Contribution of late spring Eurasian snow cover extent to Canadian winter temperatures



This study examines intercontinental linkages between late spring and early summer Eurasian snow cover extent (SCEss) anomalies and the following winter temperature anomalies over Canada for the 1972–2006 period. The structure of the second interannual mode of Canadian winter temperatures variability captures the SCEss related modulation. The North Atlantic winter atmospheric circulation changes associated with the SCEss, resembling the negative phase of the North Atlantic Oscillation (NAO), suggest a possible pathway for the SCEss influences on the Canadian winter temperatures.

Regression and composite analyses show that the SCEss relate robustly to the Canadian winter climate. Larger-than-normal SCEss is associated with below normal winter temperatures in south-central Canada and above normal temperatures over northeastern Canada. Predictive skill of Canadian winter temperatures based on a cross-validated regression model shows that the SCEss offers the predictive potential over regions of Canada where El Niño-Southern Oscillation (ENSO) related skill is weak or nonexistent. Analysis of winter extreme minimum temperatures, by a non-stationary generalized extreme value model, with the SCEss as a covariate, exhibits statistically significant changes over Canada resembling a pattern similar to that of winter mean temperatures. Wavelet analysis shows significant coherence between the SCEss and the second mode of winter temperature variability in the 8–12-year band. Copyright © 2011 Crown in the right of Canada. Published by John Wiley & Sons, Ltd

1. Introduction

The boundary forcing from sea surface temperature anomalies in the Pacific basin produces considerable interannual variations in the cold season climate of Canada. Interannual fluctuations in the Canadian temperatures and precipitation related to the El Niño-Southern Oscillation (ENSO) phenomenon as well as the Pacific-North America (PNA) pattern are evident in western Canada (Shabbar and Khandekar, 1996; Shabbar et al., 1997a). ENSO also provides a basis for seasonal climate prediction in Canada (Shabbar and Barnston, 1996; Derome et al., 2001). Besides ENSO, other circulation features associated with slow varying boundary conditions such as decadal fluctuations in the North Atlantic sea surface temperatures (Enfield et al., 2001), the North Pacific sea surface temperatures (Mantua et al., 1997) or large Northern Hemisphere SCE (SCE; Robinson, 1993) could possibly provide added predictability for the Canadian climate. The principal aim of this study is to examine the influence of the Eurasian SCE on the winter climate of Canada.

One of the most important land surface processes that can affect atmospheric variability is the interaction between large-scale snow covered regions and climate in remote locations. The Eurasian snow cover is one of the largest and most influential land surface features producing climate impact (Gong et al., 2007), possibly affecting remote locations through large-scale teleconnections. Both the observed and modelled lead-lag relationships and the possible physical mechanism related to the association have highlighted the impact of fall Eurasian SCE (e.g. Cohen and Saito, 2003; Fletcher et al., 2007). For example, studies by Cohen and Entekhabi (1999) have found fall Eurasian snow cover to be a skilful predictor of mean Northern Hemisphere winter climate.

The physical processes behind this connection are fairly well documented. The diabatic cooling associated with the extensive snow cover in Siberia produces anomalous vertically propagating energy flux (Fletcher et al., 2007), and through stratospheric–tropospheric coupling, as outlined by Baldwin and Dunkerton (1999), sea level pressure anomalies resembling the negative phase of the Arctic Oscillation (AO; Thompson and Wallace, 1998) are realized in the North Atlantic (Saito et al., 2001; Gong et al., 2003). As well, Cohen and Saito (2003) show skilful winter temperature scores over much of the eastern United States from previous fall Eurasian snow cover. Exclusive of the ENSO phenomenon, this remote influence may potentially offer additional predictive capabilities for the Canadian climate.

The Northern Hemisphere snow cover retreats rapidly from a maximum in late winter–early spring to a minimum in late summer in response to warmer temperatures (Robinson et al., 1993). The high albedo and emissivity of snow along with its low thermal conductivity cools the overlying atmosphere, potentially producing a gain or a deficit in surface radiation. A combination of extensive snow cover and the seasonal intensification of solar irradiance during late spring–early summer can potentially provide favourable conditions for land surface influence on the energy balance. Snow cover affects surface energy balance via the snow–albedo feedback and snow–hydrological feedback (Yasunari et al., 1991; Peings et al., 2010). The sustained nature of large-scale snow cover anomalies during spring and early summer can substantially affect the surface radiation and water budgets producing changes in soil moisture (Yeh et al., 1983; Groisman et al., 1994; Pielke et al., 2000). Model simulations of intercontinental teleconnections between snow water equivalent and soil moisture anomalies in Eurasia, and changes in atmospheric circulation features in North America have been shown by Barnett et al. (1988).

The importance of snow–albedo feedback has also been shown in climate change studies. While examining sensitivity in response to anthropogenic forcing in climate models, Hall and Qu (2006) demonstrate that spring snow–albedo feedback plays a central role in climate prediction over Northern Hemisphere land masses. Moreover, Saunders et al. (2003) show strong and significant correlation (r = 0.61 with p < 0.02) between summers with high Northern Hemisphere SCE and the following winters with the low index value of the North Atlantic Oscillation (NAO). Additionally, Cohen and Saito (2003) have noted that the summer Eurasian snow cover is also an effective indicator of remote subsequent winter climate. They hypothesize that extensive summer snow cover likely persists into the subsequent cold season.

By employing multivariate pattern-matching technique of singular value decomposition (SVD, Bretherton et al., 1992) in the Community Climate System Model, Ghatak et al. (2010) establish a link between declining summer sea ice in the Arctic basin and subsequent increases in the fall SCE over the surrounding landmasses. They suggest that increases in surface air temperatures and declining summer sea ice in the Arctic will lead to more snowfall over the surrounding land areas. While examining downward trend in summer sea ice, Deser et al. (2000) show declining summer Arctic sea ice is forced by an atmospheric circulation pattern in the previous late spring. Namely, positive sea level pressure (SLP) anomalies over the Arctic basin and over Eurasia, with the attendant negative SLP anomalies over the mid-latitudes during late spring, are influential in forcing the observed declining trend in summer sea ice. These results suggest that late spring and positive SLP anomalies over the Arctic and the surrounding landmass associated with Eurasian snow cover can be related to increased fall SCE over Eurasia with summer sea ice changes likely bridging these two seasonal processes. It is, therefore, conceivable that the late spring and early summer SCE could potentially be used as a long-lead indicator of subsequent winter climate anomalies over Canada.

The NAO (Jones et al., 1997) has been regarded as a preferred internal mode of extratropical atmospheric variability whose variance, for the most part, has been described as white noise on an interannual time scale (Saito and Cohen, 2003). Variations in the NAO have also been attributed to the external slow varying boundary conditions such as sea surface temperatures (Kushnir, 1994) and Eurasian snow cover (e.g. Bojariu and Gimeno, 2003). Over the last decade, a number of investigators have provided evidence of causal linkages between late spring to fall Eurasian snow cover anomalies and subsequent Northern Hemispheric winter climate through the NAO or the AO teleconnectivity (Bojariu and Gimeno, 2003; Cohen and Saito, 2003). Shabbar et al. (1997b) have shown that the climate in eastern Canada is significantly affected by the atmospheric circulation variability in the North Atlantic. A lead–lag relationship between the SCE and the NAO could provide added element of predictability.

For the most part, the timing and degree to which the Eurasian snow cover anomalies significantly affect Canadian winter temperatures remain unexplored. The goal of this observational study is to establish robust lead–lag relationship between the late spring–early summer (April–June) Eurasian SCE (hereafter, labelled as SCEss) and the low-frequency mode of variability in the observed Canadian winter temperatures, and to discuss the possible mechanism behind the linkage. We explore the strength of the variations in the Canadian winter temperatures related to the SCEss. To underline the stability of this relationship, we develop an empirical prediction model in a cross-validated framework of Canadian winter temperatures with SCEss as predictor. The non-stationarity of the extreme winter minimum temperature-SCEss connection is documented by a generalized extreme value model (GEV; Coles, 2001). Since the NAO has been shown to impact Canadian winter temperatures, the predictive capability of the NAO from the SCEss is also analysed. And finally through wavelet analysis, we examine coherence between the SCEss and the dominant structure of Canadian winter temperatures.

