Skip to main content
All projects

Wildfire monitoring, mapping & AI

Prototype on standby

FeuScope

AI-assisted automated monitoring and time-based mapping of the Saumos / Le Porge wildfire.

Visit the site

12 759

AI tasks executed

9 task families · 2 models observed

≥362,4 M

tokens documented

340.45M input · 21.99M output · cache included

≥3 875

Web searches launched by AI

Across at least 1,134 tasks

8 103

thermal observations indexed

3 source families · 7 platforms · 4 instruments

Experiment report

FeuScope

screenshot 1 / 10

Experiment report

A documented, audited and traceable system

This report separates what FeuScope actually processed, the work completed by the AI agent and figures from the monitored wildfire. The volumes come from the prototype’s consolidated archives. When detailed records for older tasks are no longer available, values are presented as minimums.

Consolidated prototype report · Replay from 22 July at 10:00 UTC to 10 August 2026 at 22:30 UTC · ADS-B positions archived through 12 August

FeuScope turns the flood of information surrounding a major wildfire into a clear, live map that is easy to follow. During the Saumos / Le Porge wildfire, it brought official alerts, affected areas, possible flare-ups, evacuations and returns, road closures, reception centres, weather, air quality, and observed plane and helicopter movements into one place. Behind the scenes, an AI agent monitored approved sources, spotted new developments, cross-checked reports, and removed duplicates before updating the map and live summary. Every item kept its timestamp and original link, so people could quickly understand what had changed and verify the source in one click.

Features

Automated monitoring with safeguards

Deterministic collectors archive and version pages, feeds, PDFs and datasets. AI extracts facts, detects changes, searches approved sources for corroboration, and prepares summaries and translations.

Mapping and timeline replay

A timeline synchronises the map, summary, figures, roads and reception centres to replay changes in hotspots, perimeters, evacuations, returns and restrictions.

Environmental and aerial context

Wind, rain, temperature, smoke and air quality provide context without claiming to predict fire spread. FeuScope also shows aircraft and helicopter movements detected through ADS-B. The signal shows where they travelled, not what mission they were performing.

Provenance and graduated validation

Each event keeps its URL, document version, excerpt and distinct dates. Media reports require corroboration or review; individuals, new sources and sensitive locations require human approval.

Public interface and supervision

Seven public views, timeline replay, French and English, and an installable PWA; an internal console tracks sources, processing cycles, reviews, tokens, duration, cost and system health.

AI and autonomy

The agent searched for sources, read documents, extracted facts and prepared updates. Detailed timings remain available for 8,776 tasks; 3,983 older tasks no longer retain individual details.

≥164 h 10

of documented cumulative AI compute

Minimum documented value · parallel execution

100 %

successful document extractions

2,263 successful extractions out of 2,263 documented

43,8 s

median duration per task

Across 8,776 tasks · P95: 3m 03s · maximum: 8m 02s

135

new sources proposed by the agent

76 organisations · 4 approved · 131 awaiting human review

Memory and traceability

Every fact remained linked to the document and excerpt that supported it. Memory distinguished new information, changes, confirmations and duplicates.

2 263

document versions archived

2,224 distinct contents · 747 contributing sources · 17.93MB of text

4 100

evidence excerpts retained

From 1,001 documents and 311 sources

1 862

operational facts linked to evidence

1,800 facts, or 96.7%, from official provenance · 212 confirmed by multiple sources

3 358

replayable transitions

1,862 new items · 740 changes · 565 confirmations · 191 duplicates

Replay and field data

Replay paired a map and a summary with every ten-minute slot. Raw archived volumes are kept separate from the lighter payload sent to the browser.

