The Svalbard Airport composite series is based on direct observations from Svalbard Airport from 1975 to the present, which we will call the principal series. Before 1975, it is a composite series consisting of many local series. The composite is established in two steps: (1) quality checking and homogenizing the local series; and (2) adjusting the local series to ensure their comparability with the principal series (the Svalbard Airport site). The Standard Normal Homogeneity Test (SNHT) introduced by Alexandersson (1986) checked the local series. This is a relative homogeneity test that requires reference stations for comparison. However, the number of reference stations is sparse in Svalbard, so taking metadata into account is even more important than when working with denser networks. During some periods the neighbouring stations are so remote that the analysis of the metadata is the only real homogeneity test. Inhomogeneities in a data series might be due to radiation screen changes, changed observational procedures, site relocations and so on. After finishing the first step, a network was available in which the series were considered to be homogeneous. In the second step, the local series were adjusted to be directly comparable with the principal series by regression analysis. The daily mean temperature of the principal series, T_s, can be written: 1 T S = α 1 T A + C 1 + e A where T_A is the daily mean temperature of a local series that partly overlaps the principal series, α₁ and C₁ are constants calculated by the least square method and e_A represents noise (residuals). The predicted daily mean temperature of the principle series, T ∧ S , is given by Eqn. 2: 2 T ∧ S = α 1 T A + C 1

Similarly, for another local series B partly overlapping series A: 3 T A = α 2 T B + C 2 + e B where T_B is the daily mean temperature of series B, α₂ and C₂ are constants and e_B represents noise (residuals). If series B does not overlap with the principal series, series B may still be used for the prediction of the principal series’ temperature. Combining Eqns. 1 and 3 gives us: 4 T S = α 1 ( α 2 T B + C 2 + e B ) + C 1 + e A = α 1 α 2 T B + α 1 C 2 + α 1 e B + e A + C 1 where α₁e_B+e_A represents the noise.

The Svalbard Airport composite series consists of monthly mean temperatures, from 1975 onwards based on direct observations, and, before that year, on regression-based predictions. These were calculated by taking the arithmetic means of the predicted daily values within each month. If the local series A overlaps with the principal series the monthly means adopted in the composite series, , are given by Eqn. 5, and for a local series that does not overlap by Eqn. 6: 5 T ∧ S m = α 1 T A m + C 1 6 T ∧ S m = α 1 α 2 T B m + α 1 C 2 + C 1 where T_Am, T_Bm are the monthly means for the local series A and the local series B, respectively.

For the whole sample, the mean values of ɛ_A and ɛ_B are zero per definition, but for a sub-sample, such as one particular month, this is not the case. If the daily mean values are not auto-correlated, the uncertainty will be reduced by a division of 5.5 (approximately the square root of the number of days in a month). Although the composite series consist of monthly means, annual and seasonal means are also necessary for climate research. In the absence of autocorrelation, the uncertainty, s_e, for the temperature in a season is: 7 s e = 1 m ∑ i = 1 m s e i 2 where s_ei is the standard deviation for month i and m is the number of months in each season (m is usually 3) or in the whole year (m=12).

Testing for trends in the data sets was performed with the non-parametric Mann-Kendall (M-K) test, which may be used without knowing the distribution of the time series as it is a rank test. Its test statistic, t, is defined by the equation 8 t = ∑ i = 1 n n i where n is the number of elements and n_i is the number of smaller elements preceding element x_i (i=1,2,.... n; Sneyers 1990). Providing that n>10, the test statistic is very nearly normally distributed under the hypothesis of randomness (the null hypothesis). Moreover, its expectation, E(t), and variance, var t, are given by the equations 9 E ( t ) = n ( n - 1 ) 4 10 var t = n ( n - 1 ) ( 2 n + 5 ) 72

The standard distribution, u(t), of the test statistic is then 11 u ( t ) = t - E ( t ) var t

A percent table of the normal distribution function can be used to decide whether the null hypothesis should be rejected or not. In order to study variations on selected time scales, the time series were smoothed by a Gaussian filter: 12 G j = ∑ i = 1 n w i j ⋅ x i ∑ i = 1 n w i j where 15 w i j = exp - ( i - j ) 2 2 s 2

G_j is the value of the smoothed series in year j, x_i is the data point in year i, w_j is the weight in year j, n is the number of years in the series and s is the standard deviation in the Gaussian distribution. The degree of smoothing is established by the value of the standard deviation in the distribution; for example, a standard deviation of three years corresponds to an approximately 10-year rectangular low-pass filter.

Climate data series may also contain regime shifts (step changes) or a blending of steps and trends (Corti et al. 1999). However, interpretations of data in terms of steps and trends depend to an extent on what results are expected, and the scientific background of any particular researcher (BACC Author Team 2008: 14). For the detection of regime shifts, we used a sequential data processing technique introduced by Rodionov (2004). The testing procedure works as follows: for each new observation a test is performed to determine the validity of the null hypothesis, H0 (the existence of a regime shift). There are three possible outcomes of the test: accept H0, reject H0, or keep on testing. Before testing it is necessary to set a cut-off length, which roughly determines the minimum length of what is accepted as a regime. Student's test was used to decide whether two subsequent regimes were significantly different.

The many local series that have contributed to the composite Svalbard Airport series are listed in Table 1, with their national station number and names, and their locations are shown in Fig. 1. They are described briefly below.

After having started measurements at a new site farther from the airport runway from 5 October 2010 onwards, parallel measurements were performed in the period 27 October 2010–8 November 2011 (Supplementary file; for calibration of the sensors see the same file). The monthly means of the temperature differences between the new and old site all fall in the interval −0.09 to +0.06°C. It is therefore concluded that the Svalbard Airport temperature series is homogeneous through the relocation and screen type shift of 5 October 2010. Further to this, the Svalbard Airport series was tested for homogeneity in the period July 1978–December 2010 by use of the SNHT method relative to the stations: 99754 Hornsund; 99842 Barentsburg; and 99910 Ny-Ålesund. The test result indicated an inhomogeneity of the Svalbard series in the test period, although without an indication of the distinct year of the shift. For the annual data, the most likely year of the shift was 2004. Compared with the reference group, the temperature at Svalbard Airport was warmer than the reference group after the shift. Metadata were examined for a possible confirmation of the shift, but with a negative result. The same screen has been used in the whole test period, and the sensors inside the screen have been tested once a year without any detected malfunction. A possible explanation of the detected inhomogeneity might be an inadequate reference group, which reflects more maritime conditions than at the airport.

In Adventdalen (only 23 km from Svalbard Airport), a permafrost station was established at Janssonhaugen (270 m a.s.l.) in May 1998 (Sollid et al. 2000; Isaksen et al. 2001, 2007). From May 2000, air temperature is available from a weather station (run by the University Centre in Svalbard) that was established 15 m north of the permafrost station. The weather station does not have complete data coverage as it lacks data for the periods June 2005–October 2006 and June 2010–April 2011. However, these are interpolated using another station in Adventdalen. The series was used as reference for the Svalbard Airport series in a SNHT test. No significant inhomogeneity was detected. On the basis of these tests, the Svalbard Airport series from 1975 to the present is considered to be homogeneous.

The nearest local series to Svalbard Airport was situated at the main settlement, Longyearbyen, at a distance of only 15 km. For the construction of the composite series, the Longyearbyen series has a key role as it overlaps both with the old local series, and with the principal series. It was tested for inhomogeneity by SNHT (Nordli et al. 1996) but no shift in the series was detected. Due to poor correlation and long distances to the neighbouring series, the SNHT test is not considered to be a strong test in this case. However, metadata studies show that the measurements have always been located approximately at the same height above the valley floor (Steffensen et al. 1996) which, in itself, is an indication of a homogeneous series.

In order to adjust the Longyearbyen series, linear regression analysis was performed with the Longyearbyen temperature as predictor, and the Svalbard Airport temperature as predictant, using the daily mean temperature during an interval of parallel measurements (November 1975 to July 1977). Analysis was performed separately for each month. The predicted Svalbard Airport temperature was given by Eqn. 2. The regression accounted for as much as 99% of the variance in large parts of winter and autumn, and in no month was it less than 90% (Table 2). Visualization of the regression is shown for March and September in Fig. 3a. For most of the year, the Longyearbyen series is warmer than Svalbard Airport so negative adjustments are needed (Fig. 4). This reflects the location of the station on a small plateau over the valley floor, which reduces the effects of inversions on the local climate in contrast to the location of the Svalbard Airport station.

The standard error of the estimates (RMSE) of the daily means is obtained directly by the regression. However, the composite series of monthly means has less uncertainty. If randomness is assumed, the uncertainty in the monthly values may be calculated from the uncertainty of the daily values by dividing by the square root of the number of days in the months. This gives an RMSE of only 0.2°C in winter and 0.1°C in summer, which is very satisfactory. These estimates might be realistic for the last period of the series, 1957–1977, in which the station was located at the same spot. However, for earlier periods, the sites differed and there might be some inhomogeneities in the series. Station history data suggest that any such inhomogeneities must be minor.

The homogeneity of the Barentsburg series was tested by SNHT for the period 1978–2010 with use of the reference stations Hornsund, Svalbard Airport, and Ny-Ålesund. For this period, the Barentsburg station was found to be homogeneous. The first part of the series has earlier been tested for inhomogeneities (Nordli et al. 1996). Their conclusion was that this part of the series was also homogeneous. For inclusion in the composite series Longyearbyen is the first choice, although Barentsburg is a good alternative for some periods where Longyearbyen lacks data. These are mainly during 1947–1956, but also for some years in the 1930s (Table 1).

The overlap with the principal series during the period 1975–2010 amounted to about 1000 days in each month. The good data coverage reduced the uncertainties in the predictions. The relative variance accounted for by the regression is largest during the period from September to May, when it was no less than 93% in any month, whereas in July it was 84% (Table 2). For May and September, scatter plots are shown as well as the regression lines from Eqn. 2 (Fig. 3a). The negative adjustment during winter, and positive during the summer, for the Barentsburg series might reflect a more maritime climate compared to the Svalbard Airport site (Fig. 4). For the Barentsburg series, the assessment of uncertainty in the monthly means is straightforward as it overlaps for about 35 years with the principal series. By performing an extra regression using monthly means (instead of daily means) RMSE is calculated directly by the regression (Table 3). It amounts to 0.7°C during late winter, whereas in summer and autumn it is only about 0.4°C.

For long periods, the Green Harbour series was the only series at Svalbard so that relative homogeneity tests such as SNHT cannot be used. There seem to be relatively good metadata from the station that indicate homogeneous conditions throughout the period of observation (Steffensen et al. 1996). The Green Harbour series is therefore considered to be homogeneous. The Green Harbour series does not overlap with the principal series, but it does overlap with the Longyearbyen series. The monthly means of the Green Harbour series could be adjusted by Eqn. 6 but not by Eqn. 5. This adds uncertainties in the prediction that might be assessed by use of the Barentsburg series.

As the Barentsburg series (local series B) also overlaps with the Longyearbyen series (local series A), predicted values of the principal series might also be inferred by Eqn. 6. The values obtained by Eqn. 5 could therefore be compared with those from Eqn. 6. It appeared that the use of Eqn. 6 increased the uncertainties by only 0.1°C or less, so the extra regression from the Longyearbyen series to the Svalbard Airport series had very little impact on the uncertainties of the monthly means. The uncertainty was assessed using monthly-mean values in a regression between Green Harbour and Longyearbyen, adding an extra uncertainty of 0.1°C (Table 3). It transpired that the uncertainty was larger for the periods predicted by Green Harbour than those for Barentsburg.

