When the ice goes

Two of those hollow dots are not nothing

A retreat date can only see a season lost at the break-up end. Emperor Island and Pfrogner Point are drawn above as no detectable change, and on this measurement that is correct — their ice does not go measurably earlier than it did. Their ice seasons are nonetheless 18.2 and 17.6 days per decade shorter, because the ice now arrives 10.6 and 14.1 days per decade later. Emperor Island is the only emperor colony known to have been lost. See the season length decomposed at all 61 colonies →

The number is Bear Peninsula

The sea around Bear Peninsula, in Marie Byrd Land, used to clear around 20 Jan — high summer, chicks long gone. It now clears around 22 Oct. That is 90 days, 6.6 times its own standard error, fitted across all 41 usable ice years — though it is the one significant colony with too few grid cells clear of land to run the land-spillover check on, which is a gap in the evidence rather than a pass.

The dashed rule is 15 December, the date this site uses for the point a chick still cannot swim. It is taken from the breeding calendar, applied identically at every colony and never fitted to anything. At Bear Peninsula the retreat has fallen from the wrong side of that line to the right side of it.

How much ice, against when it goes

One dot per colony: the trend in how much sea ice it has along the bottom, the trend in when that ice goes up the side. Filled dots are the 12 colonies where both measurements clear twice their own standard error — the only ones that can be said to agree or to disagree at all. They agree at 11 of 12.

Cruzen Island is the one that disagrees, and it is the reason the date was worth measuring. Its sea ice is increasing, at +1.86 points a decade, while its retreat runs 13 days per decade earlier. More ice through the winter and an earlier spring exit are not a contradiction — they are two different facts about the same sea, and a seasonal average cannot hold both.

All 15 colonies whose retreat date has moved

9 now lose their sea ice earlier and 6 later, each with a slope more than twice its own standard error from zero. 6 of the 9 earlier are on the Bellingshausen and Amundsen coasts and the Peninsula; 5 of the 6 later are in Victoria Land, on the Ross Sea. Same fixed axis in every panel, and the dashed rule is 15 December in each.

How often the sea opens before the chicks can swim

No fit and no threshold chosen here — a count. For each colony, the share of ice years in which the sea around it fell below its own normal before 15 December, in the first half of the record against the second. Both the date and the split year are stated in advance and applied identically everywhere. At Bear Peninsula it has gone from 7 of 17 ice years to 23 of 24.

Ranked by how far it moved rather than by how high it is, deliberately: ranked by level the top of this chart would be a list of colonies whose sea has always opened early, which is climatology and not news. 7 of the 16 colonies shown have moved the other way, and Coulman Island is the furthest of them.

Two published events, used as blind checks. One passed, one failed usefully

The one it caught. Five Bellingshausen colonies were published as catastrophic 2022 breeding failures. Ranking all 62 colonies by how early their ice went that year, against each colony's own record before 2022, puts those five at 1st, 3rd, 4th, 6th and 11th — four of the five in the top six, from a measurement that knew nothing about them.

The one it missed, and why that matters more. The Halley Bay colony failed from 2016 after its fast ice broke up. On this measurement Halley Bay comes out 13th of 54 for that year — early, and entirely unremarkable. That is not a defect to fix. The 2022 event was the sea ice going wholesale across a region, which a 25 km cell sees clearly. The 2016 event was fast ice breaking up inside one inlet, which it cannot see at all. Anything built on this archive may claim the first kind of failure and must never claim the second.

Every colony with a retreat date, ranked

Slopes are quoted ± two standard errors. A trend is called significant only above twice its standard error, and only where at least 20 ice years carry the event. The last column is the amount of April–November sea ice at the same colony, for comparison — a different measurement on the same archive. 6 of the 67 colonies on this site have too few usable ice years to fit a retreat date at all and are absent from this table.

