Notice: Wiley Online Library will be unavailable on Saturday 27th February from 09:00-14:00 GMT / 04:00-09:00 EST / 17:00-22:00 SGT for essential maintenance. Apologies for the inconvenience.
 The spatiotemporal sensitivity of Antarctic sea ice season length trends are examined using satellite-derived observations over 1979–2012. While the large-scale spatial structure of multidecadal trends has varied little during the satellite record, the magnitude of trends has undergone substantial weakening over the past decade. This weakening is particularly evident in the Ross and Bellingshausen Seas, where a ∼25–50% reduction is observed when comparing trends calculated over 1979–2012 and 1979–1999. Multidecadal trends in the Bellingshausen Sea are found to be dominated by variability over subdecadal time scales, particularly the rapid decline in season length observed between 1979 and 1989. In fact, virtually no trend is detectable when the first decade is excluded from trend calculations. In contrast, the sea ice expansion in the Ross Sea is less influenced by shorter-term variability, with trends shown to be more consistent at decadal time scales and beyond. Understanding these contrasting characteristics have implications for sea ice trend attribution.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 The Arctic and Antarctic regions have experienced contrasting annual sea ice trends over the satellite-observing era; while the total Arctic sea ice extent (SIE) has shown a pronounced circumpolar decrease, total Antarctic SIE has concurrently exhibited a weak, yet statistically significant, increase [e.g., Stroeve et al., 2007; Cavalieri and Parkinson, 2008]. These hemispheric-scale metrics, however, mask the variable nature of Antarctic sea ice trends which have, in fact, been far from spatially uniform. Specifically, there have been regionally compensating positive and negative sea ice trends in the Ross and Amundsen-Bellingshausen Seas, respectively, promoting a robust dipolar pattern that is evident across analyses of sea ice concentration (SIC) [e.g., Simpkins et al., 2012], SIE [e.g., Cavalieri and Parkinson, 2008], sea ice season length [e.g., Watkins and Simmonds, 2000; Parkinson, 2002], and day of sea ice advance/retreat [e.g., Stammerjohn et al., 2008, 2012].
 While this dipolar trend pattern is a robust feature of analyses performed across the satellite era, with little change observed in its large-scale spatial characteristics, the magnitude of the corresponding trends exhibit substantial sensitivity to the period of the record analyzed. For example, the strength of Amundsen-Bellingshausen SIE trends decreased from −9.2 ± 3.6 to −5.4 ± 1.9% decade−1between 1979–1998 and 1979–2006, respectively [Zwally et al., 2002; Cavalieri and Parkinson, 2008]. Concurrently, SIE trends in the Ross sector fell from 6.5 ± 1.1 to 4.4 ±1.7% decade−1, with similar sensitivities observed across alternative sea ice metrics [e.g., Parkinson, 2002]. This inconsistency in the magnitude of sea ice trends indicates that estimates of multidecadal change are significantly influenced by short-term variability [e.g., Cavalieri et al., 2003; Watkins and Simmonds, 2000].
 Using data over the 34year period 1979–2012, this study aims to understand the sensitivity of Antarctic sea ice trends to the chosen measurement period and, in doing so, determine the extent to which higher-frequency variability affects calculations of multidecadal trends. Particular emphasis is placed on understanding the large-scale characteristics of sea ice season length trends in the Ross and Bellingshausen Sea regions. Specifically, this study sets out to (1) compare the spatial structure of sea ice trends over the two multidecadal periods 1979–1999 and 1979–2012, (2) assess how the magnitude of trends alters by lengthening the sea ice record, and (3) quantify trends as a function of start/end date.
2 Data and Methods
 This investigation makes use of quasi-daily estimates of Antarctic SIC derived from passive microwave retrievals (Scanning Multichannel Microwave Radiometer (SMMR) and Spatial Sensor Microwave/Imagers (SMM/I)) and processed using the Goddard Space Flight Center (GSFC) Bootstrap Version 2 algorithm [Comiso, 2000]. Note that repeating analyses with NASA Team SIC yields quantitatively similar results to those presented here. The ice concentrations are gridded onto a 25×25 km mesh, and are available from the Distributed Active Archive Center at the National Snow and Ice Data Center, Boulder, Colorado (http://nsidc.org).
