0.06 Between the Clusters. 0.70 Between Two Drought Years.

A 0.70 gap.

Subtract 2023's Amazon carbon balance, 0.11 ± 0.09 billion tonnes of carbon, from 2024's 0.81 ± 0.09, and both years sit inside the drought cluster of a 15-year flight record Nature published on October 21, a swing no cluster average can see.

Luciana Gatti and colleagues report 1,017 aircraft profiles from 2010 through 2024 over 6,700,798 square kilometres, about 95 percent of Amazonia on their back-trajectories, and a 15-year balance of 0.32 ± 0.08 petagrams of carbon a year: fires 0.40 ± 0.07, net biome exchange −0.08 ± 0.12, a small sink whose bar includes a small source (INPE, the same day).

INPE's tonne-of-gas conversion is the factor recomputed here, 44/12 = 3.6667, so 0.32 × 3.6667 = 1.173 billion tonnes of CO₂, their 1.17 after rounding, and 0.40 × 3.6667 = 1.467, their 1.47 for fire.

Three clusters, then two subtractions

A k-means cut on the carbon balance, rain, temperature, deforestation, burned area, fire counts and the ocean indices, including NOAA's Oceanic Niño Index at a 0.5 °C threshold, split the 15 years into three groups.

Group 1 (2011-2014, 2017, 2018) is the wet, low-disturbance row in the table, with 6,389 square kilometres cleared a year, 55,115 burned, and a sink the paper says offsets 64 percent of its fires. Group 2 (2019-2022) matched that temperature with 6 percent less rain, 97 percent more deforestation and 25 percent more burned area. Group 3 (2010, 2015, 2016, 2023, 2024) was 13 percent drier and 0.5 °C warmer than group 1, with deforestation only 22 percent higher and burned area 92 percent higher.

Balances are Gatti et al., in billion tonnes of carbon a year. Steps, ratios and the rough standard-deviation checks are ours, using the abstract's ±0.13 for group 2. One results sentence in the paper prints ±0.09 for that group instead. Blank fire and biome cells were not given for that slice.
Slice	Total	Fire	Net biome exchange	Step we compute
15-year mean, 2010-2024	0.32 ± 0.08	0.40 ± 0.07	−0.08 ± 0.12	baseline
Group 1, wet, low disturbance	0.11 ± 0.08	0.32 ± 0.07	−0.20 ± 0.12	baseline for the clusters
Group 2, high clearing	0.43 ± 0.13			0.43 − 0.11 = 0.32
Group 3, warm and dry	0.49 ± 0.07	0.55 ± 0.10	−0.06 ± 0.10	0.49 − 0.43 = 0.06
2023 alone, inside group 3	0.11 ± 0.09			matches group 1's center
2024 alone, inside group 3	0.81 ± 0.09			0.81 − 0.11 = 0.70

From the printed centers, 0.43 minus 0.11 equals 0.32 and 0.49 minus 0.43 equals 0.06, so 0.32 / 0.06 equals 5.3 only if centers are the whole story. Treating the bars as independent standard deviations, the drought step's width is the square root of 0.13 squared plus 0.07 squared, or 0.148, and 0.06 / 0.148 equals 0.41, while the square root of 0.13 squared plus 0.08 squared is 0.153 and 0.32 / 0.153 equals 2.1, matching the paper's finding that groups 2 and 3 do not differ significantly.

Using the body's ±0.09 instead of the abstract's ±0.13 only lifts that drought-step ratio to 0.06 / 0.114, or 0.53. Applying the paper's percentages to the group-1 areas it prints, which it does not multiply out, gives 6,389 × 1.97 = 12,586 square kilometres cleared a year, 55,115 × 1.25 = 68,894 burned, and 55,115 × 1.92 = 105,821 in the drought cluster. (0.49 minus 0.11) / 0.11 equals 3.45, against a stated 325 percent, so the rounded means are what this piece uses.

One cluster, two years

2023, though filed in group 3, came in at 0.11 ± 0.09 because a wet-season sink offset 75 percent of emissions from clearing and fire that were 38 percent and 52 percent above group 1, while 2024 hit 0.81 ± 0.09, 1.2 °C hotter than group 1 and 462 millimetres drier, with burned area 182 percent higher, the year a separate mapping study calls the worst fire degradation in two decades (Bourgoin et al., Biogeosciences, 2025).

0.81 minus 0.11 equals 0.70, or 0.70 / 0.06 = 11.7 times the gap between cluster centers and 0.70 / 0.32 = 2.2 times the clearing step. On the printed intervals the square root of 0.09 squared plus 0.09 squared is 0.127, so 0.70 / 0.127 equals 5.5. 0.81 / 0.32 equals 2.53 times the 15-year mean, and 0.81 × 44/12 equals 2.97 billion tonnes of CO₂.

Lightning ignited 1.6 percent of 2023 fire events and 1.0 percent of 2024's, the only two years scored, so 100 minus those shares is 98.4 and 99.0, next to a caption's 98 percent human cause, and only where stroke detection is above 50 percent (INPE fire program).