2. Datasets

2.1. Temperatures

The rehabilitated gridded (200 km2 resolution) Canadian historical air temperature database is the end result of data homogenisation technique at Environment Canada ( = en&n = 70E82601-1). Long-term temperature time series of daily and monthly maximum, minimum and mean temperatures (Vincent and Gullett, 1999) has been specifically designed for climate variability and climate change research. Data from highly correlated neighbouring stations were used to estimate missing values. Short-term station segments were joined, in some cases, to create long-term series and to test for ‘relative homogeneity.’ Non-climatic steps due to station alterations, including changes in site exposure, location, instrumentation, observer, observing program, or a combination of the above, were considered in the identification of inhomogeneities. Regression models were used to derive correction factors, and adjustments were applied to bring each homogeneous segment into agreement with the most recent homogeneous part of the time series. The main causes of the identified discontinuities were retrieved from station maintenance reports whenever possible.

2.2. Snow cover extent

Weekly and monthly snow cover data produced by the National Oceanographic and Atmospheric Administration (NOAA) for the 1972–2006 period are used (Robinson et al., 1993). The large-scale spatial extent of Eurasian snow cover is based on visible band satellite imagery and is available from the mid-1960s. However, satellite-derived reliable snow cover data are only continuously available from 1972. Hereafter, monthly Eurasian snow cover extent data will be labelled as SCE. The SCE, representing the total snow-covered area over the Eurasian landmass, were obtained from the Rutgers University climate lab at for the 1972–2006 period. At the time of this study, snow cover data were available only through 2006.

2.3. Sea level pressure

The NAO index employed here is the standardized difference in mean sea level pressure between southwest Iceland and Gibraltar compiled by the Climatic Research Unit at the University of East Anglia∼timo/datapages/naoi.htm. This version of the NAO is a useful index of the strength of the oscillation in the North Atlantic, and is most applicable to the winter half of the year (Jones et al., 1997). Other versions of the NAO (e.g. Hurrell, 1995) were also employed in this study with very similar results. Winter composites of the North Atlantic circulation, in Section 3.4, are computed from monthly mean sea level pressure originating from the National Center for Environmental Prediction National Center for Atmospheric Research (NCEP-NCAR) reanalysis (Kalnay et al., 1996).

3. Methods and results

3.1. Empirical orthogonal function

In order to extract dominant structures of variability, an empirical orthogonal function (EOF) analysis, based on covariance matrix, has been applied to the gridded Canadian winter mean and monthly mean temperature anomalies. Temperature anomalies are relative to the 1961–1990 base period, and values at each grid location were detrended prior to the EOF analysis. The associated principal component time series (PC) are obtained by projecting temperature anomalies for each winter and individual winter months upon their respective EOF spatial patterns. The leading mode of winter temperatures (explaining 43% of the variance) has the same sign features across Canada. This mode of variability captures PNA-related structure over Canada (not shown). The three largest eigenvalues of seasonal temperatures EOF along with their sampling errors, as computed according to North et al.'s (1982) criteria of eigenvalue separation, shows that the first two EOFs are well separated from their neighbours. Eigenvalues ranked after two EOFs are not generally well separated, suggesting a limit beyond which EOF cannot identify useful signals in seasonal temperature dataset. The second mode (EOF2) of winter temperatures (Figure 1(a)) is of interest in this paper because of its previously established linkage to the NAO-like winter circulation pattern (Shabbar et al., 1997b). It explains 25% of the interannual variance and exhibits a northeast–southwest dipole pattern across Canada. The associated PC2 time series is shown below the spatial pattern of EOF2 in Figure 1(a). To explore the lead–lag relationship on interannual timescales between SCE anomaly and the large-scale features in Canadian winter temperatures, monthly SCE for the 1972–2006 period were used. Prior to the analysis, the SCE time series were detrended.

Figure 1.

(a) EOF2 of Canadian winter temperatures. Normalized PC2 time series (solid line) and inverted normalized SCEss (dashed line) are shown below. (b) Lead–lag correlation plot between monthly SCEss and monthly PC2. Significant negative correlations at the 5% level or higher are indicated by blue pixels

