5f4087f89cf71c51 AI sifts through video data to spot fires -- GCN | BLOGGERELITES

AI sifts through video data to spot fires — GCN

smoke from wildfires over hills and vineyards in Sonoma Valley, California (Serj Malomuzh/Shutterstock.com)

AI sifts through video data to spot fires

Sonoma County, Calif., is using artificial intelligence to help detect blazes before they become wildfires.

Twenty-five tower-mounted cameras are transmitting images via Amazon Web Services to Alchera, a South Korean company that specializes in visual AI algorithm development and deployment.  Alchera’s solution applies an algorithm to compare those images against the more than 10 million it’s been trained on to detect smoke plumes.

That’s a big change from the previous approach, which involved emergency dispatchers keeping an eye on the video feeds as they came in — a nearly impossible task given their other responsibilities, said Sam Wallis, Sonoma’s community alert and warning manager.

“Most of the time, the dispatchers aren’t looking at the screen,” Wallis said. “Of course, they have duties. They have to be answering the phone and dispatching.”

Now, the video from the 25 connected cameras is sent via the cloud to a location where a human in the loop — an Alchera employee — can verify that the algorithm has detected smoke.

“We put a box immediately around the smoke,” said Bow Rodgers, manager of Alchera’s U.S. business operations. “It quickly determines that it is smoke, not fog, not steam … and then quickly — within seconds — sends the information to a dispatch center that that latitude and longitude is a fire and needs to be addressed.”

Sonoma’s fire and emergency medical technician dispatch center, known as REDCOM, is the county’s primary recipient of these alerts. An alarm sounds in the facility, triggering dispatchers to manually look at the affected area with at least two cameras so they can triangulate the location and dispatch crews, Rodgers said.

Although dispatchers are the only ones who can act on the alerts, the messages go to about 32 email addresses and phone numbers for individuals at fire departments, the county’s Department of Emergency Management and the California Department of Forestry and Fire Protection (Cal Fire), Wallis said.

The county announced March 17 that the Federal Emergency Management Agency’s Hazard Mitigation Grant Program had awarded it $2.7 million for early detection improvements, including augmenting an existing system with AI monitoring. That existing system, called ALERTWildfire, was put in place by a consortium of public and private entities after the 2017 Tubbs Fire, which was one of the state’s most destructive. ALERTWildfire has 746 tower-mounted cameras throughout California.

Sonoma put its cameras on existing infrastructure such as radio towers built for emergency and regular communication. Many are located on high ground, which is crucial for spotting smoke; they have power and backup power and contain communications infrastructure needed to transmit the data, Wallis said.

Sonoma went live with the AI system in March, but Wallis quickly put it on hold. March and April are burn season for the county, which is home to many wineries, and as vintners and other agricultural businesses burned excess vegetation ahead of Sonoma’s summer fire season, the cameras sent alerts about those planned and permitted burns.

“It really quickly overwhelmed us because what we really didn’t take into account is how effective it would be,” he said. “On the first day, we had 18 fire alerts.”

The county resumed the AI detection May 1, when burns were no longer authorized. That month, dispatchers sent 14 responses to fires that the AI detected. Sonoma also began comparing how long it took the AI to detect a fire vs. a 911 call about it to come in.

“In the vast majority of cases, 911 beats the AI,” Wallis said. “That isn’t necessarily a bad thing, though. A lot of the people that start the fire are the ones that call 911.”

Another shortcoming of the camera and AI system is that it is only good for line-of-sight detection, he said. If a fire starts in a valley, the smoke has to get above the ridgeline before the camera can see it. “Very often, the local person is able to respond and call 911 before the smoke gets above it,” Wallis said. “But what we are most concerned with are fires that are started in rural areas where somebody may not have eyes on the fire. That’s where we think we’re going to get the most use out of this.”

Under the contract with Alchera, the county has 24 months to decide whether it will continue and expand the program. Wallis said he plans to evaluate its performance this and next fire season before making any long-term decisions.

California’s 2020 wildfire season was its worst on record, but on July 12, Cal Fire tweeted that almost 142,500 acres had already burned so far this year. That’s 103,000 more than in the same period (Jan. 1 to July 11) last year.

About the Author



Stephanie Kanowitz is a freelance writer based in northern Virginia.



COMMENTS