Geophysical Research Letters
  • Open Access

When will the summer Arctic be nearly sea ice free?

Authors


Abstract

[1] The observed rapid loss of thick multiyear sea ice over the last 7 years and the September 2012 Arctic sea ice extent reduction of 49% relative to the 1979–2000 climatology are inconsistent with projections of a nearly sea ice-free summer Arctic from model estimates of 2070 and beyond made just a few years ago. Three recent approaches to predictions in the scientific literature are as follows: (1) extrapolation of sea ice volume data, (2) assuming several more rapid loss events such as 2007 and 2012, and (3) climate model projections. Time horizons for a nearly sea ice-free summer for these three approaches are roughly 2020 or earlier, 2030 ± 10 years, and 2040 or later. Loss estimates from models are based on a subset of the most rapid ensemble members. It is not possible to clearly choose one approach over another as this depends on the relative weights given to data versus models. Observations and citations support the conclusion that most global climate model results in the CMIP5 archive are too conservative in their sea ice projections. Recent data and expert opinion should be considered in addition to model results to advance the very likely timing for future sea ice loss to the first half of the 21st century, with a possibility of major loss within a decade or two.

1 Introduction

[2] The large observed shifts in the current Arctic environment represent major indicators of regional and global climate change. Whether a nearly sea ice-free Arctic occurs in the first or second half of the 21st century is of great economic, social, and wildlife management interest. There is a gap, however, in understanding how to reconcile what is currently happening with sea ice in the Arctic and climate model projections of Arctic sea ice loss. September 2012 showed a reduction in sea ice extent of 49% relative to the 1979–2000 baseline of 7.0 M km2 (Figures 1 and 2a). Further, the extent of thick multiyear sea ice has been reduced by the same percentage (roughly a reduction from 4 M km2 for 2000 through 2005 to 2 M km2 in 2012 [Kwok and Untersteiner, 2011, updated; Comiso, 2012]). It is difficult to reconcile this current loss rate with climate model projection dates of summer sea ice loss of 2070 [International Panel on Climate Change (IPCC), 2007] or 2100 [Boé et al., 2009a] made just a few years ago.

Figure 1.

Arctic sea ice extent for September 2012 (white areas) at 3.6 M km2. The magenta line indicates the median climatology for 1979–2011. September 2012 represents a 49% decline. The black cross is the geographic North Pole. Credit: National Snow and Ice Data Center (NSIDC).

Figure 2.

Time series of September Arctic (a) sea ice extent from NSIDC and (b) sea ice volume as computed from PIOMAS of APL/UW. The trend line for 1979–2012 is shown in solid black with shaded areas showing 1 and 2 standard deviation from the trend. Units are in M km2 for Figure 2a and 1000 km3 for Figure 2b. When expressed in terms of percentage of change, the declining trend in the sea ice volume is larger than the sea ice extent. http://psc.apl.washington.edu/wordpress/research/projects/arctic-sea-ice-volume-anomaly/.

[3] The question, however, is not as straightforward as simply comparing data time series with model results. Global climate models (GCMs) are often run several times, referred to as ensemble members, with slightly different initial conditions to simulate a possible range of natural variability in addition to steady increasing greenhouse gas forcing. Data, in contrast, are a single realization of a range of possible climate states. Observations confound signal (global warming forcing) and noise (natural variability). Thus, it is not completely valid to compare the ensemble mean of a model or several models, which could be considered the expected value of the climate state, with the single data realization. A better approach is to look at the range of ensemble members and to determine if the data time series could be considered a possible member of the population of ensemble members. Unfortunately, there are seldom enough ensemble members to test this hypothesis. The science question becomes as follows: is the observed rapid loss of sea ice in the real world consistent with model ensemble members with the fastest rate of loss? Multiple groups (AMAP, WCRP, and various national programs), as well as the climate research community and the general public, are interested in this question for adaptation planning and as a popular indicator of climate change.

[4] When will the summer Arctic be nearly sea ice free? A first issue is the word “nearly.” It is expected that some sea ice will remain as a refuge north of the Canadian archipelago and Greenland at the end of summer. Thus, the practical limit for sea ice loss is arbitrary, but several sources have converged on 1.0 M km2 as a minimum transition point. There are three scientific approaches to the posed question in the scientific literature. The first is based on extrapolation of sea ice volume data. The second considers that it will take several more rapid loss events such as the losses in 2007 and 2012 to reach the minimum. The third approach is to base predictions on fast-track model ensemble member projections. We refer to the three approaches as trendsetters, stochasters, and modelers. Time horizons for summer sea ice loss of these three approaches turn out to be roughly 2020, 2030, and 2040, as discussed below. At present, it is not possible to completely choose one approach over another as it depends on the weight given to data, understanding of Arctic change processes, and the use and purpose of model projections. The next sections address these three approaches.

