Sea-ice area variability and trends in Arctic sectors of different morphology, 1996–2015

ABSTRACT This study presents a comparison of the sea-ice cover of the whole Arctic Ocean with two arctic sectors of different morphology: the Greenland Sea, as a typical “open sea”, and the Beaufort Sea, as a typical “closed sea”. The study refers to the period January 1996–December 2015 and makes use of the Arctic sea-ice concentration data set produced, on a daily basis, by the Institute of Environmental Physics of the University of Bremen. From the whole Arctic data set, two subsets, covering the Greenland Sea and the Beaufort Sea, were extracted. The extent of sea-ice cover was obtained by the sea-ice area (SIA) parameter, which was computed according to the conventional NASA method. Our analysis shows that the strong summer decline of the Arctic SIA in the last 20 years is not observed in the Greenland Sea (the trend of SIA minimum values is ) while it is even greater in the Beaufort Sea.

In this paper, the annual cycle of sea ice area (SIA) distribution in the whole Arctic Ocean has been analysed for the period January 1996-December 2015 and compared with what happened, in the same period, in two sectors of the Arctic of different morphology, the Greenland Sea, a typical "open sea", and the Beaufort Sea, a typical "closed sea". While being well aware that sea-ice cover depends on several physical factors (low temperatures, persistence of a surface fresh water layer, inflow of low salinity water through the Bering Strait etc.) (Serreze & Barry, 2014), the aim of this study was to investigate trend differences between two seas of different morphology of the Arctic Ocean. This result seems to confirm the hypothesis of King (2014) which tries to explain the different degrees of sea-ice decline in the Arctic and Antarctic regions: in the Arctic Ocean, where the ice is mostly confined by the surrounding continents, the summer ice is dramatically decreasing, while in Antarctica, where the sea ice is largely free to drift with the wind and ocean currents, the ice is growing.

Data and methods
The IUP data set The study was carried out using the daily Arctic sea-ice concentration (SIC) data set produced by the Institute of Environmental Physics (IUP) of the University of Bremen. SIC data were derived from the observations of SSM/I, SSMIS, AMSR-E and AMSR-2 passive microwave sensors by using the ARTIST Sea Ice (ASI) algorithm (Kaleschke et al., 2001;Spreen, Kaleschke, & Heygster, 2008), based on 89 GHz brightness temperature values distributed by NSIDC and by JAXA. The SSMIS data set was only used to fill the gap between AMSR-E and AMSR-2. All SIC data are mapped by IUP onto a stereo-polar grid having a pixel of 6:25 km. From the Arctic data set, two subsets covering the Greenland Sea, approximately from 57.1 N to 83.5 N and from 41.4 W to 14.9 E, and the Beaufort Sea, approximately from 80.4 N to 66.5 N and from 116.6 W to 184.3 W, were extracted (see Figures 1 and 2). These two regions, together with the whole Arctic Ocean, are our regions of interest (ROIs) (Figure 3).
Hereafter, the three study areas will be identified as ROI1 (the Arctic Ocean as a whole), ROI2 (the Greenland Sea) and ROI3 (the Beaufort Sea). As the sensors do not cover a small circular zone around the North Pole, for the full Arctic region, a constant value of 100% SIC was added to fill the hole. With reference to the AMSR-2 product, the surface area of the three ROIs is approximately 85, 2:6 and 2:0 M km 2 , respectively.
Building the monthly SIA time series [1996][1997][1998][1999][2000][2001][2002][2003][2004][2005][2006][2007][2008][2009][2010][2011][2012][2013][2014][2015] Given N grid cells the SIA is conventionally defined as  where SIC i and A i are, respectively, the sea-ice concentration and the surface area of the ith grid cell. In addition, the sum only takes into account the cells having SIC ! 0:15 (Zwally et al., 1983). After SIA estimation, an outlier-removing method, based on the distance of the points from the running mean of the series, was applied. The percentage of outliers for each ROI was 0.3%, 2.2% and 0.4%, respectively, with the highest values occurring for the SSM/I and SSMIS series. The next step was to put together the SIA time series derived from different satellite sensors. Due to the different spatial resolution between SSM/I and the AMSR data, a slight difference between the SIA as measured by the two sensors is to be expected.
The consistency between SSM/I and AMSR-E data sets was obtained by normalizing the SSM/I data set by means of parameters derived using a linear regression of the data from the two sensors during the period of overlap. In Comiso and Nishio (2008), the normalization was applied to the brightness temperature of each channel, but in this work, a linear regression model was simply applied to SIA: In Table 1, the regression coefficients, c 0 and c 1 , together with the mean-area difference, md, and the standard deviation, sd, between the two daily products (SSM/I and AMSR-E) for the overlapping period, that is are shown. Determination coefficients, R 2 , for three ROIs are greater than 0.99, as expected.
In Table 1, the three slope values, being close to unity, suggest that the use of the above adjustment procedure was not critical; on the contrary, the application of a linear regression for the transition period from AMSR-E to AMSR-2 is needed, as the simple insertion of SSMIS data would produce a SIA underestimation. A new linear regression, using common data for both SSMIS and AMSR-2, was applied: Results, analogous to those of Table 1, are shown in Table 2; determination coefficients, R 2 , are still larger than 0:99.

