Journal of Geophysical Research: Atmospheres

Role of Eurasian snow cover in wintertime circulation: Decadal simulations forced with satellite observations



[1] We investigate the impact of the Eurasian snow cover extent on the Northern Hemisphere winter circulation by performing a suite of ensemble simulations with the Météo-France “Arpege Climat” atmospheric general circulation model, spanning 2 decades (1979–2000). Observed snow cover derived from satellite infrared and visible imagery has been forced weekly into the model. Variability in autumn-early winter snow cover extent over eastern Eurasia is linked to circulation anomalies over the North Pacific that are influencing the North Atlantic sector in late winter through the development of the Aleutian-Icelandic Low Seesaw teleconnection. The forcing of realistic snow cover in the model augments potential predictability over eastern Eurasia and the North Pacific and improves the hindcast skill score of the Aleutian-Icelandic Low Seesaw teleconnection. Enhanced eastern Eurasia snow cover is associated with an anomalous upper-tropospheric wave train across Eurasia, anomalously high upward wave activity flux, and a displaced stratospheric polar vortex.

1. Introduction

[2] The impact of surface conditions on the atmosphere is receiving increased attention as a possible source of improved seasonal-to-decadal predictability. While the effect of the ocean surface on atmospheric predictability is widely recognized [Palmer and Shukla, 2000; Pavan et al., 2000, and references therein], the impact of land conditions such as soil moisture or snow cover and depth has only been recently explored [Koster et al., 2004; Douville, 2004; Cohen et al., 2007, and references therein]. Correlative studies between observed land conditions and atmospheric teleconnections suffer from a lack of causal relation, confounded by the fact that the land variables (e.g., snow cover) depend themselves on the atmospheric circulation. Hence the indirect feedback of snow cover, or soil moisture for that matter, on the atmospheric circulation is not easily derived from observational or standard model studies alone. Dedicated model studies are more amenable to identify such a weak coupling [see, e.g., Koster et al., 2006]. To this end, we use a suite of atmospheric general circulation model (AGCM) simulations and satellite observations to investigate the impact of the snow cover, and the Eurasian snow cover in particular, on the Northern Hemisphere (NH) wintertime large-scale circulation.

[3] Landmasses cover a large portion of the NH, and nearly one half of Eurasia and North America is extensively covered with snow in the cold season [Déry and Brown, 2007]. Snow-covered land plays a key role in the climate system, owing to the snow radiative and thermodynamical properties, such as high albedo, high emissivity and low thermal conductivity, and its effect on surface fluxes of moisture and heat. Snow covered land can hence impact climate in a variety of ways. The snow-albedo feedback plays an important role in the spring [e.g., Schlosser and Mocko, 2003] when an early seasonal retreat of the snow cover acts as a positive feedback on spring temperatures. Thermodynamical feedbacks in the surface energy balance also play a role: using a climate model, Cook et al. [2008] showed that the thermoinsulation effect of the snow cover impacts on the lower atmosphere temperature in winter. In idealized, snow-free climate model simulations, Vavrus [2007] showed that the lack of surface snow leads to low-level atmospheric warming, as expected from the snow-albedo feedback, but also to soil cooling and drying.

