Recent summer Arctic atmospheric circulation anomalies in a historical perspective

Introduction Conclusions References

leader-algorithm approaches (Fettweis et al., 2011), principal component analyses (Huth, 2000), optimization algorithms (Philipp et al., 2007) including self-organising maps (Käsmacher and Schneider, 2011;Bezeau et al., 2014;Hope et al., 2014), no method can be considered as being overall better than the others (Philipp et al., 2010).Thus, we use our CTC that has been developed for the Arctic region and especially for Greenland (Fettweis et al., 2011).This CTC has already been used to compare reanalysis datasets and General Circulation Model outputs over Greenland with the aim of detecting circulation changes (Belleflamme et al., 2013) and to analyse temperature related flow analogues over the Greenland ice sheet (Fettweis et al., 2013).
In this study, we apply the CTC developed by Fettweis

Data
We used daily SLP and Z500 data for the summer months (JJA -June, July, and August) of five reanalysis datasets: the ERA-Interim reanalysis (Dee et al., 2011)   to avoid the problem of decreasing pixel area near the pole 168 when using geographic coordinates, all reanalysis outputs 169 have been linearly interpolated to a regular grid with a spatial 170 resolution of 100 km.Our integration domain has a size of 5000 × 6000 km and covers the whole Arctic Ocean, Greenland, and the northern part of the Atlantic Ocean (Fig. 1).
Finally, monthly NAO data over the period 1871-2013 were obtained from the Climatic Research Unit (CRU).This NAO index is defined as the normalised difference between the SLP measured in the Azores (Ponta Delgada) and Iceland (Reykjavik).

Method
The SLP data from the different reanalyses were compared Europe by Belleflamme et al. (2014).This CTC is considered a leader-algorithm method (Philipp et al., 2010), because each class is defined by a reference day and a similarity threshold.After having calculated the similarity index (see below) between all pairs of days of the dataset, the day counting the most similar days (i.e. with a similarity index value above the similarity threshold) is selected as the reference day for the first type.All days considered as similar to this reference day are grouped into this type.The same procedure is repeated type by type over the remaining days of the dataset.This whole process is repeated many times for various similarity thresholds in order to optimize the classification.The similarity between the days is gauged by the Spearman rank correlation coefficient.The key feature of using correlation-based similarity indices is that they are not influenced by the average SLP of a day, but only by its spatial pattern (Philipp et al., 2007).Thus, in contrary to Fettweis et al. (2011) and Fettweis et al. (2013), who used the Euclidean distance as similarity index and Z500 to take into account the influence of the temperature on the upper level circulation, we used the Spearman rank correlation and SLP to focus exclusively on the circulation pattern.In order to minimize the influence of eventual temperature biases into the SLP retrieving computation, especially over elevated regions like Greenland (Lindsay et al., 2014), we only considered oceanic pixels when performing the SLP-based classification.For comparison, the same procedure was done using Z500, but all pixels of the domain were taken into account, since there is much less influence of the surface and its elevation at this level.This CTC is automatic, meaning that the circulation types are built by the algorithm and not predefined by the user.This implies that the circulation types obtained using different datasets will be different and thus difficult to compare.To overcome this problem, we "projected" the types of a reference dataset onto the other datasets, i.e. the types obtained for the reference dataset were imposed as predefined types for the other datasets, as proposed by Huth (2000) and implemented by Belleflamme et

20CRv2 frequency uncertainty
The uncertainty of the 20CRv2 frequencies strongly decreases between 1930 and 1950 to become insignificant over the four or five last decades for all types (Fig. 2).It is interesting to observe that this uncertainty remains relatively constant over time before 1940, at a level of around 7-11 % for the first four types, and around 4-6 % for Types 5 and 6.
Moreover, the 20 000-run ensemble mean frequency, its standard deviation, and its 10th and 90th percentiles show an evolution over time that is almost parallel to the 20CRv2 reference run.There is no smoothing of the interannual variability, which remains similar to the variability of the last decades when going back in time.Finally, the last class, which groups the unclassified days, does not show any increase towards the beginning of the 20CRv2 period, meaning that there are not more days that do not correspond to the main types before 1940 than over the three last decades .Thus, even if there is some uncertainty about the exact frequencies before 1940, there is high confidence in the magnitude and the time evolution of the circulation type frequencies.