2 Trendsetters

[5] Two groups are active in investigating sea ice volume loss [Schweiger et al., 2011; Maslowski et al., 2012]. Their main points are that sea ice volume is decreasing at a percentage rate that is faster than sea ice extent and that volume is a better variable than extent to use for sea ice loss. Schweiger and Zhang's group uses the Pan-Arctic Ice Ocean Modeling and Assimilation System (PIOMAS), which assimilates sea ice concentration and sea surface temperature and hindcasts using NCEP Reanalyses into a high-resolution sea ice model. PIOMAS results have recently been confirmed by satellite ice thickness measurements [Laxon et al., 2013]. PIOMAS monthly averaged ice volume for September 2012 was 3400 km3 (Figure 2b). This value is 72% lower than the mean over 1979–2012 and 2.0 standard deviations below the 1979–2012 trend line. September 2012 ice volume was about 800 km3 less than the prior minimum in September 2011. In contrast to the dramatic reduction in 2012 sea ice extent, the 2011 to 2012 change in sea ice volume was similar to the volume losses that occurred in the previous 2 years. The long-term trend is about −3.1 × 103 km3 decade−1. While the PIOMAS team does not directly extrapolate, the already major volume loss of 70%−80% and recent losses suggest that extrapolation into the future from the current volume amount shows that Arctic sea ice is vulnerable within the next decade.

[6] Monthly mean Arctic sea ice volumes from the NAME model and recent satellite estimates show that sea ice volume changed little during the 1980s through the mid-1990s [Maslowski et al., 2012]. After 1995, one can estimate a negative trend of −1.1 × 103 km3 yr−1 from combined model and most recent observational estimates for October–November 1996–2007. Given this estimated trend and the volume estimate for October–November of 2007 at less than 9000 km3, one can project a nearly ice-free summer Arctic Ocean before 2020 [Maslowski et al., 2012].

3 Stochasters

[7] In the recent half decade, young melt-prone sea ice has come to dominate the Arctic sea ice pack which supports the arguments of the trendsetters. However, the paper of Kay et al. [2011] suggests that there can be a modifying influence from natural variability especially for the timing of sea ice loss. They show a widening of the distribution of possible 10 and 20 year trends in sea ice extent in the Community Climate System Model 4.0 (CCSM4) model due to increased vulnerability of sea ice to large meteorological or oceanic events. Kay et al. [2011, their Figure 3] show that over a future 20 year period, sea ice loss can vary over a range of 0%–80%. Both CCSM3 and CCSM4 models show rapid ice loss events with different timing in different ensemble members [Holland et al., 2006; Vavrus et al., 2012]. The key argument of the stochasters is that it will take several rapid loss events such as that which occurred in 2007 and 2012 to reach the 1.0 M km2 sea ice extent threshold. If we select the 5 year interval that occurred between the 2007 and 2012 sea ice loss events as an expected value, then three more events puts a nearly sea ice-free timing at 2028. Serreze [2011] states that we should be looking at sea ice-free summers only a few decades from now.

Figure 3.

September sea ice extent based on 89 ensemble members from 36 CMIP5 models under the RCP8.5 (high) emission scenario. Each thin colored line represents one ensemble member from the model. The thick yellow line is the arithmetic mean of all ensemble members, and the blue line is their median value. The thick black line represents observations based on adjusted HadleyISST_ice analysis for the period 1953–1978 and NSIDC from 1979–2012. Observation data were provided by Meier, NSIDC. The horizontal black dashed line marks the 1.0 M km2 value, which indicates nearly sea ice-free summer Arctic.

[8] Holland et al. [2006] and Wang and Overland [2009, 2012] show a large range of timing of sea ice loss for different ensemble members of the same GCM. Based on a subset of available GCMs, Wang and Overland [2012] estimated the time for a nearly sea ice-free summer Arctic to be reached starting from a value of 4.5 M km2 (the observed 2007 value) ranged from 14 years to 36 years with a median of 28 years based on individual ensemble members, which puts the loss event in the 2030s with a large range. Given that most sea ice trends in models are slower than observed trends for 1979–2011 [Stroeve et al., 2012, their Figure 3; see next section], we should select a value at the earlier end of this range, i.e., 2030 ± 10 years.

