When a wildfire started in California鈥檚 Kern county recently, firefighters were able to reach the site in their fire trucks even before someone called 911. That speedy response was thanks in part to artificial intelligence that spotted the early telltale signs of smoke on a live video feed.
In a year of record-breaking wildfires in North America, California is turning to an AI tool that could speed up wildfire detection and help prevent blazes from growing out of control. The ALERTCalifornia program is testing how the AI could help six of its command centres more quickly identify and snuff out wildfires. The tests involve asking the AI to look for signs of wildfire on live streams from a state-wide network of 1032 high-definition cameras, which are deployed in pairs on poles, towers and mountaintops.
鈥淭he biggest success of this program is the fires you never heard of,鈥� says , head of the ALERTCalifornia program at the University of California, San Diego.
Advertisement
If the trial 鈥� which began in July 鈥� proves successful, the AI tool could eventually become available to more than 2500 firefighters who regularly log onto the to check the views from the cameras. Each camera can sweep the entire horizon every 2 minutes while seeing up to 95 kilometres on a clear day and 190 kilometres on a clear night. Firefighters registered as watchstanders can also manually point cameras towards locations of interest.
The AI complements watchful human eyes by continuously analysing the camera feeds for signs of smoke. It puts a red box around possible fires in the videos and includes a percentage reflecting its confidence that a wildfire has been identified. It even lets a watchstander know which other nearby cameras in the network would offer a clearer line of sight for verifying the possible fire.
Once a human watchstander has visually confirmed the presence of a wildfire, they can manually trigger an alarm that sends firefighters racing to the scene. If the watchstander decides that the AI made a mistake or if the fire is an intentional burn, they can provide that feedback to help the AI learn from the incident.
After the 2023 fire season ends, ALERTCalifornia will evaluate whether the AI helped to speed up fire responses without wasting too much time through false alerts. The California Department of Forestry and Fire Protection (Cal Fire) has already invested more than $20 million in the overall ALERTCalifornia program over the past four years.
Fire-spotting AIs could be especially useful 鈥渋n open areas where there鈥檚 no people to call Cal Fire鈥� or for monitoring long utility lines where equipment can spark fires, says , a fire adviser for the University of California Agriculture and Natural Resources. He also notes how AIs could automate analysis of live data from patrolling drones and orbiting satellites.
Many start-ups have begun offering similar AI-powered wildfire detection services in recent years. But unlike those companies, ALERTCalifornia is a publicly funded program providing free, open-source data and tools for anyone else to use 鈥� including members of the public who want to check the livestreams when deciding whether to evacuate from a threatening wildfire.
Despite the turn towards AI-powered surveillance, many US states continue to rely heavily on human lookouts in wilderness watchtowers. The current limitations of AI-powered computer vision mean that people are also needed to verify the AI alerts 鈥� although AI performance may improve with more human feedback.
鈥淭here might be some false alarms,鈥� says Carmignani. 鈥淏ut it鈥檚 better to have false alarms than seeing a fire later.鈥�
