Under global warming, Arctic sea ice has declined significantly in recent decades, with years of extremely low sea ice occurring more frequently. Recent studies suggest that teleconnections with large-scale climate patterns could induce the observed extreme sea ice loss. In this study, a probabilistic analysis of Arctic sea ice was conducted using quantile regression analysis with covariates, including time and climate indices. From temporal trends at quantile levels from 0.01 to 0.99, Arctic sea ice shows statistically significant decreases over all quantile levels, although of different magnitudes at different quantiles. At the representative extreme quantile levels of the 5th and 95th percentiles, the Arctic Oscillation (AO), the North Atlantic Oscillation (NAO), and the Pacific–North American pattern (PNA) have more significant influence on Arctic sea ice than El Niño–Southern Oscillation (ENSO), the Pacific decadal oscillation (PDO), and the Atlantic multidecadal oscillation (AMO). Positive AO as well as positive NAO contribute to low winter sea ice, and a positive PNA contributes to low summer Arctic sea ice. If, in addition to these conditions, there is concurrently positive AMO and PDO, the sea ice decrease is amplified. Teleconnections between Arctic sea ice and the climate patterns were demonstrated through a composite analysis of the climate variables. The anomalously strong anticyclonic circulation during the years of positive AO, NAO, and PNA promotes more sea ice export through Fram Strait, resulting in excessive sea ice loss. The probabilistic analyses of the teleconnections between the Arctic sea ice and climate patterns confirm the crucial role that the climate patterns and their combinations play in overall sea ice reduction, but particularly for the low and high quantiles of sea ice concentration.
Under the impact of global warming, Arctic sea ice (ASI) has been decreasing significantly in recent decades. With a warmer atmosphere, the melting season has lengthened (Stroeve et al. 2017) and the ice cover has become younger and thinner (Kwok 2018; Lindsay and Schweiger 2015; Stroeve et al. 2014). Warming over the Arctic region is twice the global mean, through Arctic amplification. Arctic amplification occurs because the warmer atmosphere induces ice loss (Serreze et al. 2009; Serreze and Barry 2011), leading to feedbacks that accelerate that ice loss, such as water vapor feedbacks (Dessler et al. 2013; Held and Soden 2000; Solomon et al. 2010), cloud feedbacks (Bony et al. 2015; Ceppi et al. 2017; Vavrus 2004; Wetherald and Manabe 1988), lapse-rate feedbacks (Bintanja et al. 2011; Feldl et al. 2017; Graversen et al. 2014), and ice-albedo feedbacks (Kashiwase et al. 2017; Landy et al. 2015). According to climate model simulations that include these feedbacks (Stroeve et al. 2012, 2016), future reduction of Arctic sea ice will be continuous and amplified (Derksen and Brown 2012; Pithan and Mauritsen 2014), with ice-free summers occurring as early as the 2030s, and an ice-free year occurring as early as the 2050s (Onarheim et al. 2018).
The circulation of ASI, characterized by the anticyclonic Beaufort Gyre (BG) and the Transpolar Drift Stream (TDS) that transports ice from the Siberian coast across the North Pole and into the North Atlantic (Serreze and Barrett 2011), is largely controlled by the surface wind field (Thorndike and Colony 1982). The local atmospheric circulation is strongly teleconnected to the climate of remote regions through climate patterns, such as the Arctic Oscillation (AO), the North Atlantic Oscillation (NAO), the Pacific–North American pattern (PNA), and El Niño–Southern Oscillation (ENSO). The AO is the first mode of wintertime sea level pressure (SLP) variability for regions north of 20°N (Thompson and Wallace 1998). The positive (negative) AO is characterized by low (high) SLP anomalies over the Arctic that lead to cyclonic (anticyclonic) atmospheric circulation anomalies (Armitage et al. 2018), an eastern (western) TDS, and a contracted (expanded) BG circulation (Kwok et al. 2013; Rigor et al. 2002). These, in turn, are found to minimize sea ice growth in winter (Hegyi and Taylor 2017). During positive NAO years, the enhanced north–south gradient in SLP over the North Atlantic has also been shown to drive greater southward ice flux through Fram Strait (Armitage et al. 2018; Hilmer and Jung 2000; Hurrell 1995; Kwok 2000; Kwok et al. 2013; Rigor et al. 2002). From the persistent surface forcing of quasi-stationary meridional thermal gradients, the NAO pattern has affected the sea ice variability at interannual time scales (Caian et al. 2018). The PNA is one of the dominant patterns of low-frequency variability in the extratropics of the Northern Hemisphere, and it was strongly teleconnected to the ASI in summer 2007 (L’Heureux et al. 2008). There was an extreme positive phase of PNA that exhibited a 500-hPa cyclonic anomaly west of the Aleutian Islands and a large anticyclonic anomaly south of Alaska. Combined, they drove warm maritime air from lower latitudes poleward, thereby warming the western Arctic (L’Heureux et al. 2008).
