The whole continent's sea ice

Name: Antarctic sea ice extent and area, daily, 1979–2024
Creator: NOAA/NSIDC

16,730 daily readings · 1979–2024 · NOAA/NSIDC CDR v4 · every trend on both gates

Everything published about Antarctic sea ice is a trend, and the trend is the least defensible thing in the record. Fitted across all 46 years, extent falls 0.06 ± 0.06 million km² per decade — 1.1 times its own standard error, which is nothing. Fitted from 1988, where the record becomes one instrument, it falls 0.16 ± 0.08, 1.9σ, which does not clear either. The shape is what matters: the ice rose to its highest year in 2014 and its lowest in 2023, nine years apart, and a straight line drawn through that says nothing at all. So this page does not draw one as its answer. It asks each of the 366 calendar dates in the year which of the 46 years holds its record — no fitting, no chosen period — and the answer is that 345 of the 366 lows were set since 2016 and 361 of the 366 highs between 2007 and 2015.

253 of 366

dates whose record low is a day in 2023

no other year in the record holds more than 38 of them. Drop the pre-1988 sensor entirely and it is 266.

345 and 361

record lows since 2016 · record highs 2007–2015

the record's extremes are not spread across 46 years. They are stacked in two adjacent windows, and nothing is fitted to say so.

1.1σ and 1.9σ

the two fitted trends, 1979–2024 and 1988–2024

-0.063 ± 0.057 and -0.161 ± 0.085 million km² per decade. Neither clears twice its own standard error, which is this site's bar — and that is the finding, not a gap in it.

Every day of the record on one calendar. The grey band is the full range of 1979–2015 — not a percentile, the actual highest and lowest reading each date saw in those 37 years — with their median through it. The 9 lines are every year since. 2023 sits below the entire band on 317 of its 365 days; that last figure is the one number on this page that depends on where the band is cut, and it is quoted here because the cut is drawn right beside it. The record's highest single day is 20.54 million km² on 2014-09-20 and its lowest is 1.80 on 2023-02-15.

The calendar changed hands

Ask a different question and no fitting is needed. For each of the 366 calendar dates in the year, which of the 46 years holds its highest reading, and which its lowest? Only 7 years hold a record low at all and only 7 hold a record high — and no year holds one of each. The highs belong to 2007–2015 and the lows to 2016 onwards, two adjacent windows out of 46 years, with 345 of the 366 lows and 361 of the 366 highs inside them.

Both gates say the same thing and the stricter one says it harder. Drop the pre-1988 sensor entirely and every one of the 366 record lows belongs to 2016 onwards and every one of the 366 record highs to 2007–2015 — 366 and 366 exactly. The 21 lows the full record puts before 2016 are the only ones this gate removes, and 9 of them fall on days the producer interpolated rather than observed. Of the 345 modern ones, 0 do.

Each block is a run of consecutive calendar dates held by one year. Filled where the holder is inside its window, hollow everywhere else — the hollow blocks are the whole of the rest of the record. 2023 holds the record low on 253 of the 366 dates in the year, and its longest unbroken stretch runs 1 May to 24 August — 116 consecutive calendar dates on which no other year in the record has ever been lower.

Year	Record lows held	Record highs held	Which end
2023	253	0	record low
2014	0	138	record high
2015	0	81	record high
2008	0	61	record high
2016	38	0	record low
2013	0	37	record high
2017	36	0	record low
2010	0	28	record high
1980	19	0	record low
2007	0	16	record high
2024	11	0	record low
2022	7	0	record low
1985	0	5	record high
1984	2	0	record low

There is no trend, and that is the finding

Both fits are drawn because this site publishes both gates on every series. Over 1979–2024 extent falls -0.063 ± 0.057 million km² per decade — 1.1 times its own standard error, so it does not clear the bar and cannot be quoted as a decline. Over 1988–2024, the period on which the archive is one instrument rather than two, it falls -0.161 ± 0.085, 1.9σ, which does not clear it either. Until 14 August 2026 this page printed that second figure as -0.171 ± 0.085, 2.02σ, and called it a trend that clears while arguing that nothing here may be quoted. It was wrong for a reason that had nothing to do with the ice, and the section below is what moved it. The argument this page makes is the same one and it no longer rests on two gates disagreeing: neither gate clears, which is why the lead figure here is a count rather than a slope.

