Skip to main content

KPBS: Can artificial intelligence learn to spot wildfires?

Link to Full Article

When you’re trying to prevent a wildfire from getting big and dangerous, you need to spot it quickly and precisely.

“Currently fires are being detected by fire lookouts, who are human experts, right? Who are trained to spot fires by looking at camera feeds or by patrolling on the ground or by vehicles,” said Mai Nguyen, an artificial intelligence researcher at UC San Diego. “And wildfires are also reported by 911 calls from residents and passersby.”

The problem with relying on people to spot fires is you can miss a lot of them, especially those late at night, and when undetected they can grow out of control.

That’s why Nguyen and her colleagues at UCSD and its San Diego Supercomputer Center (SDSC) created SmokeyNet. It’s a computer program that uses artificial intelligence to spot and report fires 24/7.

Since its beginning, the system has been linked to about 100 cameras around San Diego County. They are part of a wireless surveillance system called the High Performance Wireless Research and Education Network, or HPWREN.

And SmokeyNet has just been beefed up.

Nguyen said SmokeyNet is also mining satellite detection images and figuring in weather data, which indicate how dry and windy it is. She said the speed of detection has improved.

“It went from about five minutes with just baseline, just the camera data, to a little over four minutes.” she said.

And what about precision? If SmokeyNet says there is a fire in the backcountry, is there really a fire?

“If you talk about precision, it is about 88%. That means that about 12% of the predictions are false alarms,” she said.

If that sounds pretty good, maybe it is. But SmokeyNet was also missing about 30% of the fires that actually occurred.

With fire detection, there is no technological silver bullet because smoke is still mistakable to people and computers.

“Smoke is very difficult,” Nguyen said. “It can be very transparent. It can be very amorphous, right? It has no clear borders. It can change shape.”

Suzann Leininger is an intelligence specialist with Cal Fire San Diego and she’s been working with a similar AI system, called Alert California. She acknowledges that a lot of things look like smoke, things like clouds and fog.

She has seen the false positives.

“It could be like mid-level clouds that cast shadows over the landscape. So (AI) can detect something moving and it thinks that a smoke column is going up,” Leininger said.

Alert California is the statewide equivalent to SmokeyNet, and it also has a strong connection to UC San Diego. Neal Driscoll is the chief investigator of the Alert California and a geophysics professor at UCSD’s Scripps Institution of Oceanography.

He said AI fire detections systems, like humans, have their flaws. But it’s still a vital alarm system for fire professionals.

“The AI says ‘this, from what I know, could be smoke.’ Then you have a subject matter expert say yes or no,” Driscoll said.

And Alert California has spotted fires that would otherwise have been missed. She gives one example of a recent fire in the Laguna Mountain area that the technology caught right away.

“It was in the middle of the night. So it was dark. People were sleeping,” Leininger said. “It’s a remote area and so there weren’t a lot of people around anyway. A detection occurred … We looked at the cameras. There was a response and they were able to keep the fire very small.”

Artificial intelligence is a computer program that can learn from its mistakes. Leininger said false positives are pointed out to help teach the system to detect smoke.

But for now, the only thing that can determine if a fire has begun is a human.