ASI is teleconnected to climate patterns in a much more complex manner than a simple linear, univariate, or multivariate regression manner. All PNA, AO, and PDO are related to the anomaly of the Beaufort high (L’Heureux et al. 2008; Moore et al. 2018; Petty 2018; Serreze and Barrett 2011), which is an anticyclone centered north of Alaska that largely controls the mean circulation of the Arctic sea ice cover (L’Heureux et al. 2008; Thorndike and Colony 1982). Simultaneously strong ENSO and NAO episodes were found to be associated with anomalous sea ice extent because of strong SST anomalies and a deepened Icelandic low that led to very strong northerly winds in the Labrador Sea (Mysak et al. 1996). Liu et al. (2004) also found that in the western Arctic the reaction of sea ice to a positive AO is similar to that to El Niño, but that this reaction is opposite in the eastern Arctic. In addition, the interaction between climate oscillations further complicates the teleconnections to sea ice: the NAO responses to ENSO in the central Pacific are mostly linear, while the NAO responses to ENSO in the eastern Pacific are predominantly nonlinear (Q. Ding et al. 2017; Zhang et al. 2019). This nonstationary interaction is affected by the Atlantic multidecadal oscillation (AMO), such that the negative ENSO–NAO interaction in late winter will only be significant when ENSO and AMO are in phase (Zhang et al. 2018). Moreover, during the winter of 2009/10, ENSO was found to amplify the anomalous temperature patterns across the extratropical landmasses of the Northern Hemisphere generated by a moderate to strong AO (Cohen et al. 2010). The simultaneous occurrence of a negative PDO and La Niña events generate strong and significant NAO-like pattern anomalies, but with opposite polarity (S. Ding et al. 2017).
ASI changes nonuniformly under the influence of multiple nonuniform internal or external factors (Q. Ding et al. 2017; S. Ding et al. 2017; Ding et al. 2019; England et al. 2019; Olonscheck et al. 2019). As a time series, sea ice characteristics could be described by statistical measures such as the mean, standard deviation, skewness, and kurtosis. Quantiles, which are widely used in hydrologic frequency analysis, represent the relative magnitudes of particular values in the historical records. For example, in this study an extremely low ice cover is represented by a small quantile, while an extremely high ice cover is represented by a large quantile. Past studies about changes in sea ice are more based on a linear regression that describes the average changes of sea ice. Quantile regression replaces the conditional mean function in linear regression with a conditional quantile function (Barbosa 2008; Koenker and Bassett 1978; Koenker and Hallock 2001). It provides slopes at any arbitrary quantile so that it is possible to analyze the trend analysis of extreme ice conditions accurately. Using quantile regression to detect the trend of ASI extent, Tareghian and Rasmussen (2013) found high variability in the change of sea ice, such that low sea ice extent tends to decrease faster than an average sea ice extent. However, they did not consider spatial patterns of the changes in sea ice. In this study, regression coefficients between gridded sea ice concentration and climate indices were estimated to represent the spatial distribution of the sea ice’s responses to the climate patterns. Next, a multivariate quantile regression model was developed that allows us to project sea ice under different possible combinations of climate patterns. Based on a composite analysis of climate variables such as sea level pressure (SLP), sea surface temperature (SST), geopotential height at 500 hPa (GPH), and wind speeds (UV), atmospheric circulations associated with the teleconnections between ASI and climate patterns are discussed.
