Satellite passive microwave observations document an overall downward trend in Arctic sea ice extent and area since 1978. While the record minimum observed in September 2002 strongly reinforced this downward trend, extreme ice minima were again observed in 2003 and 2004. Although having three extreme minimum years in a row is unprecedented in the satellite record, attributing these recent trends and extremes to greenhouse gas loading must be tempered by recognition that the sea ice cover is variable from year to year in response to wind, temperature and oceanic forcings.
 As reported in a previous paper [Serreze et al., 2003] sea ice extent (the region with sea ice concentrations of at least 15%) in the Arctic reached a new record minimum in September of 2002. Ice extent in September 2002 was 4% lower than any other previous September since 1978 and 14% lower than the long-term mean [1979–2000]. Analysis of the extended HadSST1 ice record [Rayner et al., 2003] using passive microwave records, other forms of satellite data (back to 1972), as well as aircraft reconnaissance and ship reports, shows that September 2002 ice extent was probably the lowest of at least the last 50 years.
 The downward trend is characterized by strong interannual variability, with a low September ice extent in one year typically followed by some recovery the next September (the annual minimum in ice extent occurs in September). Our analysis shows that while ice extent for winter of 2002/2003 recovered to near its median position, there was a significant decrease in ice concentration within the pack by mid-summer, followed by a rapid reduction in ice extent in August and September of 2003. The ice extent again recovered to near its median position over the winter of 2003/2004, but the melt season began with a record minimum in May extent. Ice extent retreated rapidly in August and September, to a position near that of 2002. Having three extreme minimum years in a row is unprecedented in both the satellite era (1979–present), and over the past 50 years as assessed from the HadISST1 record.
2. Data and Observations
 Following Serreze et al. , our analysis relies primarily on passive microwave-derived sea ice concentration data sets available from the National Snow and Ice Data Center (NSIDC). Ice charts from the Navy/NOAA National Ice Center (NIC) corroborate the extreme retreat of the ice edge in 2002, 2003 and 2004.
 Ice concentration maps are derived using the NASA Team sea ice algorithm of Cavalieri et al. . A merged SMMR-SSM/I time series product using this algorithm from NSIDC provides the record from 1979 to 2004 [Fetterer and Knowles, 2002]. Figure 1 shows ice concentration anomalies for September 2002, 2003 and 2004, together with the median ice extent calculated over 1979–2000. There are strong similarities between the last three Septembers, in particular, a pronounced retreat of the ice margin from the shores of Siberia and Alaska. While this feature, albeit generally less well expressed, is common to other extreme minimum ice years (1990, 1993, 1995, 1998), the past three Septembers have also seen an unusual lack of ice in the Greenland Sea, a feature not observed at any other time during the satellite era.
Figure 2 shows the time series of September ice extent since 1979. The September trend is (−0.54 ± 0.11) × 106 km2 per decade at a 99% confidence interval (or −7.7% ± 3% per decade). This compares to the trend calculated by Comiso  of −6.4% per decade using data through 2000. It is clear that inclusion of the last three years has strongly reinforced the downward trend. Note also that the higher variability from the 1980s to the mid 1990s reported by Comiso  has subsequently decreased. The longest available sea ice data set, known as HadISST1 [Rayner et al., 2003] extends from 1870–2003. While records prior to the early 1950s must be interpreted with caution due to sparse data, the time series (not shown) indicates that extent in recent years is well below normal compared to any previous period. At no other time in this record have there been three consecutive years with September ice extent that is consistently as low as that seen in 2002–2004. Based on Monte-Carlo simulations randomizing the shorter, but more reliable 26 year passive microwave time series for September, the chance of having a three-year run with substantially below normal sea ice as observed in 2002–2004 is 0.8%.
Table 1 summarizes ice extent and area for the last three Septembers, the extreme low ice years of 1990, 1995, 1998, 1999 and 2000 and for the 1979–2000 mean. The total Arctic ice extent in September of 2003 and 2004 is similar to that for 1995. In the 26 years for which we have satellite data, September ice extent has been outside one standard deviation (using 1979–2000 as a base period) nine times: in 1980 and 1996 when ice extent was high, in 1990, 1995, 1999 and 2000, when extent was low, and for the last three years.