There are significant circulation type frequency differences between the 20 000-run ensemble mean and the 20CRv2 reference run before 1940.In particular, the frequency of Types 1 and 2 is strongly overestimated by the 20 000-run ensemble compared to the 20CRv2 reference run.
The 20CRv2 reference run annual frequencies turn around the 10th percentile of the 20 000-run ensemble for these two types.This is compensated by an underestimation of the frequencies of the other types by the 20 000-run ensemble whose 90th percentile frequencies are of the same order than the 20CRv2 reference run frequencies.These frequency shifts are due to the pattern of the SLP spread, which is much higher over the Arctic Ocean than over the rest of the domain for the 1871-1930 period (Fig. 3, bottom).Consequently, when adding (with a multiplying factor between 0 and 1) the SLP spread, the pattern of the SLP daily mean is changed towards a more anticyclonic pattern over the Arctic Ocean, making it similar to Type 2. In the same way, when 437 subtracting (with a multiplying factor between −1 and 0) the 438 SLP spread, the SLP daily mean becomes more similar to 439 Type 1, which presents a low pressure over the Arctic Ocean.S1).For these runs, the spread multiplied by a factor varying randomly between −1 and 1 has been added to the daily mean.
Despite an uncertainty of about 5 to 11 % for the circulation type frequencies before 1930, the magnitude and the time evolution of the frequency anomalies can be reasonably well estimated using 20CRv2 SLP.Further, this uncertainty becomes less significant after 1950, due to improved assimilated data availability and reliability.The strong impact of the number and quality of observational data on the reliability of reanalysis datasets is also highlighted by the important discrepancies between 20CRv2 and ERA-20C during the first half of the 20th century.These discrepancies can have strong impacts on the interpretation of the results.For example, the 1923-1931 warmer summers over Greenland (Chylek et al., 2006) could be attributed to anomalous atmospheric circulation conditions according to ERA-20C but not to 20CRv2.
The particular spatial pattern of the 20CRv2 spread, i.e. highest over the Arctic Ocean, causes a strong overestimation of Type 1 (low pressure over the Arctic Ocean) and Type 2 (high pressure over the Arctic Ocean) at the expense of all other types for the SLP-based 20 000-run ensemble compared to the 20CRv2 reference run.Thus, it is interesting to note that, although no systematic SLP bias is introduced through adding the spread, systematic circulation type frequency shifts appear.In a similar way, the Z500 spread also introduces artefacts in the 20 000-run ensemble.This shows the importance of accounting for the spread of the 20CRv2 data to get an estimation of the range of plausible results.
We have found the same summertime circulation anomalies as described by other authors (Fettweis et ERA-40 (1958ERA-40 ( -1978)), NCEP/NCAR (1948-2014), ERA-20C (1900-2010), and the 20CRv2 reference run and the 20CRv2 20 000-run ensemble mean (1871-2012).The corresponding solid lines represent the 10 year binomial running mean frequencies.For the 20CRv2 20 000-run ensemble, the 10th and the 90th percentiles as well as the one standard deviation interval around the mean are also given.