This paper is organized as follows: The ASI data and large-scale climate patterns are described in section 2; the technical details on quantile regression and composite analysis are described in section 3; trend detection, the influence of climate patterns on Arctic sea ice, and sea ice projections are given in section 4; a discussion of atmospheric circulations associated with the teleconnection patterns is given in section 5; and conclusions are presented in section 6.
Several datasets were used in this study to investigate the changes in ASI and its teleconnections with large-scale climate patterns. The data of monthly sea ice concentration (ASI) from 1979 to 2017 were downloaded from the near-real-time DMSP SSMIS daily polar gridded sea ice concentrations and the sea ice concentrations from Nimbus-7 SMMR and DMSP SSM/I and SSMIS passive microwave data. This dataset is stored in the NSIDC database of monthly sea ice index, version 3, with a spatial resolution of 25 km × 25 km (Fetterer et al. 2017). Monthly data are grouped into seasonal data such that summer is June–August, autumn is September–November, winter is December–February, and spring is March–May.
The AO index is the first leading mode of the mean height anomalies at 1000 hPa, and a positive AO index means a lower than normal pressure in the Arctic. The NAO index is the leading mode from the rotated principal component analysis of the monthly standardized 500-hPa height anomalies in the region of 20°–90°N with variability at interseasonal and interannual time scales (Bladé et al. 2012; Hurrell 1995; Hurrell and Deser 2010; Ogi et al. 2003). The PNA index is derived from the variability of opposite geopotential height anomalies centered between the Aleutian and the Hawaiian Islands, with positive phases coinciding with a warmer western North America and drier western Canada (Assel 1992; Leathers et al. 1991; Sheridan 2003; Gan et al. 2007). El Niño events based on the Niño-3.4 index are considered active if a 5-month running mean of SST anomalies in the Niño-3.4 region of the tropical Pacific exceeds 0.4°C for 6 months or more (Trenberth 1997). Niño-3.4 was downloaded from the Climate Prediction Center of NOAA/Earth System Research Laboratory (https://www.esrl.noaa.gov/psd/data/climateindices/list/). In addition, the Pacific decadal oscillation (PDO), the leading pattern of sea surface temperature (SST) anomalies in the North Pacific basin (typically, poleward of 20°N; Deser et al. 2016), and the Atlantic multidecadal oscillation (AMO), a near-global-scale multidecadal climate variability with alternating warm and cool phases over large parts of the Northern Hemisphere (Enfield et al. 2001), were considered in the multivariate quantile regression model.
The composites of SST, SLP, and GPH at 500 hPa and wind speed at 500 hPa (UV) were estimated to investigate the effects of climate patterns on Arctic sea ice. Datasets for the above climate variables from 1979 to 2017 are taken from the ERA-Interim monthly dataset of 0.75° × 0.75° spatial resolution (Dee et al. 2011).
a. Quantile regression
Quantile regression has been widely used in trend analysis (Cannon 2018; Fan and Chen 2016; Gao and Franzke 2017; Malik et al. 2016; Tan et al. 2018). Here, it is used to examine the temporal changes of sea ice extent and concentration at multiple quantiles. The teleconnections between sea ice and climate patterns are examined through quantile regression with climate indices as predictor variables, in which regression coefficients at the low (5th), median (50th), and high (95th) quantiles were extracted to represent the responses of extremely low, medium, and extremely high ice cover to the climate patterns. Moreover, multivariate quantile regression models are developed to project sea ice concentrations under the effects of different climate patterns.