[4] There are also indications that snow cover not only affects local atmospheric conditions, but also influences large-scale atmospheric variability. Several facets of the interaction between the Eurasian snow cover in particular and the general circulation have been investigated in the past. The spring Eurasian snow cover has been shown to remotely influence the Asian monsoon [Barnett et al., 1989; Douville and Royer, 1996]. Several model studies demonstrated that the North Pacific region, situated downstream of the Eurasian landmass, is also influenced by Eurasian snow cover. The 30-day forecast experiments by Walsh and Ross [1988] showed that the spring Eurasian snow cover impacted the Aleutian Low (AL). Using an AGCM, Walland and Simmonds [1996] showed a low-level cooling, an intensified Siberian High and a deepened AL, associated with prescribed extension of Eurasian snow cover. In a model study of the relative impact of sea surface temperatures (SSTs) and snow cover during the winter 1988–1989, Watanabe and Nitta [1998] found a similar downstream influence of anomalous snow cover on the AL. In another AGCM study, Yasunari et al. [1991] showed that extensive Eurasian springtime snow influenced the North Pacific and the North American continent from spring to summer. Clark and Serreze [2000] used satellite observations and meteorological analyses to show that extensive East Asian snow cover lead to negative height anomalies over the North Pacific. Hence, there is a robust track of observational studies and modeling experiments indicating an Eurasian snow cover impact over the North Pacific, with extensive snow cover leading to a deeper than normal AL. Furthermore, Cohen and Entekhabi [1999], Saito et al. [2001], Saito and Cohen [2003], and Gong et al. [2003b, 2004] identified an influence of the Eurasian snow cover onto the North Atlantic sector. They used observations and models to demonstrate the correlation of the autumn Eurasian snow cover and the NAO in the following winter. They proposed that, through the albedo feedback, extensive Eurasian snow cover anomalies in early autumn induce diabatic cooling, amplify the Siberian High and augment the vertical stationary wave activity flux above Siberia. In their view, the response to snow anomalies, far from being shallow, involves the propagation of stationary waves into the stratosphere [e.g., Saito et al., 2001; Fletcher et al., 2007, 2009a]. This deep influence is partially supported by Vavrus [2007], where the temperature difference between snow-free and control simulations extended into the upper troposphere. While we focus on the cold season, other studies about the remote influence of the spring-summer Eurasian snow cover extent have been carried out by Déry et al. [2005], who linked it to Canadian river discharge, and Fletcher et al. [2009b], who investigated the response in the summertime circulation in multimodel climate scenario simulations.

[5] Several studies have aimed at validating model prognostic snow variables in simulations of the current and future climate [Frei and Robinson, 1998; Déry and Wood, 2006; Roesch and Roeckner, 2006]. Albeit the geographic extent of the snow cover in climate model simulations can be realistic, the interannual variability in many models is lesser than in the satellite observations during transition seasons [Hardiman et al., 2008]. Models can also suffer from systematic biases. To correct model deficiencies, one approach consists of either prescribing or nudging idealized or observed snow variables in the model. In this category are the large-ensemble, nudging experiments using satellite observations in Gong et al. [2003a, 2003b], but these were often restricted to extreme winters. Also, Fletcher et al. [2009a] and Gong et al. [2004] used prescribed, idealized snow forcings in their studies, applying a constant depth throughout Eurasia. Schlosser and Mocko [2003] performed spring simulations from 1982 to 1998 using prescribed, analyzed snow cover and depths from an in situ network. Kumar and Yang [2003] performed decadal simulations, using prognostic or prescribed climatological snow variables, but not satellite-derived varying snow cover.

[6] It remains to be evaluated if interannual circulation anomalies could be attributed to the snow cover variability in model simulations spanning several decades, with realistically varying, satellite-derived snow cover. To this end, we performed a suite of dedicated AGCM ensemble simulations, spanning 2 decades (1979–2000). We did not perform rigorous data assimilation, but rather forced the observed snow cover extent from satellite observations onto the model, akin to a “data insertion” approach. We carried out a control experiment with prognostic snow variables. In particular, we have reexamined the connectivity between Eurasian snow cover extent and winter North Pacific circulation anomalies, as well as teleconnections into the North Atlantic, such as the Aleutian-Icelandic Low Seesaw (AIS) [Honda and Nakamura, 2001; Honda et al., 2001].

2. Simulation Design and Data Sources

[7] Our ensemble simulations with the “Arpege Climat” AGCM (V3.0) were made at horizontal resolution T63, with 31 levels and a lid at 10 hPa. The time step is 0.5 hours. We analyze the simulations over the years 1979 to 2000, after a 5-year spin-up started from 1974. Background information on the model can be found in the work of Deque et al. [1994] or Cassou and Terray [2001]. Observed SSTs and sea-ice conditions from the Reynolds data set are used for all simulations [Reynolds and Smith, 1994]. An ensemble of five members was generated by time-shifting initial conditions. Statistical significance has been assessed by a two-tailed Monte Carlo method. Monthly mean fields have been retained, and we focus on winter months (DJF). Monthly mean geopotential heights from the European Centre for Medium-Range Weather Forecasts ERA-40 reanalyses are used for verification.

[8] The Arpege Climat model comprises a land surface scheme and a physically based snow hydrology model, as described by, e.g., Douville et al. [1995a, 1995b]. First, a control simulation, labeled PCL, is run with prognostic snow cover. In the snow-forced simulation, labeled SNS, the model NH snow cover is overwritten using year-round satellite observations every 5 days.

