This study uses an extensive dataset of monthly surface air temperature (SAT) records (including previously unutilized) from high-latitude (>60°N) meteorological land stations. Most records have been updated by very recent observations (up to December 2008). Using these data, a high-latitude warming rate of 1.36°C century−1 is documented for 1875–2008—the trend is almost 2 times stronger than the Northern Hemisphere trend (0.79°C century−1), with an accelerated warming rate in the most recent decade (1.35°C decade−1). Stronger warming in high-latitude regions is a manifestation of polar amplification (PA). Changes in SAT suggest two spatial scales of PA—hemispheric and local. A new stable statistical measure of PA linking high-latitude and hemispheric temperature anomalies via a regression relationship is proposed. For 1875–2008, this measure yields PA of ∼1.62. Local PA related to the ice–albedo feedback mechanisms is autumnal and coastal, extending several hundred kilometers inland. Heat budget estimates suggest that a recent reduction of arctic ice and anomalously high SATs cannot be explained by ice–albedo feedback mechanisms alone, and the role of large-scale mechanisms of PA of global warming should not be overlooked.
Changes in the arctic climate system over the past decade were exceptional in the history of arctic observations (e.g., Lindsay et al. 2009; Belchansky et al. 2008; Meier et al. 2007; Polyakov et al. 2005; Walsh and Chapman 2001), culminating with the summer of 2007 when the arctic ice retreat broke all records (Stroeve et al. 2008; Comiso et al. 2008). Despite the fundamental importance of high-latitude changes for global climate, there are numerous gaps in our understanding of how the system functions and what forces are driving changes in the Arctic. In particular, information is limited or controversial regarding the role of enhanced high-latitude warming, called polar amplification (PA), in long-term surface air temperature (SAT) variations and modern arctic warming. Positive feedbacks are believed to lead to PA, as shown by analysis of observed SAT (e.g., Vinnikov et al. 1980; Serreze et al. 2009) and as predicted by general circulation models (e.g., Serreze and Francis 2006; Holland and Bitz 2003). For example, in the ice–albedo feedback mechanism, warming leads to a reduction of ice and snow coverage and decreasing albedo, resulting in further snow and sea ice retreat (Manabe and Stouffer 1980). The most pronounced effect of this feedback, expressed as the strongest ocean–atmosphere upward heat flux and SAT rise, may be found in areas of maximum ice retreat. Recent warm SAT anomalies are consistent with climate model projections in response to reductions in sea ice extent (Overland et al. 2008). Local PA governed by this feedback is seasonal, extending into autumn and early winter, as shown by regional coupled ice–ocean models (Lindsay and Zhang 2005) and global climate models (Murray and Walsh 2005). The exact timing of warming and its magnitude and spatial pattern are, however, sensitive to simulated sea ice retreat, which varies over a wide range from model to model.
There is a host of other important feedbacks leading to PA. Positive feedback related to ice thickness variations is believed to lead to enhanced high-latitude climate change. Model simulations showed stronger PA with thinner ice cover (Holland and Bitz 2003), confirming earlier modeling results by Rind et al. (1995). This mechanism may be viewed as a part of ice–albedo feedback since thinner ice is subject to faster melt, resulting in more openings in ice cover and a corresponding change of surface albedo (important in summer when there is a lot of sunlight) and heat release from the ocean (important during the cold season). Moreover, heat accumulated by the ocean as a result of the ice–albedo feedback mechanism is partially spent making ice thinner, thus leading in winter and autumn to an increase of SATs (Manabe and Stouffer 1980). Therefore, we will refer to mechanisms related to the change of reflective and insulating properties of the ice/ocean surface as ice–albedo feedback mechanisms. Uncertainties related to modeling both ice thickness and ice extent are substantial (ACIA 2005). Therefore, current model estimates of the quantitative contribution of this feedback to PA should be viewed with caution.
Ice–albedo feedback is widely accepted as one of the strongest mechanisms of PA of global warming (e.g., Manabe and Stouffer 1980). However, Winton (2006) has recently proposed that the ice–albedo feedbacks do not dominate the simulated PA; longwave radiation effects appear to play an equally important role. Using modeling results, Graversen and Wang (2009) suggested even a stronger, dominant, role of the longwave radiation feedback mechanism. Even with the ice–albedo feedback excluded, several GCM studies (Schneider et al. 1997, 1999; Alexeev 2003; Alexeev et al. 2005; Langen and Alexeev 2005, 2007) have found feedbacks involving changes in longwave radiation and turbulent forcing on the high-latitude surface temperature, which are as efficient as ice–albedo feedback mechanisms. These mechanisms of hemispheric-scale PA may be linked to interaction between equator-to-pole air temperature gradients and poleward atmospheric heat transports. These mechanisms will be referred to as the “advective” mechanism of PA. This is a year-round-acting mechanism, amplified during the cold season when the heat and moisture meridional transports are at a maximum. Large-scale PA may also be forced by “local” longwave radiative forcing linking “local to the Arctic” SAT anomalies and changes in surface properties to changes of atmospheric composition (cloudiness and humidity), which has its strongest expression in spring (Francis and Hunter 2007). This will be called the “radiative” mechanism. Note that arctic clouds, particularly in summer, may add strong local arctic flavor to this mechanism (Holland and Bitz 2003). However, the source of high-latitude humidity may still be nonlocal, which could make it difficult to distinguish between local and nonlocal mechanisms.
