Human influence on Arctic sea ice detectable from early 1990s onwards
Abstract
[1] Human influence has previously been identified in the observed loss of Arctic sea ice, but this hypothesis has not yet been tested with a formal optimal detection approach. By comparing observed and multi-model simulated changes in Arctic sea ice extent during 1953–2006 using an optimal fingerprinting method, we find that the anthropogenic signal first emerged in the early 1990s, indicating that human influence could have been detected even prior to the recent dramatic sea ice decline. The anthropogenic signal is also detectable for individual months from May to December, suggesting that human influence, strongest in late summer, now also extends into colder seasons.
1. Introduction
[2] Arctic sea ice has been declining since the 1980s and its reduction rate appears to be accelerating [e.g., Serreze et al., 2007; Meier et al., 2007]. It has been suggested that the recent accelerated Arctic sea ice loss may have been the result of the rapid reduction in Arctic multi-year sea ice in combination with other factors related to ice movement and atmospheric thermodynamics [Francis et al., 2005; Lindsay and Zhang, 2005; Maslanik et al., 2007b].
[3] Observed and model simulated Arctic sea ice extents (ASIEs) both show decreasing trends with generally smaller magnitude in the simulations [Hegerl et al., 2007, and references therein]. Some previous studies [Vinnikov et al., 1999; Gregory et al., 2002] have shown that observed trends in ASIE are inconsistent with climate model simulated internal variability. Gregory et al. [2002] in particular found that the observed ASIE trends for 1970–1999 are consistent with those from HadCM3 ensemble simulations under anthropogenic forcing but not with those from natural (solar and volcanic) forcing runs, implying that human influence has likely contributed at least in part to the trend in observed Arctic sea ice decline. It has also been suggested that Arctic sea ice may melt faster in the future than projected by the models [Stroeve et al., 2007]. These changes result in significant direct impacts on ecosystems and human communities throughout the Arctic region [Arctic Climate Impact Assessment (ACIA), 2005].
[4] In this study, we present the first quantification of human influence on Arctic sea ice by comparing observations with simulations from 18 climate models using an optimal fingerprinting technique. We show that anthropogenic signals are clearly evident from the early 1990s onwards, irrespective of the rapid sea ice loss in recent years.
2. Data and Methods
[5] We calculated the observed ASIE for each month from 1953 to 2006 by combining ice concentration estimates of the Hadley Centre sea ice and sea surface temperature (HadISST) dataset [Rayner et al., 2003] and a satellite passive microwave dataset following Meier et al. [2007] (observational data processing details given in auxiliary material). The satellite dataset, covering 1979 to 2006, was produced at the NASA Goddard Space Flight Center using the NASA-Team algorithm (Goddard NT) [Cavalieri et al., 1984]. ASIE is defined as the total area represented by grid points or pixels that are more than 15% covered by sea ice. Data preparation included adjustments to reduce inhomogeneities that result from different data sources. Nevertheless, in order to consider possible remaining inhomogeneity in the observed ASIE estimates between pre- and post-satellite eras [Walsh and Chapman, 2001], we repeated our detection analysis reported below based on ASIE for the satellite era only (1979–2006). Our main detection results remain robust for summer months for the shorter period analysis (Figure S1).
[6] We used climate simulations from the Coupled Model Intercomparison Project phase 3 (CMIP3) multi-model data set (Table 1). Arctic sea ice concentrations for 1953–2006 were extracted from all available multi-model 20th century experiments (20C3M) which were performed under observed anthropogenic only (greenhouse gases plus sulfate aerosols, ANT) and natural (solar and volcanic activities) plus anthropogenic (ALL) forcings up to 1999. The ANT and ALL simulations were extended to 2006 using corresponding multi-model simulations employing the Special Report on Emissions Scenarios (SRES) A1B forcing scenario. Utilizing A1B scenario runs for 2000–2006 to extend 20C3M is acceptable because, in this period, natural (solar and volcanic, NAT) forcing and the differences among various SRES scenarios are negligible. Since none of the models provides both ANT and ALL runs, the available A1B segments can be uniquely used for extending either ANT or ALL simulations. This provides 17 ANT runs from 10 models and 23 ALL runs from 8 models for 1899–2006. These ensembles of extended runs are smaller than the available 20C3M ALL and ANT ensembles due mainly to the availability of A1B runs and in part to some missing data from 20C3M runs (Table 1). Data for the earlier 1899–1952 period from the same historical simulations were used to estimate the internal variability (unforced climate variability; see auxiliary material for details).