[9] The snow cover data consist of observed, weekly snow cover fractions, provided by the U.S. National Snow and Ice Data Center (NSIDC, Boulder, Colorado) on the EASE 25-km equal-area grid. The remotely sensed data are based on interpretation of infrared and visible satellite imagery. The higher-resolution satellite snow cover data is aggregated onto the model grid, by averaging over the cells fully contained into the model grid box. Detailed information about the treatment of snow variables in the model simulations, including the insertion of satellite-derived snow cover, is given in Appendix A.

3. Snow Cover Extent Climatology and Variability

[10] In Figure 1, the October and December climatological fractional snow cover area established over the years 1979–2000 is compared in the satellite observations, and in the forced and control simulations. In October, a low snow cover bias is observed in the model over northern Eurasia, at latitudes between 50°N and 70°N. In December, the low bias extends across Central Eurasia, and is especially pronounced over the eastern part of the continent. There is also a low bias over North America, but much weaker in extent. The snow cover is more extensive in the forced simulation (SNS), and the low bias is removed. In fact, the snow cover is now more extensive than in the satellite observations, esp. in October, as will be explained below. The climatological 2-m temperature difference between SNS and PCL in December is overlaid on the SNS snow cover fractional area, showing a surface cooling of −1.5 K over Central Eurasia in the forced (SNS) simulation.

Figure 1.

(left) October and (right) December climatological snow cover fraction in (a, b) the satellite observations, (c, d) the prognostic control simulation (PCL), and (e, f) the snow-forced simulation (SNS). The climatological 2-m temperature difference between SNS and PCL in December is also shown in the bottom right map (one contour at −1.5 K).

[11] In Figure 2, the mean climatological annual cycles of the fractional snow cover area over eastern Eurasia (80°–155°E; 35°–70°N), as well as its interannual variance, are shown in the forced and prognostic simulations, and in the satellite observations. The snow cover area in SNS is more extensive than in PCL throughout the cold season (November to March), when it slightly overestimates the corresponding satellite-derived values. The overestimation is larger in October, as noted in discussing Figure 1. The interannual variance is enhanced in the two transition seasons: in the autumn-early winter, marking the beginning of the cold season snow buildup, and in the spring. These two seasonal maxima are better reproduced in SNS than in PCL simulation, albeit still weaker than in the observations. Figure 3 shows the mean October-November-December (OND) fractional snow cover over eastern Eurasia from 1979 to 2000. The year-to-year variability is much weaker in the prognostic snow cover than in the observations, but is augmented in the forced simulation, as could be inferred from Figure 2 as well. We hence anticipate that this enhanced snow cover variability in the forced simulation would also feedback on the general circulation.

Figure 2.

Climatological annual cycle of the fractional snow cover area over eastern Eurasia (80°–155°E; 35°–70°N). The three curves refer to the forced (SNS, bold line) and prognostic (PCL, long-dashed line) simulations and satellite observations (dash-dotted line). Shown are the mean over the period 1979–2000 and the interannual variance.

Figure 3.

Fractional snow cover area of eastern Eurasia (80°–155°E; 35°–70°N) in autumn-early winter (October-November-December, labeled as winter on x axis) for the satellite observations (dash-dotted line), the prognostic snow (PCL, long-dashed line), and the forced snow (SNS, bold line) simulations.

[12] In the snow-forced simulation (SNS), the snow cover closely follows, but is not identical to, the observations. The mean snow cover is slightly more extensive than in the observations, and cases of reduced snow cover are overestimated. This overestimation may arise from the fact that the model has a bias in precipitation over Eurasia and tends to rapidly reintroduce a snow cover when the latter is removed to fit satellite observations. Also, an added snow cover is slower to dissipate.

4. Potential Predictability

[13] To demonstrate the impact of enhanced interannual variations in snow cover, we calculated the potential predictability in the ensemble simulations. We write the total variance as the sum of the external variance, which characterizes the year-to-year variations in the ensemble mean and reflects the effect of boundary conditions, and of the internal variance, which characterizes the member spread and reflects internal, chaotic model variability. Potential predictability, defined as the ratio of external variance to the total (internal plus external) variance, indicates regions where the boundary forcing has a strong influence, distinguishable from the model internal chaotic dynamics.

