AI Cameras Are Spotting Wildfires Before Anyone Calls 911, Often 45 Minutes Sooner
Robauto Team · Published · Last updated
AI smoke-spotting cameras in California and Arizona are catching wildfires before the first 911 call, with people confirming each alert. In 2024, 636 California fires were seen by cameras first. Here's how it works and what brands can learn.
Wildfires are easiest to stop when they are small. The hard part has always been noticing them in time, especially at 2 a.m. on an empty ridge. Across the American West, a team effort between AI cameras and human dispatchers is closing that gap. The software watches the hills around the clock and flags anything that looks like smoke. A trained person checks it, and crews are rolling before anyone has picked up a phone to call 911.
This is one of the clearest examples of people and AI each doing what they're best at. The machine never gets tired of staring at hundreds of camera feeds. The human decides what's real and sends the trucks.
The numbers so far
California's ALERTCalifornia network, run by UC San Diego with the state's fire agency and the Chico software company DigitalPath, added AI smoke detection in 2023. In its first two months, the system alerted dispatchers to 128 incidents before any 911 call was logged, and 77 of them were confirmed fires, sometimes more than 20 minutes ahead of the first call, according to data reported by the Los Angeles Times.
It has kept working as it scaled. In 2024 there were 7,553 wildfires in the state fire agency's jurisdiction. The cameras picked up 1,668 of them, and 636 of those, about 38%, were spotted by the cameras before any person called them in, as reported by Phys.org.
| Where | What happened | Source |
|---|---|---|
| California, first 2 months (2023) | 128 early alerts before 911, 77 confirmed fires, some more than 20 minutes early | LA Times |
| California, full year 2024 | 1,668 of 7,553 fires seen by cameras; 636 (38%) caught before any 911 call | Phys.org |
| Grass Valley, CA | First 911 call came about half an hour after firefighters were already on scene; fire held to a quarter acre | CENIC |
| Auburn, CA (2:33 a.m.) | AI alert with no 911 call; crews kept it to about 30 by 30 feet | Phys.org |
| Coconino National Forest, AZ (March 2026) | Diamond Fire contained before it grew past 7 acres | KJZZ / AP |
| Arizona Public Service network | Alerts arrive about 45 minutes faster on average than the first 911 call | KJZZ / AP |
Small fires, stopped small
The stories behind those numbers are the best part. Near Grass Valley, California, the AI flagged a new fire on September 11. Firefighters were already working it when the first 911 call came in about half an hour later, and they kept it to a quarter of an acre, according to CENIC, which gave the project an innovation award. During the pilot in July 2023, the system caught a fire in the Cleveland National Forest at 3 a.m., when hardly anyone would have been awake to see the smoke. Crews put it out in less than an hour.
The cameras themselves do a lot of patient work. They sit on towers, observatories, and rooftops, turning a full circle every two minutes and taking 12 photos each pass, day and night, as Phys.org describes. On July 6, one of them posted an alert at 2:33 a.m. for a fire near Auburn. Nobody had called it in. Dispatchers checked the image, sent engines, and the crew stopped the fire before it spread past a patch about 30 by 30 feet.
Arizona joins in
In March 2026, an AI camera installed for the utility Arizona Public Service spotted what looked like smoke in the Coconino National Forest. Human analysts confirmed it wasn't cloud or dust and alerted the state forestry service. Firefighters contained the Diamond Fire before it grew past 7 acres. APS meteorologist Cindy Kobold said the system notifies the utility about 45 minutes faster on average than the first 911 call, according to KJZZ, reporting with the AP. APS had nearly 40 of these cameras active at the time, with plans for 71. Maricopa County now runs nearly 50 of its own, and it keeps a person in the loop: "A live human being then reviews that to make sure that it isn't just a dust devil and that it is a wildfire," Coleman told TechInformed.
Why the human step matters
The teams are open about the limits, and that honesty is part of why this works. The software can mistake dust storms or clouds for smoke, and it had to be taught to ignore steam from geothermal fields in Sonoma County. In dense cities, people with phones often report fires almost as soon as they start, so the cameras matter most in quiet, remote country, as Phys.org notes.
Early research showed why pairing works. A 2020 prototype study in Remote Sensing found the system usually spotted smoke within 15 minutes of ignition, with about 0.66 false alerts per camera per day, at least ten times better than an older system that produced 10 to 30 false alerts a day. The researchers pointed out that it's far easier for a person to judge one flagged image every few minutes than to stare at 65 live camera feeds.
Next up: eyes in orbit
Ground cameras can't see over every hill, so the next layer is going up. FireSat, a satellite constellation from the nonprofit Earth Fire Alliance built by Muon Space with Google Research, can detect fires as small as 5 by 5 meters, about the footprint of two parking spaces. Its test satellite spotted a small roadside fire near Medford, Oregon that other satellites missed, according to Google, and collected more than one million infrared images in its first year, per Muon Space.
The first three operational FireSats reached orbit on July 7, 2026, and sent back their first active-fire images in September. Fire agencies in the program should get data at least twice a day by the end of 2026, with an hourly global revisit targeted for 2029, according to Earth Fire Alliance.
The takeaway for brands
These systems succeed for a simple reason: the signal is clear, it arrives in a format the next person can act on right away, and a human stays in charge of the decision. The camera doesn't just say "something's wrong." It sends an image, a confidence score, and a location.
AI shopping and search agents work the same way. They act fastest on information that's clear and structured: who you are, what you sell, your prices, your policies, and how to reach a person. When that information is easy for an agent to read, you get found and recommended sooner. When it's buried or missing, the agent moves on. The opportunity is to make your site as easy for AI to read as a clear smoke alert is for a dispatcher, while keeping your people in charge of the decisions that matter.
{
"topic": "AI cameras detecting wildfires early",
"human_in_loop": true,
"california_2024": {"fires_total": 7553, "seen_by_cameras": 1668, "before_911": 636, "share_before_911": "38%"},
"california_first_2_months_2023": {"early_alerts": 128, "confirmed_fires": 77},
"arizona_aps": {"avg_minutes_faster_than_911": 45, "diamond_fire_acres_max": 7},
"firesat": {"min_fire_size_m": "5x5", "operational_sats_launched": "2026-07-07", "revisit_end_2026": "twice daily", "hourly_target": 2029},
"brand_takeaway": "Clear, structured, actionable info gets acted on first. Make your site readable by AI agents.",
"cta": "https://robauto.ai/scan"
}Want AI agents to find and understand your business as quickly as these cameras spot smoke? Scan your site free at robauto.ai/scan and see what agents can read today.
Sources
- LA Times: How California firefighters are using AI to look for wildfires
- Phys.org: AI cameras are spotting wildfires across California
- CENIC: ALERTCalifornia AI tool recognized
- KJZZ/AP: AI cameras are helping catch Arizona wildfires early
- TechInformed: Maricopa County expands AI wildfire cameras
- Remote Sensing (2020): Wildfire detection with deep learning on camera images
- Google: First images of wildfires detected by FireSat
- Muon Space: First three operational FireSat satellites
- Earth Fire Alliance: First active fire images from operational FireSat
Where can an agent get this data live?
Robauto MCP — Agent Intelligence & Brand Discovery Infrastructure. A public, stateless Streamable HTTP endpoint at https://mcp.robauto.ai/mcp, callable with no key. Setup and the full tool list are in the Robauto MCP directory.