729 sites, 56 places

The species with the most sites have the fewest places

This is the atlas's finding arriving as a measurement error bar rather than as a map. Four of the six penguins here never leave one twelfth of the coastline — and one twelfth of the coastline does not hold many independent patches of sea. The chinstrap has 369 known breeding sites and stands in eight places. The emperor, which breeds right round the continent, has 67 sites and stands in 45. Every count of sites on this site now carries its thinned count beside it, and for the four rock-nesting species this archive may be published per place and never per site: anything counted across their sites is a number about the South Shetlands, counted several hundred times.

Where the collapse happens

The frame below is the window this site already draws on the atlas — the three fullest ten-degree longitude sectors, 70°W to 40°W, fixed before anything was counted. 509 of the 728 measurable sites are inside it, and it is where the thinning does nearly all of its work. The 453 sites holding at least one of the four rock-nesting penguins stand in eight independent places; seven of those eight are in this frame, and the other one is Chinstrap Islet, in Victoria Land on the far side of the continent. Nothing here is dropped for being off the map — every site is counted in every figure on this page.

What has happened to their ice, said carefully

Measured inside the homogeneous era at both ends — the first decade after the 1988 sensor change against the last decade in the archive, so the instrument that reads low is nowhere near it — the sea-ice season at the typical rock-nesting site went from 112.9 days a year to 74.0. At the emperor's and the Adélie's sites the same two decades give 339.9 against 335.5 and 294.1 against 307.2: nothing.

And the place-weighted size of that loss cannot be quoted at all. The thinning is greedy: it keeps whichever site it reaches first, so which eight places stand for the Scotia arc is arbitrary even though the rule is not. Shuffle the order 200 times and the chinstrap change runs anywhere from -39.0 to -6.6 days. What survives the arbitrariness is the count of places and the direction — a loss in 200 of 200 shuffles at four of the four rock-nesters, and a result that crosses zero for the emperor and the Adélie. Quote the per-site figure as what it is, the typical breeding site, or quote the direction. Never quote a size off a thinned list.

A hundred and four falling sites are three places

On the gate this archive publishes — a slope that clears twice its own error bars, keeps its sign when refitted from 1988, and still clears them when a level shift at 2010 is allowed for — the ice season is getting shorter at 104 of the 453 rock-nester sites and longer at none. Drawn as 104 falling dots that would look like a continental result. Thinned, those 104 sites are three independent places, and the page has to say so in the same breath as the count.

The emperor reaches the same geography from the other direction, which is worth setting beside this because it is the species the archive can say most about. Seven of its 67 colonies carry a publishable ice-season trend, five of them falling — and the steepest of those is Emperor Island, the one emperor colony known to have been lost. How long the ice lasts at 61 emperor colonies →

The blind check found something bigger than the thing it was checking

A measurement nobody has run before needs a test it cannot pass by accident. In February 2010 iceberg B09B knocked the Mertz Glacier Tongue off the coast of George V Land; that December the same berg grounded in Commonwealth Bay, the bay froze over for good, and the Adélie colony at Cape Denison collapsed. So: score every one of the 728 sites, blind, on the change in summer sea ice from the 2001–10 decade to the 2011–20 one, in units of its own year-to-year spread, and see where a known catastrophe lands.

Cape Denison comes out 1st of 728. Summer sea ice around it went from 4.0% to 53.5% — 9.2 times its own variability, and the nine nearest sites fill the rest of the top ten.

Which turns into a warning rather than a headline. That straight line is not a description of this series. Allow the fit a single level shift in 2010 and the slope goes to -5.1 ± 5.9 points per decade and stops clearing its own error bars, with the shift itself worth +50.0 points. This site already refuses to fit a trend across the 1988 satellite change and the 2013 Landsat arrival; this is a third break of the same kind, it is larger than either, and it is real ice rather than an instrument — so nothing in the data announces it.