[14] We focus on the upper-tropospheric (250 hPa) circulation to identify remote snow-induced influences. In both model and observation-based studies, the upper-tropospheric potential predictability is normally highest in the tropical Pacific, with some extension over North America [e.g., Honda et al., 2005a]. Little of this tropical, ENSO-related enhanced potential predictability penetrates over the Euro-Atlantic sector, reflecting the large internal variability in the Atlantic jet stream exit region. The difference in potential predictability between the forced and the prognostic simulations reveals where the influence of the snow variability enhancement would manifest. This is shown for the winter-averaged 250 hPa geopotential height in Figure 4: a positive potential predictability increment of about 15% is observed over eastern Eurasia, extending over the Sea of Japan. The robustness of that enhancement was tested by calculating the time correlation of a particular member, the “truth,” with the ensemble mean of the remaining members, and iterating on the truth member. This method [Koster et al., 2006] tests the reproductibility of the model anomalies against internal chaos, and produces nearly identical results (not shown). The magnitude of the potential predictability enhancement due to snow cover is of the same order (15–20%) as the potential predictability enhancement due to soil moisture in summer [Koster et al., 2004]. In summary, in the ensemble mean snow-forced simulations, there is consistently more year-to-year upper-tropospheric variability over eastern Eurasia and the North Pacific, downstream of the snow-covered land mass. The latter hence appears as a key region for snow-induced circulation changes in the upper troposphere. We note that there are also weak positive increments on the eastern and western flanks of North America (Figure 4).

Figure 4.

Difference in winter mean geopotential height potential predictability between the forced (SNS) and prognostic (SNS) simulations at 250 hPa. Maximum positive values of 0.15 over eastern Eurasia and the Sea of Japan amount to a 15% change in potential predictability.

5. Interannual Variability of the Aleutian and Icelandic Lows and the Aleutian-Icelandic Low Seesaw

[15] We next assess whether this potential predictability enhancement translates into an improved skill for leading atmospheric teleconnections. The winter high-latitude circulation is characterized by two semipermanent oceanic lows, the Aleutian Low (AL) and the Icelandic Low (IL). We now demonstrate that their year-to-year variability is more realistic in the ensemble mean forced simulation than in the control one. Variability of the AL and IL is examined using monthly indexes, obtained from averaging sea level pressure over the regions where its interannual variability is maximum [Nakamura and Honda, 2002], the sector (40°–60°N; 125°–216°E) over the North Pacific, and the sector (55°–67°N; 317°–350°E) over the North Atlantic. Indexes are calculated and standardized for each of the winter months, and for either individual simulations or the ensemble-mean, separately. In late winter, observational and model studies have shown that the North Pacific variability is linked to Euro-Atlantic variability by a teleconnection termed the AIS [Honda and Nakamura, 2001; Honda et al., 2001, 2005a, 2005b]. The two lows fluctuate in unison in a seesaw, with a peak period in late winter (February or March). In our forced simulation, the AL-IL anticorrelation is −0.32 while it is −0.43 in ERA-40 reanalyses for the same period (Table 1). The AIS originates from early winter Pacific anomalies, which propagate downstream into the Atlantic sector over a time scale of 1–2 months, through the eastward extension of a Pacific-North America (PNA)-like pattern. In the positive phase of the seesaw, the AL is weaker than normal. The AIS has been shown to be strongly influenced by the El Niño-Southern Oscillation (ENSO) phenomenon, and to extend into the stratosphere [Nakamura and Honda, 2002; Orsolini et al., 2008]. The AIS index is defined as the difference between the standardized AL and IL [Honda and Nakamura, 2001; Honda et al., 2001]. The February indexes for the IL, AL, and the AIS are shown in Figure 5 over the period 1979–2000, for the ensemble means of both simulations and for ERA-40 reanalyses. Before discussing in detail the variability of these indexes, we first show that the characteristic AIS pattern in the upper troposphere can be found in the ensemble mean of the forced (SNS) simulation. In Figure 6, the February difference of composite anomalies (high minus low AIS index) is shown in the 250-hPa geopotential height field. In the positive phase of the AIS, positive anomalies are found over the North Pacific expanding into a PNA pattern. The wave train extends into the North Atlantic, leading to opposite anomalies in the IL sector between Greenland and Iceland. The amplitudes of the anomalies (120 geopotential meters (gpm)) are statistically significant at the 90% confidence level. The model pattern is very similar to both the observed and model pattern of Honda and Nakamura [2001] and Honda et al. [2005a], respectively).