| Model Name | ALL | ANT | A1B | MMEa | MMEb | |||
|---|---|---|---|---|---|---|---|---|
| 1 | BCCR_BCM2.0 | - | 1 | 1 | ANT | |||
| 2 | CCSM3 | 5 | - | 5 | ALL | ALL | ||
| 3 | CGCM3.1(T47) | - | 5 | 5 | ANT | ANT | ||
| 4 | CGCM3.1(T63) | - | 1 | 1 | ANT | ANT | ||
| 5 | CNRM-CM3 | - | 1 | 1 | ANT | ANT | ||
| 6 | CSIRO-Mk3.0 | - | 1 | 1 | ||||
| 7 | ECHAM5/MPI-OM | - | 3 | 3 | ANT | ANT | ||
| 8 | ECHO-G | 2 | - | 2 | ALL | ALL | ||
| 9 | GFDL-CM2.0 | 1 | - | 1 | ALL | |||
| 10 | GFDL-CM2.1 | 1 | - | 1 | ||||
| 11 | GISS-AOM | - | 2 | 2 | ANT | ANT | ||
| 12 | GISS-ER | 5 | - | 5 | ||||
| 13 | INM-CM3.0 | 1 | - | 1 | ||||
| 14 | IPSL-CM4 | - | 1 | 1 | ANT | ANT | ||
| 15 | MIROC3.2(medres) | 3 | - | 3 | ALL | ALL | ||
| 16 | MRI_CGCM2.3.2 | 5 | - | 5 | ALL | |||
| 17 | UKMO-HadCM3 | - | 1 | 1 | ANT | |||
| 18 | UKMO-HadGEM1 | - | 1 | 1 | ANT | ANT | ||
| Sum | Runs | 23 | 17 | 40 | 11 | 14 | 15 | 16 |
| Models | 8 | 10 | 18 | 4 | 7 | 4 | 9 | |
- a ANT and ALL (20C3M) simulations during 1953–1999, and SRES A1B simulations during 2000–2006. Historical simulations for 1899–1952 from the corresponding ALL and ANT runs are used to estimate internal variability. MMEa and MMEb represent constituent members of two sub-model groups with better performance in simulating Arctic sea ice climatology.
[7] We use a standard optimal detection method where the observations (y) are expressed as a sum of scaled fingerprints (X) and internal variability (
): y = Xβ + . We estimate the regression coefficient β using the total least squares method [Allen and Stott, 2003]. The fingerprint is estimated from multi-model means and the internal variability is also estimated from multi-model simulations following the approach by Gillett et al. [2002] which showed that multi-models in general give more robust detection results than a single model. Our method differs from Gillett et al. [2002] in that we use transient simulations rather than control runs for estimating internal variability (details described in auxiliary material). Dimensionality is reduced in the standard way by projecting both observations and simulations onto the 5 leading empirical orthogonal functions (EOFs) of internal variability. Consistency between model simulated and observed variability is examined using a residual consistency test [Allen and Tett, 1999]. Detection of the external forcing signal is claimed when the 90% confidence range for β lies above zero.
[8] The optimal detection method requires that the residual follows a normal distribution. The observational record is too short to assess normality reliably. However, examination of climate model simulated internal variability using graphical tools and an objective test (the Kolmogorov-Smirnov test) indicates that it is reasonable to assume that ASIE anomalies follow a normal distribution.
[9] We compared long-term fluctuations and changes in the seasonal cycle of ASIE in observations and model simulations. Comparing change in both the amplitude and shape of the annual cycle of ASIE reduces the likelihood of spurious detection due to coincidental agreement between the response to anthropogenic forcing and other factors, such as slow internal variability. Specifically, we compared non-overlapping three-year means of observed and simulated anomalies over 1953–2006 for March, June, September, and December. These months were selected because they represent the annual maximum and minimum sea ice extents that occur in March and September respectively, and the intervening transition season values. Selecting four different months (January, April, July, and October or February, May, August, and November) for detection did not change the main results given below (not shown).