3.2. Lagged snow-temperature relationship

Figure 1(b) shows the lead–lag correlation plot of monthly SCE with the time series of the second mode (PC2) of Canadian monthly temperatures for up to 24 months. The second EOF calculated individually for December, January and February shows a similar dipolar structural pattern with opposite sign loadings in the south and the north of Canada (not shown). The variance explained by each of the three winter months is close to 21%. Thus, the nature of PC2 is consistent for all three months, and averaging three months to represent winter season is a reasonable choice. The diagonal line in Figure 1(b) delineates contemporaneous correlations; to the left of the diagonal PC2 leads and to the right SCE leads. Prominent significant negative correlations are evident between the SCEss and the winter (December–February) PC2. Statistical significance of this relationship is established via the Monte Carlo procedure by randomly shuffling the SCE time series 1000 times and evaluating the observed correlation against a distribution of correlations generated randomly. The SCE-PC2 relationship is significant at p < 0.01 level for December, January and February. Lead–lag correlations were also calculated with both eastern and western Eurasia separated by the 90°E line; however, the snow cover–winter temperature relationships were quite similar (not shown). Although July SCE also shows significant correlation, it occurs when there is large uncertainty in snow cover data (Dery and Brown, 2007) and is not included in the analysis. Table I shows that the correlation coefficient between PC1 and the PNA is 0.63 (p < 0.01), while those between the SCEss and PC2 is − 0.67 (p < 0.01).

Table I. Correlation between the time series of PNA, the SCEss, fall Eurasian snow cover SCE (SON) and the leading five modes of Canadian winter temperatures. Bold entries are significant at the 1% level
PNA0.63− 0.03− 0.250.04− 0.01
SCEss0.140.67− 0.020.320.04
SCE (SON)− 0.04− 0.29− 0.310.190.02

Since the snow–albedo feedback is strong in the late spring in the Northern Hemisphere (Hall and Qu, 2006), it is reasonable to expect that any cryospheric relationship with the Canadian winter temperatures would be particularly effective. We explore this possible linkage by plotting correlation between winter temperature PC2 and the three-month running mean of the SCE in Figure 2. Statistical significance is established via the Monte Carlo procedure by randomly shuffling the SCE time series 1000 times and evaluating the observed correlation against a distribution of correlations generated randomly. The correlation sign indicates that the high values of SCEss precede winters with low index state of PC2. The strength and significance (r = − 0.67, p < 0.01) of the SCEss–PC2 relationship far exceeds that from any other lagged Eurasian snow cover period. Along with PC2, the inverted time series of the SCEss is also shown in the lower panel of Figure 1(a). The two time series are broadly similar, however, individual values can be quite different. Whereas Cohen and Entekhabi (1999) have found statistical linkages between the fall Eurasian snow cover and the United States winter temperatures, our results show that correlation between Canadian winter temperature PC2 and fall (Sept-Nov) SCE is generally weak and not statistically significant.

Figure 2.

The strength and significance of the correlation between running three-month SCE and winter PC2 time series. Correlations are calculated from detrended time series. Dashed lines indicate confidence levels as determined by the Monte Carlo method. The lagged SCE range from JFM (Jan–Feb–Mar) to DJF (Dec–Jan–Feb). The largest anti-correlation (−0.67) is attained during the AMJ (April–May–June) period

3.3. Regression model

We explore the linear relationship between the winter temperatures and SCEss by regressing winter temperatures over Canada upon the normalized SCEss time series. Shown in Figure 3(a) are temperature anomalies obtained by multiplying the regression coefficients by the standard deviation of the SCEss time series. Anomalously large SCEss is statistically related to colder-than-normal winter temperatures across southern Canada and warmer-than-normal winter temperatures across a substantial portion of northeastern Canada. Locations possessing statistical significance at p < 0.05 are indicated by hatching. Accordingly, one standardized unit change in the SCEss leads to about 1 °C change in southern Ontario and nearly 2 °C change in winter temperatures over northeastern Canada. With the polarity reversed, the spatial structure of this relationship is markedly similar to the second mode of the winter temperature pattern in Figure 1(a).

Figure 3.

(a) Spatial pattern generated by regressing winter temperatures upon normalized SCEss during 1972–2006. Values are multiplied by the standard deviation of SCEss to yield temperature anomalies in °C. Hatching indicates areas with values significant at the 5% level. (b) Coefficient of determination R2 in the regression model relating winter temperatures with previous spring SCEss is expressed as per cent

The goodness of fit of the regression pattern, as measured by the coefficient of determination, or R2, is shown in Figure 3(b). The adjusted R2 indicates that the regression model accounts for 35–60% of the winter temperature variability over eastern Canada and 15–30% over southern Ontario. The strength of the snow cover–temperature relationship by the regression model against the error in the regression model is given by the F ratio. The F ratio is 15–45% (p < 0.05) over northeastern Canada and 10–30% (p < 0.05) over south-central Canada (not shown).