years, and particularly since 2007, significant atmospheric circulation anomalies have been observed over different parts of the Arctic.Based on 500 hPa geopotential height (Z500), Fettweis et al. (2013) reported a doubled frequency of summertime anticyclones centred over Greenland, representing an increased frequency of negative NAO (North Atlantic Oscillation) conditions.This circulation anomaly impacts the climate of a major part of the Arctic region by 2 A. Belleflamme et al.: Recent summer Arctic circulation anomalies in a historical perspective have put the 2007-2012 summertime atmospheric circulation anomaly over the Arctic region in perspective with the reconstructed circulation over the instrumental period.To achieve this, we have used an atmospheric circulation type classification (CTC) to distinguish the main circulation types over the Arctic region and to analyse their frequency changes over time, as done by Ballinger et al. (2014) over the Beaufort Sea, Bezeau et al. (2014) over the Canadian Arctic Archipelago, and Fettweis et al. (2013) over Greenland.Since the aim of CTCs is to group similar circulation situations together (Huth et al., 2008; Philipp et al., 2010; Käsmacher and Schneider, 2011), this methodology allows a synthetic analysis of the atmospheric circulation over a given region at a daily scale (i.e. the characteristic time scale of synoptic circulation patterns like high pressure systems).CTCs are widely used to compare datasets (e.g.reanalyses, General Circulation Model outputs), to evaluate their ability to reproduce the observed atmospheric circulation, and to detect changes in the observed and projected atmospheric circulation (Bardossy and Caspary, 1990; Kyselý and Huth, 2006; Philipp et al., 2007; Anagnostopoulou et al., 2009; Demuzere et al., 2009; Pastor and Casado, 2012; Fettweis et al., 2013; Belleflamme et al., 2013, 2014).While a wide range of classifications has been developed to study the atmospheric circulation (e.g. et al. (2011) (described in Sect.3) to daily SLP and Z500 fields from different reanalysis datasets (detailed in Sect.2).In Sect.4.1.1,we put in perspective the summertime circulation of 2007-2012 with the circulation variability observed since 1871.The influence of the uncertainties of the past circulation on our results is discussed in Sect.4.1.2.After a comparison between the SLP and the Z500-based results in Sect.4.2, we analyse the links between the circulation type frequencies and NAO and SIE in Sect.4.3.
from the European Centre for Medium-Range Weather Forecasts (ECMWF) (spatial resolution: 0.75 • × 0.75 • ) over the 121 period 1979-2014, 122 the ERA-40 reanalysis from the ECMWF (Uppala et al., 123 2005) (spatial resolution: 1.125 • × 1.125 • ) over the pe-124 riod 1958-1978 used to extend ERA-Interim.It should 125 be noted that ERA-40 is known to have significant bi-126 ases in its vertical temperature profile (Screen and Sim-127 monds, 2011), which is used in the geopotential height 128 calculation.However, the impact of these biases on our 129 Z500-based results should be limited, since the most 130 problematic year (i.e.1997) is not included in the ERA-131 40 period considered here.132 the NCEP/NCAR reanalysis from the National Centers 133 for Environmental Prediction -National Center for At-134 mospheric Research (Kalnay et al., 1996) (spatial reso-135 lution: 2.5 • × 2.5 • ) over the period 1948-2014, 136 the ERA-20C reanalysis from the ECMWF (Poli et al., 137 2013) (spatial resolution: 1.125 • × 1.125 • ) over the pe-138 riod 1900-2010.The spread evaluating the uncertainty 139 of the ERA-20C data was not yet available when con-140 ducting this study.141 the Twentieth Century Reanalysis version 2 (20CRv2) 142 (Compo et al., 2011) from the NOAA ESRL/PSD (Na-143 tional Oceanic and Atmospheric Administration Earth 144 System Research Laboratory/Physical Sciences Divi-145 sion) (spatial resolution: 2 • × 2 • ) over the period 1871-146 2012.The 20CRv2 data are constructed as the ensemble 147 mean of 56 runs.The standard deviation (called spread) 148 of this ensemble mean is also given for each variable 149 (in our case SLP and Z500).We used it to estimate the 150 uncertainty of our 20CRv2-based results, in particular 151 before the overlapping period with the other reanalysis 152 datasets when the assimilated observations are sparse.153 In fact, the spread, and thus the uncertainty of the recon-154 structed atmospheric circulation in 20CRv2, strongly 155 depends on the number of pressure observations, which 156 is low before 1940 (Compo et al., 2011).