Figure 5.

Ensemble mean normalized indexes for the Icelandic Low (IL), Aleutian Low (AL), and the Aleutian-Icelandic Low Seesaw (AIS, or AL-IL) in February. Indexes are shown for ERA-40 reanalyses (dash-dotted line) and the forced (SNS, bold line) and prognostic (PCL, solid line) simulations. All indexes are based on the sea level pressure. Although the AIS is calculated in February, the x axis refers to the winter, labeled as of December to ease comparison with Figure 3 (e.g., winter 1989 refers to winter 1989–1990 and the AIS index in February 1990).

Figure 6.

Difference of the AIS composites (high - low AIS index) for the geopotential height at 250 hPa, showing the ensemble mean seesaw in February in the SNS simulation. Units are 103 geopotential meters (gpm). The black contour indicates the statistical significance at the 90% confidence level.

Table 1. February Hindcast Skill of the AL, IL, and AIS; Ratio of the Standard Deviations of the Model AIS to the ERA-40 AIS; and AL/IL Anticorrelation for the Two Simulationsa
  • a

    The AL/IL anticorrelation based on ERA-40 for the same period is −0.43. Statistical significance at 99% confidence levels is indicated in boldface. Abbreviations are as follows: AIS, Aleutian Low-Icelandic Low Seesaw; AL, Aleutian Low; IL, Icelandic Low; PCL, control simulation; SNS, snow-forced simulation.


[16] In the forced simulation, the year-to-year variations in the AIS index are better reproduced (Figure 5). This can be quantified using a (deterministic) skill score, defined as the correlation of the ensemble mean index with the corresponding observed index. The latter is derived from ERA-40 reanalyses. Table 1 summarizes the February skill scores. There is a significant positive skill improvement in February for the IL that reflects on the AIS. In the forced simulation, the ensemble mean skill for the AIS is higher (0.66) than in the control simulations (0.38), and also higher than in other model studies such as Honda et al. [2005a], hinting that the snow cover is indeed modulating the AIS in late winter. The enhanced snow cover variability impact on the North Atlantic sector is in line with the observed link between the autumn Eurasian snow cover and the Icelandic Low and the North Atlantic Oscillation in the following winter, pointed out by Cohen and Entekhabi [1999]. For the forced simulation, the ratio of the ensemble mean model to observed AIS standard deviations is closer to 1, indicating a higher year-to-year variability.

[17] To ensure that the hindcast skill improvement does not originate from a particular member, we calculated ensemble means based on four members, rejecting one particular member at a time, and found that the AIS indexes based on the smaller ensemble means are close to the five-member ensemble mean (not shown for brevity).

6. Relation Between the Aleutian-Icelandic Low Seesaw, Snow Cover Extent, and El Niño-Southern Oscillation

[18] A running thread through our experiments is that extensive (lessened) Eurasian snow cover modifies the atmospheric circulation downstream, over the North Pacific, by deepening (weakening) the winter AL. On the other hand, the AIS is strongly anticorrelated with the ENSO phenomenon [Honda et al., 2001, 2005a, 2005b]: when warm tropical Pacific SSTs prevail in the ENSO El Niño phase, the AL is deeper than normal. Owing to the relative shortness of the forced simulation (21 years), it is challenging to clearly separate the effects of ENSO from those resulting from snow cover anomalies, or to rule out that the snow cover extent itself merely reflects indirect ENSO influences. In fact, Yang et al. [2001] argued that the snow cover acts as an amplifier of the ENSO forcing while investigating wintertime surface climate anomalies over North America. In the forced simulation, the Aleutian Low correlation with the winter mean El Niño 3.4 index is −0.50. We expect that strong ENSO conditions (El Niño or La Niña) will have a predominant influence on the AIS development, and inspection of Figure 4 shows that nearly identical negative values of the AIS index, are found in the two simulations for the strong El Niño years 1982 and 1997, i.e., irrespective of the snow variable being forced or prognostic. The ensemble mean difference in the February AIS index between the two simulations (SNS and PCL) is smaller when the El Niño index is large, than for the ENSO-neutral years. Hence, the snow-induced modulation of the AIS is overwhelmed in strong ENSO years, and is not a by-product of ENSO.

