Open Forest Watch does not detect deforestation — Global Forest Watch, GLAD alerts, and Sentinel satellite feeds do that. What OFW adds is the reasoning layer above detection: structured hypothesis generation, multi-source evidence weighting, actor attribution, and regulatory flag assignment. This page documents how that reasoning works.
OFW ingests alerts from Global Forest Watch GLAD-S2, RADD, and FIRMS active fire feeds via their public APIs. Each alert carries a bounding box, detection date, and confidence level. OFW does not re-derive deforestation — it consumes existing detection products.
For each detected event, OFW queries Sentinel-2 (optical NDVI), Sentinel-1 (SAR backscatter), and PlanetScope (3m daily) composites for the event bounding box and evidence window. Evidence items are extracted as structured records with source, date, and supporting weight.
The HGE hypothesis engine analyses the spectral signatures, burn patterns, clearing geometry, and road network changes to generate a cause hypothesis with a confidence score. The cause taxonomy is: fire, logging, agriculture, mining, infrastructure. A secondary cause is generated when multiple mechanisms are present.
Spatial proximity analysis against land concession registries, corporate databases, supply chain flow maps (Trase.earth), and OpenStreetMap road networks generates an actor hypothesis. Named entities are flagged only when a spatial overlap with a registered parcel or concession can be established. All actor attributions carry an explicit disclaimer and confidence score.
Each event is automatically assessed against the EU Deforestation Regulation (EUDR), the Carbon Border Adjustment Mechanism (CBAM), REDD+ safeguards, CITES Appendix II forest species, ILO Convention 169 (indigenous rights), and applicable national forestry law. Flags are assigned based on biome, commodity linkage, and protected area status.
The overall confidence score is a weighted composite of evidence item weights, cause attribution confidence, and data completeness. Uncertainty bounds are derived from cloud cover fraction over the evidence window, number of independent evidence sources, and concession boundary positional uncertainty. All uncertainty inputs are logged in the report.
The structured JSON report is generated, signed with a SHA-256 evidence hash over all evidence item content, and published to the public report library. The hash provides tamper-evident provenance: any modification to the evidence set changes the hash. The report is permanently citable by event_id.
No tool in the market does this. GFW, RADD, and Planet tell you an alert occurred. OFW asks: given this alert and these independent evidence sources, what is the posterior probability that this specific cause explains the event? We use log-odds evidence fusion — the same mathematical framework used in medical diagnosis and intelligence analysis — to combine satellite alert confidence, spectral signatures, fire co-location, historical country priors, and supply chain linkage into a single calibrated probability estimate.
── Step 1: Satellite alert base probability ──────────────────────────────
P_base = 0.88 if alert_confidence = "highest" (multi-sensor agreement)
= 0.72 if alert_confidence = "high" (single sensor)
= 0.55 otherwise
── Step 2: Log-odds initialisation ──────────────────────────────────────
Λ₀ = log( P_base / (1 − P_base) )
── Step 3: Sequential evidence update (Dempster-Shafer inspired) ─────────
for each evidence_item eᵢ (excluding primary alert):
Λᵢ = Λᵢ₋₁ + α × log( wᵢ / (1 − wᵢ) )
where wᵢ = evidence weight ∈ [0.01, 0.99]
α = 0.35 (partial update coefficient, guards against overconfidence)
── Step 4: Posterior conversion ─────────────────────────────────────────
P_posterior = sigmoid(clamp(Λ_final, −4, +4))
= 1 / (1 + exp(−Λ_final))
── Step 5: Contextual corrections ───────────────────────────────────────
+ 0.04 if area_ha > 2000 ha (large events have fewer ambiguous causes)
+ 0.02 if area_ha > 500 ha
− 0.04 if area_ha < 50 ha (small events are ambiguous)
+ 0.05 if FIRMS fire co-located AND cause ∈ {fire, agriculture}
── Step 6: Uncertainty bounds ───────────────────────────────────────────
half_width = max(4, 18 − 3 × N_independent_sources)
CI_lower = max(20, score − half_width)
CI_upper = min(96, score + half_width/2)
(Asymmetric: upper bound is tighter — overconfidence is penalised more)OFW does not identify individuals. Actor hypotheses are spatial and supply-chain inferences, not legal findings. Named entities are flagged only when spatial overlap with a registered parcel can be established. All actor hypotheses carry an explicit disclaimer.
Tropical forest regions have persistent cloud cover. When optical satellite data is unavailable for more than 50% of the evidence window, the confidence score is automatically discounted and flagged in the report.
Brazil's CAR system is well-maintained. DRC, Cameroon, and Papua New Guinea have known boundary data gaps and positional uncertainties of up to 500 m. This is noted in relevant reports.
Regulatory flag assignment identifies potential applicability of frameworks (EUDR, REDD+, etc.). It is not a legal opinion. Operators should seek independent legal counsel before acting on regulatory flags.
The HGE reasoning methodology is published openly as a research paper and technical specification. The API is open for researchers to build on. This is the AlphaFold approach: open the science, close the infrastructure. Anyone can read how OFW reasons. Only Northflow runs the production engine at institutional scale. The methodology paper is available at doi:10.XXXX/ofw-methodology-v1.