How stationary is the relationship between Siberian snow and Arctic Oscillation over the 20th century?

Authors


Abstract

[1] Both observational and numerical studies suggest that fall snow cover extent over Eurasia is linked to subsequent winter variations in the predominant Northern Hemisphere teleconnection pattern, known as the Arctic Oscillation (AO). The present study uses the recent 20th Century Reanalysis to explore the snow-AO relationship over the entire 20th century for the first time. 20th Century Reanalysis is first shown to have a consistently realistic simulation of the onset of the Eurasian snow cover compared to a large number of in situ observations. It is then used to explore the snow-AO relationship over both the satellite and presatellite periods. Results show that this teleconnection is not stationary and did not emerge until the 1970s. The possible modulation of the teleconnection by the Quasi-Biennal Oscillation is then discussed, because it could have favored the influence of snow anomalies on the Arctic Oscillation in recent decades. These results have important implications for seasonal forecasting and suggest, in particular, that statistical predictions of the wintertime AO should not be based on snow predictors alone.

1 Introduction

[2] The Arctic Oscillation (AO), together with its regional manifestation over the North Atlantic (North Atlantic Oscillation, NAO), is the predominant mode of variability of the Northern Hemisphere atmospheric circulation in winter [Thompson and Wallace, 1998]. The AO has a strong influence on winter temperature and precipitation over Eurasia and North America [Thompson and Wallace, 2000]. Although it is a fundamental internal mode of the atmosphere, it is modulated by slowly evolving external forcings and/or lower boundary conditions, making its predictability a major issue in seasonal forecasting.

[3] Significant observational evidence relates the fall Eurasian/Siberian snow cover extent to the subsequent winter Arctic Oscillation over recent decades [Cohen and Entekhabi, 1999]. The proposed mechanism involves a wave mean-flow interaction through a complex stratospheric pathway, described in detail in Cohen et al. [2007]. This teleconnection is supported by idealized numerical sensitivity studies [Gong et al., 2003; Orsolini and Kvamstø, 2009; Fletcher et al., 2009; Allen and Zender, 2010; Smith et al., 2011], but is still poorly captured by free-running coupled ocean-atmosphere general circulation models [Hardiman et al., 2008]. Some authors suggest that certain model limitations can be circumvented by correcting their biases concerning the Siberian snow variability [Allen and Zender, 2011] or the mean state of the extratropical circulation [Peings et al., 2012].

[4] The ability of an October Siberian snow index to predict the surface temperature over North America and Europe has been tested in linear statistical models that outperform the hindcast scores of dynamical models over a 20 year period [Cohen and Fletcher, 2007]. According to a recent study [Cohen and Jones, 2011], the level of AO predictability is even higher when using a snow index that takes into account the rate of snow cover advance over Eurasia during October. However, this index requires daily snow cover data that have only been available since 1997. Although weekly data, available from 1966 onward, can be used as an alternative, this is still too short a period to determine the multidecadal variability and constancy in time of the snow-AO link.

[5] The present study aims to explore the snow-AO relationship throughout the 20th century by taking advantage of the recent availability of the 20th Century Reanalysis (20CR) [Compo et al., 2011]. The global tropospheric circulation, and in particular the AO/NAO variability, compares well with other reanalyses over recent decades [Compo et al., 2011; Ouzeau et al., 2011]. More surprisingly, we show in this paper that the onset of the fall Eurasian snow cover is also represented accurately in 20CR compared with in situ snow depth observations available since the late 19th century. It is thus possible to use the 20CR reanalysis to assess the stationarity of the snow-AO relationship over the entire 20th century.

2 Data and Methods

2.1 Datasets

[6] We used the 20CR reanalysis as a representation of both the AO and the fall snow cover over Northern Eurasia. 20CR is a recent ensemble atmospheric reanalysis (driven by observed sea surface temperature and sea ice) extending from 1871 to 2010 [Compo et al., 2011]. We used the 2° resolution daily gridded data set from the 56-member ensemble mean. 20CR is based on the assimilation of surface pressure observations only. In contrast to other reanalysis systems, it does not include an analysis of surface conditions, either from observations of the surface state or from atmospheric observations in the boundary layer. While 20CR proved capable of representing the historical variability of the Northern Hemisphere atmospheric circulation [Compo et al., 2011; Ouzeau et al., 2011], the present study includes the first evaluation of the 20CR snow cover over Northern Eurasia during October and November.