157
It is important to note that only SLP, sea surface tempera-158 ture (SST), and sea ice are assimilated into 20CRv2, and SLP, 159 SST, and oceanic near-surface air temperature and wind into 160 ERA-20C.The other reanalyses also assimilate satellite and 161 upper air data every 6 h.Therefore, 20CRv2 and ERA-20C 162 are a priori less reliable than the other more constrained re-163 analyses.164 Additionally, daily sea ice cover data from ERA-Interim 165 are also used over the 1980-2014 JJA period.166 Since the reanalyses have different spatial resolutions, and 167 using the automatic circulation type classification developed by Fettweis et al. (2011) and used over Greenland by Belleflamme et al. (2013) and Fettweis et al. (2013), and over al. (2013) and Belleflamme et al. (2014).Since the types are now the same for all datasets, they can easily be compared.Lindsay et al. (2014) 4 A. Belleflamme et al.: Recent summer Arctic circulation anomalies in a historical perspective ridge over Greenland.In Type 3, the depression is situated over the Greenland and Svalbard region.Type 4 is marked 280 by two depressions, the Icelandic Low and a low over the 281 Chukchi Sea, and a ridge over the Barents and Kara Seas.282 Type 5 combines the Greenland High and the Beaufort Sea 283 High.Finally, Type 6 shows a high pressure system over the 284 Arctic Ocean, while the depression is split into three parts 285 (i.e. the Icelandic Low, a low over the Kara Sea, and a low 286 over the Canadian Arctic Archipelago).
very good agreement between the frequencies of 289 the circulation types and their evolution over time for all re-290 analyses over 1958-2012, as well for SLP (Fig. 2) as for 291 Z500 (Supplement Fig. S2).Nevertheless, two notable dif-292 ferences have to be pointed out.The first difference is a sys-293 tematic overestimation of about 4-6 % of the frequency of 294 Type 3 at the expense of Type 2 by 20CRv2 compared to 295 the full constrained reanalyses (ERA-40/ERA-Interim and 296 NCEP/NCAR) for SLP.This bias is in agreement with the 297 findings of Lindsay et al. (2014), who report a positive SLP 298 bias over Asia for 20CRv2 (using monthly data), since Type

2993
is characterised by an anticyclone over the Asian part of 300 the domain, in contrary to Type 2. The second difference is 301 an overestimation of Type 2 at the expense of Type 1 and 302 to a lesser extent of Type 3 by ERA-20C over its whole pe-303 riod (1900-2010) compared to all the other reanalyses used 304 here, for the SLP-based classification.This frequency bias 305 is particularly important before 1950 compared to 20CRv2.306 In fact, ERA-20C overestimates SLP over the whole Arctic 307 Ocean, and especially over the Beaufort Sea, compared to the 308 other reanalyses (not shown).This implies that more ERA-309 20C days are considered as similar to Type 2 (Beaufort Sea 310 -Arctic Ocean High) at the expense of Type 1 (low pres-311 sure over the Arctic Ocean) and Type 3 (low pressure over 312 Greenland and the Canadian Arctic Archipelago).The ERA-313 20C SLP bias is particularly important before 1950 compared 314 to 20CRv2, but it is still present over the last decades com-315 pared to ERA-40, ERA-Interim, and NCEP/NCAR.How-316 ever, since 20CRv2 also shows systematic biases, it is not 317 possible to consider one of these two reanalyses as more re-318 liable than the other.
type frequency evolution 321 The frequencies of Type 2 (Beaufort Sea -Arctic Ocean 322 High) and Type 4 (Greenland High) are almost twice as 323 large during 2007-2012 compared with the 1871-2014 aver-324 age (Fig. 2).These frequency anomalies are similar to those 325 found by Ballinger and Sheridan (2014) and Ballinger et al. 326 (2014) for the Beaufort Sea, by Fettweis et al. (2013) for 327 Greenland, and by Bezeau et al. (2014) for the Canadian 328 Arctic Archipelago.They are compensated by a decrease in 329 frequency of Type 3 by a factor of two, and to a lesser ex-330 tent of Type 1.Both types are characterised by a low pres-331 sure system over the