3. Detection Results
[10] Figure 1 shows the observed and modeled seasonal evolution of 3-year mean ASIE anomalies for 1953–2006. The observations exhibit a decreasing trend since about 1980 that is common to all months with stronger amplitude in summer than in winter [Walsh and Chapman, 2001; Stroeve et al., 2007]. The ANT and ALL model simulations display patterns of sea ice decreases consistent with observations including the stronger trend in summer, but decreasing more slowly than observations. The temporal pattern and magnitude of change are similar in both the ANT and ALL simulations, indicating that the influence of the NAT forcing is unlikely to be detectable. In our optimal detection analysis, comparisons between observations and simulations are done in the reduced space of fingerprints represented by the 5 leading EOFs of internal variability (see above). In order to see how well the fingerprints capture the real world, we have reconstructed 3-year mean ASIE patterns by projecting the fingerprints back onto the physical space (Figure S2). Overall trend patterns in Figure S2 are very similar to those in Figure 1 although variation across the annual cycle becomes weaker.
[11] Figure 2a shows results of detection and attribution analyses for time periods beginning in 1953 and ending at various points in time from 1976 to 2006. It is clear that regression coefficients become significantly greater than zero for the time periods ending in the early 1990s and remain so from that point onward, indicating that anthropogenic influence begins to be detectable at that time. This also suggests that the detection of anthropogenic influence on Arctic sea ice does not depend on the recent dramatic and accelerated reduction in ASIE. It should be noted, however, that the regression coefficients are generally significantly greater than unity, implying that the model simulated Arctic sea ice response to ANT forcing substantially underestimates the observed changes. This is consistent with qualitative comparisons between observations and model simulations [Stroeve et al., 2007]. The model simulated sea ice response to ALL forcing can be similarly detected. Detection continues to hold when internal variability is doubled (dashed lines). We inflate variance to account for the possibility that models may underestimate internal variability and to test the robustness of detection. Models appear to simulate internal variability adequately on the time scales retained in our analysis.
[12] To examine which season/month may have contributed more to the detection results, we analyzed the 3-year mean ASIE anomalies for each of the twelve calendar months separately. We found that overall the ANT signal is robustly detectable in sea ice extent changes in individual months from May to December (Figure 2d). This indicates that human influence on sea ice extent is not only apparent in September when sea ice extent is at its lowest, but that it is also detectable in other parts of the year, extending from the beginning of the sea ice melt season in May through mid-winter. It should be noted, however, that the residual consistency test fails frequently when considering individual months, indicating that models reproduce internal variability less successfully at the scales retained in the monthly analyses.
[13] We also considered whether NAT forcing might have influenced ASIE during the latter half of the 20th century. Solar influence appears negligible due to lack of long-term trends. However, volcanic activity might have increased ASIE following eruptions through global cooling [Gregory et al., 2002], which might have been reduced by regional ‘winter’ warming associated with enhanced positive phases of the Arctic Oscillation (AO) [Hegerl et al., 2007, and references therein]. In this context, we have also attempted to separate ANT and NAT signals with a two-signal analysis using ANT and ALL simultaneously (not shown). ANT and NAT are jointly detectable in our full four-month analysis as well as in some individual month analyses. In these analyses, the ANT signal is clearly detectable and separable from the NAT signal whereas the NAT signal is not detected (i.e., the marginal confidence interval for NAT does not exclude zero) even when including high-frequency EOF modes onto which the NAT signal may project well. We speculate that this is mainly due to under-simulation of the NAT signal by the models [Stenchikov et al., 2006].