[7] The reconstructed 20CR snow cover was compared with two snow data sets. The first is the Northern Hemisphere weekly snow cover extent Version 3 product [Armstrong and Brodzik, 2005], available at the National Snow and Ice Data Center (NSIDC). This data set was used for the 1972–2006 period and interpolated on the 20CR horizontal grid for the sake of comparison. The snow fraction on each grid cell is set to 1 if at least 50% of the grid cell is snow-covered; otherwise, it is set to 0. The second data set is the Historical Soviet Daily Snow Depth (HSDSD) product [Armstrong, 2001], the only source of daily snow depth observations over Northern Eurasia starting at the end of the 19th century. Thanks to its consistent quality control, this data set makes it possible to identify the presence or absence of snow on the ground [Brun et al., 2012]. The weekly NSIDC snow cover is generally representative of the 5th day of the week [Robinson et al., 1993]. Therefore, the corresponding days are extracted from the daily 20CR/HSDSD snow data when they are compared to NSIDC.

2.2 Snow and AO Indices

[8] To represent the rate of progress of the snowpack in October, the Snow Advance Index (SAI) is computed similarly to Cohen and Jones [2011]. The SAI is the regression coefficient of the least square fit of the daily or weekly (weeks 40–44) snow cover averaged over the 35°N–60°N/40°E–180°E region (SAI domain) in October. Another snow index, the Snow Cover Index (SCI), is computed simply via the average snow cover over the same domain in October. Normalized yearly anomalies for these two indices are calculated relative to the 1971–2000 climatology. The SAI is multiplied by –1, such that a strong positive SAI value anomaly means that the snow advance was rather weak during October. Consequently, a positive (negative) correlation is expected between the SAI (SCI) and the AO index to support the inverse snow-AO relationship identified in the literature (a positive snow anomaly is followed by the negative phase of the AO, and vice versa). The SAI and SCI derived from the NSIDC and 20CR data are here referred to as SAI-NSIDC, SAI-20CR, SCI-NSIDC, and SCI-20CR, respectively.

[9] The wintertime (December-January-February, DJF) AO index from the Climate Prediction Center (CPC) is used on the 1951–2011 period (AO-CPC hereafter). This index is derived from the National Center for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) reanalysis [Kalnay et al., 1996]. For the 20CR reanalysis, the DJF AO index is derived from an Empirical Orthogonal Function analysis applied to the sea level pressure north of 20°N. The principal component of the first Empirical Orthogonal Function mode is defined as the AO index (AO-20CR). A positive (negative) AO index corresponds to low (high) pressure anomalies throughout the polar region and high (low) pressure anomalies across the subtropical and midlatitudes.

3 Results

3.1 Evaluation of the 20CR Snow Cover

[10] To evaluate the snow cover extent from both NSIDC and 20CR against the daily HSDSD snow depth observations, we built a snow detection performance index (SDP) as follows: a daily observation at a station is considered in agreement with SC-20CR or SC-NSIDC when snow depth is above (under) a given threshold and the corresponding grid cell has a snow fraction of 1. We used 5 cm as the threshold value but the results do not change significantly when the threshold value varies between 2.5 and 10 cm. For a given month and a given geographical domain (or a given station), the SDP is defined as the percentage of SC-20CR or SC-NSIDC values in agreement with daily ground-based observations. Figure 1a shows the performance of SC-20CR in October for the period 1891–1994 at individual stations (405,992 observations, Figure S1 of the Supporting Information). Most stations (77%) exhibit an SDP higher than 80%. Figure 1b compares SC-20CR SDP with SC-NSIDC SDP over the common dates during the period 1972–1994 (21629 observations). SC-20CR clearly outperforms SC-NSIDC, especially over the transition area where snow covers the ground during half of October on average (Figure 1c). Figure 2a shows the historical evolution of the October SDP averaged over the available HSDSD stations within the SAI domain. Despite the varying number of stations (red dashed line), the SC-20CR performance is consistently very high and always higher than its SC-NSIDC counterpart. November exhibits similar results (Figures 2b and S2 of the Supporting Information). Limiting the SDP calculations to a subset of stations with the longest record during the period 1972–1994 does not significantly change the results (Figure S3 of the Supporting Information). In summary, the 20CR reanalysis exhibits an unprecedented ability to represent the daily advance of snow cover over Northern Eurasia in October and November throughout the 20th Century. Detailed examination of the daily snow depth simulated by 20CR at individual stations shows that this performance stems from a realistic representation of the synoptic weather systems that induce the first cold waves and snowfalls over Eurasia in autumn (see Figure S4 of the Supporting Information for an example). This is confirmed by Figure S5 of the Supporting Information, which shows that 20CR snow cover compares very well with the 1922–1997 reconstruction of the October snow cover anomalies averaged over Western Eurasia [Brown, 2000].