And it lands on the right side of a boundary this site drew last week rather than breaking it. A 25 km satellite product can see a stretch of coast change and cannot see one bay. What is detected here is not the fast ice inside Commonwealth Bay — one bay, invisible at this resolution — it is the regional summer ice along 200 km of coast, which is what calving a 78 km glacier tongue did to it. The colony died of the bay; the measurement is entitled to the coast, and it is worded that way. There is no penguin count here and there will not be one.

The summer window at Cape Denison holds a median of 11 days of ice a year across the whole record, which is under the 30 this archive requires before it will fit anything — so no summer trend is published there at all. What is published is the change, and a change from almost no summer ice to half a cover is not a trend fitted on nothing.

The control is not a Mertz detector, and this is how you can tell

Every summer trend at every site now carries a level shift at 2010 beside the 1988 sensor control, and a slope that stops clearing its error bars once the shift is allowed for may not be quoted. It would be neat to say that flag finds the Mertz event. It does not — it fires at 49 of the 728 sites, across eight regions. What marks that coast out is size and consequence: within 82 km of Cape Denison the shift runs 35.6 to 50.0 points and the slope reverses sign at 9 of 9; everywhere else it runs 0.4 to 20.3 points and reverses at 0 of 40. The gap between 20.3 and 35.6 points is empty.

This check was pointed at 2010 because a human already knew what happened there. Every fitted series on this site has since been made to nominate its own break year with nothing supplied, which is the version of this test that can find something nobody has named. Which year did the ice change? →

One rule was written down in advance and written on the wrong axis

Before any of this was run, a rule was fixed to stop empty measurements being published: a site whose median ice year carries fewer than 30 days of ice has nothing to measure. All 728 sites cleared it, so as declared it excluded nothing — while the summer window at 365 of the 369 chinstrap sites holds under a month of ice, and fitting it anyway produced 116 chinstrap summer trends that passed both the significance and the sensor gate on a window with essentially nothing in it. The rule now applies to whichever window is being fitted rather than to the year around it, which is the same rule rotated onto the thing being fitted. It removes the summer window at 365 of 369 chinstrap sites, 113 of 287 Adélie sites, 140 of 142 gentoo sites, 44 of 44 macaroni sites and 8 of 8 king sites — and at none of the 67 emperor colonies.

This is recorded as a correction rather than quietly fixed, because the failure is a general one and this site has now made it in three disguises: asking which statistic a known bias reaches, asking which axis a smoother moves a number along, and now asking which window a gate was declared on. A pre-registered gate has to be applied to the thing being fitted.

How this was measured

For each site the ocean cells within 100 km are fixed once and never re-chosen, and the daily sea-ice concentration is averaged over them for every day of the 1979–2024 satellite record — 16,802 daily grids, read out of 46 annual files. A site needs at least 8 ocean cells to be measurable, which excludes exactly one of the 729 known sites: Gipps Ice Rise, with 7. An ice year runs 15 February to 14 February and has to be 95% complete to be used at all.

Four windows were fixed before the first number was seen: the whole ice year, winter (1 June to 30 September), summer (1 December to 14 February) and the count of days at or above 15% concentration. Three gates then decide whether a trend may be quoted: it must clear twice its own standard error; it must keep its sign and its significance when refitted from 1988, because the earlier sensor reads low at the coast; and it must still clear its error bars when a level shift at 2010 is allowed for. An empty window is not fitted at all. Only trends that pass all of that appear anywhere on this site, and the counts that fail are published beside them.

Two limits, and both are load-bearing. At 25 km this is the sea around a breeding site and never the fast ice under it — the thing an emperor stands on is invisible at this resolution. And the archive ends 2024-12-31, so it is a climate record rather than a current reading; the live sea-ice figure on the wire is a different product entirely. There are still no penguin counts anywhere on this site: counting a colony from orbit does not work yet, and a number we cannot defend is worse than none.

Where the penguins breed, all 729 sites → · The sea ice around the 67 emperor colonies → · All 728 records in one frame → · The full site directory →