figure 11). However, the results for Mt. Washington (1910 m elevation) suggest a nonlinear relationship. As the next highest station is Frostburg at 661 m, an analysis of stations between 661 m and 1910 m in future studies could help to clarify whether there is a particular elevation at which trends in snow losses may begin to reverse. Studies in the Western United States indicate that the largest relative snow losses occur in areas at lower elevation with warmer midwinter temperatures, while there are slower losses or even gains possible at higher elevations (Mote et al., 2005). In addition, Mote et al. (2005) found that in some colder regions, snow trends only weakly depend on midwinter temperature. Instead it is precipitation trends that are the dominant factor in snow trends. This is consistent with colder parts of the Northeast where expected increases in precipitation may offset some of the snow loss occurring from increased temperatures.

Figure 11.

Change in maximum snow depth (%) 1970s versus 2080s by elevation (m). Mt. Washington is not included since its elevation is an outlier.


The main aim of this study was to provide a rigorous and easily interpretable methodology for downscaling seasonal snow cover characteristics from climate model output. This methodology was then used to downscale projections of snow cover over the Northeast United States and future potential changes to snow cover were examined. The methodology performed well using station observations as inputs over a large geographic range that included a variety of climates. Downscaling to gridded snow depth observations, although feasible was less accurate. The methodology also provided a better match to the observed snow cover characteristics than the methodology used in previous climate change work across the region (i.e. NECIA), particularly in relation to the length of the persistent snow season.

All the snow cover metrics investigated in this study reveal patterns that are consistent with declining future snow cover. Overall these results show gradual decreases in snow depth across the Northeast over time with little change in snow depths in the early part of the snow season and significant decreases in depths toward the end of the snow season. However, within this analysis there were variations that reveal the potential in coming decades for increases in snow cover at locations that are in regions of normally high snow accumulation.

The results largely confirm the findings of previous studies (Martin et al., 1997; Mote et al., 2005; Ghan and Shippert, 2006; Hayhoe et al., 2007; Campbell et al., 2010) and future projections of snow cover showed features that were consistent with a warming world. These results indicate while there is likely to be an overall decrease in the amount of snow cover, there are likely to be regional variations, particularly in the earlier part of the 21st century. Interestingly, the locations with potential for increases in snow cover tended to occur mainly at the snowiest stations. Although the projected increase in winter temperature will decrease the number of cold days, for most of the Northeast, even toward the end of the century there will still be many days with the potential for snow and the projected increases in precipitation may offset some of the warming induced snow decreases.


This work was partially supported by the Northeast Regional Climate Center under NOAA contracts EA133E07CN0090, NA10OAR4310179, and the New York State Energy Research and Development Authority. We thank Katherine Hayhoe and Jeff Van Dorn for providing the downscaled GCM temperature and snow depth data and Dave Robinson and Thomas Estilow for making the gridded snow dataset available. We also appreciate the comments of three anonymous reviewers, which greatly improved our analyses.