Figure 1.

(a) 20CR snow detection performance: percentage of October days with snow/no snow in both 20CR snow cover and HSDSD data (threshold 5 cm) over 1881–1994. (b) Difference in % between the 20CR and NSIDC snow detection performance over 1972–1994. (c) Observed snow frequency in % of October days over 1972–1994, defined as the ratio of HSDSD data higher than 5 cm.

Figure 2.

Time evolution of the snow detection performance for 20CR and NSIDC averaged over the available HSDSD stations. The monthly snow frequency averaged over available stations is indicated. The snow detection and frequency are expressed in % on the left axis and the number of HSDSD stations is indicated on the right axis.

3.2 Snow-AO Relationship in 20CR

[11] In order to compare our results with Cohen and Jones [2011], it is necessary to determine whether the SAI-20CR index compares well with its SAI-NSIDC counterpart over the overlap period and whether it is significantly correlated with the AO. Figure 3a shows the time series of SAI-20CR with those of SAI-NSIDC and AO-CPC from 1973/1974 to 2006/2007. The SAI-20CR snow index is computed from weekly data to be compared with SAI-NSIDC (see section 2a). First of all, a significant correlation is found between SAI-NSIDC and AO-CPC, in line with Cohen and Jones [2011]. Our correlation is somewhat weaker (R = 0.58 instead of 0.63), but the period of analysis also differs slightly (1973/1974–2006/2007 period instead of 1973/1974–2010/2011). Over the common period, the correlation between the two SAI is quite high (R = 0.70, p < 0.01), which underlines the realism of the 20CR snow data. Finally, the correlation between the AO-CPC circulation index and SAI-20CR is also significant (R = 0.45, p < 0.05), although lower than that obtained with SAI-NSIDC.

Figure 3.

(a) Timeseries and correlations for the following indices: AO-CPC; SAI-NSIDC; SAI-20CR from weekly data over the 1973/1974–2006/2007 period. Stars indicate the significance of correlations: **p < 0.01; *p < 0.05. (b) Correlations on a 21 year moving window: AO-CPC vs SAI-20CR; AO-20CR vs SAI-20CR; AO-CPC vs SAI NSIDC; AO-CPC vs SCI-20CR; AO-20CR vs SCI-20CR; AO-CPC vs SCI NSIDC. The 95% confidence level for correlations is indicated by the horizontal dashed lines. SAI-20CR is computed from daily data.

[12] Given the steadiness of the 20CR snow quality over the entire 20th century (section 3a), and the good agreement between SAI-20CR and SAI-NSIDC over recent decades, the study of the snow-AO relationship was extended back in time (until 1891). Figure 3b shows the correlation between two AO indices (AO-CPC and AO-20CR) and SAI-20CR, computed over a 21 year moving window (black lines). Over the common period, these sliding correlations are very close, depending on the use of the CPC or 20CR AO index. It gives us confidence in using 20CR to construct an AO index over the whole 20th century. The main result in Figure 3b is that the snow-AO relationship is not stationary. While the correlation significance is higher or close to the 95% confidence level over recent decades (in line with SAI-NSIDC, solid red line), it is not significant before the 1970s and even changes sign. The same conclusion is reached when using the SCI instead of the SAI (blue and dashed red lines). In summary, these results suggest that the significant relationship between the fall Siberian snow and the winter AO has emerged recently and is not a consistent feature over the 1891–2010 period.

