California’s firefighting agency is using artificial intelligence to scan camera feeds for signs of wildfire. During its pilot, the system detected 77 fires before dispatch centers received 911 calls, offering crews an earlier chance to investigate and respond.
Watching more than 1,000 cameras
Cal Fire’s system analyzes images from a network of 1,039 high-definition cameras, many of them positioned on mountaintops. The cameras were already used by human operators looking for smoke, a task that can be repetitive and tiring across such a large network.
The AI processes “billions of megapixels” of images every minute. The cameras cover approximately 90 percent of California’s fire-prone areas, giving the system a broad view while still leaving some locations outside its reach.
The program started in June in six of Cal Fire’s command centers. It is expected to expand to all 21 command centers in September. Its pilot detections represent about a 40 percent success rate, according to The New York Times.
Earlier alerts still need human review
When the system identifies what may be smoke, it alerts firefighters. Human operators then need to assess the warning. Engineers at DigitalPath, the California-based company that created the software, have been manually checking each fire flagged by the AI.
That review matters because many things in a camera image can resemble smoke. Fog, haze, dust raised by tractors and steam from geothermal plants have all produced false positives. Ethan Higgins, a chief architect of the software, told The New York Times: “You wouldn’t believe how many things look like smoke.”
The system can only spot fires that are visible to its cameras. It also does not determine on its own what a fire means or whether it calls for a response. The alert is an early signal for people to verify, not a replacement for their judgment.
What early detection could change
Finding a fire before the public reports it can give responders an opportunity to tackle it while it is still manageable. Phillip SeLegue, Cal Fire’s staff chief of intelligence, told The New York Times that the system has improved response times and helped identify some fires early.
For Neal Driscoll, a leader of the Cal Fire AI project, the most meaningful results may be fires that stay small enough to be extinguished quickly and never become widely known. That measure depends on more than an alert: crews must be able to verify the sighting and act on it.
Some experienced fire operators remain doubtful that AI can understand the circumstances surrounding every fire. Andrew Emerick, duty chief for Cal Fire’s northern region, pointed to fires deliberately set for agricultural purposes as an example of context that matters. “I don’t think this robot is ever going to take my job,” he said.
One part of a wider detection effort
Camera analysis joins other ways Cal Fire learns about fires, including 911 calls from residents and Fireguard, a partnership with the United States military that uses classified spy satellites, drones and other aircraft to detect fires. Each method offers another source of information for locating a blaze.
Cal Fire reported 4,792 wildfires so far this year, below the five-year average of 5,422 for the summer months. A lower count does not make any fire harmless: a blaze that grows out of control can still be devastating. The AI system’s value will depend on whether earlier alerts help responders catch fires before they spread, while human operators continue to check what the cameras show.