Table 1. Arctic Sea Ice Extents (×106 km2) for September for the Six Minimum Years since 1979 and (in Parentheses) Number of Standard Deviations From the 1979–2000 Meana
Total Arctic Ice Extent
Total Arctic Ice-Covered Area
Computation of the ice area excludes the region near the pole not imaged by the sensor. This area is larger for SMMR than for SSM/I, consequently the area for SMMR was used to remain consistent between sensors.
1979–2000 mean (std dev)
3. Factors Contributing to the Decrease in Ice Cover
 Why has there been an overall downward trend in Arctic sea ice, and, perhaps more importantly, why have the last three Septembers shown such extreme low ice cover?
 Warming in the Arctic has been widely examined using analysis of surface air temperature (SAT). Although computed trends are quite sensitive to the time period examined, the data source, and how the data are analyzed, it is clear that sub-arctic land areas have significantly warmed since about 1970 [Serreze et al., 2000]. The largest warming is observed during winter and spring. Information for Arctic Ocean SAT comes primarily from Russian North Pole measurements (1950–1991), collected at a series of manned drifting camps, arrays of drifting buoys maintained since 1979 by the International Arctic Buoy Program (IABP) and various satellite retrievals (since the early 1980s). From North Pole records, Martin et al.  calculate a significant increase in Arctic Ocean SAT during May and June between 1961–1990. Rigor et al.  examined the period 1979–1997 via fields that combine land station records with data from North Pole stations and IABP buoys. Positive ocean trends are most pronounced and widespread during spring. Comiso  examined trends from 1981 onwards based on retrievals from Advanced Very High Resolution Radiometer (AVHRR) data. An update through 2003, with improved retrieval algorithms (presented at AGU annual meeting, December 2004), shows Arctic Ocean warming in spring and autumn. While these studies are certainly not in complete agreement, they point to an earlier melt onset in spring and lengthening of the melt season, favoring less total ice cover at summer's end. Comiso  calculates an increase in the melt season of 10–17 days per decade. Further support for extended melt comes from Belchansky et al.  based on passive microwave satellite retrievals.
 The sea ice cover can be broadly divided into the perennial pack in the central Arctic Ocean and the surrounding seasonal ice zone. Seasonal ice melts completely in summer at rates that can reach 4 cm per day [Maykut, 1986]. In the perennial ice zone, the ice does not completely melt. That which remains at summer's end thickens through the following autumn and winter. With every subsequent year, the perennial ice follows a growth/ablation cycle until an equilibrium thickness is reached where melt season ablation equals winter ice growth. Unlike thinner ice, for which growth or ablation rates respond quickly to changes in SAT, thicker, snow-covered multiyear ice of the perennial zone responds more slowly, ablating by only about 0.5 m over a season.
 Although the concepts of perennial and seasonal ice zones, and of equilibrium thickness, are simplifications that ignore ice dynamics, they are useful in helping to understand the response of sea ice to thermal forcing. Climatic conditions have not been constant, and we expect that the extended ice melt season discussed above has reduced the equilibrium thickness. During the last three minimum ice years, temperatures at 925 hPa (assessed from NCEP/NCAR reanalysis data) were 1–2°C above normal over much of the Arctic, with particularly high temperatures along the Siberian and Alaskan coasts and the East Greenland Sea, pointing to less ice growth in autumn and winter, and more summer melt.
 Atmospheric variability has certainly played a role. From about 1970 through the mid 1990s, there was a general upward tendency in the Arctic Oscillation (AO) index during winter. The AO is a primary mode of atmospheric variability in the northern hemisphere and has pronounced influences in the Arctic. Rigor et al.  demonstrate that as the winter AO became more positive, altered wind fields fostered more offshore ice motion and ice divergence along the Siberian and Alaskan coastal sectors, leading to an anomalous coverage of thin, first-year ice in spring. Thinner ice requires less energy to melt in summer and promotes larger heat fluxes to the atmosphere in spring, leading to increased spring SATs. Since thin, first-year ice has a relatively high salinity, the freezing point is depressed, meaning earlier spring melt. Both processes favor less ice at summer's end.
 The Rigor et al.  model does not neatly explain the extreme September minima of the past three years. In recent years, the winter AO has retreated toward a more neutral state. Serreze et al.  argue that the extreme ice losses of 2002 can be explained in part by the strongly cyclonic circulation of the atmosphere during summer of that year. When the atmospheric circulation (hence sea ice circulation) is cyclonic, the ice cover tends to diverge, resulting in more open water. While by itself, ice divergence spreads the existing ice over a larger area, Serreze et al.  suggest that enhanced absorption of solar energy in the resulting dark open water areas between ice floes led to stronger melt, reducing ice extent. Furthermore, cyclonic atmospheric circulation led to a weakening, or reversal of the wind-driven component of the sea ice flux out of the Arctic through Fram Strait (between Svalbard and Greenland), helping to explain the lack of ice in the Greenland Sea. Similarly strong cyclonic conditions were observed in summer of 2003, but not in 2004.
