Situation Report: Amazon Fire, Governance Topology, and the 2026 Season Onset

Capybara Center · Early Warning System (EWS) · Alejandro Feged-Rivadeneira, Tecnológico de Monterrey · 19 July 2026

Bottom line. As of week 29 (mid-July 2026) the Amazon basin is in the pre-onset shoulder of its dry season. Basin-wide active fire is tracking close to 2025 and about 40 percent of the extreme 2024 level, and the southern-arc burning ramp that usually breaks near week 30 has not yet started. Our analysis of 2012 to 2024 shows that the popular reading, that Indigenous territories and protected areas suppress fire, is largely an artifact of where those lands sit rather than how they are governed, and that the effect reverses in the dry southern forests. The 2024 record drought registered as an extreme topological anomaly on a rising trend, not a permanent shift, and the same signature is what a weekly monitor detects as each season breaks. Soil moisture, the limiting factor for fire here, now gives that monitor roughly four to six weeks of lead.

~11,300

fire clusters basin-wide, 2026 year-to-date (through wk 29) — near 2025, ~40% of 2024

2024

topological onset year: burned area doubled to ~19 Mha; fire-size inequality spiked in every regime

~55%

of the apparent Indigenous-territory fire "protection" (Guaviare) attributable to landscape position, not management

1,715

governance polygons linking Brazil-nut resilience to fire exposure and boundary geometry

~5 weeks

lead time: southern-basin soil moisture forecasts fire ~4 to 6 weeks ahead (strongest in moisture-limited dry forest)

1. Current situation: the 2026 pre-onset window

Using NASA FIRMS active-fire detections aggregated to weeks from January 2024 through 13 July 2026, and tagged by governance regime, the current season is quiet relative to recent extremes. Year-to-date through week 29, the basin has recorded roughly 11,300 weekly fire clusters, almost identical to 2025 (about 11,500) and far below 2024 (about 27,200). The cumulative trajectory sits on the flat pre-onset shoulder. The southern basin below three degrees south, which carries the arc-of-deforestation fire season, has not yet entered its ramp: that ramp historically begins around week 30 to 32 (early August). Early ignitions in the run-up are led, as expected, by the agricultural frontier, while Indigenous and protected lands remain near zero.

Figure 1. Weekly fire monitor (FIRMS, 2024 to 2026). 2026 (red) tracks 2025 and sits well below the 2024 extreme; the southern-arc ramp has not started; the frontier leads pre-onset ignitions.

2. Method: necessary conditions and the topology of fire

Wildfire occurrence is not reliably predictable at the ignition level (roughly 95 percent of Amazon ignitions are human, and the smallest fires fall below satellite detection). Rather than predict where a fire starts, the Capybara approach detects where the necessary conditions for fire co-occur, and then studies the geometry and topology of the resulting fields. A cell is receptive only if its weakest necessary condition is met, following Liebig's law of the minimum. A robust and repeated result across our regions is that fuel is never the limiting factor in these humid systems: receptivity is governed by moisture at the regional scale and by slope locally. This matters for what follows, and for the forward-looking recommendation in section 6.

On top of receptivity we compute the topology of the fire and fuel fields using persistent homology and Mapper: Betti-0 counts the number of separate fire components (mosaic fragmentation), and Betti-1 counts enclosed unburned refugia (holes in the burn matrix). Unlike a Gini coefficient, these measures see spatial arrangement, so two landscapes with identical fire-size inequality but different organization are cleanly separated.

The Mapper skeleton makes this concrete. Built on the Capybara susceptibility field S(x), the 1-skeleton (its edges) and 2-skeleton (the shaded triangles, or 2-simplices) trace the topological backbone of the susceptible network. For the whole basin, that network is not the forest interior, which sits at a uniform moderate susceptibility and carries no skeleton at all, but the arc of deforestation along the southern and eastern margins, resolved into a set of high-susceptibility corridors.

Figure 2. Mapper 1-skeleton (left) and 2-skeleton with 2-simplices (right) over the Capybara susceptibility S(x) for the whole Amazon. The intact interior (green) carries no skeleton; the high-susceptibility network (warm colors) traces the arc of deforestation and fragments into distinct components.

3. Governance and fire: mostly position, and location dependent

Analyzing 234,006 fires across the basin from 2012 to 2024, the naive test of the fire-stick-farming hypothesis fails in an informative way. At the aspatial level, Indigenous territories carry the heaviest fire-size tail, not the lightest (Gini 0.855 versus 0.796 on the frontier), and this holds in all thirteen years. The likely reason is that the coarse burned-area sensor (about a 21 hectare floor) is blind to the small cultural burns that define a fine managed mosaic, so it captures only the large escaped fires. This is a limitation of the instrument rather than a refutation of the mechanism, and it is precisely why a spatial and finer-resolution approach is needed.

The decisive step is to control for a polygon's position in the basin's fire field. We built a fire-history-free spread-centrality from exogenous dryness and computed, per governance unit, whether apparent regime effects survive matching on that position. They largely do not, and they are strongly location dependent.