Maximum and minimum annual values
The SIA monthly time series, together with the regression lines for maximum and minimum values, are shown in Figure 4 for the three ROIs.
The slope values β max;ROI and β min;ROI (ROI ¼ 1; . . . ; 3) with their standard errors, se max=min;ROI , are reported in Table 3. The two-sided null hypothesis H 0 : β max=min;ROI ¼ 0 was tested with the t-statistic, t ¼ ðβ max=min;ROI =se max=min;ROI Þ. The null hypothesis should be rejected if the corresponding p-value is smaller than a selected significance level.
Observation of Table 3 leads to the following considerations: • All maximum and minimum values, except Greenland Sea minimums, show a statistically significant trend at significance level α ¼ 0:05. • The variability of maximum annual values for ROI2 is greater than for ROI1 and ROI3, particularly before 2003 as can be seen in Figure 4. • For all ROIs, minimum values are always observed in August or September; most of them, for ROI1 and ROI3, occur in September while those for ROI2 mainly occur in August (Table 4). Note that for ROI3, the   Table 4). • In 2000 and in 2012, i.e. in correspondence with SIA minimum, maximum values in the Greenland Sea were observed in May.

Monthly SIA anomalies
In order to minimize the intra-annual seasonal variations, the monthly anomaly trends, obtained by subtracting the monthly climatological averages from each monthly average (Comiso & Nishio, 2008), were computed and are shown in Figure 5, with Table 3 reporting their slope values. With regard to Figure 5 and Table 3, and comparing the results to those of other authors, it is possible to say: • Negative trends of monthly anomalies are all statistically significant at virtually any statistical significant level α.   different physical processes for the 2007 and 2012 events is clear: in 2012, the cause may have been the arrival of a strong storm over the central Arctic, which produced the separation and the melting of a large expanse of ice (Parkinson & Comiso, 2013); for the 2007 decline, instead, the contributions of cloud and radiation anomalies were decisive (Kay, L'Ecuyer, Gettelman, Stephens, & O'Dell, 2008). • For each investigated region, monthly SIA anomalies were compared to the monthly surface air temperature anomalies. The comparison for the three ROIs is shown in Figure 5 where, to improve clarity, a moving average smoothing was applied and the linear trend was removed. Surface air temperature data were retrieved from the NCEP-NCAR R1 archive (Kalnay et al., 1996). Large temperature anomalies are observed in correspondence of the large negative SIA anomalies, as for the Beaufort Sea region (Figure 5) in 1998(Figure 5) in , 2007(Figure 5) in and 2012. Besides, it is interesting to note the strong anticorrelation for ROI2, not so apparent for ROI3, due to the low anomaly values during the 2002-2005 period; ROI2, in the same period, shows a high annual relative range ( Figure 5). Correlation coefficients between SIA and temperature anomalies for each regions are À 0:40, À 0:29 and À 0:31, respectively, and are all statistically significant at significance level α ¼ 0:05. • Despite the different analysed periods, a comparison of SIA value trends obtained by other authors is shown in Table 5, where column 5 reports the SIA trend values for 1979-2010 and 1979-1996, derived from 32 years of SSM/I-SSMIS observations by Cavalieri and Parkinson (2012); a rough estimation of the trends for the period 1996-2010 gives for ROI1, À 70:7 km 2 year À1 , for ROI2, À 5:4 3 km 2 year À1 , values very close to the one reported in this paper. In , the  trend for ROI1 in the period 1996-2007 appears higher than the one reported by Parkinson and Cavalieri (2008) for the period 1979-2006. • Correlations between monthly SIA anomalies and monthly Southern Oscillation Index (SOI) for the three regions are À 0:14, À 0:08 and À 0:14, respectively; p-values suggest a significant statistical correlation between SIA and SOI for ROI1 and ROI3.

Discussion and conclusions
Some conclusions based on the results of the previous section are • Trend comparison (Table 5) confirms a considerable compatibility between SIA computation using NSIDC and IUP data sets, although obtained from independent observations; the time series of the other authors in Table 5 are based on the Bootstrap (Comiso, 2012) and NASA Team (Cavalieri & Parkinson, 2012) algorithms, respectively, which both use the 19 and 37-GHz channels, whereas, as already stated, the ASI algorithm used here uses the 89-GHz brightness temperatures. • All trends, except for minimum values in the Greenland Sea (0:7 AE 2:010 3 km 2 year À1 ), are negative: with the null hypothesis of trend's absence, all p-values are much lower than the significance level α ¼ 0:05. The apparent slightly increasing SIA in the Greenland Sea is most likely due to the continuous oceanic outflow through the Fram Strait. • For ROI2, summer minimum values are occurring about one month earlier (August) than ROI1 and ROI3 (September). • In ROI3, the maximum values are often persistent for 6-7 months.
The relevant result of this study is provided by Figure 4 which shows, for the 20 years under analysis, an increase of SIA in the Greenland Sea, in contrast to what has occurred in the Arctic Ocean as a whole and, to a greater extent, in the Beaufort Sea. Studies on the ice loss in the Beaufort Sea (Hutchings & Rigor, 2012;Kwok & Cunningham, 2010;Steele, Dickinson, Zhang, & Lindsay, 2015) have shown how extreme cyclonic circulations of the recent years have produced a decrease of the ice age and how this resulted in ice within the central Beaufort Gyre being less than 5 years old and hence more easily subject to melting. For the Greenland Sea, it has been demonstrated that the large flux of ice from the Arctic Basin through the Fram Strait, with its injection of freshwater to the North Atlantic (De Steur et al., 2009;Kwok, Cunningham, & Pang, 2004), leads to an increased SIA and to a rapid decline in the ice area of the Arctic. Even if formulated for the Antarctic Ocean, the hypothesis of King (2014) may apply to Arctic regions of different morphology. For the reasons explained, the Beaufort Sea, a closed sea, has seen a marked decline of ice in the recent years while in the Greenland Sea, an open sea, the SIA tends to grow.