During the period September 1946–November 1947, none of the adjusted series mentioned so far have data coverage. The only exception is the Isfjord Radio series, which is not an ideal series to be used as it is much influenced by open water off the coast of Spitsbergen. The series overlaps with the principal series by only one year, so the Longyearbyen temperature series was used for the prediction. Due to the maritime influence on the station, large negative adjustments had to be used in winter (about −3°C; Fig. 4), whereas positive adjustments were needed during summer. The uncertainties were assessed using the same procedure as for the Green Harbour series (Tables 3, 4).

As new stations were established on the old sites Akseløya, Crozierpynten, Svarttangen and Sørkappøya, their data series overlap with the principal series during the period 2010–12. Equation 5 was therefore used for the adjustment of the series. For daily values it turned out that the prediction for most of the year accounted for roughly 90% of the variance (highest for Akseløya and lowest for Crozierpynten; see Table 2). For summer, however, the prediction was very poor. The sample size was increased from 60 to 180 daily means by joining data from the three months of June, July and August. Nevertheless the predictions turned out to be of little use except for Akseløya. Scatter plots with regression lines for March and September are shown in Fig. 3b.

The old stations lie outside the Isfjorden area with quite long distances to the Svalbard Airport site: Akseløya is 70 km, Svarttangen 150 km, Crozierpynten 180 km and Sørkappøya 200 km. It can be argued that one day is too small a time window for the travelling weather systems, so over such distances they might be out of phase. A system might be at Sørkappøya one day, and at Svalbard Airport the next day. By increasing the time window to pentads, the same weather system could more easily be caught in the same pentad. Regression analyses were performed with pentads but hardly any improvement was seen (as the sample size was small deleted residuals were compared). The only exception might be Crozierpynten for summer where the variance accounted for amounted to 50% with pentads versus 36% with daily means.

During winter, autumn, and spring, the RMSE of the predicted daily means was about 1.5°C for Akseløya, 2.5°C for Crozierpynten, 2.0°C for Svarttangen and 2.0°C for Sørkappøya (Table 2). The adjustments differ between the series and months (Fig. 4). As long as the fjords are covered with ice, there are only small adjustments needed for the Akseløya series. Akseløya is cooled by the fjord during summer, whereas in autumn it is warmed by the open water in the fjord. It is also seen that the northernmost station Crozierpynten is colder than Svalbard Airport during the whole year. For Sørkappøya, the adjustments reflect its maritime influence with negative adjustments during the cold period of the year, and positive during summer. For Svarttangen (Fig. 4) the sign of the adjustments are opposite to those at Sørkappøya, that is, it is positive during the cold part of the year and negative during summer.

During the Second World War, and also during five winters in the period 1898–1911, no observations were made in Svalbard, so the only possibility for filling data gaps is by interpolation. Monthly-mean temperatures for the grid point 78°N, 16°E (Compo et al. 2011) were used for this. Regression lines for each month were established, with the composite series as predictand, and the grid point as predictor. Data before 1912 stem from series outside the Isfjorden area so they might be less accurate than the rest of the series. They were therefore omitted in the regression, which used the data for the period 1912–2008, which is complete except for the Second World War gap.

The variability of the monthly mean of the grid point was much lower than for the station data, in particular during winter, leading to a regression coefficient during winter larger than 2 (Table 2). In the season September through February, 60–82% of the variance was accounted for by the regression depending on the month, whereas in the season March through August only 34–56% was accounted for. During winter the RMSE in the monthly means was nearly 3°C, much higher than the uncertainty in the station-based values in the Svalbard Airport composite series.

Based on the assumption that the RMSE for monthly-mean temperatures are really white noise, the RMSE of the seasonal and annual means may be calculated by Eqn. 7. These are given in Table 4 for the series where regression analyses on monthly means are performed. Comparison with Table 3 shows how much uncertainties are reduced. The RMSE for the seasonal means is not higher than 0.5°C for winter and spring and 0.3°C for summer.