To gain further insight into how strongly coupled the temperature patterns shown in Figure 3(a) are to the SCEss upon which they are based, a time series of the temperature pattern index is constructed. The temperature pattern index is defined by projecting the temperature regression pattern (Figure 3(a)) upon the temperature anomaly field in each winter from the 1972 to 2006 period. Figure 4 shows that the temporal correspondence between the SCEss and the associated winter temperature pattern indices is remarkably similar. The correlation coefficient between the two time series is 0.71 (p < 0.01).

Figure 4.

Time series of normalized SCEea (solid curve) and the associated winter temperature anomaly pattern index (dashed curve; see text for details). The two time series track each other with a correlation of 0.71 significant at the 1% level

To evaluate the association between the SCEss and the following winter Canadian temperatures (not necessarily constrained by linear relationship), we compute the difference in winter temperatures between two five-year composites, which are comprised of the five extreme lows (normalized values less than − 1 standard deviation) and five extreme highs (normalized values greater than 1 standard deviation) of the SCEss for the 1972–2006 period. The five extreme high years are 1976, 1979, 1981, 1996 and 1998. The five extreme low years are 1972, 1975, 1982, 1988 and 1990. Figure 5 presents the result for the (high–low) temperature composite difference at each grid location. With the sign reversed, this pattern bears a strong resemblance to the winter temperature EOF2 (Figure 1(a)). A broad area of positive temperature anomalies is evident over northeastern Canada, whereas the south-central region displays negative temperature anomalies. The statistical significance, as determined by the 2-sided Student t-test, is indicated by hatching. Although these relationships do not alone imply any cause-and-effect relationship; nevertheless, collectively they provide evidence of a strong statistical linkage between the late spring SCEss and the following winter Canadian temperatures.

Figure 5.

Winter average composite difference in temperature anomaly in °C based on years with five highest minus five lowest values of SCEss. Extreme high (low) values are those above (below) one standardized unit of SCEss. Significant differences are assessed using a 2-tailed Student's t-test. Values significant at the 5% are indicated by hatching

3.4. Winter temperature prediction

Predictive capabilities of the SCEss are examined by a simple regression model. The SCEss over the 1972–2006 period is used as a predictor to forecast the following winter gridded temperature anomalies in Canada. Both the SCEss and Canadian winter temperatures were linearly detrended prior to entering the model. To guard against artificial skill in a relatively small size (35 years) of predictor and predictands, forecast model performance was evaluated in a cross-validation framework with one year held out sequentially. The forecasts are then verified on the withheld year. The accuracy of the model in the hindcast mode was evaluated by the Pearson product-moment correlation, referred to as anomaly correlation here. Skill was evaluated by correlating temperature forecasts with the observed values. The overall skill is determined by averaging predictive skill over the 35 models. Figure 6 shows that the model has positive skill in south-central and northeastern Canada. Values greater than 0.35 are statistically significant at p = 0.05. Locally, cross-validation skill is much higher at 0.6 over the lower Great Lakes and 0.7 over extreme northern Quebec. The overall skill pattern is very similar in structure to the regression of winter temperatures on the SCEss as shown in Figure 3(a).

Figure 6.

Cross-validated anomaly correlation values between hindcasts of regression model and observed winter surface temperature anomalies for the winters of 1972–2006. Values greater than 0.35, indicated by darker colours, are statistically significant at the 5% level

3.5. Snow cover and extreme minimum temperature

We also examine the effect of the SCEss on the extreme minimum winter temperatures in Canada, defined as the lowest night-time temperature attained during the winter (Tmin). A non-stationary GEV model with time varying covariate is used to assess changes in Tmin in response to variations in SCEss covariate. The GEV has a distribution of the form

equation image(1)

