Methodology

Two sources that do not measure the same thing

These two counts are never added together. One declared fire can match forty hotspots, or none. On this site they stay in two separate columns, in two separate colours.

Why comparisons between years use a single satellite

The VIIRS constellation grew during the period covered:

Three satellites fly over France more often than one. For the same amount of fire, 2026 therefore mechanically counts more detections than 2012. Publishing a raw curve over the whole constellation would manufacture a trend that does not exist.

Every comparison between years on this site is computed on Suomi NPP alone, the only continuous series from 2012 to today. The map of the current year does use all three satellites: there we want the best possible coverage, not a comparison.

Permanent heat sources

Flares, blast furnaces, incinerators and cement works burn all year round and appear in the satellite data like any other fire. Untreated, the recurrence map is saturated by Fos-sur-Mer, Gonfreville-l'Orcher and Dunkirk, and the forest signal disappears.

A cell active in more than six distinct months and more than three years is flagged automatically, then checked by hand before being excluded. Automatic detection alone never removes data. A button on the map brings them back.

How hotspots become a fire

On the active fire pages, hotspots are grouped: two detections belong to the same fire if they occupy the same 5 km cell or a neighbouring one, and are less than 1 day apart. A group of fewer than 4 detections is not kept: an isolated pixel is not a wildfire.

This grouping errs in both directions. Two neighbouring fires can be merged into one; a fire cut by an unobserved area can be counted twice. It replaces no official source.

A final filter excludes industrial installations. A flare or a blast furnace heats the same place every day: without a guard, the grouping turns it into a ten-day "wildfire".

Our first rule looked at duration and power. It correctly excluded Fos-sur-Mer and Grande-Synthe, but also the fires at Corte and Albertacce, which were the two largest fires in Corsica in July 2026. A mountain fire smouldering for two weeks in steep undergrowth has exactly the same signature as a flare. The site was hiding what it exists to show.

What separates them is that a fire's footprint advances. An installation burns the same pixels every day; a fire, even a slow one, gains new ones. A group is therefore only excluded if all three conditions hold together: detected on at least 5 different days, less than 15 MW per point, and less than 15 % of its footprint gained in the second half of the event. Corte was at 34 %, Fos-sur-Mer at 0 %. When in doubt we display: hiding a real fire is worse than letting a flare through. The detections stay in the database either way.

The danger announced for tomorrow

This is the only data on this site that speaks about the future, and it is not ours. Météo-France publishes the Météo des forêts twice a day at 5 pm during the fire season, from early June to late September, for the next day and the day after. It combines temperature, humidity, wind and dryness of the soil and the vegetation.

It does not say a fire will start. It says that if one starts, conditions will help it spread. The four levels and their labels are reproduced word for word: it is not for us to reword a danger scale.

The bulletin is published by department. It appears as is on department pages, and on municipality pages with a note that it does not go down to that grain. Out of season there is nothing to show, and a bulletin older than 3 days is no longer served: a September level shown in January would be worse than nothing.

The layers on a fire map

They are not equal, and the order below runs from the most solid to the most fragile.

• The hotspots are measurements: a sensor saw a thermal anomaly, at a known time.
• The rain is an observation: a radar measures the drops. The public tiles we use stop at zoom 7, so the image is enlarged and blurry up close.
• The wind and the smoke come from European models, AROME and CAMS, for the current hour. These are calculations, made by teams whose job it is, and we report them without extending them.
• The aircraft are observations: the aircraft broadcast their position over ADS-B, in the clear, like every plane. We capture nothing and follow nobody, we read a public feed. Two limits: reception close to the ground is poor, so an absence on the map does not mean an absence of aircraft; and the trace history stops at twenty-four hours.
• The computed plume is ours, and it is the most fragile layer. Each hotspot emits from the time it was seen; that smoke is carried by the wind over a distance equal to its speed times the elapsed time, widening by about 12° on each side. The calculation ignores terrain, convection above the fire, local breezes and air stability, and it applies a single wind to the whole area. It stops at the present moment: we do not extend it by a minute, because a shift in the wind would make the whole track wrong.

None of these layers says what to do. For instructions, the reference is the prefecture, the fire service and FR-Alert.

The current year is provisional

NASA first publishes a real-time version, then a corrected version several months later. The corrected one removes some detections, notably false alarms. The current year therefore mixes both: its early months are corrected, the most recent ones are not yet.

Consequence: comparing the current year with past years slightly favours it. This is written on the home page next to the figure, and the monthly resync closes the gap as it goes.

The approximations we accept

• The department filter on aggregated layers assigns each cell to the department most represented in its detections. At the border, the imprecision reaches half a cell. The counters on department pages are computed point by point, and are therefore exact.
• Recurrence covers the 14 complete years from 2012 to 2025. The current year is excluded until it is over.
• The season shown is the one that concentrates 80 %% of the year's hotspots, from the 10th to the 90th percentile. A first version counted the days above a detection threshold: it produced seasons running from 1 January to 31 December, because the satellite also sees agricultural burning all year. The percentile depends on no arbitrary threshold and compares from one year to the next.
• Satellite data from the last five months is provisional. It is replaced by the corrected version through the monthly resyncs.