A key point for discussion is the uncertainties regarding variations in monthly, seasonal and annual mean temperatures, and the homogeneity of the local series. Inhomogeneous series may lead to biased composite series. In this regard, the Barentsburg series provided crucial information for the current reconstruction as it overlaps for some 40 years with the principal series. It became clear that the uncertainties in the monthly means during autumn and winter were slightly larger than 0.5°C, whereas during summer they were slightly lower than 0.5°C (Tables 3, 4). For the other old series, the addition of further information is not straightforward due to a lack of overlap with the principal series, so Eqn. 6 had to be used (see methods section). The question that needs to be addressed is therefore how much additional uncertainty was added by using Eqn. 6 instead of Eqn. 5. The Barentsburg series provided an excellent possibility to test this as it was possible to use both Eqns. 5 and 6. Comparison of the results showed that Eqn. 6 added very little additional uncertainty to the prediction, about 0.1°C.

Green Harbour is situated only 2.5 km from Barentsburg, and Longyearbyen only 15 km from Svalbard Airport, so it was expected that the regression between Green Harbour and Longyearbyen should lead to approximately the same uncertainty as that between Barentsburg and Svalbard Airport. It appeared, however, that the uncertainties in predicted monthly means by Green Harbour were somewhat larger than those from Barentsburg. This might be due to different time windows for the two stations as Green Harbour was closed before Barentsburg started. When Green Harbour station was active, that is, before 1930, Longyearbyen station was run by a private mining company, which might have carried out less accurate measurements than those done by the Norwegian Meteorological Institute when it took over in the 1930s.

There seem to be no changes in the Longyearbyen series during the period 1957–1977, that is, the most recent part of the Longyearbyen series, but it is likely that the Longyearbyen series is not strictly homogeneous for the whole period (although only slightly different sites and almost the same altitudes should imply that there would only be minor inhomogeneities. For annual values the RMSE is only 0.2°C for the stations in the composite series (Table 4), which is almost as good as the measuring accuracy of old thermometers. Therefore, while there are no RMSE errors that may hamper trend studies, there may be inhomogeneities. This is a problem in Arctic regions where there are very few or no reference stations. In order to redress this difficulty it is necessary to use metadata as much as is possible and feasible.

It is somewhat paradoxical that for the purpose of the current reconstruction the summer seasonal mean temperature predictions for the period 1899–1911 have been discarded, but the same values have been included in the predictions for annual means. The reason for not including them in the composite summer series relates to the fact that the regression sum of squares was almost as large as the total sum of squares so that predictions by regression analysis were considered to be of little use (see Table 2: Crozierpynten, Sørkappøya and Svarttangen). It is fortunate that the RMSEs (Table 2) for the summer months are smaller than for winter. In the case of Crozierpynten, Sørkappøya and Svarttangen, it was roughly one half of that in winter. Therefore, the main source of uncertainty in the annual means does not stem from the summer means despite the low variance accounted for by the summer regression.

In the later years of the period of observations, the Gaussian filtered curves show higher values than otherwise (also if the insignificant parts are omitted). This was also confirmed by the test for regime shifts as a regime warmer than all others in the period occurred at the latter end of the curves. However, it is important to note that there are settings in Rodionov's test programme that have to be chosen. For the level of significance p=0.05 was chosen, whereas Rodionov's default value is p=0.1. By changing the settings, information will be acquired regarding how robust the results are, so testing with p=0.1 was also performed. The result was close to the same: the only change was that the cold 1960s for winter was now defined as one regime separated from the 1972–2004 period. Even with p=0.01, a very strong constraint for regime shifts, the results were nearly the same so it is concluded that the results presented in Fig. 6 are robust. Another important setting is the cut-off length, which was set to 10 years. This means, a priori, that a regime of shorter duration than 10 years would not be accepted for the purpose of this reconstruction. However, in the testing procedure the onset of a regime is adjusted so that slightly shorter regimes than the cut-off length might be possible. Here, the cold period in the 1910s was accepted as a distinct regime although its length in the annual and winter series turned out to be only eight and seven years, respectively.