7. Eastern Eurasia Snow Composites

[19] The next step consists of determining circulation anomalies associated with anomalous eastern Eurasian snow cover in autumn and early winter (OND). To this end, the model time series of fractional snow cover over eastern Eurasia (80°–155°E; 35°–75°N) (as in Figure 3) was standardized to construct a snow index. Geopotential height difference composites between high and low snow index (normalized indexes higher or lower than 0.67) have then been calculated for the snow-forced simulation. The number of years used in the composite is small, autumns of 1985, 1993, 1998 and 1999 in the positive index case, and of 1980, 1988 and 1994 in the negative index case, and the statistical significance of snow composites is low, not exceeding the 60% confidence level for the geopotential composites, and 70% level for the wave flux composites. It is nevertheless of interest to address the consistency of the dynamical anomalies induced by eastern Eurasian snow cover variability. We verified that the snow index compositing is neutral with respect to the ENSO phase.

[20] At 250 hPa in December, a pair of opposite anomalies is found over Eurasia. It consists of a negative geopotential anomaly (40 gpm) over the North Pacific, on the southern flank of the AL and downstream of the anomalously cold snow-covered region (Figure 7), and a positive geopotential anomaly upstream, over Central Eurasia. The composited anomalies reflect a wave train over Eurasia that is in good quantitative agreement with Fletcher et al. [2009a, Figure 2]. The dominant dipole across Eurasia is also consistent with their explanation in terms of negative (positive) vorticity anomaly upstream (downstream) of the maximum surface cooling over Central Eurasia (see their Figure 1f). These early winter (December) anomalies in the upper troposphere over the northwest Pacific then expand eastward across the central North Pacific throughout winter, as revealed by one-point correlation maps (not shown).

Figure 7.

Composite difference of the ensemble mean geopotential height (units are 103 gpm) based on the early winter (OND) eastern Eurasian snow index in the forced SNS simulation at 250 hPa in December, and at 30 hPa in January; the two shaded contours indicate the statistical significance at the 50% and 60% confidence levels.

[21] Anomalous stationary wave propagation characteristics associated to eastern Eurasian snow cover were further diagnosed using the wave activity flux (WAF), or Plumb vector [Saito et al., 2001]. The Plumb vector vertical component is proportional to the meridional eddy heat flux. The snow index composite difference of the upward WAF at 250 hPa (Figure 8) shows an anomalously high upward flux in December stretching from Central Asia to the North Pacific (80°E to 200°E), associated with the anomalous wave train shown on Figure 7. This enhanced upward vertical propagation influences the stratosphere: geopotential height composites calculated at 30 hPa in January (Figure 7) indicate positive geopotential anomalies at high latitudes with a strong wave-1 component, i.e., a polar vortex displacement toward northern Eurasia.

Figure 8.

Difference of the ensemble mean composites of the vertical component of the wave activity flux at 250 hPa in December based on the eastern Eurasia snow index for the SNS simulation. Units are 102 meters squared per seconds squared. The shaded contours indicate statistical significance at the 70% confidence level.

8. Discussion

[22] Several findings of Cohen and Entekhabi [1999], Saito et al. [2001], Gong et al. [2004], and Hardiman et al. [2008], are consistent with our study: with extensive (lessened) Eurasian autumn snow cover, the late-winter AIS negative (positive) phase is indeed associated with enhanced upward wave activity flux across Eurasia in midlatitudes, and projects onto a negative (positive) NAO phase. On the other hand, we did not find an anticorrelation between autumn Eurasian snow indexes and the winter NAO in our simulations, while such an anticorrelation exists between the satellite-derived Eurasian snow cover index and the ERA-40-based NAO index (−0.51). A comparison of various climate model simulations of Hardiman et al. [2008], including a version of the AGCM used here, indeed revealed the lack of such a relation in all models. One possible explanation is that, while Eurasian snow cover exerts an influence on the IL, the models might not capture the midlatitude component of the NAO, that is, the Atlantic High variability. This issue needs further investigation.