Look at what the lines are drawn through. The five lowest years of the 46 are 2023, 2024, 2017, 2022 and 2019 and the four highest are 2014, 2013, 2015 and 2008 — every one of the lows since 2017 and every one of the highs between 2008 and 2015. Mean extent stood at 12.99 million km² in its highest year and 10.12 in its lowest, 22.1% apart, nine years later. A line fitted through a rise and then a fall reports the difference between two accidents of where the record starts and stops.

Annual mean extent, 1979–2024, with both fitted lines. The shaded years at the left are the every-other-day Nimbus-7 SMMR era, which the second fit excludes. The eight hollow markers are what the raw annual mean said at the eight years missing a block of days.

An incomplete year is not a shorter year

8 of these 46 years are missing days, and the missing days are not scattered through the year — they are contiguous blocks. 1987 loses 3 Dec to 31 Dec, 29 days of it. 1988 loses 1 Jan to 13 Jan. 1984 loses 12 Aug to 24 Aug. Until 14 August 2026 this record's annual mean was taken over whatever days were present.

Antarctic sea ice runs from about 3 million km² around 20 Feb to 19 around 22 Sep, so an average over the days that happen to be there is an average over a different part of the seasonal cycle — not a noisier estimate of the same one. Drop the 13 days 1988 is missing and 13 of the year's lowest readings have gone with them, and the answer comes out too high. It reads +0.177 million km² high in 1988 and -0.255 low in 1984, each of them larger than the decadal trend the whole record is fitted to, and both sit at the start of the 1988 window where a straight line is pulled hardest.

1988's whole year. The 13 missing days are the shaded block at the left; the orange line through it is what the correction puts there — the day-of-year average of the 38 complete years, shifted by what this particular year was running at either side of the gap. The two dashed rules are the two annual means. Nothing in the pixels marks the gap; it is simply 13 readings that were never taken, at the flattest, lowest part of the year.

The estimate was scored against a known answer

An estimate that cannot be checked is a guess with arithmetic on it. So each of the eight real gaps was cut out of each of the 38 complete years in turn — the climatology refitted without that year each time — and both methods asked to put back the annual mean that had just been destroyed. At the worst block, 1984's 13 missing days of August, the median error goes from 0.234 million km² to 0.004.

Every one of the eight gap blocks, scored on 38 years where the answer was known. The correction wins at all eight — and note that the damage is not about the length of the gap. The longest block, 1987's 29 missing days, costs just over half what 1984's 13 do. In the average year those 29 days sit at 10.3 million km² against the year's own mean of 11.8, and the 13 shorter ones at 18.1. What a gap costs is how far its days stand from the annual mean, times how many of them there are — which is where in the season they fall, and is the whole finding.

What it costs this site: two published verdicts

Both move the same way, and both make this page's existing argument stronger. The 1988-onward trend in extent goes from just over this site's bar to just under it, and the ice area does the same. This page already refused to quote a trend, on the grounds that its two gates disagreed with each other. Now neither of them clears, and the refusal stops being a technicality. Both numbers are printed because the raw one is what this site published before today.

Series	Window	As published before		Composition-corrected		Verdict
extent	1979–2024	-0.063 ± 0.058	1.09σ	-0.063 ± 0.057	1.11σ	unchanged
extent	1988–2024	-0.171 ± 0.085	2.02σ	-0.161 ± 0.085	1.90σ	verdict moves
area	1979–2024	-0.040 ± 0.054	0.74σ	-0.038 ± 0.053	0.72σ	unchanged
area	1988–2024	-0.156 ± 0.077	2.04σ	-0.146 ± 0.077	1.91σ	verdict moves
compactness	1979–2024	+0.00105 ± 0.00066	1.60σ	+0.00122 ± 0.00065	1.88σ	unchanged
compactness	1988–2024	-0.00115 ± 0.00081	1.42σ	-0.00101 ± 0.00082	1.23σ	unchanged

This is not the interpolated days below, and the distinction is the whole point. A fill puts a value in the slot; a straight line drawn between two readings leaves an average exactly alone, which was measured and is why the section under this one changes no number here. A gap removes the slot, and what it costs depends entirely on where in the season it falls. Same archive, same years, opposite exposure. The eight years are listed with their gaps in the table at the foot of this page.