where µ, σ and ξ are the location, scale and shape parameters respectively. Both the stationary GEV0(µ, σ, ξ) without covariate and non-stationary GEV1t = µ0 + r1SCEss, σ, ξ) models with covariate are constructed. Assuming that the shape parameters do not change significantly in response to SCEss changes, we examine the effect of the time-varying covariate on the location parameter µ. This parameter portrays information about the shift in the mean of the distribution. The GEV1 model was trained with the SCEss as covariate for each Tmin grid point. The goodness-of-fit of the GEV model and the significance of the relationship were assessed by the ‘likelihood ratio’ test. The validity of GEV1 over GEV0 is tested by forming log-likelihood L1 and L0 respectively, and the test statistics 2(L1–L0) are assessed against the χ2 distribution at the 5% level of significance. The field significance (Livezey and Chen, 1983) of the regression parameters is represented by the percentage of grid points where the GEV1 model is favoured over GEV0 model at the 5% level of significance. Overall, SCEss impact is significant at 31% of grid locations when the effect of the SCEss is accounted for in the location parameter. These grid locations mainly cover the northeastern and south-central regions of Canada. Figure 7 shows that the mean of Tmin is significantly affected (p < 0.05) by changes in the late spring SCEss. Over northeastern Canada, the magnitude of Tmin is significantly increased, and significant decreases are observed over an area from eastern Manitoba to the lower St Lawrence Valley. These results show that the primary impact of the late spring SCEss on the extreme minimum temperatures is very similar to that of the mean temperatures.

Figure 7.

The estimated coefficients linking the GEV location parameters (r1) for the extreme winter minimum temperatures with SCEss in units of °C/106 km2. Hatching indicates locations where the relationship is significant at the 5% level based on the likelihood ratio test

3.6. Spring snow cover and the winter NAO

A number of previous studies have reported observed and modelled lead–lag relationships between SCE from summer to fall and the NAO (Cohen and Entekhabi, 1999; Bojariu and Gimeno, 2003). Since the NAO has been identified as an important teleconnection pathway connecting Eurasian snow cover with the Northern Hemisphere winter climate (e.g. Gong et al., 2007; Fletcher et al., 2007), we examine the effect of the SCEss on the following winter atmospheric circulation in the North Atlantic sector. Figure 8 shows the difference in the composite of winter SLP between the extreme positive and negative phases of the SCEss. Here all values of the SCEss above (below) one standardized unit are classified in the positive (negative) phase. The resulting pattern shows a broad area of the positive anomalies encompassing the North Atlantic sector with a band of concomitant negative anomalies across the Atlantic basin near 45°N. This circulation pattern is remarkably similar to the negative phase of the NAO/AO pattern.

Figure 8.

Winter average composite difference SLP based on years with five highest less five lowest values of SCEss. Extreme positive (negative) values are those above (below) one standardized unit of the SCEss. Contour interval is 1 mb. Image provided by the NOAA-ESRL Physical Sciences Division, Boulder, Colorado, from their Web site at

Next we establish a statistical relationship between the SCEss and the winter NAO by developing a forecast regression model in a cross-validation framework. Both time series are detrended and the influence of autocorrelation is minimized for the 1972–2006 period prior to entering the model. With one year held in abeyance, 35 regression models are developed in sequence in the hindcast mode. The forecasts are then verified on the withheld year. The overall skill is determined by averaging predictive skill over the 35 models. The test statistics used are the correlation coefficient (r), the percentage variance explained (PVE). Mean Absolute Error (MAE) is considered a superior measure of average error compared to the commonly used root mean squared error (Willmott and Matsuura, 2005). The MAE skill score (MAESS), expressed as percentage improvement over a climatological forecast is given by

equation image(2)

where, fi is the forecast, oi is the observed value and overbar denotes climatology defined over 34 years of model development data. Table II displays the average predictive skill of the NAO from the SCEss. The model correlation skill is 0.55 (p < 0.05) with nearly 30% of the variance explained. The skill improvement in mean absolute error over climatology is near 19%.

Table II. Hindcast predictive skill of winter NAO from the SCEss
Skill measureCross validated (1972–2006)
Correlation (r)0.55
PVE (%)29.2
MAESS (%)19

For the 1972–2006 period, our results are supported by the findings of Bojariu and Gimeno (2003) showing significant correlation between the winter NAO and the SCE beginning in the preceding May for the 1973–2003 period. The large measurement uncertainty in summer Eurasian SCE (Dery and Brown, 2007) underscores the importance of the relationship between the summer SCE and the following winter NAO. These results show that a substantial portion of the winter NAO variability can be accounted for by the SCEss.

4. Discussion

The main aim of this study was to explore a statistical association between SCE and subsequent winter temperature anomalies over Canada. We found a strong and statistically significant association between the late spring–early summer SCEss and the subsequent Canadian winter temperatures than any other lagged snow cover period for the 1972–2006 period. Although the direct physical mechanism for the SCEss–winter climate connection is not fully understood, a number of observational and modelling studies have provided clues for intermediate steps relating pan-Arctic spring snow anomalies and the subsequent Northern Hemisphere climate.