[9] Stochasters are further supported by recent papers that suggest that there is no tipping point associated with sea ice loss, again based on modeling studies [Amstrup et al., 2010; Armour et al., 2011; Ridley et al., 2011]. Tietsche et al. [2011] suggest that anomalous loss of Arctic sea ice during a single summer is reversible as the ice-albedo feedback is alleviated by large-scale recovery mechanisms. That is, continued sea ice loss requires continued increases in green house gases. However, consensus is not universal as adequately representing cloud feedbacks in GCMs may be placing too much faith in them [Lenton, 2012].

[10] Thus, it is suggested that the stochasters would require 20 years or more after 2007 or around 2030 with a wide range of uncertainty to have several rapid ice loss events occur and to reach nearly sea ice-free conditions. While not unreasonable, stochasters are the most ad hoc of the three approaches.

4 Modelers

[11] GCMs are major quantitative tools available to provide future climate projections based on physical laws that control the dynamic and thermodynamic processes of the atmosphere, ocean, land, and sea ice. Recently, modeling groups around the world have improved their GCMs and made their results available to the wider scientific community through the archive at the Program for Climate Model Diagnosis and Intercomparison (PCMDI). This constitutes the fifth phase of the Coupled Model Intercomparison Project (CMIP5) following the successful third phase (CMIP3). Typically, results from more than 20 models are available. All models show loss of sea ice as greenhouse gas concentrations increase and that Arctic warms faster than lower latitudes. Multiple model simulations are particularly useful in assessing uncertainty due to differences in model structure, natural variability, and different greenhouse gas emission scenarios [Hodson et al., 2012].

[12] A first major difficulty is the wide spread of model hindcast results; they vary by model, location, variable, and evaluation metric (Figure 3) [Overland et al., 2011; Kwok, 2011]. Figure 3 is based on the high greenhouse gas emission RCP8.5 scenario [Moss et al., 2010]. A second major difficulty is that 80% of 56 CMIP5 ensemble members have trends for 1979–2011 that are of less magnitude than the 2 standard deviations bound for the observations [Stroeve et al., 2012, their Figure 3]. Thus, there is no ideal all-purpose model for the Arctic. It is difficult to pin down the reasons for these two difficulties [Walsh et al., 2008]. For example, DeWeaver et al. [2008], Eisenman et al. [2008], Hodson et al. [2012], and Holland et al. [2012] note that the Arctic radiation budget results from complex balances and tradeoffs between sea ice amounts, albedo parameterization, and cloud properties. Another issue is that real-world Arctic conditions (sea ice, snow cover) are evolving substantially faster than ensemble means of models [Stroeve et al., 2012; Derksen and Brown, 2012]. The time series of the grand mean of CMIP5 ensemble members based on the high greenhouse gas emission RCP8.5 scenario for September sea ice (yellow line in Figure 3) never reaches the nearly sea ice-free definition of 1.0 M km2 by 2100. Winton [2011] shows that climate models underestimate the sensitivity of Arctic sea ice cover to global temperature change. Further, Boé et al. [2009b] conclude that GCMs' Arctic response to anthropogenic forcing is generally too small. Thus, there is ground to consider that models provide a range of projections based on their individual assumptions, rather than providing a collective definitive Arctic climate prediction on the timing of sea ice loss.

[13] Pavlova et al. [2011] note that the multimodel ensemble mean is closer to the data curve for the late twentieth and early 21st centuries for CMIP5 relative to CMIP3 results. However, the spread of hindcasts and future trajectories remains large in CMIP5 models (Figure 3) [also see Massonnet et al., 2012, their Figure 1]. Boé et al. [2010], Hodson et al. [2012], and Massonnet et al. [2012], among others, note that the rate of sea ice loss in models depends on the amount of sea ice present. Thus, there is concern with projections from models that do not simulate the amount of observed sea ice near the end of the twentieth century.