Quantile regression is derived from the ordinary linear regression (OLR) model denoted as Yi = βt + γ, with Yi as the dependent variable, time t as the independent variable in temporal trend analysis, which could also be the time series of a climate pattern index, and β and γ as the slope and the y intercept estimated from the OLR model [e.g., Y = f(β, γ, t)]. The parameters β and γ were estimated from the traditional least squares method by minimizing, which is essentially estimating the mean of Y conditioned on t, E[y|t]. However, quantile regression replaces the target function E(y|t) with the quantile of Y conditioned on t, denoted as Q[yτ|t]. For a quantile τ, the quantile regression model is written as Y = g(βτ, γτ, t), where βτ, γτ are the quantile slope and intercept, respectively, which can be estimated by minimizing
where ρτ is the tilted absolute value function (Koenker and Hallock 2001), βτ and γτ are the quantile regression coefficient vectors to be estimated, and τ is the quantile level. The details of the algorithm can be found in Koenker and D’Orey (1987), Tan and Shao (2017), Tan et al. (2018), and Yu et al. (2003), and it has been implemented in the R package “quantreg” (Koenker 2018) used in this study. A detailed description of quantile regression can be found in Barbosa (2008), Cade and Noon (2003), and Koenker and Hallock (2001).
b. Composite analysis
As climate patterns dominate the regional climate variability of a specific area, composite analysis is used to investigate the relationship, if any, between sea ice concentration and climate patterns. In this study, composite values are calculated as the difference of the climate variables (SLP, SST, GPH, and UV) in the extreme positive and negative phases of climate patterns (AO, NAO, PNA, and Niño-3.4). As an example, the composite of SLP is computed from the average observations over the five lowest AO years and the five highest AO years, and are denoted SLP−AO and SLP+AO, respectively. The composite value of SLP, δSLPmax = SLP+AO − SLP−AO, represents the effects of AO on SLP. Bootstrap resampling is used to generate the empirical distribution of the composite values. By resampling δSLPmax 5000 times from the sea ice concentration time series of each grid, we can obtain an empirical distribution of δSLPmax, with the cumulative probability distribution function denoted as F(δSLPmax). If δSLPmax ≤ F−1(0.025) or δSLPmax ≤ F−1(0.975), the composite of SLP with respect to AO is statistically significant at the fifth level, in which F−1(0.025) and F−1(0.975) denote the 2.5th and 97.5th quantile of δSLPmax at the fifth significance level, respectively. Detailed descriptions and the formulas can be found in Tan et al. (2016) and Zhang et al. (2010).
a. Changes in Arctic sea ice
Figure 1 shows the probability density functions and the regression coefficients of seasonal ASI. The winter and spring ASI are distributed more symmetrically than summer and autumn. In other words, the probability density functions of winter and spring sea ice extent (Figs. 1a1 and 1d1) have more bell curve characteristics but those of summer and autumn (Figs. 1b1 and 1c1) tend to skew toward low sea ice extent. The slopes of the regression lines for summer and autumn sea ice extent at low and high quantile levels are steeper than they are at medium quantile levels (Figs. 1b3 and 1c3). Comparing the seasons, the average regression coefficients of summer and autumn are much larger than those of winter and spring (Figs. 1a2, 1b2, 1c2, and 1d2).
The regression coefficients of the gridded sea ice concentration at the 5th, 50th, and 95th quantiles, which respectively represent the extremely low, the medium, and the extremely high ice covers, show the spatial characteristics of changes in ASI. There are significant decreasing trends over the margins of the Arctic Ocean (Fig. 2), where the ice tends to be thinner and younger. The boreal summer sea ice over the Beaufort Sea westward to the Kara Sea shows decreasing trends over the study period (Figs. 2f,g). Also during boreal summer, sea ice at the 95th quantile has smaller areas showing a decreasing trend (Figs. 2j,k) compared with the 50th quantile (Figs. 2f,g), while that at the 5th quantile has larger areas and regression coefficients (Figs. 2b,c).
During spring and winter, areas with decreasing ice concentrations are small and concentrated in the Barents Sea and the Sea of Okhotsk (Figs. 2e,h). The total areas with a decreasing trend in sea ice concentration do not differ much between the quantiles. However, decreasing trends of the 5th quantile in the Barents Sea were as large as 40% per decade, larger than those during the boreal summer. Based on trends of the sea ice extent, the decline of summer sea ice would be more likely to occur faster in the Chukchi Sea. Also, winter sea ice concentration in the Barents Sea is more likely to be much lower in ice concentrations than in other regions, which is in agreement with previous findings (Cavalieri and Parkinson 2012; Kay et al. 2011; Wang and Overland 2009).
b. Effects of teleconnections with climate patterns on Arctic sea ice
Figures 3–6 show teleconnection trends by quantile levels and climate indices for seasonal sea ice concentrations. The difference in sea ice conditions between the 5th (95th) and the 50th quantiles show the difference in how sea ice covers responds to the influence of climate patterns under extremely low (high), and mean sea ice conditions.