Arctic.Over the record, Types 2 and 332 4 never experienced such high frequencies over several con-333 secutive summers since 1871.Between 2007 and 2012, two334summers for Type 2 and five summers for Type 4 presented 335 a higher frequency than the 90th percentile frequency, mean-336 ing a return period of about 10 years (Table1).However, 337 20CRv2 suggests that similar circulation type frequencies 338 were observed before 1880 while the uncertainty is very 339 high over that period.Moreover, Type 4 shows some sum-340 mers with anomalously high frequencies between 1891 and 341 1896.Despite the frequency biases described above, ERA-342 20C shows many summers with exceptionally high frequen-343 cies of Type 2 between 1923-1931.Thus, on the basis of 344 ERA-20C, the anomalously warm conditions and the associ-345 ated high surface mass loss rates observed over the Green-346 land ice sheet over that period (Chylek et al., 2006; Fettweis 347 et al., 2008) could be attributed to atmospheric circulation 348 anomalies.In contrast, the 20CRv2 circulation type frequen-349 cies do not present any anomalies over the 1923-1931 pe-350 riod.Finally, atmospheric circulation conditions similar to 351 2007-2012 are observed around 1957-1960 for all reanal-352 yses used here.Nevertheless, these four periods were shorter 353 than 2007-2012 and not marked by as many anomalous sum-354 mers, except for the 1891-1896 period for 20CRv2 Type 4 355 and the 1923-1931 period for ERA-20C Type 2. The anoma-356 lies of the frequencies of Types 2 and 4 (+20 %), and Type 3 357 (−15 to −20 %), are much higher than their interannual fre-358 quency variability (with a standard deviation over the 1871-359 2012 period for the 20CRv2 20 000-run ensemble mean of 360 about 7.7 %, 9.7 %, and 8.2 % for Types 2, 3, and 4 respec-361 tively).The frequency anomalies of the other types are of the 362 same order than their interannual variability (with a standard 363 deviation of about 9.4 %, 5 %, and 5.6 % for Types 1, 5, and 6 364 respectively).The exceptional frequency anomalies of 2007-365 2012 could suggest that they are related to global warming.366 However, the 2013 summer shows opposite extremes.On the 367 other side, the circulation type frequencies of the 2014 sum-368 mer are of the same order than the 2007-2012 average.This 369 suggests that, even if the 2007-2012 circulation anomalies 370 might be related to global warming, this link is not straight-371 forward, and the natural variability could largely exceed the 372 global warming induced signal.373 The circulation type frequency anomalies are not due to 374 changes in the persistence (i.e. the duration of consecutive 375 days grouped in the same type).In fact, there is a persistence 376 increase for Types 2 and 4, and a decrease for Type 3 over 377 2007-2012 with regard to the overall average (not shown).378 But a more detailed analysis shows that these persistence 379 changes are artefacts due to the frequency anomalies.Note 380 that the 20CRv2 20 000-run ensemble persistence cannot be 381 used for a persistence analysis.Since the spread is added with 382 a multiplying factor determined randomly for each day, the continuity of the atmospheric circulation over time, i.e. the transitions between the circulation types and the succession of the types themselves, is not preserved.The analysis of the 20CRv2 reference run monthly circulation type frequencies shows that the 2007-2012 frequency anomalies affect all three months (JJA).In this way, the 2007-2012 period differs from the other anomalous periods(1871-1880, 1891-1896, 1923-1931, and 1957-1960).For example, the positive frequency anomaly of Type 2 over 1958-1960 is due to high frequencies during August and to a lesser extent during June.For the year 1957, Type 4 shows frequencies far above normal for June and July, but not for August.The 1871-1880 period has many summers with above normal frequencies for Type 4, but no systematic frequency anomaly lasting a few summers can be detected for one particular month.