[23] The stratospheric pathway proposed by these authors would in fact be consistent with the quasi-horizontal propagation associated to the AIS discussed here: in the negative phase of the AIS, enhanced wave propagation into the stratosphere [Nakamura and Honda, 2002] would lead to a more disturbed polar vortex, and the downward propagation of stratospheric anomalies would reinforce the negative phase of the AIS, as well as the NAO, and increase heights over the Arctic. The ability of our model to demonstrate the occurrence of a stratospheric downward-propagating influence is limited by the fact that only monthly mean fields have been retained, and that the model lid is at 10 hPa.

[24] While other model or observational studies (cited in Introduction) had found a seasonal influence of the Eurasian snow cover onto the North Pacific circulation, our analysis of decadal simulations demonstrate that the autumn-early winter eastern Eurasian snow cover distinctly influences the year-to-year variability over the North Pacific from the surface to the stratosphere (extensive snow cover leading to a deepened Aleutian Low), even impacting the North Atlantic in late winter. This improves the hindcast of the Aleutian and Icelandic lows, and the AIS teleconnection. The snow cover influences the stationary planetary waves and their upward propagation, as demonstrated in previous studies [e.g., Saito et al., 2001], reinforcing the trough over eastern Eurasia and the North Pacific.

[25] The degree to which the snow cover weakly feeds back on the atmospheric circulation from the surface and throughout the troposphere and stratosphere is the result of complex interactions between many model parameterizations, and not a single prescribed model input. This is analogous to soil moisture feedbacks [Koster et al., 2006]. These feedbacks start from radiative and thermodynamical processes but may well, later, involve large-scale dynamical feedbacks. Eastern Eurasia is a region where the Asian jet establishes a near-zonal waveguide for propagating Rossby waves arising from thermal anomalies, and is also a region of cyclogenesis. Dynamical feedbacks from transient baroclinic eddies influenced by near-zonal snow cover anomalies could also be at play [Corti et al., 2000; Clark and Serreze, 2000].

9. Summary and Future Prospects

[26] We have performed a suite of 20-year simulations with Arpege Climat (V3.0) model, forced with satellite-derived snow cover data to test the impact of snow cover on large-scale NH interannual variability. Our focus lies on the major wintertime surface circulation features, the Aleutian or the Icelandic Lows, and the development of the AIS, which represents a late-winter coupling of their covariability induced via planetary wave horizontal propagation [Honda and Nakamura, 2001; Honda et al., 2001]. Hence, we emphasize the upper-air height response over the Aleutian and Icelandic sectors, away from the continental regions where snow is deposited. Snow-induced North Pacific circulation anomalies and their downstream influence onto the North Atlantic had not been investigated before in decadal simulations forced by satellite observations of the snow cover.

[27] The main results are as follows:

[28] 1. The model has too weak interannual variability in the Eurasian autumn snow cover, esp. in October, compared to satellite data. The forcing with satellite observations remedies that lack of variability.

[29] 2. The enhanced snow cover variability leads to a potential predictability increment of 15% over a prognostic simulation, located over eastern Eurasia extending over the Sea of Japan. The magnitude of the potential predictability enhancement due to snow cover is of the same order (15–20%) as the potential predictability enhancement due to soil moisture in summer [Koster et al., 2004], albeit in different areas.

[30] 3. This enhanced potential predictability translates into a higher February skill score of the AIS index calculated over the period 1979–2000: 0.66 for the snow-forced simulation vs. 0.38 in the prognostic snow simulation. It is considerably improved compared to earlier studies [e.g., Honda et al., 2005a; Orsolini et al., 2008]. In summary, the AIS is better hindcasted with a more realistic, more variable snow cover. The hindcast skill improvement in late winter also applies to the Icelandic Low. This is consistent with the relation between Eurasian autumn snow cover and the Icelandic Low-North Atlantic Oscillation pointed out by Cohen and Entekhabi [1999].

[31] 4. Composites based on an eastern Eurasia snow index reveal a dipole of upper-tropospheric anomalies across Eurasia, deepening the trough over eastern Eurasia, accompanied by enhanced upward wave activity fluxes. This also leads to a displacement of the stratospheric polar vortex toward Eurasia.