Experiments with climate models demonstrated that the reduction of the Northern Hemisphere (NH) snow cover may lead to approximately the same increase of radiative heating as that due to reduction of sea ice cover (Winton 2006). Analysis of 1972–2006 satellite observations demonstrated substantial reduction of the NH snow cover, particularly in spring (1.26 × 106 km2/35yr, Dery and Brown 2007). The maximum contribution of the positive snow–albedo feedback mechanism should be in spring. We note, however, that the impact of this feedback on PA is not quantified. Moreover, the documented earlier snowmelt by itself may be a result of an increase of incoming longwave radiation due to an increase of SAT and concentration of water vapor in the atmosphere.
Identification of enhanced high-latitude warming using long-term climate trends (i.e., Vinnikov et al. 1980) is complicated by strong arctic intrinsic variability dominated by multidecadal fluctuations (e.g., Polyakov and Johnson 2000; Polyakov et al. 2008). For example, Polyakov et al. (2002) found similar Arctic and Northern Hemispheric air temperature trends during the twentieth century, thus lacking manifestation of enhanced high-latitude warming. However, the exceptional temperature rise observed in the twenty-first century sheds new light on the relative role of PA. We will demonstrate that the ratio of long-term Northern Polar Area (NPA) and NH SAT trends as a measure of PA, widely used in previous studies, may suffer from serious shortcomings. This research assesses enhanced arctic warming via analysis of an extensive collection of meteorological observations including previously unutilized ones and station data that have been updated by very recent (until December 2008) observations. Building upon these measurements, we have evaluated the impacts of PA on large-scale long-term changes of arctic SATs using a new measure of PA.
2. Data and methods
The observational database for the NPA consists of monthly surface air temperature (SAT) records from 441 meteorological stations (Fig. 1). Following ACIA (2005), we consider the NPA as the area north of 60°N. Since interpolation at the southern boundary sometimes requires a wider domain, our collection of data includes observations limited by 59°N. There are a few exceptions for areas with poor data coverage, where we used observations from the region limited by 58°–59°N. In addition, our database includes two very long meteorological records from northern Scotland (57°–58°N). The only marine meteorological station used in our database is the Weather Ship Mike, which provides information from the central Norwegian Sea. Most records have been updated by very recent (until December 2008) observations.
The observational database used in this study combines the following datasets:
monthly mean SATs from the Climate Research Unit, University of East Anglia (CRU UEA) provided by P. Jones;
monthly data from the Web site of the NOAA National Climatic Data Center (NCDC) and NOAA Global Historical Climatology Network (GHCN), http://www.ncdc.noaa.gov/oa/climate/ghcn-monthly/index.php;
monthly data from the Web site of the Goddard Institute for Space Studies (GISS), http://data.giss.nasa.gov/gistemp/;
monthly data from the University Corporation for Atmospheric Research (UCAR) Web site, http://dss.ucar.edu/datasets/ds570.0/data/;
The Environment Canada, National Climate Data and Information Archive, available online at http://climate.weatheroffice.ec.gc.ca/climateData/;
The Environmental Working Group Arctic Meteorology and Climate Atlas (Fetterer and Radionov 2000), available online at http://nsidc.org/data/docs/noaa/g01938_ewg_arctic_met_atlas/;
National Snow and Ice Data Center (NSIDC) meteorological data from the Russian Arctic, 1961–2000, available online at http://nsidc.org/data/docs/noaa/g02141_esdimmet/index.html;
monthly SAT for several Russian meteorological stations from the All-Russian Institute of HydroMeteorological Information, World Data Center (RIHMI-WDC) Web site, http://meteo.ru;
monthly and daily data from Alaskan meteorological stations provided by D. Atkinson and M. Shulski;
hourly data from the NOAA Web site, ftp://ftp.ncdc.noaa.gov/pub/data/noaa/;
daily observations from Russian stations provided by P. Groisman;
daily data from European stations provided by The Royal Netherlands Meteorological Institute (KNMI) Web site, http://eca.knmi.nl/dailydata/datadictionaryall.php;
hourly data from the Web site of the Russian Academy of Science Space Research Institute, Moscow, Russia, http://meteo.infospace.ru/;
hourly data from the Web site “Raspisanie Pogodi,” http://rp5.ru;
monthly data from the Web site of the Icelandic Meteorological Office, http://www.vedur.is/;
monthly data from the Finnish Meteorological Institute Web site, http://www.fmi.fi/en/;
monthly data from the Web site Rimfrost, http://rimfrost.no;
monthly data from the meteorological station Abisko (Sweden) for 2000–07 provided by C. Jonasson, Abisko Scientific Research Station;
monthly data from several Scandinavian stations updated using the NordKlim project archive Web site, http://www.smhi.se/hfa_coord/nordklim/;
monthly data for several Scandinavian and Greenland stations updated or corrected using archive created under the Nordic Arctic Research Program project Web site, http://projects.dnmi.no/~narp/; and
data from several Greenland stations from the Danish Meteorological Institute Web site, http://www.dmi.dk/dmi/tr00-18-data_files.zip.