 To account for the varying temporal resolution of satellite retrievals, SIC data taken every other day are examined from February 1979 to February 2013. Consistent with Stammerjohn et al.[2008, 2012], the sea ice year is therefore defined to begin and end in mid-February, the approximate climatological SIE minimum. Consequently, the reference year relates to the time during which the ice advanced, such that analyses span 1979 (February 1979 to February 1980) to 2012 (February 2012 to February 2013). Using these data, three distinctive, yet interrelated, metrics are defined for each ice season and for each grid point: (1) season length, defined as the total number of days where SIC is at least 15%, linearly scaled to a 365day year, (2) day of ice advance, the day when SIC first exceeds 15% for six consecutive days, and (3) day of ice retreat, the day when SIC first falls below 15% for the remainder of the ice season. As shown by Parkinson , subsequent results are largely insensitive to the choice of the ice concentration cutoff, and thus the 15% threshold is used to facilitate comparison with previous studies. Due to several months of missing data, sea ice fields in 1987 were replaced with the climatological mean; note that subsequent results are largely unaffected by this substitution and remain consistent with results wherein 1987 was excluded from the analysis. The statistical significance of trends is assessed at the 95% confidence level using two-tailed Student t tests following Santer et al. . Throughout the manuscript, subdecadal, decadal, and multidecadal refer to periods of 1–9, 10–19, and >20years, respectively.
 The top panels of Figure 1 display the climatological mean sea ice season length, day of sea ice advance, and day of sea ice retreat calculated over 1979–2012 (see supporting information Figure S1 for the equivalent plot, including differences, calculated over 1979–1999). In all cases, pronounced meridional gradients are observed, reflecting the strong annual cycle of Antarctic sea ice growth and decay (see Parkinson for further discussion of climatological conditions). Considerable interannual variability is also observed across all three metrics, as indicated by the 30day standard deviation isopleth for season length and the 15day standard deviation contour for ice advance/retreat (Figure 1, upper panels, black contours). This variability is largely constrained to the ice periphery (regions of shorter season length/later advance/earlier retreat), where the reduced consolidation of the ice pack makes it more vulnerable to intraseasonal/interannual variability in the atmosphere and ocean. Nevertheless, within the Bellingshausen Sea, all three metrics reveal pronounced variability that extends from the ice periphery to the Antarctic coastline.
 Despite the apparent year-to-year variability and local differences in climatological conditions, the lower panels of Figure 1 reveal marked large-scale and spatially coherent trends in season length, ice advance, and ice retreat over the two multidecadal periods 1979–1999 and 1979–2012; these time periods correspond approximately to the analyses of Parkinson  and Stammerjohn et al. , respectively, and are used to highlight how published estimates of sea ice trends have changed as the sea ice record has lengthened. Nevertheless, dividing the record into two equal periods yields similar motivating results to those presented in Figure 1(see supporting information Figure S2). Typically, trends in season length are found to be associated equally with trends in advance/retreat, although several regional asymmetries are evident; at the northern edge of the Ross Sea, for example, the trend toward a longer ice season appears to be tied more closely to a later retreat. In agreement with past studies, Figure 1 highlights that all metrics exhibit a heterogenous spatial pattern of change, dominated by contiguous opposing trends between the broader Amundsen-Bellingshausen (∼55–100°W) and Ross Sea (∼205°W–155°E) regions. Although comparisons of trends over the two multidecadal periods reveal localized differences, the large-scale spatial features of this dipolar pattern are shown to persist, as illustrated by the consistent structure of statistically significant trends (gray hatching).
 While the spatial structure of trends appear reasonably similar, Figure 1 illustrates that many regions exhibit a substantial decline in the magnitude of trends when considering the longer record (compare middle and lower panels). These changes are further highlighted by Figure 2, which shows a histogram of season length trends (the subsequent primary metric of analysis) as a function of magnitude and area. Figure 2 reveals a considerable narrowing of the distribution curve between 1979–1999 (gray bars) and 1979–2012 (green bars), corresponding to a decrease in the variance (i.e., spread), and an associated increase in the peak about the mean. Accordingly, a two-tailed Student t test indicates that the average trends are significantly different at the 95% confidence level.
 The largest differences in the magnitude of sea ice trends, however, are observed at the distribution tails (Figure 2). As seen in Figure 1, these regions largely correspond to the localized Bellingshausen and Ross Seas, as outlined by the contours labeled “B” and “R”, respectively, which delineate regions exhibiting statistically significant trends of at least ±10days decade−1over both time periods. Inset within Figure 2(top left) is a scatterplot of trends at all grid points over 1979–1999 and 1979–2012 (black dots), with those within the Bellingshausen (“B”) and Ross (“R”) Sea regions colored red and blue, respectively. Between the two time periods, marked decreases are evident in the magnitude of sea ice trends within the Bellingshausen region. Nearly all grid points track above the 1:1 line and, in many instances, track above the 2:1 line, signifying that a 13year extension of the sea ice record typically reduces multidecadal trend estimates by ∼25–50% in this location. Due to the weaker distribution of sea ice trends in the Ross region (blue dots), changes here appear less pronounced; nevertheless, it is apparent that trend magnitudes have also decreased by ∼25% at many grid points. While the spatial pattern of sea ice change therefore appears broadly consistent across the satellite era, Figures 1 and 2 highlight a considerable decline in the strength of multidecadal sea ice trends over the past decade, particularly in the localized Ross and Bellingshausen Seas evaluated here.
