How a 1980s satellite invented an Antarctic sea ice trend.

Here are the winter peak sea ice concentrations at Davies Bay, on the Victoria Land coast, for the first five years of the satellite record: 86.9, 88.2, 87.6, 89.1, 88.1. And here are the first five after 1988: 99.9, 99.3, 99.9, 99.5, 99.6, followed by near 100% for the next thirty-seven years running.

Climate does not do that. An instrument change does. The sea at a coastal cell in August is solid every single year whatever else is happening to the planet, so the winter peak there is pinned at the ceiling and has no physical room for a trend. A multi-point step in it is not weather. It is the sensor.

The sensor is Nimbus-7 SMMR, which flew until mid-1987 and produced the first eight years of the passive microwave record. It reads low. This post is about what that cost us, how we proved it, and the rule that came out of it, which is worth more than the result.

Five low years at the far left of a long line

The bias is only 3 to 5 concentration points on average. That sounds harmless. It is not, because of where those years sit. A straight line fitted through 46 years of data pivots on its middle, so the earliest and latest points carry the most weight. Push the first five to nine years down by four points and the fitted line tilts upwards, and the tilt is entirely manufactured.

You can watch it happen. Across all 67 emperor penguin colonies we measure, the size of a colony's early sensor deficit predicts the size of its fitted rise with a correlation of +0.50. Across the 20 colonies whose later winter peak saturates, where the deficit is provably instrumental because there is no room for anything else, r = +0.91. The cut is stated on the physics rather than chosen for its answer, and the sweep is published so you can see it holds wherever the line is drawn.

Grouped the other way round it is just as blunt. The 12 colonies with a deficit of 6 points or more carry a mean fitted trend of +2.10% per decade and 11 of the 12 are rising. The 29 colonies with a deficit under 3 points average +0.19% per decade.

What it cost this site

A headline. We had published that sea ice around emperor colonies had risen at 26 of 67 colonies and fallen at 8. That "risen at 26" is mostly the sensor.

There are two honest corrections and they disagree about strength, so both are published and the conservative one leads. Dropping the pre-1988 era outright assumes nothing and costs nine years of record. Fitting an explicit step term keeps every year but assumes the offset is a clean level shift, and a step at one end of a series is partly interchangeable with the slope, so it can eat a real change as easily as a false one.

Five of the eight declines survive against five of the twenty-six rises. That asymmetry is exactly what a low-biased early era predicts, because the bias inflates rises and suppresses falls, which makes declines the safer half of this dataset. The count of rising colonies is not a defensible number from this archive and we no longer quote it at all.

The core finding survives and gets cleaner. Sea ice around emperor colonies has not broadly declined, and the loss is regional rather than continental. On the homogeneous window 57 of 67 colonies show no detectable trend at all, which is a stronger way of saying the same thing than "26 rose" ever was. That is what the colony map draws.

Why it is coastal, which explains everything

The bias is not a flat calibration offset. Binned by distance from land, the August deficit runs 4.17 points inside 50 km, 1.02 points at 50 to 100 km, 0.34 at 200 to 400, and 0.17 beyond 800. Monotonic over four orders of distance, with 2.25 points in the marginal ice zone where the pack meets open water.

The mechanism is footprint. SMMR's real resolution was far coarser than the 25 km grid the data is delivered on, so a cell next to a coastline carries some of the coastline in it. Every emperor colony cell set is coastal by construction, which is why the colony numbers were badly hurt. At hemispheric scale the same step is 0.34 of a point, ten times smaller.

Why the live number on our front page is untouched

Because it is a different kind of number, and this is the part worth carrying away. Sea ice extent is not a measurement of how much ice a patch of sea holds. It is a count of cells across a 15% line, at an edge where concentration falls from 100 to 0 within a cell or two. Shift every concentration reading in the record by 3 points and extent moves by at most 0.2 million square kilometres and usually by nothing you can measure.

Tested directly on the ranking we publish daily: 2023 is the lowest year of the record on the full 45 years and the lowest of 36 with the early era dropped. 2024 is second either way. 2022 is fourth either way. A fault that wrecks one kind of statistic leaves the other alone, so check which kind of number you are holding before deciding whether a known fault reaches it. Our daily sea ice reading and the whole 1979 to 2024 record both rest on that.

The rule, which is the useful part

We nearly missed this. The 1988 test was first run on one series, the date the ice retreats in spring, where it passed at 15 of 15. Passing that cleanly made it look like a formality, so it was not run on the neighbouring series. Run on the freeze-up date, it kills three quarters of the result and reverses the sign of the three largest trends on the continent.

A sensitivity test that passes on the one series you ran it on is not a passed test. Run every gate on every series, and treat the disagreement between them as the finding rather than as an inconvenience. Every trend this site publishes now carries a robust flag beside its significance flag, and where the two disagree we say so on the page instead of quietly picking the version we like.

There is a second fault in the same years, incidentally, and one gate happens to cover both. Before 1988 the satellite flew every other day and the archive fills the missing days rather than leaving them out, so 1,555 pre-1988 days are drawn straight through a gap and nothing in the pixels says so. Dropping the era removes 92% of them. That is luck rather than design, which is why the cut-off never gets relaxed.

Everything above comes out of the free, keyless NSIDC daily sea ice extent record and the NOAA and NSIDC concentration climate data record behind it. Both are open, both are on the same public archive, and anybody can rerun this.