One of the major previously unutilized and not digitized data sources in our study was the Climatologic Reference Book of the USSR, a series published in the 1930s–60s (Rubinshtein et al. 1930, 1933; Climatologic Reference Book of the USSR 1954a,b,c,d, 1956, 1957, 1962). These books contain, among other information, monthly SATs for Russian meteorological stations for separate years from the beginning of instrumental observations. Many stations in the European part of Russia (Russian North) and in West Siberia were established in the middle or end of the nineteenth century. The longest record is from the St. Petersburg station, which was established in 1743. The use of historical Russian archives provided spatiotemporal data coverage for this region of the NPA comparable to that for the Scandinavian region (Fig. 1). Daily and hourly data were processed using data quality control procedure outlined in appendix A.
The record from each station was reduced to monthly anomalies relative to 1961–2000. Annual SAT anomalies were computed only when at least six monthly values were available. These data were used to compose annual and seasonal time series of SAT anomalies for the entire NPA. To minimize the effect of spatially inhomogeneous data coverage, we used a technique similar to the climate anomaly method (CAM) (Jones et al. 1999). In this method, the total area is divided into boxes, and anomaly time series are averaged within each box. The resulting averaged time series for each grid box are averaged again to obtain a single “global” time series. In our case, individual records from land stations over 27 regions were averaged to produce regional time series. Each of the averaging areas was limited by 10° along latitude and 20° along longitude. The resulting records were averaged to produce the final record for the entire NPA. Upon comparison, the difference between the two resulting time series (obtained from simple averaging of all available records and by the CAM) is insignificant (for SAT, the correlation is 0.90, and the standard deviations (SDs), which are not statistically distinguishable, are equal to 0.77). Trends in the SAT time series were evaluated by the least squares best-fit method.
Monthly winds, net atmospheric heat fluxes, and ice concentrations were provided by the National Centers for Environmental Prediction (NCEP)–National Center for Atmospheric Research (NCAR) Reanalysis and obtained from the NCAR data server (Kalnay et al. 1996). (These data are used for preparation of Fig. 10 only.) The CRUTEM3v dataset combines SAT over land (Jones and Moberg 2003; Brohan et al. 2006) and is used to obtain NH SAT time series (available online at www.cru.uea.ac.uk).
The major emphasis of this study is an analysis of observations augmented with a model analysis that focuses on the long-term SAT variations. For this purpose, we used 42 available model runs for 1871–1999 from a suite of fully coupled global climate models. These models were used in the Coupled Model Intercomparison Project (CMIP3) (Meehl et al. 2007).
3. Northern Polar Area surface air temperatures
Composite annual and seasonal time series of SAT anomalies for the NPA are shown in Fig. 2. NPA temperatures were exceptionally high, reaching 1.72°C in 2003, 2.11°C in 2005, and 2.18°C in 2007, the warmest years in the history of instrumental observations. These years were also associated with the minimum Arctic ice cover. Over the last decade, Arctic warming became stronger and seasonally more uniform compared with warming in the 1930s–40s (Fig. 3). The earlier warming was expressed more in the autumn and winter, with 5–6-yr-long cooling/warming events, whereas since the late 1980s the SAT increase in spring was as strong as in winter and autumn (Figs. 2 and 3). There is, however, substantial spatiotemporal heterogeneity: for example, in winter 2008 SAT in Europe was exceptionally high but a cooler spring and summer made 2008 overall a cooler year compared to 2007 (Fig. 4).