 To examine the evolution of trends in the Bellingshausen and Ross regions, time series of season length are calculated by averaging within the encompassing black contours of Figure 1, as defined previously. While individual grid-point analyses will invariably differ due to contrasts in local conditions and associated year-to-year variability (Figure 1, top panels), correlations of these area-averaged time series with season length at all other points reveal significant correlations throughout the broader Bellingshausen and Ross regions (supporting information Figure S3). These time series therefore reflect the large-scale variability and change observed in these areas of interest, as expected given the contiguous, and spatially coherent, nature of the trends observed here. Furthermore, we find that subsequent results are robust to changes in the thresholds defining these regions (e.g., using limits of 5 or 20days decade−1, thus extending/constricting their spatial extent).
 Figure 3a displays the time series (black line) for the Bellingshausen region and illustrates a principal reduction in season length over the observational record. Embedded within this long-term decline are considerable variations at interannual and subdecadal time scales. The latter is most clearly evident between 1979 and 1989, when a dominant decline in season length of ∼120days is observed. This strong reduction is followed by a period of reduced change, wherein season length fluctuates at ∼230days and decadal-scale oscillations become more apparent. Consistent with the strong variations over the early part of the record, the magnitude of trends (shown as bars and calculated from 1979 to the location of the bar) exhibit a marked and coherent decline when extending the end date of analysis forward in time, illustrating a confounding impact of higher-frequency variability on long-term trend estimates; note that the bars shown in Figures 3a and 3c correspond with the time periods used to examine sea ice trends in previous literature and thus act as markers to track estimates of sea ice change through time.
 To further determine the extent to which trend magnitudes are impacted by higher-frequency variability, trends are calculated as a function of both start and end date for all periods greater than 5years in Figure 3b. Given the short data record used to examine subdecadal to decadal variability, trends over these periods should not be overinterpreted and are shown simply to highlight their potential impact on trends at multidecadal time scales. Figure 3b reveals marked variability in the sign of subdecadal tendencies. Consistent with Figure 3a, the onset of the Bellingshausen record is associated with a pronounced negative period, with magnitudes that typically exceed those experienced at any other time. Consequently, decadal and multidecadal trends are strongly influenced by the early part of the time series, as highlighted by the systematic decrease in the magnitude of trends as the first 10years are excluded from the calculations. In fact, virtually no trend is detectable when the first decade is removed from the analysis, emphasizing that changes over subdecadal time scales are dominating the interpretation of multidecadal trends in the Bellingshausen Sea. Similarly, the observed autocorrelation function (Figure 3b, inset, black line) reveals an exponential decay as a function of lag. These characteristics are consistent with those of a theoretical AR1 process (dashed red line), highlighting a lack of consistent low-frequency (multidecadal) variability in the Bellingshausen time series.
 Complementary analyses are also performed for the Ross region and reveal contrasting characteristics to those observed for the Bellingshausen. The time series (Figure 3c) displays a fairly consistent increase in season length over 1979–2012. Accordingly, the magnitude of trends (bars) display weakened sensitivity to the end date of analysis (cf. Figures 3a and 3c, note the contrasting y axes), as illustrated by significant trends (bars with thick borders) which vary by 9days (17 to 26days decade−1) for the “R” region, compared to 30days (−22 to −52days decade−1) for the Bellingshausen. Similarly, Figure 3d demonstrates that the magnitude of Ross Sea trends exhibit greater consistency at decadal time scales and beyond. However, subtle modulation by higher-frequency variability is still evident, as indicated by decadal trend magnitudes peaking when extending through to ∼1996 and the relative reduction observed thereafter. Nevertheless, the observed autocorrelation function (Figure 3d, inset) reveals a higher degree of persistence than that of a theoretical AR1 process. These features highlight a strong low-frequency component to the Ross time series and thereby provide supporting evidence for a fairly consistent increase in season length over the observational record.
4 Summary and Discussion
 Using sea ice season length as the primary metric of analysis, this study has investigated the spatial and temporal variability of Antarctic sea ice trends over the satellite-observing era. The key conclusions from this study include the following:
 While the large-scale spatial structure of Antarctic sea ice trends persist over multidecadal time periods, the magnitude of these trends have weakened considerably as the satellite record has progressed, particularly in the localized Bellingshausen and Ross Seas and, thus the spatially dominant sea ice dipole (Figures 1 and 2).
 A rapid decline in sea ice season length is observed between 1979 and 1989 in the Bellingshausen Sea (“B”) region, with relative stability exhibited thereafter. As such, multidecadal sea ice trends in this location are dominated, to the first order, by subdecadal variability at the onset of satellite records, with trends becoming negligible when this period is excluded from the analysis (Figures 3a and 3b and supporting information Figure S2).