Rigor and Wallace  give a different view. The winter AO was in an especially positive state from about 1989–1995. Their basic argument is that during this period, the altered wind field acted to “flush” much of the thick multiyear ice out of the Arctic through Fram Strait, leaving more thin ice. During the summers of 2002 and 2003, this thinner ice circulated into Alaskan coastal waters where very pronounced melt occurred. Fowler et al.'s  calculation of ice age in the Arctic Basin similarly suggests that most of the perennial ice pack has shifted from ice older than three years to ice that survives only two melt seasons or less. In summary, the sea ice system may have retained a memory of the previous high AO state.
 A common feature of climate model projections is that largely due to albedo feedbacks involving snow and sea ice, the effects of greenhouse gas (GHG) loading will be observed first and will be most pronounced in the Arctic [Holland and Bitz, 2003]. Sea ice loss is a near universal feature of these projections. While the rate of decline varies widely between models, some predict a complete loss of summer ice cover by about 2070 [Arctic Council and IASC, 2004]. At face value, the observed trend in Arctic sea ice is consistent with climate model projections for the 21st century, and there certainly seems to be some direct response of the sea ice to recent warming, seen in the increasing length of the melt season [Comiso, 2003]. However, this view must be tempered by recognition that decadal and multi-decadal climate variability is especially pronounced in the Arctic. Available records, albeit primarily from inland and coastal stations, document a rise in Arctic SATs from about 1920 to 1940 that was just as large as the recent warming, followed by a period of cooling. While the recent warming seems to be quite different, appearing as part of a more coherent global, rather than regional signal [Arctic Council and IASC, 2004], such low-frequency variability can make it very difficult to separate natural climate fluctuations from those due to GHG loading [Polyakov and Johnson, 2000; Overland et al., 2005].
 To further complicate the picture, Rigor et al. , Rigor and Wallace  as well as other studies [e.g., Maslanik et al., 1998; Holloway and Sou, 2002] point to the influences of individual anomalies and longer term changes in the atmospheric circulation as they influence the sea ice circulation, thickness and/or thermodynamic state. Maslowski et al.  also argue that an enhanced inflow of warm, Atlantic-derived waters at depth has promoted a stronger upward oceanic heat flux, discouraging ice growth. Subsequent work [Maslowski et al., 2001] suggests inflow of warm Pacific waters through Bering Strait. As a general statement, it appears that dynamic as well as thermodynamic forcing, over a period of years, have favored a relative increase in the areal extent of thin seasonal ice, favoring low September ice extent.
 A key part of the puzzle regards the low-frequency variability in the AO that seems strongly linked with the sea ice decline. Multi-year to decadal variability is a fundamental characteristic of internal atmospheric dynamics, with lower-frequency variability involving ocean coupling. The recent decline in the AO is consistent with natural variability. However, some modeling evidence suggests that external forcings may favor a generally more positive AO state. Cooling of the stratosphere, in response to GHG loading or losses of stratospheric ozone, can lead to “spin up” of the polar stratospheric vortex, favoring a positive shift in the AO [e.g., Gillett et al., 2003]. Another idea, [Hurrell et al., 2004; Hoerling et al., 2004], is that a slow increase in sea surface temperatures in the tropical Indian Ocean has helped to “bump” the AO toward a preferred positive state. They further argue that the ocean warming may itself represent a non-linear response to GHG loading.
 It is increasingly difficult to argue against the notion that the effects of greenhouse gas loading will eventually be clearly expressed in the Arctic. Are we already seeing these effects on the sea ice cover or is natural variability still the dominant driver? Could the extreme ice minima of 2002–2004 represent the crossing of a threshold - that because of general warming, the spring ice cover has now thinned to the point that large areas of the pack ice no longer survive the summer melt period? While these are intriguing questions, it may well be that they can only be answered by the passage of time.
 This study was supported by NASA contracts NNG04GO51G and NNG04GH04G and NSF grants OPP-0242125, OPP-0229651. We thank the reviewers for their constructive comments.