Racing Models, Not Opinions: How We Ran Wildfire ML R&D for Pyronear

Fri, 12 Jun 2026 00:00:00 +0000

In 2024, we wrote about building an early forest fire detector with the NGO Pyronear: cameras on antenna towers, a YOLO object detector (a fast model that draws boxes around what it finds in a single image) running on a Raspberry Pi, and real fires detected from 35 kilometers away. This post is the next chapter — and it is less about a model than about a method.

Over the past months we helped Pyronear answer one question: can a model that watches smoke evolve over time cut false alarms without missing fires? The answer turned out to be yes — the winning model raises 4× fewer false alarms than the production baseline while catching slightly more fires. But the part worth writing about is how we got there: a literature survey distilled into a shortlist, a standardized experiment harness, and a leaderboard where five candidate models — including the current production system — raced on the same frozen test set.

Smoke Is a Behavior: Inside Pyronear's Temporal Wildfire Detection Model

Fri, 12 Jun 2026 00:00:00 +0000

In the previous post we told the story of how the new smoke verifier for Pyronear was selected: a literature survey, five candidate models, one leaderboard, one winner with 4× fewer false alarms. This post opens up the winner itself — the bbox-tube temporal model (bbox for the bounding boxes a detector draws around candidate smoke, tube for the way those boxes are linked across frames) now released as a versioned package — and walks through how it judges a sequence of camera frames, stage by stage, with figures computed from real wildfire sequences.

Temporal Models on EarthToolsMaker

Racing Models, Not Opinions: How We Ran Wildfire ML R&D for Pyronear

Smoke Is a Behavior: Inside Pyronear's Temporal Wildfire Detection Model