Accelerated warming of the NPA in recent decades is captured by SAT trends (Fig. 2, Table 1). For example, the trend over the last 50 years (1959–2008) was 0.36°C decade−1, increasing to 0.64°C decade−1 since 1979. NPA SAT trends over longer periods of time were also very strong, up to 1.36°C century−1 for 1875–2008, which is almost two times greater than NH SAT trends (0.79°C century−1). However, since 1901 NPA trends decreased to 1.20°C century−1 (Table 1). Strong multidecadal variability (MDV) affects the behavior of NPA SAT trends, which allowed Polyakov et al. (2002) to suggest that MDV dominates the SAT fluctuations. The inherent difficulty in differentiating between trends and strong long-term fluctuations makes quantitative analysis of the contribution of trend versus MDV difficult. However, the Arctic warming in the last decade was exceptionally strong, reaching 1.35°C decade−1, thus casting less doubt about the existence of a long-term externally forced warming trend. Spatial patterns of the SAT trends based on annual data are mostly dominated by positive values (particularly over northern Europe) for both longer (1875–2008 and 1901–2008) and shorter (1979–2008) periods. Negative SAT trends appear for some seasons over limited areas only, mostly in the Pacific sector, probably due to processes governed by the Pacific decadal oscillation. Stronger warming in the maritime region, particularly in autumn and spring (Fig. 5), may be considered as an expression of PA. Understanding the importance of careful evaluation of statistical significance of trends, we propose a new nonparametric approach, which was used in this study (see appendix B for details).
4. Polar amplification
Trends of winter–spring NPA SATs over extended periods of time were generally stronger than in summer–autumn (Fig. 2) and were enhanced compared with the NH trends (Table 1). However, in the recent ∼15–20 years, autumnal maritime (<200 km inland) SAT anomalies were amplified (Fig. 5). This change of the seasonal/interannual PA pattern suggests at least two different spatial scales of mechanisms of PA: large-scale (hemispheric) and local (limited by NPA).
a. Large-scale PA
A strong contribution of large-scale (hemispheric) PA to the spatial pattern of global warming is evident from, for example, Fig. 6, showing amplification of zonally averaged SAT variations from tropical through high-latitude regions. A similar figure, but for high-latitude (>50°N) zonal bands only, was presented by Serreze and Francis (2006, their Fig. 8), who found that the warming in the 1930s was confined to the high-latitude Arctic, whereas the warming that started in the 1970s has more of a “global” expression. Extending the Serreze and Francis Fig. 8 southward (cf. our Fig. 6) provides a new perspective on high-latitude warming. Figure 6 suggests that SATs north of 55°N lead low-latitude SATs, implying the important role of long-term variability in the North Atlantic forced by high-latitude fluctuations. We note, however, that because of the limited quality of observations and weak signal this suggestion should be viewed with caution.
A direct comparison of Arctic and NH SAT trends used in earlier studies to define a twofold magnitude of the PA (e.g., Vinnikov et al. 1980) may be compromised owing to strong variability, particularly of MDV (Polyakov et al. 2002). For example, the ratio of NPA and NH trends computed for 1924–2000 is atr = 0.24, which is difficult to accept as a measure of accelerated NPA warming. We propose a new quantitative measure of PA linking NPA and NH temperature anomalies T ′NPA(t) and T ′NH(t) via a linear regression equation T ′NPA(t) = arT ′NH + br. The new measure of PA yields ar = 1.52 for 1959–2000, ar = 1.62 for 1875–2008, and ar = 1.45 for 1901–2008, and these new annual and seasonal regression-based estimates of PA are more stable compared with the estimates based on the ratio of trends atr (Table 2). Particularly striking is the difference between ar and atr in autumn, with a vast spread of atr values and a narrow 6% band of ar change. Such a drastic difference may be explained by peculiarities of the statistical distribution of atr.
To illustrate this point, we used a suite of 42 CMIP model runs covering the period from 1871 through 1999. Figure 7a shows model-based ar and atr with a much stronger spread of atr values compared with ar as quantified by a twofold sample estimate of the standard deviation (SD) of atr relative to the SD of ar. Dependence of NPA and NH trends with their twofold amplification in the NPA is shown in Fig. 7b. We note that observational and modeling estimates agree well. Some model experiments produce very weak NH trends, resulting in extreme values of atr that have no physical meaning and cannot be used to characterize quality of model simulations. Further analysis of the nature of this phenomenon is possible if we assume that the NPA and NH are statistically independent and that both have normal distributions with zero means. In this case, NPA and NH SAT trends will be subject to a Gaussian statistical distribution with zero means. Their ratio, atr, will be subject to a Cauchy probability distribution, statistical moments of which are not defined (i.e., the so-called heavy tail distribution, Cramer 1999). Therefore, atr as a measure of PA may lead to unrealistic estimates. Figure 7c shows a time series of SDs derived from an ensemble of model-based ar and atr. This figure suggests that ar provides stable estimates of PA, whereas the wide spread of atr precludes the use of atr as a meaningful measure of PA. However, we expect that with a warmer climate, the difference between atr and ar may diminish because an increasingly higher portion of T ′NPA(t) and T ′NH(t) variance will be yielded by trends.