Figure 3. Three zooms, three answers. In the wet northwest (Guaviare) Indigenous territories appear strongly protective, but about 55 percent of that is position and only a 0.27-fold residual survives control. In the cerrado transition (Xingu) the effect is roughly neutral with a hidden crossover (managed burning shows up at high fire pressure). In the dry Chiquitano forests of Bolivia the sign reverses and Indigenous lands burn more. There is no single protective effect.

The heterogeneity is not noise, it is structure, and the susceptibility topology itself shows why. Reading the Mapper skeleton over S(x) region by region, Guaviare is a fragmented high-susceptibility ring (Betti-0 about 25 separate components) wrapped around a large wet, low-susceptibility core, which is exactly where the Indigenous territories sit, so their low fire is mostly a matter of that quiet interior. Xingu is a connected susceptible network in the cerrado transition (Betti-0 about 4), the setting in which managed burning becomes visible at high pressure. Bolivia is one dense, pervasive high-susceptibility mass, the dry Chiquitano, where designation offers little geometric refuge and Indigenous lands burn more, not less.

Figure 4. Mapper 1-/2-skeleton over susceptibility for the three zooms. The topology of the susceptible network differs by region (a fragmented ring in Guaviare, a connected web in Xingu, a solid mass in Bolivia) and mirrors the heterogeneous governance-fire outcome.

Region	Susceptibility topology	IT burn vs frontier (crude → position-adjusted)	Reading
Guaviare (wet NW)	fragmented ring around wet core (b0≈25)	0.05× → 0.27×	protective, ~55% position
Xingu (cerrado)	connected susceptible web (b0≈4)	0.55× → 0.52×	neutral, with managed-burn crossover
Bolivia (Chiquitano)	one dense high-S mass (b0≈20)	1.63× → 1.99×	reversed: IT burn more

The methodological contribution is to treat topology on both sides of the causal question: as the outcome (mosaic grain and refugia) and as a control for spatial spillover (fire crosses boundaries, which violates the usual independence assumption). A network spread-centrality is a principled way to absorb that interference.

4. Brazil-nut resilience: the same lesson, and a boundary signature

Brazil nut (Bertholletia excelsa) is a fire-sensitive, long-lived keystone, which makes fire the natural bridge to the resilience of these landscapes. Merging the Composite Resilience Index for 1,715 governance polygons with fire exposure and position reproduces the position story: differences in resilience between governance types are roughly 40 percent explained by where the land sits, echoing both the fire result and the propensity-matched finding of the resilience study. The robust internal signature is geometric: boundary fractal irregularity predicts within-polygon resilience instability (partial correlation 0.11 after controlling for area, p about 1e-5), consistent with edge-driven degradation. Honestly, at the coarse polygon-mean scale, fire exposure itself is not yet a strong direct predictor of resilience once polygon size is accounted for, which points to finer fire data and fire-sensitive sub-indicators as the next step.

Figure 5. Brazil-nut resilience merged with fire-governance topology. Governance effects are largely position; boundary geometry is the robust internal signature; the direct fire-to-resilience link is confounded by polygon size at this resolution.

5. Temporal onset: 2024 as an extreme anomaly on a rising trend

Across 2012 to 2024, fire in the Amazon is climate-pulsed, not steady, and 2024 stands apart. The record 2023 to 2024 drought roughly doubled burned area to about 19 million hectares and pushed fire-size inequality to its series maximum in every regime. Read topologically, the burn mosaic reorganized that year: it fragmented into more clusters and trapped more interior refugia across the basin, and three independent topological summaries jumped together by about two standard deviations, in 2024 alone. A simple test suite is careful about what that means. A Grubbs test flags 2024 as a significant outlier in burned area (p about 0.003, a robust five standard deviations above the median), and a Mann-Kendall test finds a significant rising trend across the whole period that persists even when 2024 is removed (p = 0.02). But the near-real-time record shows 2025 and 2026 falling back to about 40 percent of the 2024 level, so this was an extreme, drought-driven pulse on a climbing baseline, not a persistent shift to a new state. Notably, the usually quiet regimes, Indigenous territories and national parks, surged the most relative to their own baselines, meaning the extreme year eroded exactly the relative advantage they hold in normal years.

Figure 6. Temporal topology, 2012 to 2024. The CROCKER panels and the onset detector isolate 2024 as a coordinated topological anomaly; panel D shows protected and Indigenous regimes reorganizing the most in the drought.

2024 was also not a bolt from the blue, it was the peak of a multi-year drought. Reading the soil-moisture record, the accumulated deficit turned negative around 2020 to 2021 and deepened to a trough in 2024, so the extreme fire year capped a 2021 to 2024 drying. Annual burned area is predicted both by the concurrent season's dryness (Spearman -0.83) and by the accumulated 24-month soil-moisture deficit (-0.72), and 2024 is a larger outlier on the accumulated measure (about three standard deviations dry on the trailing 12 months) than on the single season (about two). The drought had memory, which helps explain both the intensity and why the usually protected lands lost their buffer.

Figure 7. Accumulated drought. The accumulated soil-moisture deficit (left) troughs in 2024 after a multi-year decline; accumulated dryness predicts annual burned area across 2013 to 2024 (right, 2024 circled).

