High-resolution cameras scanning Colorado forests once every minute, 24-7 — linked to a smoke-detecting artificial intelligence algorithm residing in a computer cloud — are popping up on Colorado high points from Lookout Mountain near Golden to Telluride’s Ajax Peak.
“Colorado has had five years of the largest fires in its history,” Telluride Fire Chief John Bennett said. “This is a new tool to help with early detection, allowing us to try to get to fires before they become big fires.”
The early detection system was developed and is operated by San Francisco-based Pano AI, and by the end of 2023 there will be 40 Pano installations in Colorado, including ones already in Telluride, Beaver Creek, Vail, Aspen and Boulder.
The biggest investment in the technology is being made by Xcel Energy, the state’s largest electricity provider, which Tuesday announced that it has committed to installing a total of 21 stations, each with two cameras, by year’s end.
“The risk of wildfire continues to evolve … driven by climate change,” Robert Kenney, the CEO of Xcel Energy’s Colorado subsidiary, said at a news conference at Arvada Fire Station 9. Fire season, he said, has turned into “a year-round battle.”
At a cost of $50,000 a year for each two-camera unit, the five-year contract with Pano AI is valued at an estimated $5.25 million. The cost, Kenney said, will be recovered in customer rates.
Among the other Xcel Energy locations getting cameras are Breckenridge, Silverthorne and Hay Canyon, east of Rangely.
The deployment of the Pano system, Kenney said, is just part of the company’s wildfire mitigation plan, which includes managing vegetation around power lines, increased inspections, covering bare conductors and using helicopters and laser radar to check line clearances and for overloaded poles.
Xcel Energy has been under pressure since the 2021 Marshall fire, the most destructive in Colorado history razing more than 1,000 Boulder County homes and businesses and inflicting an estimated $2 billion in damage.
County officials said one cause of the fire was an Xcel Energy distribution line that had come loose in the high winds. Winds also drove the wildfire. The utility disputes the finding but is already facing lawsuits from homeowners and businesses.
On its third-quarter earnings call on Oct. 27, Xcel Energy CEO Bob Frenzel was repeatedly asked by stock analysts about the Marshall fire and wildfire risks. He said there are now 675 plaintiffs and he expects a litigation calendar sometimes early next year.
As for Xcel Energy’s wildfire planning, he said, “we are looking at more capital investment moving forward.”
In 2022, state regulators approved $23.5 million in wildfire mitigation investments as part of Xcel Energy’s electric rate case.
One of the camera units will be placed in the Marshall area, another will be located near Fort Collins and one near Golden, but fire officials said the greatest benefit will likely be in locating fires too remote to be initially detected by people.
The system is a far cry from the old fire tower with its solitary fire watcher armed with binoculars and a mechanical Osborne Fire Finder.
Pano AI’s high-resolution cameras rotate 360 degrees every minute and can see 20 miles out. They tend to be located about 10 miles apart so that a fire location can be triangulated, said Arvind Satyam, Pano AI’s chief commercial officer.
At night the cameras switch to near infrared to seek out heat signatures in the forest.
Although Pano AI integrates its data with satellite imagery, Satyam said satellites generate images less frequently and are less precise on location. “This is a way to ground truth what we are seeing,” he said.
“On a clear day I can see all the way into Utah with those cameras,” Telluride Fire Chief Bennett said. The fire district has had the cameras in operation since March.
The goal is detecting the first wisps of smoke, fixing a location and determining what resources are needed.
Once the cameras pick up an image, the algorithm assesses whether it is smoke. If it cries “smoke,” it is double-checked by a human in the company’s “intelligence center” and an alert is sent to the relevant fire agencies and utilities.
In Oregon, for example, the system has sent out alerts 15 to 20 minutes before a 911 fire call was even placed, Satyam said.
AI didn’t understand changing aspen leaves. Or snow.
The algorithm uses texture, movement and radiance to make a binary decision “smoke/ not smoke.”
The decision is binary, but it is hardly simple.
The initial “learning” by the algorithm was done on the West Coast and when it did its first pilot runs in 2021 in the Aspen Fire Protection District, with four camera arrays, it had a whole bunch more learning to do.
“It was freaking out when it saw fall colors or snow,” Aspen Fire Chief Rick Balentine said. Similarly, the algorithm has to learn: A plume of smoke or quarry dust?
Slowly the algorithm began to integrate those images into its database. The system now compares an image from the cameras to millions of logged images.
Satyam said Pano is looking to add additional intelligence, so that the algorithm might be able to assess events such as lightning strikes.
It has been a quiet fire season without any major fires in the Aspen district, still Balentine said “we found it quite helpful in monitoring controlled-burns.”
And while there were no fires in Balentine’s patch, his cameras did pick up a fire over the mountain range in the neighboring Crested Butte Fire Protection District.
This fall the Carbondale & Rural Fire Protection District, which is roughly 30 miles from Aspen, also installed a Pano AI system. “Soon we could have a whole network in the Roaring Fork Valley,” Balentine said.