Findings · Indonesia Wildfire Evidence Report

Research findings · estimates with uncertainty

What the completed analyses found

This page contains the fitted results and their robustness checks. Read estimates as associations or predictive evidence unless a section explicitly states otherwise.

Primary finding · fire followed by land-cover change

Are fire-positive cells followed by mapped forest loss?

The registered comparison follows the same exact daily 1:4 sets into MapBiomas annual land cover. The primary outcome is loss of at least 10% of the forest still present before the index event, measured one year later.

Estimated association

Current evidence state5.9 pp95% CI 4.5 to 7.3 pp · p<0.001

The registered result is complete. Read the estimate and uncertainty as an association, not proof of intent.

Among 7,138 complete matched sets, the unadjusted ≥10% forest-loss probability was 13.7% for fire-positive cells and 4.4% for matched fire-negative cells. Adjustment narrows the difference to 5.9 percentage points.

Complete model sets7,138

Model cells11,758

Sets with outcome contrast1,732

Primary event years2015–2023

MapBiomas support1990–2024

Local annual raster2014

Publication diagnostics: selection and baseline trajectory

The primary model retained 7,138 of 12,178 temporally eligible matched sets; 41.4% were excluded by complete forest/observation-support rules. Included and excluded locations differ materially, so the estimate applies to the retained Kalimantan analysis population.

A pre-exposure negative-control interval was also positive: 2.31 pp (95% CI 1.13 to 3.50 pp). This signals pre-existing land-change trajectory or residual confounding. It does not erase temporal ordering, but it rules out a causal reading of the main estimate.

Registered robustness checks

The association remains positive at the 5%, 10%, and 20% loss definitions and at one-, two-, and three-year follow-up.

Definition	Adjusted difference	95% CI	p-value
≥5% loss · 1 year	10.78 pp	9.07 to 12.49 pp	<0.001
≥10% loss · 1 year (primary)	5.89 pp	4.52 to 7.25 pp	<0.001
≥20% loss · 1 year	2.76 pp	1.74 to 3.77 pp	<0.001
≥10% loss · 2 years	5.93 pp	4.26 to 7.59 pp	<0.001
≥10% loss · 3 years	6.10 pp	4.42 to 7.78 pp	<0.001

Where the mapped forest transitioned

These exploratory outcomes mean at least 10% of pre-index forest was mapped as the destination one year later. Holm p-values adjust the estimable destination family.

Destination	Adjusted difference	95% CI	Holm p / support
Nonforest Natural	4.683 pp	3.384 to 5.982 pp	<0.001
Rice Paddy	Not estimated	—	1 varying sets
Oil Palm	0.336 pp	0.087 to 0.584 pp	0.016
Pulpwood Plantation	Not estimated	—	6 varying sets
Other Agriculture	0.304 pp	0.007 to 0.601 pp	0.045
Mining	Not estimated	—	12 varying sets
Urban	Not estimated	—	0 varying sets
Other Nonvegetated	0.862 pp	0.468 to 1.256 pp	<0.001
Aquaculture	Not estimated	—	5 varying sets
Water	Not estimated	—	6 varying sets

What this answers—and what remains unresolved

The registered Kalimantan analysis finds that fire-positive cells were more often followed by mapped forest loss. The Indonesia province map below is descriptive context, not the fitted model domain. The oil-palm destination association is small and cannot identify deliberate ignition, actor, ownership, legality, or profit.

Why this still cannot prove deliberate plantation burning

A subsequent oil-palm class does not show that the fire was deliberately set to create a plantation.
The analysis cannot identify an individual, company, concession holder, or beneficiary.
The analysis cannot measure government mitigation, enforcement, restoration, or negligence without dated intervention data.
Kalimantan results do not establish a global pattern; global generalization requires harmonized replication.

Mechanism result · peat and pre-fire dryness

Does drier peat show a stronger fire-detection gradient?

The primary model compares cells with ≥50% versus <50% mapped peat extent inside exact daily 1:4 matched sets, while adjusting for rainfall, VPD, wind, vegetation, forest fraction, and soil moisture measured before detection.

Inconclusive result

Peat ≥50% × one-SD drier soil0.87×95% CI 0.69–1.08 · p=0.209

The interval includes no interaction, so this analysis does not establish that drier root-zone soil changes the peat-associated detection-odds gradient.

Because the interval crosses 1, this is not evidence of a reliable increase or decrease. “Inconclusive” is not the same as proving no effect.

Matched sets14,090

Fire-positive cells14,090

Mixed-peat sets2,572

Held-out top-137%

Held-out log loss1.468

Uniform log loss1.609

Required robustness checks

A significant sensitivity cannot replace the frozen ≥50% primary result.

Check	Interaction OR	95% CI	p
Exclude fallback-history dates	0.86	0.69–1.08	0.201
Peat threshold ≥25%	0.80	0.65–0.99	0.038
Peat threshold ≥75%	1.04	0.85–1.28	0.711
Locked 2024–2025 association	0.65	0.44–0.97	0.033

What this result does and does not mean

Odds ratios are within matched sets and describe detectable fire association, not absolute fire probability or burned area.
A non-significant result is inconclusive, not proof of no effect.
Static peat extent is not peat moisture, drainage state, peat depth, or current land cover.
No result from this track identifies deliberate burning, plantation expansion, profit, government effort, or human-access causality.
Global generalization requires a separately harmonized observation-denominator analysis.
One matched set was excluded before fitting because a control contained the CHIRPS missing sentinel −9999; no value was imputed or changed to zero.
The held-out model beats uniform ranking on conditional log loss, but that predictive signal does not make its coefficients causal.

Predictive robustness · prior-year satellite context

Does Earth AI add information beyond named weather and land variables?

Each 1-km cell receives a 64-dimensional AlphaEarth summary from the calendar year before its fire opportunity. The model was selected on 2018–2022 spatial folds, rehearsed on 2023, and evaluated once on locked 2024–2025 data.

Locked test passed

Conditional log-loss improvement0.20195% bootstrap interval 0.171–0.231

Prior-year satellite embeddings add reproducible out-of-time ranking information beyond the explicit covariates in these matched Kalimantan cells.

Lower log loss is better. The combined model ranked the true fire-positive cell more accurately, but an embedding dimension is not an identified physical or human mechanism.

Locked sets1,913

Explicit log loss1.467

Combined log loss1.266

Explicit top-136.8%

Combined top-146.7%

Embedding years2017-2024

Out-of-time model comparison

The named-variable model remains visible so the opaque embedding is never presented without a transparent baseline.

Model	Log loss ↓	Top-1	MRR
Explicit weather, vegetation, forest and peat	1.467	36.8%	0.605
Prior-year Earth AI embedding only	1.271	45.8%	0.672
Combined	1.266	46.7%	0.677

Leakage safeguards and interpretation boundary

The embedding year is always event year minus one; same-year AlphaEarth and post-fire features failed the automated gate.
Spatial folds purge training sets that share a recurring cell with the held-out fold.
Source: Google Satellite Embedding / AlphaEarth Foundations, CC-BY-4.0. Google Satellite Embedding dataset, produced by Google.
This is predictive validation, not causal evidence about deliberate burning, plantations, actors, motives, government performance, or the mechanism represented by any embedding dimension.