Asked blind whether it breaks, and it will not say

Everything above is an argument that a straight line is the wrong shape for this record. Here is the same argument made by something that could have disagreed. Each of the 14 annual series here — extent, ice area and compactness, at the mean, the peak and the minimum, on both gates — was asked to nominate the year it changed, against two different alternatives to a line: a jump, and a bend. It was told nothing about satellites, icebergs or the calendar. 28 tests in all.

Not one of the 28 clears the threshold for having asked 28 times. That is the verdict and it is not a technicality: with 46 annual numbers there is not enough in this record to establish a break, whichever shape you look for. Nothing on this site may say the Antarctic sea-ice record breaks in any year.

What is worth publishing is what happens underneath the bar. 22 of the 28 fire at an ordinary uncorrected 5%, and they are unanimous about where. Every one of the 8 jump tests that fires picks 2015 or 2016. Every one of the 14 bend tests picks 2012 to 2014 — four years earlier, from a window 30 years wide that both were free to search. And every single bend test reverses direction there: mean extent runs +0.19 million km² per decade before 2014 and -1.74 after. A rise, then a fall, which is exactly the shape the two figures above are drawn to show and exactly the shape a straight line destroys.

Every test as one dot, at the year it nominated. The grey rules under each lane are the years it was free to choose from — a break within eight fitted years of either end cannot be found, so the search window is narrower than the record, and it still leaves three decades nothing chose. The two lanes pile up four years apart, and a record with a real turning point in it gets found in the same place whichever way you look for it.

Two things this test is not allowed to be turned into. It is not evidence that the record is fine — 22 of 28 firing at 5% is a great deal of smoke, and the limit here is how many years exist, not how big the change is. And it is not a licence to quote 2015 or 2016 as the year Antarctic sea ice changed, however often that year appears above. Both alternatives were corrected separately and the joint threshold is published beside every row; where both fire, neither is preferred. Picking the larger of two statistics after looking at them is not a test.

The scan was run on the composition-corrected annual means from the section above, and then run again on the raw ones as a check that the correction is not what produces the answer. It is not: the two agree on the nominated year at 11 of the 12 tests re-run, the one disagreement is a single year at compactness, and nothing clears the threshold on either version. And the same question cannot be asked of this site's ice velocities at all: the longest fitted series there is 13 years against the 25 this test needs, so not one of the 22 glacier series has a single candidate year to test. A break test on ten points is not a weaker test — it is not a test. The scan of everything else on this site →

How much of this record was never observed

The archive publishes a value for every day of every year and roughly half the early ones were never measured. Nimbus-7 SMMR ran every other day until 1987; the producer fills the gaps cell by cell and nothing in the pixels says which days those are. There are 1,617 such days in the record, 1,555 of them before 1988 — and the only witness is a flag stored beside the data. They read +0.033 million km² high against their own neighbours.

A second fault sits in the same years and the same cut-off removes both. The pre-1988 sensor reads several points low on concentration, which is why the 1988 gate exists at all. It removes 92% of the interpolated days as a side effect. Days that could not see 90% of the consolidated winter pack are dropped rather than patched, which is why four years in the middle of the 1980s are visibly short below.

Every bar is one year against a full 366 days. The orange portion is the part the producer interpolated rather than observed; the shortfall at the top of a bar is days dropped for insufficient coverage.

The count above was attacked on purpose, and it held

A record is an extreme, and averaging destroys extremes. Half of every day before 1988 in this archive was never observed, so the early years have had their extremes shaved off and the modern ones have not — and any bias from that runs in exactly the direction of the finding above, making the recent record lows look more dominant than they are. That is the strongest objection to this page, and it deserves a measurement rather than a reassurance.

So: take a modern year, where the archive really is daily; throw away every other day of it; fill the gaps the way the producer does; recount the whole calendar. Repeat for every modern year at both parities — 44 attempts. The decimated years keep 100.7% of their record lows and 99.0% of their record highs, and 0.39 of the 732 calendar slots change hands per attempt. The filling displaces a day by 0.023 million km² against a between-year spread of 0.656 on the same date — 3.6% of the gap that decides the contest.