1) Arctic Oscillation
Figure 3 shows the regression coefficients of sea ice concentration at the 5th, 50th, and 95th quantiles with seasonal AO indices as the covariate. During boreal winter–spring, AO has a negative relationship with the sea ice at the margins of the Arctic Ocean. Compared with the sea ice of the 50th quantile (Figs. 3e,h), both the 5th and 95th quantiles of sea ice concentration have larger areas that show a stronger relationship with AO, especially in the Sea of Okhotsk (Figs. 3i,l), Davis Strait (Figs. 3a,d,i,h), Hudson Bay (Fig. 3d), and the Barents Sea (Figs. 3a,e,i). These regions are strongly affected by the extreme positive AO in winter, effects that the ordinary linear regression analysis underestimated (Figs. 3e–h). The effects of AO on the autumn sea ice concentration are also found to be strongly related to AO at the 5th and the 95th quantiles (Figs. 3c,k) more than at the 50th (Fig. 3g). The 50th quantile of sea ice concentration has large areas positively correlated to the AO index in autumn (Fig. 3g) because both the AO index and sea ice experienced downward trends in the boreal summer–autumn of recent decades. However, the 5th quantiles of sea ice in the Kara Sea and the Chukchi Sea are still negatively correlated to AO in summer (Fig. 3b).
2) North Atlantic Oscillation
In Fig. 4, it was found that the responses of Arctic sea ice to NAO are different in the western and eastern Arctic Ocean at some quantiles. In boreal winter–spring when the NAO is strongest, a large area of sea ice at the 5th quantile in the Barents and Chukchi Seas has a strong negative relationship with the NAO, and that in Davis Strait it has a strong positive relationship (Figs. 4a,d) than at the 50th quantile (Figs. 4e,h). Sea ice at the 95th quantile of the Sea of Okhotsk and the Barents Sea has a significant decreasing trend with the NAO index. Again, the results show that modeling the sea ice concentration with the NAO index as the covariate would underestimate the decreasing trend in these regions. A large area of summer sea ice has a positive relationship with the NAO index (Fig. 4f), which may be partly due to the downward trend of the summer NAO itself in recent decades. In autumn, the relationship between NAO and sea ice at low quantile sea ice is negative in the East Siberian Sea but positive in the other regions (Fig. 4c).
3) Pacific–North American pattern
Figure 5 shows the regression coefficients between Arctic sea ice with PNA as the covariate. From the coefficients of the 50th quantile, it is evident that a large area of summer and autumn sea ice from the Beaufort Sea westward to the north of Barents Sea has a negative relationship with the PNA (Figs. 5f,g), and this result agrees with the findings of L’Heureux et al. (2008). In the 5th quantile, the decreasing trend in the Chukchi Sea is strong, but the trends are positive in the Beaufort and Laptev Seas (Figs. 5b,c). These summer and autumn spatially distributed regression coefficients imply that a strong positive PNA phase may lead to summer and autumn sea ice loss in the Chukchi Sea, but not so in the Beaufort and Laptev Seas. In winter, sea ice in the Barents Sea has a strong negative relationship with PNA at the 5th and 95th quantiles (Figs. 1d,l) but does not have any significant relationship with PNA at the 50th quantile (Fig. 5h).
