Reading guide · Gironde, July 2026

How to read these numbers

Six pages of method compressed into one: what the scale means, what the reference is, which figures are firm, and which are floors.

1 · the scaleEverything is a departure, not a value

The map does not show how wet the vegetation is. It shows how far it sits from its own normal for that week of the year, in standard deviations (σ). A pixel at 0σ is behaving exactly as it usually does in that week; −2σ means it is drier than it has been in roughly 39 out of 40 comparable observations.

This matters because it removes geography and season from the picture. Ground that is bone dry every July does not light up. Only ground that is dry for itself does.

0 to −1σ

Within normal variation. Not a signal.

−1 to −2σ

Below normal. Expected on ~16% of area in any ordinary year.

−2 to −3σ

Clearly abnormal. Expected on ~2% of area.

Below record

Drier than any of the eight earlier years on that week.

The single most important comparison on the page is not the σ value itself but the share of area past a threshold, against what an ordinary year would give. Inside the burn scar, 59% of the ground was below −1σ on 21 July. In an ordinary year that figure would be about 16%. That ratio — roughly 3.7× — is the finding.

2 · the referenceWhat “normal” is built from

Element	What we used
Record length	2018–2025 (8 years)
Observations pooled	264 cloud-screened acquisitions
Matching rule	Same calendar date ± 20 days, across all eight years
Statistic	Per-pixel median and robust spread (MAD), not the mean
Minimum support	A pixel needs ≥ 6 valid observations or it is left blank
Grid	40 m, 1425 × 1702 cells

Two consequences worth holding onto. First, the normal is per pixel — a pine stand is compared with itself, not with the region. Second, the pooling window is symmetric: the reference for 21 July draws on observations from 1 July to 10 August in earlier years. An asymmetric window built only from earlier dates would have made every summer reading look drier than it is.

3 · the headline figuresWhat each one does and does not say

Figure	Value	Read it as
Share below normal	59%	Of the ground that later burned, this share was under −1σ on 21 July. Firm.
Record dryness	17%	Below anything seen in the eight-year record for that week. Firm, and the strongest form of the signal.
Mean departure	−1.73σ	Average across the scar. Against −1.28σ on nearby wildland that did not burn — the burned ground was the drier patch inside a dry region.
Burn scar area	21 850 ha	A floor, not a measurement of the fire. See section 5.
Ignition distance	0 m	The reported ignition point falls inside the scar we detected without being told where the fire was. A location check, not an accuracy figure.

4 · by land coverWhere the signal was strongest

Cover	Area	Below −1σ	Mean
Forest	147 398 ha	50.9%	−1.17σ
Heath / shrub	100 846 ha	58.5%	−1.52σ
Cropland	48 363 ha	53.3%	−1.45σ

Heath and shrub dried hardest — which is the operationally relevant result, because that is the fast-igniting surface fuel under the pine. The drought layer on the map deliberately shows forest and heath only. Cropland is excluded from the map because a harvested field is not drought stress, even though the departure statistic handles harvest correctly by construction.

Land cover here is derived from the imagery itself — from how each pixel's greenness behaves through the season — not taken from an external database. Pine plantation holds high greenness with little seasonal swing; cropland swings hard; built and bare ground stay low.

5 · the cross-checkWhat the burn scar proves, and what it doesn't

What it proves

The burn scar is derived only from the change in infrared reflectance between 21 July and the two post-fire passes. No fire product, perimeter or external map went into it. It lands on the right ground: the reported ignition point falls inside it, and its composition — 63.5% forest, 26.3% heath, 10.1% cropland — matches the pine-and-heath landscape the fire ran through.

What it does not prove

The area is not an independent confirmation of the official 42 000 ha, and should never be presented as one. Three separate reasons pull the two figures apart:

Cloud. Usable post-fire observations covered 57% of the grid, and 81% of the detected fire footprint. Ground we could not see cannot be counted.
Different definitions. The official figure is the area within the fire perimeter, which includes unburned islands. Ours counts pixels that actually changed.
Timing. The last usable pass was 31 July, and it was largely clouded over the southern tile.

So the honest statement is: 21 850 ha seen directly, about 27 100 ha after correcting for cloud inside the detected footprint, against 42 000 ha reported. Our number is a lower bound.

6 · the nine-year stripThe cleanest single comparison

Nine thumbnails, same three weeks of July, same colour scale, one per year. The mean canopy moisture over forest and heath:

2018	2019	2020	2021	2022	2023	2024	2025	2026
.272	.253	.268	.280	.241	.261	.299	.263	.226

2026 is the lowest of the nine, and below 2022 — the year of the previous large Gironde fires. This needs no statistics to read, which is why it is the strongest single exhibit on the page. Note that 2024 had only 53% cloud-free coverage, so treat that column as the least reliable.

7 · guardrailsClaims that must never be made

Do not say

“We predicted this fire.” This is a retrospective measurement. The ignition was reported as accidental; dryness is a condition, not a cause, and the overwhelming majority of dry vegetation never burns.
“Our burn area matches the official figure.” It does not, and the gap is explained above.
“This measures fuel moisture.” The scale is a departure from local normal, not a calibrated moisture percentage, and it has not been checked against field samples.
“This replaces the clearance assessment.” This is the condition dimension. Height, corridor geometry and intervention packages are a separate exercise.

Safe to say

On the ground that later burned, the vegetation was measurably drier than its own eight-year normal, and the drying was visible weeks before ignition.
The signal separated burned from unburned wildland in the same area (−1.73σ against −1.28σ).
Everything shown is computed by us from raw satellite time series, with no third-party fire, burned-area or land-cover product.

8 · honest limitsCarry these with the numbers

Eight years is a normal, not a climatology. A −1.7σ departure against eight years is a weaker statement than the same figure against forty.
Dense canopy saturates. Under closed pine, the sensor mostly sees the crown; the understory that actually carries fire is only partly visible. The heath figures are the more direct read on surface fuel.
Cloud is the binding constraint after the event, not before it. The pre-fire assessment date of 21 July was essentially cloud-free, which is why the drought half of the analysis is solid and the burn half is a floor.
One index, one sensor family. No radar, no thermal, no ground measurements. Adding them would sharpen the estimate and is the obvious next step.
Irrigated ground reads wet. That is correct behaviour, not an error — but it means cropland figures should not be read as regional drought.