The reason is which way the damage points. Smoothing blurs a series along time. A record for a given calendar date is a contest between years at one fixed moment, which is the other axis. It is the same shape of argument that protects the front page's rank from the pre-1988 sensor: ask which kind of number you are holding before deciding whether a known fault reaches it.

What smoothing does reach

Every figure below was measured against a known answer — the modern record smoothed, then compared with the truth that smoothing had just destroyed, over the 18 ice years that are complete and carry no filled day.

the number is	what temporal smoothing does to it
a crossing date	moves it, by the smoother's own error divided by how steeply the ice is falling when it crosses
an average	nothing, and not approximately — zero to five decimal places, as a straight line drawn between two readings must be
a peak	clips it, one way only: 0.0093 million km² off a 19.1 million km² maximum, 0.049%
a rank between years at a fixed date	3.6% of the spread that decides it

Left: how far each smoother moves the day the ice falls back through a threshold, against how wide that smoother is. The lines do not sit on top of one another — a date read off a flat part of the curve is hurt several times more than one read off a steep part. Right: the same numbers divided by the crossing slope, which is what the law says should account for that difference. It largely does, and not perfectly: the spread across thresholds narrows at four of the five smoothers, from 6.9× to 2.2× at a monthly mean where the errors are biggest, and widens at the 3-day mean where it was already smallest. A real law with real scatter rather than a clean collapse. Both axes logarithmic.

So the archive's own filling is the smallest member of a family and a monthly average is the largest. The filling costs a date at most 0.09 days. A monthly average costs 1.6 at the steepest crossing tested and 9.2 at the flattest — and at the flattest it does not merely move the date: 25% of the crossings stop existing. The damage scales with the width of the smoother at an exponent of 1.12, r² 0.97, from one day to a month, which is what makes these one fault at five sizes rather than five faults. The standing rule it leaves: never read a date off a monthly product. This site reads monthly data in exactly one place — the sea ice around each emperor colony — and reads only amounts from it, which the table above says is the one statistic smoothing leaves alone. Every date on this site comes from the daily record instead.

Extent and area are not two measurements wearing one name

Extent counts every stretch of sea holding at least 15% ice; area counts the ice itself. Elsewhere on this site two measurements sharing a name turned out to mean opposite things — grounded ice against its floating shelf — so the same trap was looked for here. It is not there. Their ratio, the mean concentration inside the ice edge, trends +0.00122 ± 0.00065 per decade over the full record and -0.00101 ± 0.00082 from 1988: neither clears its own error bars and the two disagree in sign. The ice inside the edge is no more and no less tightly packed than it was, and the front page quoting extent without saying so is not a defect.

Ice area divided by ice extent, by year, with both fitted lines. A published null result, drawn.

What this record is not

The record ends 2024-12-31. It is a climate record, not a feed.

The live figure at the top of the wire comes from a different product — the NSIDC Sea Ice Index — and is current to within about two days. This one is a gridded climate data record: it ends 2024-12-31, and it reads about 2% above the Index because it uses a different land mask and a different area grid. The two must never be spliced or plotted on one axis, and nothing on this site does.

The same archive, colony by colony

This is the continent behind every emperor colony page on this site. The same passive-microwave record, read at 25 km around 67 breeding sites rather than summed over a hemisphere, gives three different measurements — how much sea ice a colony has, when that ice goes, and how long it lasts — and the continental null result above is exactly why those are worth having. A hemisphere with no trend is made of coasts that are losing ice and coasts that are gaining it.

The ice around the emperors

how much · 67 colonies mapped

When the ice goes

the retreat date · 61 colonies

How long the ice lasts

season length, decomposed · 61 colonies

All 46 years

The raw annual mean is the average of the days that passed the coverage gate; the corrected one is the average over a fixed 365-slot calendar with the missing block estimated, and area and compactness are the corrected figures too. Rank is over corrected extent, 1 being the lowest of the 46. 8 of the 46 years carry a gap and the last column says where in the season it falls.