The steadiness of model-based ar (Fig. 7d) suggests that the relationship between NH and NPA SATs is preserved regardless of whether the preindustrial period or the last decades of the strongest anthropogenic warming are considered. Polar amplification may be described as an intrinsic mode of climate variability (Alexeev et al. 2005; Langen and Alexeev 2007); therefore, human-induced climate change will be projected on this mode without altering its spatial structure, in agreement with the concept of nonlinear climate dynamics perspective proposed by Palmer (1999). In that case, climate change and its variations at various time scales will produce similar estimates of PA.
Large-scale PA derived from observational data is evident from Table 2 with annual 1875–2008 ar = 1.62 ± 0.10, close to the 1.56 estimated from ensemble-average PA derived from a suite of 42 CMIP model runs. There is a seasonal cycle with maximum PA in autumn and spring (Table 2, Fig. 5). The former can be explained by local ice–albedo feedback mechanisms; however the latter cannot. Indeed, this seasonal cycle of PA is consistent with a large-scale winter and springtime “radiative” mechanism of PA and a year-round “advective” mechanism of PA, which is strongest in the cold season. The snow–albedo feedback mechanism may also be a plausible explanation for the springtime PA. During the last 15–20 years, PA during the cold season was very strong (Figs. 3 and 8). For example, in 2006–08 high-latitude winter–spring SATs were exceptionally high (Fig. 4). Unfortunately, observational SAT data preclude us from separating the contribution of “advective” and “radiative” mechanisms into the large-scale PA.
To ensure that our estimates of large-scale PA are not “contaminated” by effects of local ice–albedo feedbacks, we repeated our calculations of ar presented in Table 2 by using NPA records from meteorological stations located at least 250 km away from the coast (Table 3). In general, the differences between these two calculations are small, within 15%. However, there is a substantial, almost twofold, increase of PA during the cold season. That may be a manifestation of a positive feedback mechanism among the SAT, available precipitable water, and incoming longwave radiation, which is strongest in regions with the lowest SATs. The inland meteorological stations have the lowest winter and spring SATs. Since available precipitable water in the atmospheric column depends heavily on air temperature, and incoming longwave radiation is proportional to logarithm of precipitable water content (Zhang et al. 2001), the sensitivity of incoming longwave radiation to SAT change is higher for inland regions.
b. Local PA
PA at the local, Arctic-wide, scale is related to ice–albedo feedback mechanisms that are evident in maritime areas in autumn when the ice cover is seasonally at a minimum (Figs. 4, 5, 9, and 10; e.g., Manabe and Stouffer 1994). For example, autumnal 1979–2008 SAT trends are amplified at the coastal stations in Siberia and the Canadian Archipelago (Fig. 5). SAT anomalies in autumns during recent years (2006–08) are strongest in the maritime NPA (Figs. 4 and 9). Strong linkage of ice-edge dynamics and air–sea heat fluxes (and SAT, not shown; e.g., Overland et al. 2008) is evident from Fig. 10. The SAT increase due to ice–albedo feedback may be traced up to several hundred kilometers from the ice edge into the interior of the Siberian region, as evident from Fig. 9. Note also that, according to our dataset, the ice–albedo feedback mechanism is apparent during extended warm periods only (i.e., in the 1930s–40s and recent decades) and is not evident during cold periods (Fig. 9). However, we note that this may be partially due to a geometrical effect when the limited NPA used for our analysis does not include dynamics of the sea ice edge expanding far beyond the NPA.
Serreze et al. (2009), using NCEP reanalysis data, also found a dominant role of autumnal NPA warming in recent decades. However, our data suggest a relatively uniform spread of NPA warming over seasons in recent decades [cf. Fig. 3 from Serreze et al. (2009) and our Fig. 8]. Moreover, NPA SAT trends computed over extended time periods are stronger in winter and spring (Table 1). For example, over the last 50 years, the spring SAT trend was 4.67°C/100 yr, which is ∼30% stronger than the autumn SAT trend. Only over the last three decades has the autumnal SAT trend exceeded slightly the winter and spring trends (Table 1). These results suggest that on the hemispheric scale the relative role of local ice–albedo feedback mechanisms may be moderate compared with the contribution of large-scale mechanisms of PA.