It is generally accepted that seasonally intensifying solar irradiance along with maximum attainable SCE during late spring–early summer provide the most favourable conditions for land surface influence on the surface energy balance (e.g. Groisman et al., 1994; Hall and Qu, 2006; Peings et al., 2010). It has also been suggested by a number of studies that stationary waves forced by orographic features in Eurasia in conjunction with snow-forced anomalies during fall initiate a teleconnection involving tropospheric–stratospheric interaction that subsequently modulates atmospheric circulation in the North Atlantic resembling the negative phase of the NAO (Cohen and Saito, 2003; Gong et al., 2007).

On the basis of the results of previously published papers, here we suggest a possible sequence of cryospheric events that, to some degree, provides an explanation for our findings. The observational study by Deser et al. (2000) suggests links between years with below normal summer sea ice, and previous late spring with positive SLP anomalies over the Arctic and Eurasia. The diabatic cooling associated with positive SLP anomalies is conducive to the maintenance of the previous winter's extensive snow cover over the Eurasian landmass. The modelling study by Ghatak et al. (2010) shows that below normal summer sea ice leads to above normal fall Eurasian snow cover. Additional support for this connection is provided by Honda et al. (2009). They highlight mechanisms which may generate pathways linking summer sea ice and fall snow cover. Specifically, stationary Rossby waves thermally generated by decreasing summer sea ice strengthen SLP over Siberia producing cold conditions and ample snow cover over Eurasia. Further evidence of winter circulation patterns and preceding summer sea ice conditions is provided by Francis et al. (2009). Their analysis of satellite-based sea ice extent and atmospheric observations shows that varying sea ice conditions are associated with large-scale winter circulation changes beyond the Arctic boundaries. They attribute these changes to increased cloudiness and the weakening of the polar jet stream.

The persistent nature of the SCEss with a high degree of autocorrelation, as demonstrated by Dery et al. (2005), could also be a contributing factor. Yasunari et al. (1991) suggest that the snow–albedo feedback and snow–hydrological feedback affect the surface energy balance via large-scale persistent snow cover. The sustained nature of these processes, together with sea ice–atmosphere interaction (noted above), most likely influences hemispheric-scale circulation, supporting the role of the SCEss in affecting local as well as lagged remote climate, including variability in winter temperatures over Canada. Another interpretation is that persistence in the late spring snow cover likely causes lower-than-normal summer surface air temperatures over Eurasia that give rise to rapid development of snow cover during fall (e.g. Dery et al., 2005). This plausible interpretation along with fall Eurasian snow cover–winter NAO results of Cohen and Entekhabi (1999) lends support to the statistical relationship shown in Figure 3.

The NAO appears to provide one possible mechanism linking SCEss and winter temperatures over Canada. Indeed, the relationship between the NAO and the Canadian winter climate has been well documented (Shabbar et al., 1997b; Bonsal et al., 2001). These studies have demonstrated a distinct regional influence on Canadian winter temperatures with the NAO index affecting the northeastern and south-central regions of Canada.

Our own analysis in Section 3.6 establishes a statistical association between SCEss and the following winter NAO index. Regression model developed in a cross-validated framework shows that nearly 30% of the winter NAO can be accounted by the SCEss (Table II). Also, the composite difference in winter SLP anomalies based upon the extreme years of SCEea shows a pattern that is strongly reminiscent of the negative polarity of the NAO (Figure 8). It is recognized that there are influences other than SCEss or the fall Eurasian SCE on the NAO. As noted in the Introduction, Kushnir (1994) proposes that the North Atlantic Ocean, in particular local SSTs, have a regulating effect on the NAO variability. Additionally, a modelling study by Wu and Gordon (2002) shows that North Atlantic Ocean dynamics, mainly through thermohaline circulation, play an active role in low-frequency modulation of the NAO at decadal and interdecadal timescales. These results suggest that mechanisms other than the NAO may also affect the SCEss–winter temperature relationship. While examining predictive capabilities of fall Eurasian snow cover, Cohen and Saito (2003) also report that the fall Eurasian snow cover is more highly correlated than the winter NAO/AO with winter surface temperatures in the eastern United States.