Wood et al. (2010) reconstructed air temperature back to 1802 for the northern Atlantic area (60°–90°N, 60°W–45°E) using the CRUTEM3v data set and four long-term series as predictors (from south-west Greenland, Iceland, Archangelsk in Russia and Tornedalen in Sweden) with a calibration period of 1950–1979. Comparing their results with ours (our Fig. 5; their Fig. 1) roughly the same pattern is seen in both figures; however, the trend seems to be somewhat smaller in their reconstruction than at Svalbard Airport. This is to be expected as their area extends as far south as 60°N, where less warming has occurred.

Sea ice has a large impact on air temperature, particularly during winter (Benestad et al. 2002; Screen et al. 2012; Semenov & Latif 2012). Open water is a heat source for the air. If the shores are ice-free or have little ice near one of the stations, and at another station the shores are ice-covered, then the differences between the temperatures at these two stations will be different from a situation where the shores at both stations are ice-free or ice-covered. Ideally, different ice classes would have been taken into account, but this has not been possible due to the lack of knowledge regarding local ice conditions in former times. Fortunately, it seems that the ice cover at Grønfjorden and the rest of Isfjorden do not develop very differently so that, in most cases, both fjords are either open or covered (Nilsen et al. 2008) with a possibility of slightly more open water in Grønfjorden as it lies near to the Atlantic Water off west Spitsbergen. As daily values were used in the regression equations it may be assumed that the lowest temperatures in winter and spring primarily represent ice-covered sea, whereas the highest temperatures mostly represent open sea.

With regard to the new automatic weather stations erected in the locations of the old expeditions, knowledge of sea ice is crucial; however, data are only available for two winters. It is fortunate therefore that these two winters were very different with regard to sea-ice conditions as the first one had a long period with sea ice, whereas in the second one sea ice was almost absent (see Supplementary file). It is also to be expected that the regression lines give a realistic adjustment for both ice-free and ice-covered conditions during the period 1898–1911.

Reanalysis data for the gridpoint (78°.0°N, 16°.0°E) lying in Reindalen, 30 km to the south of Svalbard Airport, has, on average, a bias that is much too warm, in particular during winter, where the difference is about 8°C in the months January through March. For the summer months, the gridpoint is about 2°C too cold in July and August. The gridded temperature is probably influenced by the open water off west Spitsbergen as it behaves more like an oceanic station than Svalbard Airport. As expected, the uncertainty of the interpolations are much larger than for the predictions based on local observations, up to 3°C for winter months, and about 1°C for summer (Table 3). For the assessment of the uncertainty of seasonal and annual means Eqn. 7 may be used. For the annual means, the uncertainty was estimated to be 0.6°C. Interpolations were used for the gaps during the Second World War, and for five years in the period 1898–1911. For most of the regressions, the data coverage is sparse, but not for all of them. It seems that the linear assumption is valid. The method of linear regression was therefore kept as a standard method for the entire reconstruction (see Fig. 3). It was not considered necessary to group the observations as is suggested, for example, by Della-Marta & Wanner (2006) in order to account for non-linearity.

The observations by hunting expeditions have been used very little in climatological analyses of the Svalbard climate until now. Clearly, this is due to the non-continuous character of the observations. After one season in a certain area the hunters tended to move to a new location as the wildlife became decimated. It might also be assumed that as soon as the first permanent observation station was established on Spitsbergen, meteorological interest concentrated on this. An exception is an article by Birkeland (1920), which uses the Green Harbour series as well as the shorter series. He admitted that the data were sparse: “the most one can say is that, in general, the temperature seems to decrease from south to north and from west to east” (Birkeland 1920: 288; our translation). The new observations reveal a complicated picture. Of particular interest are the high winter temperatures at the mouth of Isfjorden, lower temperatures in the far north, and late temperature increase in the spring in the south-east (see Fig. 4). Generally, observations have been lacking on the eastern coast of Spitsbergen and on the eastern islands, but in the last two years this has been overcome. More comprehensive and detailed studies of east-west variations of the Svalbard climate are now available (Przybylak et al. in press) based on data from the new stations.