What 0.32 weighs

EPA's passenger-vehicle page assumes 8,887 grams of CO₂ per gallon, 11,500 miles and 22.2 miles per gallon, and 8,887 × 11,500 / 22.2 / 1,000,000 equals 4.604 metric tonnes a year, the figure they round to 4.6, so the basin mean of 1.173 billion tonnes of CO₂ is 1.173 billion / 4.604 million = 255 million car-years, and 2024's 2.97 billion tonnes is 645 million.

Over the methods' 6,700,798 square kilometres, 0.32 × 10¹⁵ grams divided by 6,700,798 × 10⁶ square metres equals 47.76 grams of carbon, times 44/12 is 175.1 grams of CO₂, or 1.75 tonnes per hectare, so one EPA car-year matches 4.604 / 1.75 = 2.63 hectares.

Friedlingstein and the Global Carbon Project put 2024 fossil CO₂, after cement carbonation, at 10.3 ± 0.5 billion tonnes of carbon, so 0.32 / 10.3 equals 3.1 percent, and their land-use term of 1.3 ± 0.7 makes 0.32 / 1.3 = 0.25 at the center, with 0.32 / 2.0 = 0.16 and 0.32 / 0.6 = 0.53 at the bar's ends (Earth System Science Data, 2026). 2024's Amazon year over that center is 0.81 / 1.3 = 0.62, a column beside a bookkeeping model, not a subtraction.

Fifteen times the mean is 0.32 × 15 = 4.8 billion tonnes of carbon, or 17.6 billion tonnes of CO₂. Against the 123 ± 23 billion tonnes an assessment chapter assigns to vegetation and soils, that is 4.8 / 123 = 3.9 percent of the stock, and 123 / 0.32 = 384 years, an identity, not a countdown (Malhi et al., 2021).

Quiet-year sink over quiet-year fire is 0.20 / 0.32 = 0.625, against the paper's 64 percent, and the full record is 0.08 / 0.40 = 0.20, about a fifth, inside a bar that includes no sink.

Earlier flights already found a southeastern source once fire was removed, and a 2023 paper tied a further rise mainly to weaker enforcement, while intact plots can still sink under a smoky column (Gatti et al., 2021; Gatti et al., 2023; Brienen et al., 2015; Hubau et al., 2020).

Annual clearing fell 30 percent in 2023 and 47 percent in 2024 relative to 2019-2022, and 2024 still set the carbon record, with PRODES, the authors say, classing 38 percent of that deforestation as fire, on a cumulative loss of 807,301 square kilometres (TerraBrasilis; MapBiomas Fire, cited at 89 percent accuracy; NOAA flasks for the coastal background).

What You Can Do

If a carbon claim depends on an Amazon sink, ask which year and whether fire sits inside the tonne, because the 15-year net biome exchange of −0.08 ± 0.12 billion tonnes of carbon includes zero, and a credit built on intact-forest plots is not this aircraft column.

Watch PRODES clearing and MapBiomas burned area together: 2023 and 2024 are the case where clearing fell and the balance did not. The next signal is whether a wet, low-fire year returns near 0.11 or whether quiet years now sit near 0.43, and the profiles for that test are already public (PANGAEA). One American car for every 2.6 hectares is a map scale, and the lever in these flights is fire and clearing policy.

Limitations

These steps use published centers. Group 2's fire and biome terms were not printed where this piece looked, the ±0.13 versus ±0.09 split is unresolved on the fetched page, and the combined standard deviations assume an independence the authors do not claim.

Clustering took the carbon balance as an input, missing months were filled by a random-forest imputer, sampling stops at 4.4 kilometres and loses fire-season carbon because convection already lifts CO₂ above the coastal background at the top of the profile, and a single CO/CO₂ ratio likely understates 2023 and 2024 as the forest dries, so the recorded 0.81 may be low.

Lightning covers two years, 2015 and 2016 flew with long gaps and still sit in group 3, and Botia and colleagues' flask-versus-spectrometer correction, which this team refused because a water-vapor fit explained 13 percent of the variance, would move the levels if it were applied across the record (Botia et al., 2025).

The 384-year stock ratio and the land-use comparisons are scales, not a closed budget. The 2024 global budget carried a large negative imbalance. The 8 percent rainfall drop is Gatti's 1979-2024 series; deforestation reducing rain is a separate result (Leite-Filho et al., 2021).

The Strongest Case Against This Analysis

Reading 0.06 as a verdict that drought barely matters is the misuse the cluster centers invite, because burned area in the drought group was 92 percent above the quiet years while deforestation was only 22 percent higher, fire carbon was 0.55 ± 0.10 against 0.32 ± 0.07, and 2024 alone was 0.81, so climate stress changes how much human-set fire escapes even when the cluster totals overlap.

Flores and colleagues estimate that by 2050 between 10 and 47 percent of these forests face compounding disturbances that can flip local ecosystems, a range too wide to be a date and untouched by a non-significant gap between two means (Flores et al., Nature, 2024). Intact plots can remain a weakened sink while this column, pasture and smoke included, is a source. Both can be true. The flights show the air gained carbon, and that the worst recent year was fire during drought.

The Bottom Line

Fifteen years of flights put the basin at 0.32 billion tonnes of carbon a year, about 1.17 billion tonnes of CO₂, 3.1 percent of 2024's fossil total and 255 million car-years. The drought cluster sits 0.06 above the clearing years, inside the error bars, and 2023 versus 2024, inside that cluster, differs by 0.70: drought picked which human fire year became the record.