How the AI decides
Three models, each simple enough to audit. Every output carries the model version, and that version is anchored with the evidence.
Group pixels into fires
DBSCAN clusters detections that touch in space and time (1.5 km, with 24 hours counted like 1.5 km). The earliest overpass marks the ignition; the brightest pixel breaks ties. Fires already burning when the record starts are marked censored, and their confidence is halved because the true start is unknown.
Decide who it belongs to
For every polygon a fire touches: did it start there or spread in? Only a start counts. Confidence multiplies persistence (more pixels, more certain), distance from the boundary (a VIIRS pixel is 375 m wide), and how much of the fire stayed inside. The escrow ignores anything below 0.70 and always allows a dispute.
Forecast next week's fires
Gradient-boosted trees on 2019–2023 fire seasons, tuned on 2024 and tested on 2025, which the model never saw.
A satellite detection inside a licensed boundary is evidence for investigation, not proof of responsibility. Concession boundaries are Indonesian Ministry of Forestry data (2012) published by Global Forest Watch and may be out of date.