Indonesia Wildfire Analysis: Evidence-Bounded Kalimantan Fire & Land-Cover Research

A gated, reproducible geospatial research program that separates validated evidence from unsupported human-fire narratives. It combines frozen study registrations, exact daily matched opportunity sets, out-of-time evaluation, and coordinate-free public reporting. The completed Phase 3 estimates a +5.89 percentage-point association between fire detection and losing at least 10% of pre-index natural forest within one year (95% CI: +4.52 to +7.25) across 7,138 complete exact matched sets. A positive pre-exposure negative control and incomplete support rule out causal, intent, ownership, legality, profit, or government-performance claims. The aggregate report and source repository are public; raw coordinates remain withheld. A separate prior-year AlphaEarth ablation adds out-of-time predictive information only.

Proof points

  • 7,138 Complete Sets
  • +5.89pp Association
  • 1,913 Locked Test Sets

Technologies

Python, Google Earth Engine, MapBiomas, ERA5-Land, AlphaEarth, Next.js, CesiumJS, Vercel

Links

Problem

Open wildfire hotspots and descriptive maps are often treated as evidence of human intent or land transformation even when their observation denominator, exposure timing, and causal design are missing. This project had to make the known evidence useful without overstating what it could show.

Approach

Built a staged research system around frozen registrations, deterministic quality gates, exact matched opportunity sets, source hashes, and reproducible analysis scripts. Kept the public report coordinate-free, partitioned descriptive context from fitted findings, and added a bounded report assistant that validates citations and numerical claims against compact public evidence packs.

Results

Phase 3 found a +5.89pp one-year forest-loss association after fire detection in 7,138 complete matched sets, with the direction retained in registered sensitivities. The project also exposed the limits that prevent causal storytelling: 41.4% of eligible sets lacked complete support and the pre-exposure negative control was positive (+2.31pp). The primary environmental interaction remained inconclusive, while the AlphaEarth extension showed out-of-time predictive gain without causal attribution.

Indonesia Wildfire Analysis: Portfolio Research Report

1. Executive Summary

Indonesia Wildfire Analysis is a gated, reproducible research program for Kalimantan wildfire evidence. Its central design choice is to prevent incomplete sources, hotspot counts, or descriptive maps from being promoted into claims about people, intent, land ownership, or policy performance.

The completed registered Phase 3 analysis finds that fire-positive cells had an adjusted +5.89 percentage-point probability of losing at least 10% of pre-index natural forest by the one-year map (95% CI: +4.52 to +7.25) across 7,138 complete exact matched sets. This is a temporal association, not causal evidence. A positive pre-exposure negative control (+2.31 points) and 41.4% loss of temporally eligible matched-set support make a causal interpretation inappropriate.

2. Research Design

  • Frozen registrations specify outcomes, estimators, exclusions, uncertainty, and sensitivity analyses before outcome extraction.
  • Exact daily 1:4 matched fire-opportunity sets distinguish valid observed negatives from missing or unprocessed observations.
  • Public reporting is coordinate-free: raw hotspot points, private cell coordinates, and the prediction surface are intentionally withheld.
  • The evidence report separates fitted findings, descriptive maps, methods, provenance, and claim boundaries.

3. Results Kept in Their Proper Scope

  • Fire-to-land-cover association: +5.89 percentage points for one-year loss of at least 10% of pre-index natural forest; registered threshold and horizon sensitivities retain the positive direction.
  • Environmental interaction: the primary peat-by-dryness interaction is inconclusive (odds ratio 0.866, 95% CI 0.692–1.084, p=0.209); it does not establish a stronger or weaker dryness gradient.
  • Out-of-time prediction: adding prior-year 64-dimensional AlphaEarth summaries improved locked 2024–2025 conditional log loss by 0.2006 and top-1 recall from 36.8% to 46.7% across 1,913 matched sets. This shows predictive information only, not what embedding dimensions mean or why fire occurred.

4. Claim Boundaries

The research does not establish deliberate burning, responsible actors, ownership, legality, profit, plantation conversion, government performance, or an Indonesia-wide causal estimate. The human-accessibility and intent tracks remain not identifiable without a validated dated exposure or intervention design.

5. Public Surface

The production Next.js report is live on Vercel. It uses aggregate, coordinate-free evidence and a bounded explainer that validates citations and numeric claims before display. The underlying research repository is public; raw coordinate data remains withheld.