440
Fig. 3 (top), the average SLP spread over the Arctic region 446

4494. 2
Geopotential height at 500 hPa 450 The detected frequency changes are similar to those of SLP.451 The Greenland High and the Beaufort Sea High (Types 2 and 452 5) were almost twice as frequent over 2007-2012 than over 453 the whole 1871-2014 period (Supplement Fig. S2, Table

454For
Type 2 (Greenland High), two of the four other high 455 frequency periods found for SLP are also detected: 1871-456 1880 and 1891-1896.The 1957-1960 period is not excep-457 tional on the basis of Z500.This is in agreement with the 458 findings of Bezeau et al. (2014), who showed on the basis 459 of NCEP/NCAR Z500 that the 2007-2012 period reached 460 record values since 1948, despite more frequent anticyclones 461 over the Canadian Arctic Archipelago before 1960.The 462 1923-1931 period, which is only exceptional on the basis of 463 ERA-20C SLP, is not anomalous at the Z500 level.464 As for SLP, the Z500 spread plays an important role in 465 the frequency distribution before 1940, when it is the highest 466 (Supplement Fig. S3, top).Before 1940, the frequencies of 467 the 20CRv2 reference run and to a lesser extent of the 20 000-468 run ensemble mean are much higher for Type 1 (about 20 %) 469 and Type 2 (about 10 %) with respect to the second half of 470 the 20th century.This is compensated by particularly low fre-471 quencies of Types 3, 5, and 6, and to a lesser extent of Type 472 4.These frequency shifts are probably due to the uncertain-473 ties in the 20CRv2 data before 1940 since they are lower for 474 the 20 000-run ensemble compared to the 20CRv2 reference 475 run.Moreover, as for SLP, the frequency differences between 476 the 20CRv2 reference run and the 20 000-run ensemble mean 477 can be explained by the pattern of the Z500 spread, which is 478 very close to the SLP spread pattern (Supplement Fig. S3, 479 bottom).When subtracting the spread, the circulation tends 480 to become more cyclonic over the Arctic Ocean, favouring 481 the shift of days into Types 1, 2, and 4. In the same way, 482 adding the spread gives a more anticyclonic character to the 483 circulation.This favours Types 3, 5, and 6.But Types 1 and 484 2 count for about 80 % of all days before 1940.Therefore, 485 subtracting the spread has only a limited impact on the fre-486 quency distribution, since it favours Types 1 and 2, which 487 already contain most of the days.On the opposite, adding 488 the spread at the expense of Types 1 and 2 induces much 6 A. Belleflamme et al.: Recent summer Arctic circulation anomalies in a historical perspective more frequency changes, since more days can be shifted into another type.491 4.3 Links with other variables 492 4.3.1 North Atlantic Oscillation 493 The 30 year running correlation between the circulation type 494 frequencies and the JJA CRU NAO index (calculated as the 495 average of the JJA monthly CRU NAO index values) shows 496 a very good agreement between the different reanalyses on 497 the basis of SLP and Z500.Further, the almost similar evo-498 lution of this correlation for the 20CRv2 reference run and 499 the 20 000-run ensemble mean confirms that taking into ac-500 count the spread does not impact the frequency variations 501 over time.502 For the SLP-based classification, Types 2 and 4 both show 503 negative correlations with NAO, while only the correlation 504 of Type 3 is always positive (Fig. 4).The link between NAO 505 and the three other types is not as clear.The association of 506 negative NAO phases with high frequencies of Types 2 and 4 507 is even more evident when adding both frequencies together, 508 the average correlation between NAO and the 20CRv2 ref-509 erence run frequencies over 1871-2012 being of −0.24 for 510 Type 2, −0.32 for Type 4, and −0.38 for Types 2 + 4. More-511 over, this association becomes more marked when the 2007-512 2012 anomalous summers are integrated into the calcula-513 tions (i.e. starting in 1992), suggesting a stronger link over 514 that period.This is in agreement with Fettweis et al. (2013) 515 and Hanna et al. (2014a), who linked the recent circulation 516 anomalies and the negative NAO anomalies.For most other 517 periods with circulation anomalies (1871-1880, 1891-1896, 518 and 1957-1960), there is no clear link between NAO, which 519 is less exceptional than over 2007-2012, and the frequency 520 of Types 2 and 4.Only the 1923-1931 period shows a clear 521 negative correlation between NAO and the frequency of Type 522 2 for ERA-20C.523 The Z500-based results are basically the same as for SLP.524 Type 2 shows a negative correlation, while the correlation 525 between NAO and the frequency of Type 1 is always pos-526 itive (Supplement Fig. S4).Nevertheless, the correlation of 527 Type 5 is not as clear.While it is negative since 1960, the 528 20CRv2 reference run and the 20 000-run ensemble mean 529 show divergent values before 1940 suggesting that the un-530 certainty due to the 20CRv2 spread has more impact on our 531 results for Z500 than for SLP.This is certainly reinforced by 532 the underestimation of the frequency of Type 5 before 1940.533 Again, the link between NAO and Types 2 and 5 is stronger 534 since 1992, when the 2007-2012 period is integrated into the 535 30 year running correlation.