ENSO, which originates from the tropical Pacific, has a weaker relationship with Arctic sea ice than other climate patterns (Fig. 6). From the west of the Beaufort Sea westward to the East Siberian Sea, sea ice is positively related to Niño-3.4 at low quantiles in summer and autumn (Figs. 6b,c), while at the 95th quantile the coastal sea ice is found to be negatively related to Niño-3.4. There is very little relationship between sea ice and ENSO at the 50th quantile, regardless of season.
c. Effects of combined climate patterns on Arctic sea ice
ASI is teleconnected to climate patterns in a much more complex manner than a simple linear univariate correlation. From the above univariate quantile regressions, we found that the AO and NAO have similar relationships with ASI, while the relationships with PNA and ENSO are progressively weaker. Therefore, a multivariable quantile regression model was used to project the effects of combined climate patterns on the extremely low sea ice concentrations. Since the AMO and PDO are found to be nonlinearly related to the Arctic (Zhang et al. 2018), the model also considers the AMO and PDO as in phase with each other according to their phase relationship (Li et al. 2016; McCabe et al. 2004). The 95th and 5th quantiles of the climate indices were chosen arbitrarily to represent the extremely positive and negative phases of the climate patterns. The projected anomalies of sea ice concentration under the extreme phases of climate patterns, minus the long-term averages (1979–2018), are presented in Fig. 7.
The spring and winter sea ice concentration anomalies under +AO/+NAO are found to be lower than the multiyear average (Figs. 7a1,b1) as expected. When AMO and PDO are both in the extremely positive phase at the same time with +AO/+NAO, the sea ice concentration is lower (Figs. 7a2,b2). Similar conditions in sea ice concentrations are found during boreal summer with active PNA and ENSO. The summer and autumn sea ice concentration is slightly lower than the multiyear average with extremely positive PNA and El Niño (Figs. 7c1,d1). Combined with simultaneously positive AMO and PDO, the summer and autumn sea ice concentration tends to be lower than with only a positive PNA or a positive El Niño (Figs. 7c2,d2). These findings indicate that the extremely positive AMO and PDO can further enhance the negative relationships between winter ASI with AO/NAO, and for summer ASI with PNA/ENSO.
Since strong relationships between large-scale climate variability and ASI quantiles have here been quantified, the seasonal influence of climate patterns under positive and negative phases on ASI was further investigated. The atmospheric circulation mechanisms that may give rise to the teleconnection patterns observed between ASI and AO, NAO, PNA, and ENSO over 1978–2017 were analyzed by a composite analysis of climate variables, respectively.
a. Atmospheric circulation mechanisms for the teleconnection with AO
The composite results show that the sea level pressure over Asia, the western Siberian Sea, and the Chukchi Sea tends to be lower during positive AO boreal winters, resulting in a northerly wind that brings cold and dry air to Asia (Fig. 8b1).
Although AO has a relatively weak temporal variability in winters, quantile regression shows that there is a complex spatially distributed response of ASI to AO in summer (cf. Figs. 3f,g). The composite analysis of climate variability also demonstrates this teleconnection. Lower SST and GPH over the North Pacific and North Atlantic Ocean during a positive AO phase favor the growth of ASI. On the other hand, SLP over the Arctic is higher during negative AO than during positive AO, and the former favors clear-sky conditions. Less cloud cover allows more downwelling shortwave radiative fluxes that enhance surface (and potentially basal) melting of ASI (Kay et al. 2008). The wind field composite shows dominant anticlockwise wind anomalies over the Arctic Ocean during positive AO autumns that could promote sea ice export through Fram Strait, resulting in less Arctic sea ice.
b. Atmospheric circulation mechanisms for the teleconnection with NAO
As predicted by the quantile regression, seasonal sea ice over the Labrador Sea and north of Greenland is higher under positive NAO (cf. Figs. 4a,d) due to the enhanced southerly winds (Figs. 9a1,b1), which is consistent with the findings of Hurrell and Deser (2010). Sea ice in the Barents Sea is lower during strong positive NAO winters and springs (cf. Figs. 4a,d,e,h,i,l). Ottersen and Stenseth (2001) found that the NAO, SST in the Barents Sea, and inflow from the Atlantic are interrelated as proposed by Ådlandsvik and Loeng (1991): under a positive NAO phase, there is a strong Icelandic low (Fig. 9a1) and enhanced cyclonic circulation that increases the inflow of warm water into the Barents Sea (Figs. 9a2,b2), which in turn enhances upwelling heat fluxes and induces low SLP. However, under a positive NAO summer, the SLP over the Arctic region tends to be higher (Fig. 9c1), which leads to anticyclonic circulation anomalies over the Arctic Ocean that favor a low sea ice extent through ice export (Serreze et al. 2016).