Year	Extent, raw	Extent, corrected	Area	Compactness	Rank	Days used	Interpolated	Missing block
1979	11.898	11.898	9.928	0.834	28	365	181
1980	11.434	11.458	9.548	0.833	9	366	183
1981	11.680	11.680	9.770	0.836	15	365	182
1982	11.876	11.876	9.982	0.840	26	365	188
1983	11.658	11.658	9.715	0.833	13	365	182
1984	11.493	11.748	9.832	0.837	21	353	176	12 Aug – 24 Aug
1985	11.811	11.886	10.022	0.843	27	360	181	5 Aug – 9 Aug
1986	11.400	11.412	9.511	0.833	7	358	182	4 Dec – 10 Dec
1987	11.853	11.742	9.843	0.838	20	336	100	3 Dec – 31 Dec
1988	12.027	11.849	9.974	0.842	25	353	8	1 Jan – 13 Jan
1989	11.665	11.665	9.875	0.847	14	365	4
1990	11.658	11.649	9.892	0.849	12	364	12	26 Dec – 26 Dec
1991	11.808	11.808	9.950	0.843	23	365	0
1992	11.665	11.690	9.820	0.840	18	366	4
1993	11.687	11.687	9.899	0.847	16	365	1
1994	12.035	12.035	10.244	0.851	34	365	3
1995	12.060	12.060	10.168	0.843	37	365	11
1996	12.017	12.041	10.135	0.842	35	366	1
1997	11.646	11.646	9.853	0.846	11	365	0
1998	11.985	11.985	10.110	0.844	32	365	2
1999	12.026	12.026	10.174	0.846	33	365	0
2000	12.021	12.045	10.242	0.850	36	366	1
2001	11.924	11.924	10.020	0.840	29	365	0
2002	11.488	11.488	9.723	0.846	10	365	0
2003	12.203	12.203	10.328	0.846	38	365	0
2004	12.205	12.229	10.352	0.847	39	366	0
2005	11.942	11.942	10.060	0.842	31	365	0
2006	11.689	11.689	9.899	0.847	17	365	0
2007	11.925	11.925	10.166	0.852	30	365	0
2008	12.499	12.487	10.557	0.845	43	364	6	24 Mar – 25 Mar
2009	12.267	12.267	10.415	0.849	41	365	0
2010	12.305	12.305	10.388	0.844	42	365	1
2011	11.718	11.718	9.963	0.850	19	365	0
2012	12.212	12.236	10.417	0.851	40	366	0
2013	12.721	12.721	10.916	0.858	45	365	0
2014	12.986	12.986	11.071	0.853	46	365	0
2015	12.626	12.626	10.634	0.842	44	365	1
2016	11.421	11.444	9.542	0.834	8	366	0
2017	10.925	10.925	9.198	0.842	3	365	0
2018	11.205	11.205	9.406	0.839	6	365	0
2019	11.062	11.062	9.413	0.851	5	365	0
2020	11.795	11.820	9.994	0.846	24	366	0
2021	11.804	11.804	9.878	0.837	22	365	2
2022	10.928	10.928	9.151	0.837	4	365	0
2023	10.119	10.119	8.443	0.834	1	365	0
2024	10.621	10.681	8.974	0.840	2	364	5	17 Sep – 18 Sep

Method

Extent is the geodesic area of every cell at or above 15% concentration on the Hughes 1980 ellipsoid; area is the same cells weighted by concentration. Every trend is fitted twice, over the full record and over 1988 onwards alone, and is called robust only where both clear twice their own standard error with the same sign.

Source: NOAA/NSIDC Climate Data Record of Passive Microwave Sea Ice Concentration v4, southern hemisphere, daily, annual aggregates. Read from https://noaa-cdr-sea-ice-concentration-pds.s3.amazonaws.com/data/final/south/aggregate/ — one file per year, 16,730 daily grids surviving the coverage gate, 1979-01-01 to 2024-12-31. Cell areas are computed on the Hughes 1980 ellipsoid and run from about 444 to 664 km² across the grid, because assuming a flat 625 would be 6% wrong. Nothing on this page is hard-coded: every figure above is counted at build time from the daily record, and the two fitted slopes are recomputed here and checked against the ones the pipeline published before the page will build at all.