Published

Data choices — arctic-sea-ice story

Rufus Pollock
Rufus Pollock

Arctic sea ice shrank to 4.60 million km² on 12 September 2026, tied for 10th lowest in 48 years of satellite records, according to NSIDC. All 20 of the lowest minimums came in 2007–2026; the winter maximum, 14.29 million km² in 2026, tied with 2025's for the lowest.

Data choices — arctic-sea-ice story

Source. NSIDC Sea Ice Index, Version 4 (G02135), Northern Hemisphere: the daily extent file for every charted and quoted number, the daily 1981–2010 climatology for the 2026 season chart, and three NSIDC files used only as cross-checks (the annual minimum/maximum rankings workbook and the September and March monthly means). It is the index NSIDC uses for its annual minimum and maximum announcements. Not used: the EUMETSAT OSI SAF sea ice index, quoted once via the Met Office as an attributed alternative ranking.

Annual minimum and maximum. Following NSIDC, each day's extent is averaged with the four days before it (a 5-day trailing mean) and a calendar year's minimum and maximum are the lowest and highest of those means. Years 1979–2026; 1978 starts on 26 October and has neither. The daily file runs to 6 November 2026, more than 50 days after the 2026 minimum on 12 September (calculate.mjs refuses to run if the minimum is within 30 days of the end of the file).

Gaps. Until 20 August 1987 the satellite record is every other day, and there is a gap from 3 December 1987 to 12 January 1988. The 5-day mean averages whichever daily values fall in the window (two or three in the every-other-day years). NSIDC evidently handles those early years differently: compared with NSIDC's own table of annual 5-day minima and maxima (nsidc-rankings-nh-5day.csv, extracted from the rankings workbook), values match to within 0.001 million km² from 1988 to 2025 but differ by up to 0.055 in 1979–1987. The 1979–2006 minimum trend is −0.54 per decade on either set of values, and the 20 lowest minima to 2025 are the same years. Missing days are never filled or plotted as zero.

Trends. Ordinary least squares on annual values, reported in million km² per decade with a two-sided 95% confidence interval (t distribution, n − 2 degrees of freedom, critical values hard-coded in calculate.mjs). This ignores autocorrelation, which can change the intervals; it is a description, not a test of a forced trend. The break year 2007 was chosen because it is where the record shows the step: every minimum from 2007 is below every minimum before it. Choosing a break from the data biases any before/after comparison, so the post reports the trend from other start years and the full-record trend too, and attributes the significance claim about the pause to a peer-reviewed study (England et al. 2025), whose break year (2005) and measure (September monthly extent, with area also analysed) differ. The 2007 start gives the flattest trend of the start years 2004–2008 (−0.36, −0.26, −0.20, −0.03 and −0.09 per decade to 2026); from 2002, before the step, it is −0.49 and significant; over 1979–2026 it is −0.73.

Instrument change. From 1 January 2025, Version 4 uses AMSR2 input; earlier days are unchanged from Version 3. NSIDC's user notice (23 July 2025) says AMSR2 places the ice edge slightly inward, giving Arctic extents smaller by "less than about 0.1 million sq km", and that intercalibration is preliminary. NSIDC's Special Report 28 (July 2025) is more precise: for 1 January–30 June 2025, Arctic differences averaged 0.05 million km², with maximum differences below 0.2, and about 80% of daily differences were negative (AMSR2 lower). Differences vary by day and sign, so there is no simple correction. The post adds 0.1 million km² to both AMSR2 years (2025 and 2026) only as an illustrative sensitivity test, not a correction or uncertainty bound: 2026 would then be 14th by numerical ordering. The data is not adjusted.

Ties. NSIDC treats extents within 0.04 million km² (40,000 km²) as ties. statistics.json lists the years within that margin of 2026: for the minimum, 2008, 2010 and 2025 (tied 10th, as NSIDC announced); for the maximum, 2025 (tied lowest).

Monthly cross-check. NSIDC's September monthly mean extent gives the same summer picture as the 5-day annual minimum: −0.55 million km² per decade in 1979–2006, −0.03 in 2007–2026; 2026's September mean, 4.81, ranks 13th lowest. The March monthly means order the two recent winters the other way round from the 5-day maxima: March 2025 is the lowest March, March 2026 the second. The 1979–2006 and 2007–2026 March trends are −0.42 and −0.33 per decade.

Climatology chart. The 1981–2010 file is by day of year, 1 to 366. For 2026, not a leap year, day of year n is matched to day n of the climatology, so dates after 28 February are compared with values about one day later in the calendar for leap years in 1981–2010. The effect on a daily median is negligible against the 1.83 million km² gap quoted.

Size. All source files are archived here: about 2 MB in total, 1.9 MB of it the daily file.