To gain additional confidence in the SCEss–Canadian winter temperature relationship, we examine the coherence between the two time series by Morlet wavelet analysis (Torrence and Compo, 1998). Wavelet coherence provides information about the relationship in time frequency domain, and a consistent or slow varying phase lag is suggestive of a physical relationship between the two time series (Grinsted et al., 2004). Monte Carlo methods were used to assess the statistical significance against the first-order autoregressive red noise background. Since the wavelet is not completely localized, there are boundary or edge artifacts in the analysis. In the cone region of influence, however, discontinuity due to the edge effects is minimal.

Figure 9 shows the squared wavelet coherence between the detrended SCEss and detrended PC2 time series. Regions possessing statistical significance at p < 0.05 are delineated by thick black lines. The left pointing arrows show anti-phase relationship between the two parameters (Grinsted et al., 2004). Significant coherence in the cone of influence occurs in the 8–12-year band during the 1984–1995 period. After 1999, significant coherence can be found in a wider 4–12-year band where boundary effects may also be present. The presence of higher-frequency signal may be due to the changes from 190.5 km resolution manually digitized weekly snow charts to a much higher 24 km resolution daily snow products (Brown et al., 2010). In contrast, wavelet coherence for the fall Eurasian snow cover and PC2 time series shows significant coherence in a rather restricted 8-year band during a shorter 1985–1990 period (not shown).

Figure 9.

Squared wavelet coherence between the normalized SCEss and PC2 time series. Statistical coherence is assessed by the Monte Carlo method against the first-order autoregressive red noise background. The 5% significance level is shown as a thick contour. The left pointing arrows characterize anti-phase behaviour between the two time series

We acknowledge that the statistical nature of the analysis undertaken here does not explain direct physical mechanism relating SCEss with the Canadian winter temperatures. On the basis of several robust statistical analyses presented here, however, we conclude that the late spring–early summer Eurasian SCE exerts a strong influence on the following winter temperatures over northeastern and south-central Canada. Careful sensitivity analysis with coupled ocean-atmosphere dynamical models, emulating realistic cryospheric processes, will be required to fully understand the nature of the association and determine causal links.

5. Summary

In this study, SCEss was found to be an important potential contributor to the interannual variability of the Canadian winter temperatures. The strength and significance of this association far exceeds that from any other lagged Eurasian snow cover lagged period and that this lead–lag relationship is most likely influenced by the impact of the SCEss on the following winter NAO. The SCEss impact on the winter temperature anomalies reveals a northeast–southwest dipole pattern across Canada that projects onto the second mode of winter temperature interannual variability. The finding that the SCEss has the most important subsequent impacts on the Canadian climate variations for 1972–2006 is consistent with previous studies (e.g. Barnett et al., 1988; Groisman et al., 1994; Hall and Qu, 2006; Peings et al., 2010). As noted previously, the largest attainable Eurasian SCE during late spring–early summer along with seasonal intensification of solar irradiance provides maximum potential for albedo feedback and creates the most favourable conditions for land surface influence on the surface energy balance.

A regression model based on SCEss shows that anomously large SCEss is significantly related to colder-than-normal winter temperatures over a large portion of south-central Canada, and warmer-than-normal temperatures over an extended region of northeastern Canada. The goodness-of-fit of the regression pattern indicates that a substantial portion of winter temperature variability in the eastern half of Canada is explained by the previous spring Eurasian SCE. Further affirmation of this relationship is provided by a close association (r = 0.71, p < 0.01) between the SCEss and the temperature pattern index derived from the temperature regression pattern. As well, a composite of winter temperature anomalies based upon extreme years in the SCEss record yields a pattern that is nearly identical to that obtained from the regression procedure.

A forecast model of winter temperatures, based on a cross-validation framework, with SCEss as the predictor provides a potential for skilful seasonal prediction in the eastern and south-central regions of Canada for the following winter. Cross-validated correlation skill is near 0.35 over the Great Lakes and the lower St Lawrence Valley, and increases to near 0.6–0.7 over a vast expanse of northeastern Canada. Application of the GEV model with time-varying SCEss as covariate shows that the extreme winter minimum temperatures in Canada are also significantly affected by the SCEss. Changes in the winter extreme minimum temperatures show a pattern similar to that of mean temperatures. In summary, improvement in wintertime temperature skill due to the variability in the SCEss is encouraging for seasonal temperature prediction prospects for the south-central and northeastern regions of Canada where ENSO-related signal is generally weak or nonexistent.