Until this week this site said the sea ice around five emperor penguin colonies is robustly increasing. It now says four, and prints the old number next to it. The colony that left the list is Mertz Glacier, in George V Land, and it left because we asked every straight line on the site a question we had never asked before: if this record changed level in one particular year rather than drifting, which year was it?
Mertz Glacier answered 2011, out of twenty-one candidate years, having been told nothing about icebergs, satellites or anything else. That is the year after the iceberg B09B grounded in Commonwealth Bay, 173 km along the same coast.
The number, before and after
The measurement is the mean sea ice concentration within 100 km of the colony, April to November, the months an emperor colony needs a platform. Fitted from 1988 onward as a plain straight line it rises. Allow the same series one level shift, at the year the scan picked for itself, and it falls in both halves.
| fit | slope, points/decade | step, points |
|---|---|---|
| straight line | +1.31 ± 0.47 | none |
| line with one step at 2011 | −2.04 ± 0.57 | +8.78 ± 1.25 |
The sign reverses. The step is nearly seven times the size of the decadal slope it replaces. Before 2011 the ice there sat between 63 and 76 per cent every year for thirty-two years. From 2011 it sits between 71 and 82. Nothing in that record is drifting. It jumped once and stayed.
A straight line through a step will always report a trend, and it will report the wrong direction whenever the drift inside each level runs against the jump. That is what happened here, and it is the reason the colony page and the emperor colony ice map both changed this week rather than being quietly edited.
How it was found without anybody knowing to look
This project already refuses to fit a straight line across three particular years. 1988, when the satellite carrying the sea ice sensor changed and the new one read several points higher near coasts. 2013, when Landsat 8 arrived and the ice velocity product jumped. 2010, when B09B took the Mertz Glacier Tongue off and then parked in a bay.
All three were found the same way: a human knew the event had happened and wrote a check for it. That method can never find a fourth. So every one of the 5,686 fitted measurements on this site, at all 67 emperor colonies and all 728 measurable breeding sites, on every window and every family, was made to nominate its own worst break year against a Monte Carlo null with no break in it. Ninety seconds of computation, no internet, nothing supplied about the world.
Most of the answer is reassuring, and that matters more than the correction. Out of 842 fitted series at the emperor colonies, four carry a level shift that survives the correction for having looked at twenty-odd years at every one of thousands of places. Out of 4,844 at the breeding sites, none do. All four of the four are 1987 or 1988, which is the satellite change this site already declines to fit across. Taken as a body, the straight lines here are not sudden changes in disguise.
The check that makes it more than a coincidence
The scan was never told where the Mertz Glacier Tongue was. So ask it a question it can fail: of the fitted measurements that flag a break anywhere within 400 km of Cape Denison, how many pick 2010 or 2011?
Thirty-two out of thirty-two. Everywhere else on the continent, one out of a hundred and fifty. That is the strongest positive control this project has run, and it is what turns "a line fits worse than a step" into "something physical happened on that coast in those two summers".
There is a second signature, and it is the sort of detail that separates a real break from a fitting artefact. A level shift left unfitted manufactures correlation between one year's residual and the next, because the early years all sit on one side of the line and the late years on the other. At Mertz that year-to-year residual correlation is +0.49 with a plain line and exactly zero once the step is allowed. Real persistence does not do that. Across all 5,686 series the ones flagged as breaks run +0.83 before the step and +0.09 after, and the ones flagged as clean run +0.08 and +0.00.
Four things this does not license
The scan found more than one thing and most of them are things we are declining to say.
Point Géologie is not the iceberg. The other George V Land colony also breaks, but at 2007, three years early, so its breaks may not be attributed to B09B and are kept separate in the data file for exactly that reason.
The 2007 cluster is a coincidence of the calendar. A hundred and eleven measurements appear to change level in 2007 against about five expected by chance, which looks like a discovery until you thin them. They are two independent places, 5,645 km apart, counted a hundred and eleven times. Every break year on this site now publishes how many independent places it is and how far apart they are, and a year with no geography is not an event.
No mechanism is claimed for the shape of it. The step is positive around Commonwealth Bay and negative to the west in Wilkes and Adélie Land. That dipole is published as a measurement and nothing more.
One place is unexplained and stays unexplained. Barrier Bay and West Ice Shelf, which are 75.2 km apart and are therefore one place measured twice rather than two colonies, both step down 9 to 14 points in 1994. It is not the satellite, not an instrument, and the year turns up in scattered places over a thousand kilometres away, so it has no geography and no cause may be invented for it. It is the only unexplained candidate the whole continent produced, and it is written down as an open question rather than as a result.
Why publish this rather than fix it quietly
Because the retraction is the more useful artefact. A site that publishes numbers has to show what happens when one of them fails, and the failure here is instructive in a way the original figure never was: the difference between a trend and a step is invisible to the eye, invisible to a significance test, and worth the whole sign of the answer.
The full working, including the calibration table showing what these gates really do and the blind spot that stops the scan seeing 1988 in one of the two archives, is on the breaks page. Every colony whose own published figures the scan puts at risk carries the verdict on its own page, because a correction that lives only on a summary page is one most readers never see. The underlying record is the free NOAA and NSIDC passive microwave sea ice climate data record, which anyone can download and refit.