6. Weekly topology and the early-warning outlook

The weekly persistence vineyard is the sub-annual version of that detector. Each dry season the fire mosaic fragments (Betti-0 rises) and traps refugia (Betti-1 rises) at fine scales; the 2024 signature is enormous, 2025 is a muted echo, and the current 2026 season is still dark. As the season breaks, this panel brightens in real time, which provides a topological onset signal with lead time over simple counts.

Figure 8. Weekly persistence vineyard (FIRMS, 2024 to 2026). The fine-scale bands light up each dry season; 2026 sits pre-onset at the right edge and will be the first thing to move when the ramp begins.

Operational outlook. Watch weeks 30 to 34 in the southern basin. The monitor can be refreshed weekly from the same export, and the onset detector flags the ramp as it happens rather than after the year closes.

Caveats. Historical burned-area (MCD64A1) has an approximately 21 hectare detection floor and misses small cultural burns. The weekly signal uses MODIS FIRMS at 1 km and counts connected fire clusters, not raw detections; a switch to VIIRS at 375 m would improve early, small-fire sensitivity. The spread-centrality control uses a precipitation-based dryness proxy rather than a full fire-history-free potential field, and all inside-versus-outside comparisons are correlational. Figures and results are reproducible from the archived scripts.

7. Soil moisture as a leading indicator (result)

Because moisture, not fuel, is the limiting factor for fire receptivity in these systems, the moisture field is the natural place to look for an early warning that leads fire. We tested it directly with weekly ERA5-Land soil moisture (2012 to 2026) against the FIRMS fire record, computing the topology of the dry set (Betti-0 dry patches, Betti-1 moist refugia, and the percolation of the dry field) alongside the bulk signal. In the southern basin, soil moisture leads fire by about four to six weeks: weekly fire correlates with soil moisture most strongly not at zero lag but at a lag near five weeks (r about -0.84), so how dry the soil is today forecasts fire roughly five weeks out. The dry-set percolation, the topological measure, leads longer still (about eight weeks) but more weakly at this ten-kilometre resolution, where the bulk moisture is the stronger signal. The onset instant is coincident; it is the moisture level and trajectory that lead. Independently of any fire data, 2024 had the lowest dry-season soil moisture of the entire record, the drought signature confirmed straight from the soil.

Figure 9. Soil-moisture dryness as a leading indicator (southern basin). In 2024 and 2025 the soil dries weeks before fire ramps; the cross-correlation peaks at a lag near five weeks; 2024 was the driest dry season of 2012 to 2026. Two seasons of overlapping weekly fire, so indicative rather than definitive.

The signal is not uniform, and that is the point: it is strongest exactly where fire is moisture-limited. Split by region, the five-week lead is strong in Bolivia's dry Chiquitano forests (r about -0.82) but collapses to weak and coincident in the wetter, more human-ignition-driven Guaviare and Xingu. Soil moisture forecasts fire where moisture is the binding constraint, and not where deforestation and ignition timing dominate, the same necessary-conditions logic that runs through this whole report. As of week 28, the 2026 southern basin is tracking near its climatological normal, not primed like 2024, so the soil is not yet signalling a large early-August fire.

Figure 10. 2026 nowcast (southern basin). Soil moisture and its dryness anomaly are near normal, not the sustained deficit of 2024; the fire season is pre-onset. Given the roughly five-week lead, watch weeks 30 to 34.

Forward work. The next step is to strengthen this from two seasons to the full record by reconstructing historical burn dates, to cross-check with observed SMAP soil moisture, and to watch the Mapper 1- and 2-skeletons progress through each season, a temporal skeleton showing the susceptible network assemble and percolate week by week, which is the natural moving-picture version of the static maps above.

FAIR Digital Object — Capybara Center EWS
title: Amazon Fire, Governance Topology, and the 2026 Season Onset (situation report)
creator: Feged-Rivadeneira, A. (Tecnológico de Monterrey)
date: 2026-07-19
temporal: historical 2012-2024 (MCD64A1); weekly soil moisture 2012-2026 (ERA5-Land); fire monitoring 2024-01 to 2026-07-13 (FIRMS)
spatial: Amazon basin (RAISG governance polygons)
sources: MODIS MCD64A1; NASA FIRMS active fire; ERA5-Land soil moisture; Dynamic World; CHIRPS; SRTM; RAISG IT/PA boundaries; Composite Resilience Index (Lozano, Feged-Rivadeneira, Escobedo, Thomas 2026); Capybara v6.0 susceptibility
methods: Liebig-minimum receptivity; persistent homology (Betti-0/1) and Mapper; dryness-personalized spread-centrality control; propensity/position matching; weekly persistence vineyard; moisture-fire lead-lag cross-correlation
data-sha256: 1512c66dd7ac8784899f52010a01f31fe449b1fb01b5665791bc17b16a327ee5
license: CC BY 4.0

Capybara Center situation report. Prepared for the Early Warning System project. Institutional affiliation: Tecnológico de Monterrey. Reproducible scripts and per-unit tables archived with the project.