Sea ice concentration over the eastern Siberian and the Chukchi Seas tends to be lower during positive NAO autumns (cf. Figs. 4c,g,k) and is attributed to the near-surface anticyclonic wind over the Arctic. Since the summers of 2007, low-level circulation over the Arctic has been much more anticyclonic than in prior years for unknown reasons (Ogi and Wallace 2012). The effects of the NAO on boreal summer sea ice are complicated because there are multiple mechanisms at play. This complex interplay warrants further research.
c. Atmospheric circulation mechanisms for the teleconnection with PNA
The composite analysis of the seasonal wind fields shows a strong anticyclone over the Laptev and Barents Seas during positive PNA years (Figs. 10a1–d1) that drives the Arctic sea ice to the Atlantic, resulting in reduced ASI. A positive PNA also enhances poleward shifting waves with alternating centers of anomalous pressure that turn northeastward over the North Pacific Ocean, over western-central Canada, and then southeastward over central-eastern North America. This circulation pattern increases heat intrusion to the Arctic (L’Heureux et al. 2008), which may lead to the earlier onset of ASI melting by ~2–3 days per decade (Wang et al. 2013). The strong high pressure over the Chukchi Sea (Fig. 10d1) during positive PNA autumns leads to strong southerly winds through the Bering Strait, resulting in less Arctic sea ice and giving rise to a negative relationship between ASI and the PNA index (cf. Figs. 5c,g,k).
d. Atmospheric circulation mechanisms for the teleconnection with ENSO
The quantile regression coefficients between ASI and Niño-3.4 are relatively modest compared to those for AO, NAO, and PNA, which may be partly because ENSO needs more time to teleconnect with the ASI (Park et al. 2015). Furthermore, the atmospheric circulation is significantly different during El Niño and La Niña episodes, and the average Arctic SST during winter tends to be colder during El Niño but warmer during La Niña events (Fig. 11a2), as shown by Lee (2012).
During strong El Niño winters, there is an enhanced low pressure over Iceland and a strong anticyclone over the Laptev and Barents Seas (Fig. 11a1). Liu et al. (2004) found that ENSO has similar effects on the eastern Arctic as the AO, in that sea ice expands during the negative phase (La Niña) and shrinks during the positive phase (El Niño), while the western Arctic sea ice shrinks during La Niña but expands during El Niño. In autumn, the strong low pressure over the Barents Sea and high pressure over southern Greenland favor a strong southward wind through Fram Strait (Fig. 11d1), which may accelerate the export of sea ice to the North Atlantic.
This study extends the application of quantile regression to the problem of Arctic sea ice variability and the covariability of sea ice with climate patterns. Through this new analysis technique, the trends of Arctic sea ice are determined to be that the sea ice of low quantiles decreases faster than the average, especially in most areas of the Beaufort Sea westward to the Kara Sea. Arctic sea ice of low quantiles is also found to have stronger teleconnections with climate patterns than the average. A projection of Arctic sea ice through a multivariate quantile regression model demonstrates that particular combined climate patterns have a stronger influence on Arctic sea ice than an individual climate pattern. The physical mechanisms behind these teleconnections were investigated through a composite analysis of climate variables (SLP, SST, GPH, UV). The significantly different climate variables under the influence of extreme positive and negative phases of the respective climate indices reveal the existence of forcing mechanisms behind the teleconnections between Arctic sea ice and the climate patterns. These findings aim to improve the understanding of Arctic sea ice variability, and its complex relationship with large-scale climate patterns and, hopefully, benefit the prediction of Arctic sea ice.
The first author was partly funded by the China Scholarship Council (CSC) of P.R. China, and the University of Alberta. We also acknowledge support from the Natural Sciences and Engineering Research Council of Canada (NSERC). Monthly time series of AO, NAO, PNA, and Niño-3.4 were downloaded from the Climate Prediction Center of NOAA on Earth System Research Laboratory website (https://www.esrl.noaa.gov/psd/data/climateindices/list/).