4. Sensitivity Tests
[14] Arctic sea ice variations are also influenced by the AO through associated changes in ice circulation [Rigor et al., 2002; Liu et al., 2004]. This is relevant for our study because a secular change occurred in the AO from about 1970 through the mid-1990s [Thompson and Wallace, 2001]. In order to test if our detection results are sensitive to long-term AO changes, we repeated our optimal detection analyses using sea ice extent series from which the AO influence has been removed. We first regressed the observed monthly ASIE time series onto the January-March mean AO index defined as the leading principal component of Northern Hemisphere monthly mean geopotential height anomalies at 1000-hPa [Thompson and Wallace, 2001]. The monthly AO indices were obtained from the NOAA's Climate Prediction Center. Next we subtract out the regressed components from the original time series to produce AO-residual observations for every month separately. Comparison of the ASIE fluctuations from the raw and AO-residual observations (not shown) indicates that the secular AO change does not substantially affect the long-term variations of the Arctic basin sea ice extent, in accord with previous studies [Liu et al., 2004; Lindsay and Zhang, 2005; Meier et al., 2007]. Similarly, our detection results are robust to the exclusion of the AO influence on observations (Figure S3). This is consistent with previous studies which suggest that, since the late 1990s, the AO influence on sea ice extent has been reduced whereas thermodynamic factors such as changes in downward longwave and ice-albedo feedback have become more important [Francis et al., 2005; Lindsay and Zhang, 2005]. However, it should be noted that the prolonged period of high AO index through the 1990s could have flushed multi-year ice out of the Arctic, and been responsible for part of the thinning that may not have been fully captured by the year to year regression.
[15] Sea ice simulation has improved because of increases in atmospheric and oceanic resolutions and the implementation of improved sea ice dynamics and thermodynamics in some climate models. Models that contributed to the IPCC Fourth Assessment Report have been found to simulate key features of Arctic sea ice variations including the seasonal cycle reasonably well [Randall et al., 2007]. However, shortcomings remain; thus most previous studies have selected subsets of models according to model performance in simulating observed climatology [Zhang and Walsh, 2006; Stroeve et al., 2007; Overland and Wang, 2007]. In this context, we tested the sensitivity of our detection results to the use of model simulated signals estimated from two sub-groups of models (selected from the 18 models analyzed above) that satisfactorily simulate some aspects of the observed seasonal sea ice climatology. One group (MMEa, multi-model ensemble “a”) consists of models whose climatological Arctic sea ice area lies within the ±20% of the observed summer values and within ±30% of winter values during 1979–99 (model list obtained from Overland and Wang [2007]). There are 4 ALL and 7 ANT MMEa models. The second group (MMEb) consists of models whose climatological ASIE values lie within ±20% of observed September values during 1953–1995 (model list from Stroeve et al. [2007]). It contains 4 ALL and 9 ANT models (Table 1). Note that these ensembles, and their ALL and ANT sub-ensembles, are not identical because of differences in months and evaluation periods used in the two studies. We found that our detection results are largely insensitive to the use of different model samples (Figures 2b and 2c).
5. Concluding Remarks
[16] Our results indicate that human influence on the changes in ASIE can be robustly detected since the early 1990s. Consistent with previous studies [Stroeve et al., 2007], we also find that most climate models substantially underestimate the observed Arctic sea ice loss, although some individual model simulations exhibit sea ice loss comparable to observed [e.g., Gregory et al., 2002]. Poor representation of sub-grid scale ice thickness, atmospheric and ocean heat transport, the surface energy budget, and other structural errors in models are among critical factors for the under-simulation [ACIA, 2005; Ridley et al., 2007; Sorteberg et al., 2007]. Regional atmospheric circulation changes that affect Arctic sea ice motion differently from the AO may also be important [Maslanik et al., 2007a]. Quantifying the relative contributions from dynamic and thermodynamic factors will be an imperative future work on human influences on the Arctic sea ice loss. Results presented here strengthen the accumulating evidence of human influence on the Arctic, together with the recent findings based on changes in the Arctic hydrological cycle [Peterson et al., 2006; Min et al., 2008]. The implications of these changes for Arctic ecosystems, human exploitation of the Arctic, both for resource extraction and as a transportation corridor, as well as possible geopolitical implications, are profound.
Acknowledgments
[17] We thank two anonymous reviewers for constructive comments. We acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) and the WCRP's Working Group on Coupled Modelling (WGCM) for their roles in making available the WCRP CMIP3 multi-model dataset. Seung-Ki Min was supported by the Canadian IPY (International Polar Year) program.