5364.3.2Sea ice extent 537There is a link between the circulation anomalies and the 538 summertime sea ice extent (SIE) loss.As indicated by the 539 correlation between the ERA-Interim SIE loss and the ERA-540 Interim SLP-based circulation type frequencies over 1980-541 2014, SIE loss is only favoured by Types 2 (r = −0.47)and 4 542 (r = −0.52),while Types 1 (r = 0.50) and 3 (r = 0.48) tend 543 to mitigate it.Types 5 (r = 0.17) and 6 (r = 0.22) do not have 544 important impacts on the SIE loss.When considering the sum 545 of the frequencies of Types 2 and 4, the relation appears to 546 be even clearer with a correlation of −0.65.This means that 547 the frequency increase of Types 2 and 4 could partly explain 548 the summertime record Arctic SIE loss observed over the last 549 decade.550 For the Z500-based classification, only Type 5 (r = 551 −0.54) can be clearly related to enhanced SIE loss and 552 Type 1 (r = 0.57) to mitigated SIE loss.For the remaining 553 types, the correlation varies between −0.20 and −0.04.As 554 said above, Type 5 combines the Beaufort Sea High and 555 the Greenland High.Type 2 shows only a slight ridge over 556 Greenland and a depression centred over the Arctic Ocean, 557 far away from the Russian coast.Thus, the conditions favour-558 ing sea ice export through the Fram Strait and the Barents Sea 559 are not met for this type, as confirmed by its poor correlation 560 (r = −0.20)with SIE.561 Our results seem to confirm those of Wang et al. (2009) 562 and Overland et al. (2012), who showed that the record Arc-563 tic sea ice loss observed over the last years can partly be 564 attributed to more frequent positive Arctic Dipole Anomaly 565 (DA) phases.In fact, positive DA phases are characterised 566 by a higher occurrence of a high pressure system over the 567 Canadian Arctic Archipelago and Greenland and a low pres-568 sure system over the Kara and Laptev Seas (Wu et al., 2006).569 Thus, at first glance, the SLP-based Types 2 and 4 can both 570 be associated to a positive DA phase, while the other types, 571 and in particular Types 1 and 3, can be related to a nega-572 tive DA phase.During positive DA phases, the sea ice export 573 from the Arctic basin through the Fram Strait and the Bar-574 ents Sea is strongly enhanced, which is particularly effective 575 for important sea ice loss during summer (Wang et al., 2009).576 Further, our results agree with those of Simmonds and Keay 577 (2009) and Screen et al. (2011), who have shown that SIE in 578 September is lower in years characterised by a weaker than 579 normal summertime Arctic cyclonic activity, which induces 580 a higher average SLP over the region.an automatic circulation type classification to 583 study the anomalies in the summertime (JJA) atmospheric 584 circulation based on (i) the sea level pressure and (ii) the 585 500 hPa geopotential height over the Arctic region over the 586 1871-2014 period.Three reanalysis datasets were used over 587 the second half of the 20th century (ERA-Interim as refer-588 ence, ERA-40, and NCEP/NCAR).The 20CRv2 and ERA-589 20C reanalyses were used over the 1871-2012 and 1900-590 2010 periods respectively, to evaluate if circulation anoma-lies similar to 2007-2012 could already have occurred.Further, since 20CRv2 data are given as a 56-member ensemble mean with its standard deviation (spread), 20 000 runs have been done to take into account the 20CRv2 uncertainty.
