← All stories

Unverified. We have not been able to confirm the details of this account.

Reported The Amplifier Brilliant

AI smoke spotter flags 1,036 possible smoke columns near Barcelona in summer trial

A summer trial in Collserola Natural Park, near Barcelona, used an AI system called Pantera to spot smoke from two watchtowers. Organisers report 1,036 possible smoke columns and 150 confirmed incidents, most of them outside the park, with an alert reaching staff 3.3 minutes after detection on average. Every alert was checked by people. The figures come from the trial's own funder, with no independent review. Builders: tune for local conditions and keep humans checking alerts. Readers: watch for independent results before treating this as proven.

Bright, airy abstract of soft green rolling hills under a pale morning sky, with two slim tower lines on a ridge sending wide arcs of light across the landscape. No people, logos or text.
Illustration generated with GenerateImage.
Watch for

Watch for an independent evaluation, any miss-rate or accuracy figures, a decision to keep or expand Pantera in Collserola, and use in other parks.

An AI system watched for smoke over a natural park near Barcelona this summer. It flagged 1,036 possible smoke columns, and 150 were confirmed as incidents, organisers report.

The results cover 1 July to 17 September 2026 in Collserola Natural Park. The results came out in a press release on 14 October from Mobile World Capital Barcelona, or MWCapital. MWCapital funded the trial. It ran it with the park's managing consortium and the Barcelona Metropolitan Area authority.

The system is called Pantera. It is made by Umgrauemeio, a Brazilian company, and has already been used in the Amazon. The park said in June that cameras were fitted on two of its watchtowers for the test.

Here is how it works. Software studies the camera images and looks for patterns that suggest smoke. It runs a second check before raising an alert. Then a specialist team looks at the image. Only after a person confirms it does the park's team get the warning. On average, it took 3.3 minutes from detection to the alert going out.

One finding stands out. Of the 150 confirmed incidents, 103 were outside the park's boundaries. That is 68.7%. Some were spotted as far as 43.5 kilometres from the watchtower. So the cameras saw well beyond the park itself.

There was a learning curve. At first the system raised many alerts that did not matter. It needed a testing period to adjust to local features such as reflections, city lights and industrial emissions. During that period, these unhelpful alerts fell by 86%, from 207 a week to 30 a week.

Pantera does not replace the people. The park already keeps watch from eight towers in and around the area. Raimon Roda, who runs the park consortium, said Pantera "complements" that system.

Now the caveats. These numbers come from the trial's funder, not an independent study. The release gives no figure for how many fires the system missed. It gives no accuracy rate. And it does not claim that any fire was stopped because of Pantera. In June, the park said the trial would measure things like detection accuracy and false alarms. Only part of that has been published so far.

The trial was a prize. Pantera won the Barcelona Horizon Award, part of the MWCapital Awards, which back new technology tested in real places around Barcelona. Organisers say the results will help judge whether it could grow or be used elsewhere.

The pattern here is an amplifier. AI did not take over the watch. It helped the watch cover a wider area, while people kept the final say on every alert.

What to sit with

A camera that kept watch helped, but people still made every call.

Worth arguing about

  • Would you trust a fire-watch system without knowing how many fires it missed?

What’s at stake

  • An AI smoke spotter helped a park team watch a much wider area.
  • People checked every alert before it went out.

What this gets right

  • Most confirmed incidents were outside the park, so it extended the watch.
  • Unhelpful alerts fell sharply once it was tuned to local conditions.

What stays messy

  • At first it raised far too many alerts that didn't matter.

Easy to miss

  • The results come from the trial's own funder, with no independent check.
  • No miss rate or accuracy figure has been published.

One thing to try What if every trial like this had to publish what it missed, not just what it found?

Has an AI ever done something you couldn't quite believe?

Good or bad. The disaster, the near-miss, the one that made you laugh, the one that turned out better than you'd any right to expect.

Tell us what happened

A person reads every submission. We publish your first name and your city — never your email or anything else. You can take it down whenever you like.