[14] There are four major evaluations of sea ice projections in the set of CMIP5 GCMs: Pavlova et al. [2011], Stroeve et al. [2012], Wang and Overland [2012], Massonnet et al. [2012], and one detailed review for the CCSM4 model [Vavrus et al., 2012] and for the EC-Earth model [Koenigk et al., 2012]. The median value for each year of all available CMIP5 ensemble members (blue line in Figure 3) reaches the nearly sea ice-free condition near 2060 based on a nearly sea ice-free definition of 1.0 M km2. However, given the large observed rate of sea ice loss, we are primarily interested in model ensemble members with the most rapid sea ice loss. The ensemble members of seven models which track closely to recent observed sea ice extents [Wang and Overland, 2012] had their earliest nearly sea ice-free dates occurring in 2027, 2033, 2035, 2045, 2048, 2049, and 2060, with a mean of 2042. Some individual ensemble members in Figure 3 reach the nearly sea ice-free threshold at earlier dates, but many of these ensemble members start with unrealistic low sea ice extents for the late twentieth century. Several of the ensemble members of CCSM4 reach the sea ice loss threshold near 2060; this was 10 years later than their previous model CCSM3.The EC-Earth model also becomes nearly sea ice free near 2060, but the authors suggest shifting this to 2040 based on the model's overestimate of the amount of sea ice in the twentieth century. Thus, we put the early limit for sea ice loss based on GCM projections near 2040.

[15] This paper should not be used as an argument against further modeling, but quite the opposite. The Arctic community needs credible quantitative climate projections with multiple ensemble members. As noted above, the spread in Figure 3 is not only due to sea ice physics but is related to treatment of clouds, radiation, and atmospheric and ocean dynamics. For the next round of model results, CMIP6, the major goal should be reduction of model uncertainty. Perhaps more model intercomparisons would be a way forward rather than results provided from a large number of modeling centers produced under short time schedules.

5 Discussion and Conclusions

[16] We have investigated three approaches to predicting 21st century summer Arctic sea ice loss as represented by trendsetters, stochasters, and modelers. At present, it is not possible to completely choose one approach over another as all approaches have strengths and weaknesses. Models are quantitative, based on physical understanding, and can provide estimates of uncertainty. They all predict an eventual sea ice-free Arctic based on increases in greenhouse gas forcing. Modelers' projections for a nearly sea ice-free Arctic summer are centered on 2060, with a composite of the earliest removals by selected models near 2040. It is not clear that the observed rapid sea ice loss is represented in the range of model GCM results. Extrapolating current sea ice volume trends seems to capture the influence of the recent rapid loss of multiyear sea ice, yet it will be hard to remove the last sea ice near the North Pole; in 2007, removal of this sea ice required a strong atmospheric and sea ice advection event [Zhang et al., 2008].

[17] Direct extrapolation of sea ice volume, by trendsetters, gives loss projections of 2016 [Maslowski et al., 2012] (Peter Wadhams, 2012, personal communication), which may be minimizing the potential effects of year-to-year variability. Stochasters acknowledge current conditions and the range of projections suggested by model results yet point to the lack of being able to forecast the next rapid sea ice loss event. They are saved in part as it will possibly take several such events to reach the nearly sea ice-free threshold, thus adding some averaging to the final date prediction (hence stochastic). Observations and citations in this article support the conclusion that current rapid Arctic change, especially loss of multiyear sea ice, is likely out of sample for most CMIP5 models. Thus, time horizons for summer sea ice loss of these three approaches turns out to be roughly 2020, 2030, and 2040 respectively for trendsetters, stochasters, and modelers. Predictions depend on the weight given to data, understanding of Arctic change processes, and the use of model projections. It is reasonable to conclude that Arctic sea ice loss is very likely to occur in the first rather than the second half of the 21st century, with a possibility of loss within a decade or two.

[18] The title of this paper is certainly one of the major questions of interest to Arctic and non-Arctic science and management communities. Large shifts in the Arctic environment represent major observed indicators of global climate change. Available evidence suggests that scientists have been conservative in their climate projections, with a late bias in dates for change [Brysse et al., 2012]. Ignoring the rate of observed loss of multiyear Arctic sea ice in favor of multimodel results primarily from GCMs may be a further example. The possibility of a nearly sea ice-free Arctic within the next two decades, in addition to the precautionary principle, supports the Duarte et al. [2012] conclusion that society should start managing for the reality of climate change in the Arctic.

Acknowledgments

[19] The work is supported by NOAA Arctic Research Project of the Climate Program Office and by the Office of Naval Research, Code 322. Impetus for this article came from a workshop held in Seattle during October 2012, sponsored by the Arctic Monitoring and Assessment Program (AMAP); it was also a contribution to WCRP Polar Prediction and the IASC Atmospheric Working Group. We appreciate the exchange of ideas from the workshop and subsequent discussions. We acknowledge the work of PCMDI and the various modeling centers for CMIP5. This publication is partially funded by the Joint Institute for the Study of the Atmosphere and Ocean (JISAO) under NOAA Cooperative Agreement NA10OAR4320148, contribution no. 2071. PMEL contribution number 3971.

Ancillary