It is of great interest to see how trends inferred from the composite Svalbard Airport series differ from other Arctic temperature series. Przybylak (2007) analysed seasonal and annual trends from the warm epoch in the mid-1930s to 2005 for many long-term Arctic series, resulting in very little detection of significant (p<5%) positive trends. In the Norwegian Arctic the only significant trend detected was for mean spring temperatures at Bjørnøya. Seven years have now passed since that paper was written, so the two series at Bjørnøya and Jan Mayen were revisited and the recent period 2006–2012 was also added to the series. A consistent picture of the Norwegian Arctic then appeared with more significant trends detected than in the earlier paper (Przybylak 2007). However, the trends for winter and autumn still remain insignificant when the starting year is in the warm period of the 1930s.

When the data are compared with Arctic areas lying outside the Norwegian Arctic, it may be seen that, in the period 1936–2005, the trend at Svalbard Airport was larger than in Greenland and the western Euro-Asiatic part of the Arctic, but lower than in the rest of the Arctic (see Przybylak 2007; his Table 2). In winter, Svalbard Airport exhibited one of the largest negative trends in the Arctic, whereas in spring and summer it showed one of the largest positive trends. In the shorter period, 1951–2005, trends of average annual temperature in Spitsbergen were slightly higher than in the rest of the Atlantic Arctic and western part of the Russian Arctic. The southern part of Greenland and the Baffin Bay region revealed significantly less warming and even cooling in some areas, while the Pacific side of the Arctic experienced markedly greater warming (Przybylak 2007; his Fig. 6). Temperature trends in winter and summer showed the same spatial patterns to those described for the annual values.

Hanna et al. (2012) present an updated analysis of monthly-mean temperature from Greenland coastal weather stations and from a long-running site on the Greenland Ice Sheet. They demonstrate very strong recent warming along the west coast of Greenland, especially during winter (locally>10°C since 1991), and rather weaker warming on the east Greenland coast, which is influenced by different oceanographic/sea-ice and meteorological synoptic conditions to the rest of Greenland. The greatest temperature changes in all seasons are seen on Greenland's west coast and on the ice sheet. Recent climate analyses were undertaken at the reindeer herding areas in northern Eurasia from northern Norway/Finland to the Bering Strait (Vikhamar-Schuler et al. 2010). The material was sparse as several stations were discarded from the long-term trend analysis, 1901–2008, due to many gaps in the series. The trends for annual means in this large area are all positive, but smaller than for Svalbard Airport, and only one was significant (Yakutsk).

The temperature increase on Svalbard Airport during the last 30–40 years shows the greatest increase among all areas in Europe, so the question can be posed: what is the reason for this? As demonstrated in this article, the temperature of Svalbard exhibits pronounced large-scale variations, which have been extensively discussed in the literature. Several causal factors have been suggested, in particular variations in sea-surface temperatures over periods longer than decades (Benestad et al. 2002; Polyakov et al. 2003; Johannessen et al. 2004). Debate also continues on the relative roles of local factors, such as sea-ice reduction, versus remote factors driving or amplifying Arctic warming (see, e.g., Screen et al. 2012). However, these authors conclude that local sea-ice concentration and sea-surface temperature changes explain a large portion of the observed warming. Direct radiative forcing due to changes in greenhouse gasses and total solar irradiance has mainly contributed to Arctic tropospheric warming in summer.

It has also been suggested that there has been a persistent change in early summer Arctic wind patterns during the last six years, 2007–2012, relative to previous decades (Overland & Serreze 2012). This pattern, which has previously been recognized as the Arctic Dipole, is characterized by relatively low sea-level pressure over the Siberian Arctic with high pressure over the Beaufort Sea, extending across northern North America and over Greenland. Coupled impacts of the new persistent pattern are increased sea-ice loss in summer, long-lived positive temperature anomalies and ice-sheet loss in west Greenland, and a possible increase in Arctic–Subarctic weather linkages through higher-amplitude upper-level flow (Overland & Serreze 2012). Thus, circulation changes may also have enhanced Arctic sea-ice melting in addition to warming due to increased greenhouse-gas concentrations.