We further explore the role of ice–albedo feedback mechanisms in recent NPA warming over the Arctic Ocean. Oceanographic observations carried out in September 2007 showed a substantial—up to 2°C and more at some locations—water temperature increase in the upper h = 25 m layer in the Siberian sector of the Arctic Ocean, which has never been ice-free in the history of observations [Fig. 11; for data description, see Polyakov et al. (2009), manuscript submitted to J. Phys. Oceanogr.]. This upper ocean warming is used to estimate effects of atmospheric forcing due to a change of surface radiative properties. Based on these oceanographic observations, we estimate that the observed upper-ocean warming in summer 2007 led to an additional accumulation of heat of
where Θ is potential temperature, Θf is the freezing temperature at 0 db, ρw ≈ 1030 kg m−3 is water density, and cp = 4186 J/(kg K) is the specific heat of water in the area S ≈ 2.4 × 106 km2 covered by the observations. Dynamical mechanisms such as wind-driven ice compression toward the western Arctic, enhanced transport of ice with the Transpolar Drift through Fram Strait, and ice export to Baffin Bay were the major cause for the dramatic summer 2007 ice loss (e.g., Nghiem et al. 2007). Therefore the portion of heat uptake used to melt ice in the summer 2007 opening is neglected in this analysis. Note also that an increase of the upper-layer thickness in our calculation of Q led to a negligible increase in computed atmospheric heat uptake by the upper ocean. Normalized by S, this Q is equivalent to ∼283 MJ m−2 of heat content Q̃. This is ∼65% of the solar heating of 440 MJ m−2 estimated by Steele et al. (2008) based on satellite surface temperature observations in 2007, which covered the entire ice-free area of the Arctic Ocean. This apparent difference of estimates of Q̃ may be partially due to the fact that satellites measure surface skin temperature, which in summer is higher than the temperature of the bulk of the upper ocean layer. One more potential reason for this difference is a limited area of oceanographic surveys compared with broader satellite-based coverage.
We compared the anomalous upper-ocean heat uptake caused by a change of surface albedo with the advective annual horizontal atmospheric heat transport through 60°N, which is estimated to be ∼9.5 × 1022 J, with variations defined by a standard deviation (SD) of ∼1.9 × 1021 J. These estimates were derived from Fig. 4 in Trenberth and Caron (2001) by multiplying power expressed in watts by the length of the year expressed in seconds. This SD is almost three times more than the total amount of heat accumulated in the ice-free area in summer 2007. Despite its relatively small magnitude, the atmospheric effect of oceanic warming caused by ice reduction may be important due to strong near-surface atmospheric stratification leading to a large increase of atmospheric temperature near the surface (Deser et al. 2010). We note, however, that the autumnal PA can be traced over a limited (<500 km) area with a fast decay of SAT anomalies from the area of maximum ice retreat (Fig. 9), in good agreement with Ogi and Wallace (2007). Thus, we conclude that the recent dramatic reduction of arctic ice as well as anomalously high NPA SATs cannot be explained by the ice–albedo feedback mechanism alone. This conclusion agrees well with Graversen and Wang’s (2009) findings. Using modeling experiments, Graversen and Wang showed that the ice–albedo feedback mechanism is responsible for only 15% of PA and that the impacts of water vapor and cloudiness dominate the PA of global warming.
This study uses an extensive dataset of monthly SAT records from meteorological land stations. Most records have been updated by very recent (until December 2008) observations. The dataset includes previously unutilized sources, mostly from European Russia. The use of historical Russian archives provided spatiotemporal data coverage for this region of the NPA comparable to that for the Scandinavian region. Using this dataset, we document a large-scale year-round NPA warming rate of 1.36°C century−1 for 1875–2008, almost two times stronger than the NH trend (0.79°C century−1). The rate of warming increased strongly in recent decades with exceptionally high SATs reaching 1.72°C in 2003, 2.11°C in 2005, and 2.18°C in 2007, the warmest years in the history of instrumental observations.
Stronger warming in high-latitude regions, particularly in autumn and spring, is a manifestation of PA. The pattern of long-term SAT changes suggests two different spatial scales of mechanisms of PA: large-scale (hemispheric) and local (limited by NPA). Large-scale PA may be linked to both year-round hemispheric-in-scale atmospheric heat transports, which change the composition and vertical structure of the arctic atmosphere, and winter and springtime longwave radiative forcing linked to clouds and water vapor content. A combined effect of large-scale PA is evaluated using the new quantitative measure of PA ar, linking NPA and NH temperature anomalies via a linear regression equation. For 1875–2008 this measure yields PA of 1.62 ± 0.10. This estimate of large-scale PA of global warming is very similar to the ensemble-average estimate of PA of 1.56 derived from a suite of 42 CMIP model runs (see caption for Fig. 7d).
