Fires: how these numbers are made

The fires channel tracks two things, on two instruments that measure different quantities and are never converted into each other:

• Detections. Satellite thermal anomalies, daily, for 97 countries.
• Burnt area. Hectares mapped as burned, weekly, from Copernicus.

Every figure is one country compared against that same country's own history in the same week of the year. We do not compare countries to each other, and a page that appears to is misreading the numbers.

This page is written for someone deciding whether to cite us. The thresholds it quotes are emitted in data/events.json under method, so the page reads them rather than restating them and cannot drift from the code that applies them.

What is measured

Detections. NASA FIRMS, VIIRS aboard Suomi-NPP, 375 m, near-real-time for the current week and the science-quality archive for prior years. Low-confidence pixels are dropped. Each detection is assigned to a country by point-in-polygon, not by bounding box, so a fire near a border counts for the country it is in.

The trailing complete week. Seven days ending yesterday. A day is only counted once the archive has closed it, roughly three hours after midnight UTC, so the newest figure is always one day behind today. That lag is the instrument, not a delay in our pipeline.

Burnt area. Copernicus EFFIS where it has coverage, GWIS elsewhere. Cumulative hectares since 1 January, published weekly on a Wednesday and available the following morning. Between releases the figure ages up to seven days, which is why hectares and detections carry separate dates on every page.

How a week is judged

A country's week is compared against the same seven calendar days in each prior year, using the same sensor throughout. From that we take:

• rank, its position among those years, which is robust to one outlier year;
• multiple, the count against the mean of those years;
• z, the count against their standard deviation.

Rank is the sturdier reading and the multiple is the loudest. Where they disagree, the pages lead on rank. A country whose record year was exceptional will show a modest multiple against a genuinely extreme rank, and the reverse is true for a country with a flat history.

A noise floor applies before any of it. Below the floor emitted as noise_floor_detections a country cannot qualify however large its ratio, because a multiple built on a handful of pixels is arithmetic rather than a finding.

Standing limits

These are the reasons a number on this channel might mislead you. They are the part of this page worth reading.

The instrument does not see what is burning. It sees a thermal anomaly, a time and a radiative power. Everything we say about what is alight is an inference from where the heat sits, never a measurement.

Cropland is a ratio, not a share. We sample a published 500 m crop mask at each detection and compare it against the share of that country which is cropland. Cuba's detections sat on farmland 28% of the time in the week to 29 August, against 11.5% of Cuba being farmland, which is the 2.4 times the page reports. The share alone would have told you nothing.

The ratio is withheld unless enough detections actually fall on cropland, and only when the reading says they concentrate there. Peru read five times enriched on about twenty pixels, and New Zealand twice on one, which is arithmetic rather than a finding. The floor is fifty.

It applies in one direction only, because the evidence does. A reading that says fires sit on farmland MORE often than chance rests entirely on the detections that are there, so a handful cannot support it. A reading that says LESS often is evidenced by that same small count: Belgium's one detection on cropland, out of many, is what "less often" means. Flooring it would delete the finding rather than protect it. Those readings are held to fifty total detections instead.

Not assessed is not zero. Two different withholdings exist and both render as "not assessed": the mask has no data for that country at all, and the mask was unavailable to the machine that built the payload. In neither case have we looked, and neither is evidence about farmland.

Some countries are not burning, they are flaring. Gas flares and industrial heat appear as thermal anomalies every night of the year. We exclude a country whose week is nocturnal, low-powered, recurring in the same cells and flat. The flatness test matters: Algeria met the first three thresholds while a genuine record burned underneath, and a real fire is a curve where a flare is a line.

A country not in the roster is a gap in our coverage, not a quiet country. Absence from these pages says nothing about whether a place is burning. It says we have not built a baseline for it.

Averages can include years nobody measured. Copernicus counts years before its own coverage began as zero. We drop a leading run of empty years before computing any "times the average", because Algeria's mean was otherwise divided across three years that were never observed. Interior zeros are kept: a quiet year after coverage begins is real.

Records and multiples fail differently. A record ignores zeros and survives all of the above. A multiple depends on a denominator and does not. Where a claim can be made either way, the pages make it on the record.

Sources

Version history

1.0, 2026-08-30. First publication. Written after a fortnight in which most of the limits above were found the hard way, several of them by other people reading our numbers rather than by us.

1.1, 2026-08-30. The cropland floor was tested against every detection in a country rather than against the detections on cropland, so an enriched reading, whose whole evidence is the latter, was gated on a number unrelated to it. Six countries stated a reading they should have withheld; one, Tunisia at 1.48 times on forty, was on the public page. Corrected the same day, and the correction is the asymmetry described above. Found because our design chat asked to see the floor checked against the whole payload rather than against the three countries we had named as examples, which were the three most likely to pass.

The thresholds, as the code applies them

Read from the channel's own emitted method block rather than typed here, so this table cannot drift from the pages. The key is printed beside each value so a figure on a page can be traced to the field that set it.