Comment l’IA permet de détecter plus rapidement les incendies

In 2022, 72,000 hectares burned in France. Nine out of ten fires are of human origin (construction sites, agricultural activities, electric cables, cigarette butts, barbecues, vehicle fires, etc.).

In France, the majority of fires are spotted by hikers or motorists who alert the firefighters. But, as they are not trained for this risk, they often transmit information that is not very detailed, which penalizes the firefighters to precisely locate the place and adapt the means.

To be more responsive while limiting false alerts as much as possible, the Pyronear association has developed a solution based on AI. It is capable of detecting the start of a forest fire in just a few minutes.

Precious time, because after around ten minutes, it becomes very difficult to stop a forest fire! To be treated effectively in areas where the fire risk is high, a fire must have covered less than one hectare when first responders begin to fight it.

Images of Chile

Initially, these volunteers supplied their algorithm with American data (130 cameras are freely accessible), because in France no data is public.

“Our model analyzes the images and if it indicates that there is a fire, we keep the image, because there are two possibilities. Either there really is a fire or it’s a false positive. In both cases, this is interesting data which is annotated to optimize our solution”explains Mateo Lostanlen, one of the co-founders of the association.

For this engineer, AI can be very useful in quickly spotting a plume of smoke signifying the start of a fire. He became convinced of the value of AI while previously working for SquareMind, a company that automatically detects skin cancer using photos of moles.

Today, the association can rely on images from nine towers, or 36 cameras (Ardèche, Gironde, etc.). It can also train its model using images from Chile and Catalonia, following respective partnerships with researchers and firefighters (who have 17 cameras). “It is essential for our model to be confronted with many types of terrain”specifies Mateo Lostanlen.

After three years, the results are positive. “Since this summer, we have been satisfied with our results, thanks in particular to the Ardèche firefighters who placed their trust in us. We didn’t miss a single fire. But, we still have a few too many false positives. For the moment, this is not very penalizing, because it only concerns two tours in Ardèche. But, when our system covers an entire department, it will be impossible to manage many false positives”he emphasizes.

Prediction

Two avenues for improvement are planned. “We are working to develop a protocol to validate an alert more quickly and this winter, we will study temporal detection. Currently we take a still image to make a prediction. But even with the naked eye it is difficult to distinguish between a smoke and a cloud that is low. However, if we look at a series of images, it is easier to make this prediction”specifies Mateo Lostanlen.

Benefiting from public and private support, this association has joined the Citizen Initiatives Accelerator program of the interministerial digital department. It plans to provide a set of public fire data once a year. This dataset will be a combination of all the association’s sources (France, United States, Chile, Spain). Wishing to remain an association, Pyronear plans to hire a team from next January in order to better manage relations with its various partners and lead the community of volunteers.

Other variations could be considered such as the detection of floods, heavy floods or even avalanches.