al., 2013; Ballinger et al., 2014), i.e. a doubling in frequency of the Beaufort Sea -Arctic Ocean High and of the Greenland High over 2007-2012.Only four other periods(1871-1880,     1891-1896, 1923-1931, and 1957-1960)  of similar circulation anomalies were detected but the successions of summers with such anomalies are shorter than in the 2000's.These anomalies all largely exceed the interannual variability of the circulation type frequencies.Nevertheless, it is not possible to attribute the circulation anomalies over 2007-2012 to global warming.First, these anomalies are observed over a too short period, so that they could simply be an exceptionally strong deviance from the average circulation.In this way, the 2013 summer was marked by opposite frequency extremes, positive NAO index values, a low melt of the Greenland ice sheet, and lower Arctic sea ice decline compared to 2007-2012.Our findings corroborate those of Rajewicz and Marshall (2014), who state that the 2013 JJA mean Z500 over Greenland was significantly lower than the average over the last seven decades, which contrasts with the strong positive anomaly of the preceding summers.The opposite extreme anomalies between 2012 (positive anomaly) and 2013 (negative anomaly) have also been highlighted by 647 Hanna et al. (2014b) on the basis of the Greenland Block-648 ing Index.Secondly, as said above, similar circulation con-649 ditions were observed before 1880 and around 1891-1896, 650 when the Arctic climate was likely to have been much colder 651 than now.Further, Ding et al. (2014) suggest that the geopo-652 tential height increase observed over north-east Canada and 653 Greenland, as well as the negative NAO trend, could be due 654 to SST changes in the tropical Pacific that induce changes in 655 the Rossby wave train affecting the North-American region.656 Since the tropical SST changes are not reproduced by Gen-657 eral Circulation Models under current greenhouse gas con-658 centrations, Ding et al. (2014) conclude that these changes 659 are due to the natural variability of the climatic system.On 660 the other side, Wu et al. (2014) showed that the progres-661 sive intensification of the Beaufort Sea High over 1979-2005 662 can only be reproduced by climate models by including the 663 observed greenhouse gas concentration increase.Moreover, 664 Screen et al. (2012) have shown that various forcings are 665 needed to explain the observed Arctic warming: while Arctic 666 sea ice and associated SST changes, as well as remote SST 667 changes (corroborating the conclusions of Ding et al. (2014)) 668 are the main drivers of the winter warming, the summertime 669 temperature increase could mainly be due to increased ra-670 diative forcing, suggesting a role of global warming.Mat-671 sumura et al. (2014) have found a significant relation be-672 tween the earlier spring snowmelt over the Eurasian conti-673 nent and the enhanced summertime Arctic anticyclonic cir-674 culation.The earlier snowmelt could induce a negative SLP 675 anomaly over Eurasia, which is compensated by an SLP in-676 crease over the western part of the Arctic region.Finally, 677 Bezeau et al. (2014) conclude that the anomalous anticy-678 clonic patterns over the Arctic over 2007-2012 are due to 679 combined effects of sea ice loss, snow extent reduction, and 680 enhanced meridional heat advection.Thus, while it is widely 681 admitted that the Arctic region experiences a strong warming 682 since some years (Screen et al., 2012), the complexity of the 683 climate of this region due to its multiple internal and external 684 forcings and feedbacks does not allow us to solve the ques-685 tion whether the 2007-2012 circulation anomaly is (mainly) 686 due to global warming or to natural variability.687 Our findings corroborate those of Ballinger et al. (2014) 688 and Overland et al. (2012), who found that the Beaufort Sea 689 High is associated with anticyclonic conditions over Green-690 land.This is particularly clear for the Z500-based classifi-691 cation, where Type 5 combines both the Beaufort Sea High 692 and the Greenland High.Moreover, the circulation type fre-693 quency anomalies observed over the 2007-2012 period on 694 the basis of SLP and Z500 are linked with the observed neg-695 ative NAO trend (Hanna et al., 2014a).Thus, our results 696 seem to be in agreement with the hypothesis of Overland Fig. 1.The SLP-based reference circulation types over the 1980-2012 (JJA) period for ERA-Interim are represented by the solid black isobars (in hPa).The SLP anomaly (in colours) is calculated as the difference between the class mean SLP and the seasonal mean SLP over 1980-2012.The average frequency of each type is also given.

Fig. 3 .
Fig.3.Top: the average SLP spread and its standard deviation are calculated as the seasonal (JJA) average 20CRv2 spread and its standard deviation over the oceanic pixels of our domain.The maximum SLP spread is the value of the oceanic pixel showing the highest seasonal (JJA) average spread of each year.Bottom: the SLP spread is calculated as the average 20CRv2 spread over the 1871-1930 summers (JJA), left, and over the 1950-2012 summers, right.

Fig. 4 .
Fig.4.The 30 year running correlation is calculated between the annual (JJA) SLP-based frequencies of each type and the JJA CRU NAO index.For the "Types 2 + 4" correlation, the frequencies of Types 2 and 4 have been summed before computing the correlation.