Polar amplification at the local scale is related to ice–albedo feedback mechanisms, which are evident in autumn in maritime/coastal areas up to several hundred kilometers from the ice edge into the interior of continents. These mechanisms are apparent during extended warm periods only (i.e., in the 1930s–40s and recent decades) and are not evident during cold periods. Heat budget estimates used to elaborate the role of ice–albedo feedback mechanisms suggest that the recent dramatic reduction of arctic ice as well as anomalously high NPA SATs cannot be explained by these mechanisms alone. We note, however, that further diminishing arctic summer ice cover will probably enhance the role of these mechanisms in shaping high-latitude climate change, consistent with the modeling results by Deser et al. (2010). Using modeling experiments, Deser et al. showed that the projected reduction of ice cover may result in increased heat transfer during winter and fall from the relatively warm ocean to the cold atmosphere, leading to a large increase of atmospheric temperature near the surface.
The attribution of PA of global warming to particular feedback mechanisms is a nontrivial task owing to gaps in observational data, short instrumental records, and strong high-latitude variability, particularly multidecadal variability. An important caveat is the poorly defined character of these mechanisms. For example, PA caused by natural variability and forced change may be different due to nonuniform geographical distribution of anthropogenic aerosol forcing. Anthropogenic climate change may be amplified or masked by natural variations, and these modes of variability can be separated only when the mechanisms governing them are better understood. There are substantial differences among estimates of radiative feedbacks associated with water vapor, lapse rate, clouds, and sea ice derived from general circulation models (Bony et al. 2006). Stephens (2005) provided a critical overview of the effects of clouds on the radiation budget of the global climate system. Stephens suggested that one of the key problems of present-day models lies in the imbedded cloud parameterizations and argued that “in view of the complex nature of the climate system, and the cumbersome problems encountered in diagnosing feedbacks, understanding cloud feedback will be gleaned neither from observations nor proved from simple theoretical argument alone. The blueprint for progress must follow a more arduous path that requires a carefully orchestrated and systematic combination of model and observations” (p. 269). Therefore, advances in modeling and theory as well as continued observations are required in order to develop a deeper understanding of the nature of PA of global warming, which is a necessary step for developing a predictive system of future climate change.
In summary, we formulate the major results of this study:
Using a new extensive dataset of Arctic SATs, including previously unutilized historical Russian data, we documented a high-latitude warming rate of 1.36°C century−1 for 1875–2008, with an exceptionally strong warming rate in the recent decade (1.35°C decade−1).
A new stable statistical measure of polar amplification (PA) linking high-latitude and hemispheric annual temperature anomalies via a regression relationship is proposed. For 1875–2008, this measure yields PA of ∼1.62.
Local PA related to the ice–albedo feedback mechanisms is autumnal and coastal, extending several hundred kilometers inland. A recent reduction of Arctic Ocean sea ice and anomalously high SATs cannot be explained by the ice–albedo feedback mechanisms alone, and the role of large-scale mechanisms of PA of global warming may be important.
This study was supported by JAMSTEC (RB, IP, VA), NSF (IP, VA), NASA (IP), and Russian Foundation for Basic Research (RB, Grant RFBR 08-05-00569-a) grants. We thank D. Atkinson, P. Jones, P. Groisman, and M. Shulski for data; U. Bhatt, R. Graversen, and J. Walsh for useful comments; T. Pavlova for climate models time series; and J. Moss for help with the graphics. We acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) and the WCRP’s Working Group on Coupled Modelling (WGCM) for their roles in making available the WCRP CMIP3 multi-model dataset. Support of this dataset is provided by the Office of Science, U.S. Department of Energy.
Statistical Procedure of Data Quality Control
Available hourly or daily data were averaged to obtain monthly values. Statistical control of data quality was based on two methods. The first one is based on statistical distribution of random variable ν:
which is a normalized anomaly of random variable x (in our case, hourly or daily SATs; see also Cramer 1999), where x is the mean and S(x) is the square root deviation. The density distribution of random variable ν may be described by the following relationship:
where n is the number of observations and Γ is the gamma function. The following procedure of error estimate has been used. For each month, station data were used to calculate x and S(x). The random variable ν is obtained by Eq. (A1) and each ν is compared to the critical value νcr, which is obtained by solving numerically the following equation:
where α is the level of statistical significance. Note that expression (A2) for density distribution was obtained based on an assumption of independence of samples. Therefore, this criterion does not always distinguish and remove errors from meteorological data, particularly because these data are subject to daily and annual cycles.