How this is measured, and what it is not

What it is. The day sea-ice concentration over a fixed set of ocean cells within 100 km of the colony crosses a threshold, per ice year. The convention is Stammerjohn's: the ice year runs 15 February to 14 February so one growth-and-decay cycle falls inside one label, and retreat is the first day after the ice advanced from which concentration stays below the threshold for the rest of the year. The threshold quoted throughout this page is each colony's own whole-record mean, fixed. A trend is called significant only above twice its standard error, and only where at least 20 ice years carry the event.

It is not fast ice. At 25 km the land-attached ice emperors breed on is narrower than one cell. This is the sea around the colony.

It is not current. The record ends 2024-12-31, so the last complete ice year is 2023. It is an archive, not a feed.

Two thresholds were measured and they are not two estimates of one thing. The 15% and climatological thresholds agree in sign at only 18 of the 31 colonies where both carry a trend. They are different events, not two estimates of one, and must never be quoted as though they were.

Two sensitivity tests, reported rather than tuned away. Refitting on only the grid cells with no land neighbour keeps the sign at 14 of the 14 significant colonies it can be run on; the 1 it cannot be run on has too few cells clear of land to test, which is a gap in the evidence and not a pass. Refitting on 1988–2023 alone — dropping the every-other-day sensor era and the four ice years lost to outages entirely — keeps the sign at 15 of 15.

Before mid-1987 the SMMR sensor flew every other day and the CDR fills the gap by temporal interpolation, so a date in those years is resolved to about a day and no better. The December 1987 - January 1988 sensor outage takes out ice year 1987 entirely; it is dropped rather than patched.

Half of the early record was never observed

Every date on this page rests on a daily archive that is not daily. Nimbus-7 SMMR flew every other day until mid-1987 and the producer fills the missing days cell by cell, publishing a value for all 365 — 37% of pre-1988 colony-days here are interpolated rather than observed, against 0.8% since. Nothing in the concentration values reveals them: they are not copies of their neighbours, they carry no odd-and-even signature, and they wobble day to day exactly as much as real readings. The only witness is a provenance flag stored beside the data.

The obvious test — drop the filled days and refit — turns out to measure nothing, because 92.78% of the fills are the plain mean of the day before and the day after, which across a one-day gap is a straight line: removing them and drawing the line back in reproduces them. What works instead is to decimate the modern record to the same every-other-day cadence, refill it the producer's way, and compare against the truth just destroyed. Across 4,059 colony-years the retreat date comes back exactly right 67% of the time and within a day 88%, drifting 1.1 days early on average.

Carried into the trends — exactly rather than by simulation, because a fitted slope is a weighted sum of the dates — it is worth a median 0.2 days per decade against a median fitted slope of 3.0. So the interpolated days are not what makes two thirds of these trends undefendable; that is the pre-1988 calibration step, which is a different fault. And the two break different colonies: this one misses where the concentration curve crosses the threshold flatly, not where the sensor step bites. Split the colonies at the median crossing slope and the flatter half misses by more than a week at 4.5% of colony-years against the steeper half's 2.3%. Every one of the 61 colony pages prints its own figure for this beside that colony's fitted standard error, rather than folding one into the other.

And still no count of the birds themselves

This page is about ice, not about penguins. There are no population figures anywhere on this site: detecting colony size from satellite imagery does not work yet — our own detector correlates at r = −0.16 against ground counts across eight colonies. What the archive can carry is the sea around a colony, measured two ways: how much of it there is, and, here, when it goes.

The continent behind these colonies

The same archive summed over the whole hemisphere rather than read at a colony says something that sounds like a contradiction and is not. Across all 46 years Antarctic sea-ice extent trends -0.063 ± 0.057 million km² per decade — 1.1 times its own standard error, and from 1988 onwards 1.9, so there is no trend this record can defend on either gate. And 345 of the 366 calendar dates in the year have their lowest reading of all 46 since 2016. A hemisphere with no trend in it is made of coasts moving in opposite directions, which is exactly what these colonies are. See the whole record →