 Sea ice season length in the Ross Sea (“R”) region has undergone a fairly persistent increase over 1979–2012. Thus, contrary to the characteristics of the Bellingshausen region, trends here are more consistent at decadal time scales and beyond, although subtle modulation by higher-frequency variability is still evident (Figures 3c and 3d and supporting information Figure S2).
 Further to the well-established spatial heterogeneity (Figure 1), considerable contrasts are apparent in the temporal characteristics of Antarctic sea ice trends (cf. Figure 3), as also seen when evaluating the location of consecutive decadal-mean season length isopleths (supporting information Figure S4).
 We note that analyses of regional sea ice extent (supporting information Figure S5) reveal similar trend characteristics to those outlined above, highlighting broad-scale robustness across alternative sea ice metrics.
 While this manuscript does not attempt to diagnose the physical mechanisms causing Antarctic sea ice trends, understanding the temporal characteristics outlined above will have implications for attribution studies. Specifically, the contrasting characteristics between the dipole regions may indicate that separate processes with distinct temporal scales, rather than a large-scale common forcing, are driving the regional sea ice trends around Antarctica. Discerning these mechanisms, however, is complicated by the many confounding factors impacting Antarctic sea ice variability on various temporal/spatial scales. For example, atmospheric/oceanic circulation [e.g., Holland et al., 2006; Simpkins et al., 2012; Stammerjohn et al., 2008], associated ice export [Holland and Kwok, 2012], freshwater forcing [e.g., Bintanja et al., 2013; Zhang, 2007], and natural sea ice variability [e.g., Mahlstein et al., 2013; Meier et al., 2013; Polvani and Smith, 2013; Zunz et al., 2013] all drive complex atmosphere-ocean-ice interactions around Antarctica that complicate trend attribution.
 Further complications also arise due to intrinsic sea ice variability. For example, the near-complete removal of multiyear ice from the Bellingshausen region in the late 1980s/early 1990s [Jacobs and Comiso, 1993, 1997] would have brought about substantial changes to the nature of the ice pack in this location, forcing a change from a predominantly perennial to seasonal ice system (e.g., Figures 3a and 3b). Facilitated by a thinner ice cover, such shifts would tend to increase the susceptibility of the ice pack to external forcings (e.g., wind-driven ice export [Holland and Kwok, 2012]) and, similarly, to feedback mechanisms [Stammerjohn et al., 2012]. Both ice-albedo and ocean-thermal feedbacks will act to accelerate (decelerate) spring melt-back (autumn freeze onset) and thus maintain the now-seasonal nature of the ice pack in the Bellingshausen Sea. Changes in the opposite sense may likewise contribute to the observed trend characteristics in the Ross region; a more consolidated ice pack, for example, will exhibit less sensitivity to change and thus act to continually increase season length through associated feedback mechanisms.
 Regardless of the underlying mechanisms, this study highlights that our understanding of long-term Antarctic sea ice trends, particularly in the localized Bellingshausen Sea, is heavily influenced by higher-frequency (decadal-scale) variability. Analyses of earlier satellite records reveal a similar confounding impact when assessing hemispheric SIE. For example, Cavalieri et al.  note a substantial decline in SIE between 1973 and 1977, believed to be part of a long-term reduction from the 1960s [Meier et al., 2013]; consequently, multidecadal sea ice trends reverse sign when this period is included, reiterating the importance that short-term variability plays in long-term trend calculations. The results presented here thus add credence to the growing number of model-based studies which suggest contemporary Antarctic sea ice trends may be a consequence of natural internal variability in the coupled atmosphere-ocean-ice system [Mahlstein et al., 2013; Polvani and Smith, 2013; Zunz et al., 2013]. Thus, as Mahlstein et al.  note, it may not be wholly unexpected that CMIP5 models cannot replicate the observed Antarctic SIE expansion. The strong influence of natural variability, and the resultant difficulty in deriving robust estimates of contemporary trends at various spatial/temporal scales, therefore suggests that 34years of passive microwave observations are insufficient to accurately quantify Antarctic sea ice trends. Analysis of earlier satellite records, continued satellite observations, and improved proxy records thus offer the opportunity to put contemporary sea ice trends in the context of natural variability and, in doing so, offer a scope to determine anthropogenic signatures as well as provide a useful tool for climate model validation.
 The authors thank three reviewers, M. Raphael for helpful discussions, and J. Comiso for supplying the 2012–2013 SIC data. This work was supported by a UNSW University International Postgraduate Award, the Australian Research Council, and the ARC Centre of Excellence for Climate System Science.
 The Editor thanks three anonymous reviewers for their assistance in evaluating this paper.