The second method uses Dixon’s criteria (Dixon 1983). This method is based on comparison and removal of extreme values from the available record and is less sensitive to autocorrelation and seasonality in the time series. Station hourly or daily meteorological observations are subject to quality control for each month. First, two maximal xmax,1 and xmax,2 and two minimal xmin,1 and xmin,2 values (xmax,1 ≥ xmax,2 and xmin,1 ≤ xmin,2) are found in each monthly segment of the time series. Second, Dixon’s statistics are defined as
If random variables dmax and dmin exceed the tabulated critical value dcr, then the corresponding x is considered as an erroneous value and removed from the record. The procedure is repeated until all remaining values satisfy Dixon’s criteria. Analysis of method performance showed that the values identified by Dixon’s method as erroneous were, indeed, erroneous. Effects of urban warming are not considered in this study. It is known, however, that these effects are insignificant for long-term spatially averaged SAT (Jones et al. 1989; Parker 2004; Peterson and Vose 1997).
Procedure for Estimating the Statistical Significance of Trends Using a Nonparametric Approach
Let us consider a discrete sampling of a process Y(t) with N available values (i.e., Yi, i = 1, N). The linear dependence of Y(t) of time may be described by the following model:
where the linear function a + bt approximates systematic change of the parameter Y(t) and the random term ɛ(t) describes natural internal variability of the system as well as random external forcing. Time t is defined over the interval [−T/2, T/2]. The coefficient a corresponds to the mean value of Y(t) and the coefficient b defines trend; a and b may be found by the following formulas:
where the angle brackets correspond to time averaging.
Statistical significance of a linear trend may be estimated by an algorithm based on an estimation of variance of the sample trend and on introducing statistics with a distribution close to the Student’s t distribution. We note that variance of both sample mean and sample linear trend is asymptotically defined by the spectral density at zero frequency (Bloomfield 1992; Bloomfield and Nychka 1992; Von Storch and Zwiers 1999):
where S(ω) is the spectral density of a process at frequency ω and the period T is defined by the time series length as T = (N − 1)Δt in which time step Δt is assumed to be equal to 1. Our proposed method of estimating statistical significance of trends is based on the use of statistics similar to Student’s t statistics Sb = b/σb, where σb is estimated using sample spectral density at zero frequency. Similar statistics were used for estimates of sample mean (SX = X/σX) and the sample trend by Brillinger (1981), Zwiers and Thiebaux (1987), and Fomby and Vogelsang (2002). Estimates of sample mean and variance of sample mean based on spectral density S(0) are statistically independent (Zwiers and Thiebaux 1987). However, estimates of sample trend b and spectral density S(0) are not [b and S(0) are always correlated], which may lead to false conclusions. To resolve this problem, we propose to detrend the sample prior to statistical analysis. In this case, sample trend b and spectral density at zero frequency S(0) are practically always uncorrelated, which allows us to use statistics similar to Student’s t. Detrending, however, distorts the sample spectrum, which is corrected in the algorithm using correction factors. One more effect of this procedure is reduction of effective degrees of freedom by one. The following are the proposed algorithm steps.
1) Compute the mean value and reduce time series to anomalies Zi = Yi − Y.
4) Estimate spectral density of the random function X(t) at zero frequency. According to recommendations from Jenkins and Watts (1968), the maximum number of lags M of the autocorrelation function for estimation of spectral density should be 0.1 of the time series length N (i.e., M = N/10). The effective degree of freedom Ne,s for spectral density at zero frequency is Ne,s = 5.
7) Include statistics Sb = b/σb, which have a Student’s t statistical distribution in the case of the normally distributed original time series.
8) Because of detrending of the original time series of anomalies Z(t), the effective number of degrees of freedom for the statistics Sb is decreased by 1: Neff = Ne,s − 1. Therefore, using the recommended M = 0.1N, which corresponds to Ne,s = 5, we have Neff = 4.
9) Using critical values of the Student’s distribution tα, formulate the decisive rule: in the case of |Sb| > tα, the null hypothesis (i.e., the hypothesis of the nonexistence of trend) is rejected with probability 1 − α.
Thus, using computed statistics Sb, by (B16) we may obtain the probability of null hypothesis rejection.
The proposed algorithm for estimating statistical significance of linear trends is based on an estimate of sample spectral density at zero frequency of the preliminary detrended time series. This approach leads to a noncorrelated sample trend and its variance and allows statistics similar to Student’s t.
Corresponding author address: Igor Polyakov, International Arctic Research Center, University of Alaska Fairbanks, P.O. Box 757335, Fairbanks, AK 99775. Email: firstname.lastname@example.org