4 Discussion and Conclusion

[13] The 20CR reanalysis has been shown to represent the daily advance of snow cover over Northern Eurasia in October and November with a surprisingly high and steady accuracy over the whole 20th century, given the assimilation of surface pressure only. We therefore have confidence in the reconstructed indices for both snow and AO, especially as they compare well with observations and/or other reanalyses over the overlap periods.The nonstationarity of the snow-AO relationship found in 20CR does not arise from the multi-decadal variability of the snow forcing since the SAI standard deviation is relatively constant over the whole 20th century (not shown). It may be due to the non-stationary response of the extratropical circulation to Siberian snow forcing. Figure 4 supports such a hypothesis by using the NCEP/NCAR reanalysis for atmospheric fields (more appropriate than 20CR for representing the middle atmosphere) over 1948–2010. It shows the regression on the 20CR-SAI (here multiplied by –1 so that the patterns are associated with a positive snow anomaly) of the DJF zonal mean zonal wind and of the DJF wave activity flux (WAF) computed according to Plumb's formulation [Plumb, 1985]. Two periods are distinguished, one with and one without a significant correlation between SAI and AO (1948–1975 and 1976–2010, in line with Figure 3b). The WAF describes the propagation of stationary Rossby waves; its divergence (convergence) indicates an acceleration (deceleration) of the zonal mean flow. Over 1976–2010, the well-documented snow-AO mechanism is found [Saito et al., 2001; Gong et al., 2003]; positive snow anomalies enhance the stationary wave activity from the surface into the lower stratosphere, and this wave momentum deposition weakens the polar vortex, leading to an AO response that propagates downward in winter. Strikingly, this mechanism is not at work during the 1948–1975 period when an equatorward propagation of waves is visible, with no significant zonal wind anomalies in the polar stratosphere. We argue that such contrasted behavior might arise from an interaction with the Quasi-Biennial Oscillation (QBO) in the equatorial stratosphere. Indeed, the QBO is known to have a significant impact on the polar stratospheric vortex in winter, with more waves reflected toward the pole and a warmer, weaker and more easily disturbed polar vortex during the eastward QBO [Holton and Tan, 1980; Marshall and Scaife, 2009]. Furthermore, over recent decades (1976–2010) positive (negative) snow anomalies over Siberia were associated predominantly with the eastward (westward) phase of the QBO (Figure 4). This correspondence between the QBO and snow anomalies could have favored the poleward (equatorward) propagation of extratropical planetary waves and the negative (positive) phase of the successive AO. When using a reconstructed QBO index [Brönnimann et al., 2007], four years (1976/1977, 1992/1993, 2000/2001, and 2009/2010) with snow and favorable QBO anomalies (more than one standard deviation) are found over 1976–2010 and all correspond to an AO anomaly (more than one standard deviation) whose sign is in line with the discussed mechanism (see Figure S6 of the Supporting Information). In contrast, no single year with both snow and favorable QBO anomalies is found from 1948 to 1975.

Figure 4.

Regression of the DJF zonal mean zonal wind anomalies (black contours, interval is 0.5 m/s), of the WAF (arrows, scale in m2/s2) and of the WAF divergence (red contours, interval is 4.10–5 m/s2) onto the 20CR-SAI, for (a) 1948–1975 and (b) 1976–2010. Zonal wind anomalies that are significant at the 95% confidence level are shaded. For display, the vertical component of the WAF vectors was multiplied by 100 and both WAF and divergence were multiplied by the square root of p (p = pressure/1000 hPa). The SAI is here multiplied by –1 such that these patterns are associated with a positive snow anomaly over Siberia.

[14] In summary, the main conclusion of this work is that the snow-AO teleconnection is not stationary over the 20th century. While further studies will be necessary to understand the multiple drivers of the AO variability and their possible interactions better, our analysis suggests that the QBO in the equatorial stratosphere may have played a role in the modulation of the snow-AO relationship. This result is consistent with recent numerical experiments showing that the AO response to snow perturbation is sensitive to the representation of the equatorial stratosphere [Peings et al., 2012]. Given the small number of years with consistent snow and QBO forcing on the AO, this hypothesis must however be confirmed by additional numerical sensitivity experiments in which the QBO phase is prescribed and/or by the analysis of multidecadal simulations from models that show a reasonable simulation of both QBO and Eurasian snow cover.

Acknowledgments

[15] We are thankful to Gilbert P. Compo for providing us with the 20CR snow data, to S. Brönnimann for providing us with the reconstructed QBO time series and to D. Saint-Martin for insightful discussions. Support for the Twentieth Century Reanalysis Project dataset is provided by the U.S. Department of Energy, Office of Science Innovative and Novel Computational Impact on Theory and Experiment (DOE INCITE) program and Office of Biological and Environmental Research (BER), and by the National Oceanic and Atmospheric Administration Climate Program Office.

Ancillary

Advertisement