[32] Changes in the terrestrial snow cover have been occurring in recent decades. The satellite record from visible imagery indicates a negative trend in NH average snow cover extent in all seasons but autumn; the trend is largest in spring and summer [U.N. Environment Programme, 2007; Déry and Brown, 2007]. Projecting into the future, the decreasing trend in NH snow cover is expected to continue: coupled ocean-atmosphere simulations with the GFDL-CM2 model under increasing greenhouse gas emissions scenarios in Déry and Wood [2006] indicate that the NH annual mean snow cover extent will have decreased by 12 to 26% by 2100 compared to its centennial mean. The importance of the Eurasian snow cover highlighted in this study suggests that the fate of the seasonal snow cover at high northern latitudes be continuously monitored, not only for accurately assessing and projecting climate change and its impacts on the water cycle and resources, but also to benefit seasonal forecasting.

Appendix A:: Treatment of Snow Variables in the Model Simulations

[33] The Arpege Climat AGCM (V3.0) [Deque et al., 1994] comprises a land surface scheme (ISBA, interaction among soil, biosphere and atmosphere) to provide boundary conditions for moisture and temperature. This model uses an improved version of ISBA that takes into account the snow parameterization from Douville et al. [1995a, 1995b], and includes computation of several prognostic variables: surface temperature, daily mean surface temperature, surface volumetric water content, deep volumetric water content, reservoir of rain intercepted by the canopy, snow water equivalent Wn, snow albedo and snow density. The albedo of snow on ground varies between 0.85 and 0.50. To represent the effect of snow aging, the snow albedo exhibits exponential decrease in time if the snowmelt rate is positive and a weak linear decrease for nonmelting conditions. Snowfall refreshes the albedo back to its maximum value when it exceeds a threshold. The computation of surface albedo takes into account the masking effect of vegetation. The snow density is assumed to be constant with depth, and to increase exponentially to a maximum value with an e-folding time of about 4 days. After snowfall, the snow density is recalculated as weighted average of the previous day density and that of the new snow. The snow depth is a diagnostic variable determined from snow water equivalent Wn and snow density. The continuity equation for Wn contains solid precipitation as a source term and snow sublimation and melt as sink terms. The fractional model snow cover is derived from Wn through an empirical formula involving surface roughness and canopy, and is the snow cover output discussed in this article.

[34] The purpose of the snow data insertion procedure is to update (or correct) the simulated Wn at a given time in order to obtain minimum difference between the satellite-derived and the simulated Wn values. Because we do not carry out multivariate data assimilation of the snow variables, the following approach does not lead to a consistent surface energy balance.

[35] In the snow-forced simulation (SNS), Wn is overwritten every 5 days by a value, based on satellite-derived fractional snow cover f. In order to establish a robust link between Wn and f it is useful to introduce a climatological snow mass value equation image. In establishing this value, we have used the observed long-term monthly mean snow cover, equation image. In columns where the observed equation image ≈ 1, equation image is simply the long-term monthly mean simulated snow water equivalent. For the relatively few columns where 0 < equation image <1, equation image gets its value from the nearest grid square with an already established equation image (a column with unity equation image).

[36] Every 5 days, a simplified nudging procedure is carried out. At each time step, we correct the simulated snow water equivalent to an observationally based value as follows:

equation image

[37] In case 0 < f <1, then Wnu = equation image · f/equation image (where 0 < equation image). The latter case occurs only in a minor fraction of the observed data points containing snow. In between the 5-day apart nudging procedures, snow variables are evolving prognostically.

[38] In summary, this technique replaces simulated snow cover whenever the model snow cover differs from observations. This is done in any grid square over the NH and at regular time intervals (5 days). In case the model simulates snow, but snow free conditions are observed, we impose snow free conditions by removing the simulated snow amount. Oppositely, if the model simulates snow free conditions, or fractional snow cover in a grid cell while the observed area is fully covered with snow, we overwrite the model value with the critical-climatological snow amount. In case both the model and the observations exhibit fractional snow cover, the model value is corrected according to the simple linear relationship given above.


[39] This study was financed by the Norwegian Research Council under projects MACESIZ and NORCLIM. S. Ma is gratefully acknowledged for postprocessing the model data, and H. Nakamura, M. Honda, H. Douville, and J. Wettstein are acknowledged for useful suggestions. We acknowledge the U.S. National Snow and Ice Data Center in Boulder, Colorado, for snow cover data provision.