She left with her purse. Nothing more.

In October 2017, Scott Schellinger’s mother lost her home in the Tubbs Fire when it ravaged Napa and Sonoma counties for 23 days. His mother’s home caught fire mere hours after the blaze started and was one of more than 5,600 structures destroyed, making it the second most destructive fire in California’s history.

The Santa Rosa native recalled a phone call where he instructed his mom to pay attention because the flames were eight miles away.

“It’s not that big a deal,” he said. Twenty minutes later, his childhood home was on fire, and his mother, Sari, quickly evacuated with no time to gather her belongings.

Miles away, Schellinger’s neighborhood was evacuated within ten minutes of his mother’s as near hurricane-level winds pushed the fire more than 12 miles within the first three hours of ignition. In total, the fire scorched nearly 37,000 acres of wildland, forests and homes and killed 22 people, making it California’s fourth deadliest fire on record.

Stories like Schellinger’s played a large role in the passage of a bill in 2021 that created the Office of Wildfire Technology Research and Development at CalFire, the California Natural Resources Agency’s fire department that protects more than 31 million acres of the state’s wildlands.

As of September, the department has successfully rolled out the use of artificial intelligence to assist with fire detection. The program has been implemented at all 21 CalFire command centers. Since the system’s beta testing phase began in early 2023 at just six command centers, CalFire says 40% of fires have been detected by AI before a 911 call came in.

CalFire relies on two main sources of information for its wildfire detection, according to Marcus Hernandez, deputy chief of CalFire's Office of Wildfire Technology: 911 calls from civilians who spot a fire and a network of more than 1,000 cameras across the state.

The cameras, which are owned and maintained by UC San Diego’s program ALERTCalifornia, are typically located in the mountains or on hilltops so they have the highest vantage point possible to see over California’s vast terrain. The cameras can typically see up to 70 miles during the day and 110 miles at night, according to CalFire.

Operators watch screens in the command center. (Photo courtesy of CalFire)

Before AI, an operator sat at a desk in each command center and remotely turned cameras within their control, hoping to catch a glimpse of smoke on their screens. However, with some command centers in charge of more than 20 cameras, Hernandez said it was essentially impossible for one operator (or even multiple) to catch everything and get the coverage necessary for highly effective early detection.

That’s where AI comes in.

Now, every two minutes, each camera scans its surroundings while the AI model searches for “anomalies,” as Hernadez put it, in the footage that could be indicative of smoke from a fire.

Nathan Menth, a former fire captain at the Grass Valley Command Center, said AI was integral to his unit’s rapid responses to fires.

“The earlier the response, or the earlier that we get a response to go out,” Menth said, “that absolutely will have a positive impact in the suppression or the extinguishment of that fire.”

Click for bigger version

The AI system highlights potential smoke with a red rectangular box. At the same time, it triangulates one or more additional cameras to help CalFire officials pinpoint the fire’s exact location and spread path. The system also indicates how confident it is that each anomaly it detects is truly an unwanted fire by showing a percentage on the screen. That percentage allows CalFire officials to gauge how quickly the AI is learning.

This system allows CalFire officials to have the same “situational awareness” as if operators were still manually turning the cameras, according to Menth.

Turning the cameras to detect fires or smoke columns can be a bit of a guessing game when humans do it, according to Chris Africa, a fire captain at the Grass Valley Command Center, which primarily oversees the Nevada Yuba Placer Unit.

“With that information alone right there, that’s minimizing us having to turn these cameras,” Africa said.

In the past, when humans did it themselves, some fires could have gone undetected for long periods of time until a 911 call came in, according to Menth. When fires happen in remote areas with few people around, things can turn disastrous quickly if CalFire does not detect it and send out resources fast enough.

With CalFire’s goal of keeping 95% of fires to 10 acres or less, every second counts.

“That early recognition, early response and early suppression efforts, it will have a dramatic effect on keeping those fires small,” Menth said.

Although the system is improving every week, it’s still just barely out of its beta testing stage, which means it’s not infallible, according to Hernandez, Menth and Africa.

In fact, Hernandez openly admitted the system still struggles at times.

“It needs help in the night; it needs help with clouds; it needs help with smudges,” Hernandez said, explaining that there are times when the system will detect morning fog or dust kicked up by trucks and incorrectly alert CalFire officials to a potential fire.

These imperfections are a main reason why CalFire officials believe the new technology will likely never take away operators’ jobs for good.

“It’s just used as another tool to use to become more efficient,” Africa said. “There’s still the human nature needed to receive the information the cameras are giving and then putting it into our dispatch systems.”

The system currently can only detect what it thinks may be fires, it cannot understand fire paths or how large a fire is, meaning the operators and other CalFire personnel still have to judge these elements and react to them.

Because the Grass Valley Command Center processes more than 85,000 emergency calls a year, making it one of the busiest command centers in all of California, Menth said he is grateful for AI taking some of the burden off of the operators.

“With the implementation of artificial intelligence, it allows us our staffing hours back, but still maintain that same level, or even have a greater level, of situational awareness,” Menth said.

Essentially, AI is giving CalFire officials more time to focus on other necessary efforts in fire response while helping reduce overall response times.

Schellinger said he doesn't know if the AI system would have prevented the Tubbs Fire because of the day’s windy conditions creating a “perfect storm” of swift disaster, but he said he appreciates CalFire’s effort to stop smaller fires that grow larger more gradually, which could save people and property.

“I think there are lots of positive consequences to paying more attention to our environment,” Schellinger said. “The more time people spend thinking about this, the better off we’re all going to be.”