2 812

replay time slots

One slot every 10 minutes · 5,624 map and summary files

8 103

raw thermal observations

7,471 NASA FIRMS · 561 Copernicus active fires · 71 MTG observations

9 605

Copernicus geographic objects

4 map states · 11 additional FR-Alert sectors

23 974

ADS-B positions archived

19,864 aeroplane positions · 4,110 helicopter positions · 457 journeys · 163 ICAO identifiers

Monitoring, summaries and human control

The system recorded 3,346 editorial decisions, prepared 652 drafts, sent 2,583 items to review and tracked 1,254 corroboration cases.

617

media-monitoring items

341 videos · 44h 30m · 322 transcripts ready

90,6 s

median time to transcript

94.4% of videos ready · 159 within 5 minutes of publication

254

summaries recorded

246 valid bilingual summaries · 8 invalidated · 7 incident days

66

human decisions

2 people · 40 approvals · 25 rejections · 1 item marked obsolete

Monitored case: Saumos / Le Porge

These dated and sourced figures describe the wildfire, not FeuScope’s performance. Monitoring also included 45 municipality records, 14 reception centres, more than 6,000 beds and 500 accommodation offers.

42 000 ha

affected by the wildfire

≥224 000

people evacuated

More than 224,000 cumulative evacuations · 208,000 returns authorised by 2 August · 32 municipalities by 4 August

2 750

firefighters mobilised

1,500 military · 1,440 police and gendarmes · 450 volunteers · 18 aerial assets

240

homes destroyed

88 firefighters injured · no civilian deaths recorded in the retained report

Sources and validation

Memory listed 1,811 references from 722 organisations. Of these, 652 could be collected automatically, including 649 HTML pages and 3 PDFs, while 1,159 required manual handling. In total, 747 sources actually contributed to the corpus. Of 1,862 facts, 1,798 were validated automatically, 29 manually and 35 remained pending.

Mapping and environment

Replay froze 472 operational snapshots representing 144 genuinely distinct states. To follow conditions, FeuScope archived 525 hourly AROME horizons over two grids, or 44,100 cell-hour forecasts, together with 18,144 CAMS cell-hours and 362 distinct weather observations. The optimised public payload contained 7,497 detections and 4,354 generalised geometries, fewer than the raw archived volumes.

Videos and information freshness

Across 341 videos representing 44h 30m, 322 transcripts were ready. Of these, 159 were obtained within five minutes of publication and 297 within five minutes of discovery by FeuScope. The system retained 1,002 transcript corrections across 331 documents and recorded 15,763 retrieval or verification attempts.

Aeroplane and helicopter tracking with ADS-B

ADS-B is a radio signal through which aeroplanes and some helicopters broadcast their position. FeuScope archived 23,974 positions between 22 July and 12 August, then grouped them into 457 journeys linked to 163 technical ICAO identifiers, 181 call signs and 155 observed registrations. The data shows movements detected around the area, but cannot establish that the aircraft were involved in wildfire operations.

Main sources

  • Copernicus EMS
  • NASA FIRMS / VIIRS
  • EUMETSAT / MTG
  • FR-Alert
  • Préfecture de la Gironde
  • SDIS 33
  • ARS Nouvelle-Aquitaine
  • Bordeaux Métropole
  • Météo-France / AROME
  • CAMS / Open-Meteo
  • ADSB.lol

Technologies

  • React 19
  • TypeScript
  • Next.js / vinext
  • MapLibre GL
  • Node.js
  • PostgreSQL / PostGIS
  • Docker
  • Nginx
  • Cloudflare
  • OpenAI

FeuScope is an information and research prototype. It does not replace alerts and instructions from authorities and emergency services, or their operational maps and tools. A thermal anomaly does not precisely locate the fire front, and weather, smoke or air-quality data does not predict its spread. Compute time, token and Web-search volumes are minimums because 3,983 older tasks no longer retain individual details. The 96.7% rate describes the official provenance of facts, not their accuracy. Raw archived volumes differ from the simplified objects sent to the browser. Finally, ADS-B shows observed movements without confirming the mission of aeroplanes or helicopters.

1 / 10