IEvery morning, California’s top firefighters get a forecast for the day in terms of wildfires: when the wind will shift, how dry the ground is and a host of other factors that could start or spread a fire.

Nowadays, an extra step has been added to the routine: checking a machine’s opinion.

“If we know that there is going to be increased fire activity or more intense fire activity in a certain area, we can use (an AI program) to validate that that area is indeed purple,” meaning the highest alert level, Chief Phillip SeLegue explained.

If the AI ​​approves, “we will supplement any response we initiate with additional resources.”

Just how dramatically AI can change firefighting is already evident in the response to wildfires. In early July, SeLegue was battling a fire in the Los Padres National Forest, north of Los Angeles.

A few years ago, when a 911 call came in reporting a fire starting or spreading, an analyst would rush to predict the path, “go in there and collect all these different characteristics. What’s the fuel? What’s the weather?” SeLegue said, listing a half-dozen different factors. This took “anywhere from a couple of hours to several hours,” depending on the person doing it.

Now, that entire process is automated. Each 911 fire alert, once dispatched, also generates an AI prediction in “about 18 to 20 seconds,” SeLegue said, with unlimited new reports created on demand.

“We probably used them … 12 to 14 times just this morning,” he said the day after the Los Padres fire.

After the catastrophic wildfire seasons of recent years, the pressure is on to stay ahead of the fires. Predicting their spread is becoming increasingly important, but also an even more difficult task: predicting days in advance when a major fire is likely to break out.

First, you need to know how much flammable vegetation there is on large areas of land, and second, how dry it is. The third factor, and the hardest to predict, is a spark of ignition, which can be something man-made, like a cigarette butt or a lightning strike. And all that data has to be collected every day, because the weather forecast changes.

Humans can do this. But AI seems to be able to do it better, crunching massive data sets to predict wildfires with reasonable accuracy, even a week or even 10 days before they start.

“We can’t use the typical algorithms to do this kind of analysis. The amount of data is huge… you need more power,” said Adrián Cardil.

Cardil is a scientist at Technosylva, a company that supplies California with the AI-based program that SeLegue uses, and similar versions in several other U.S. states, along with Chile, Spain, the Netherlands and other places.

His team first had to create an accurate portrait of California’s scrub and timber forests. Lidar, a very high-resolution sensor operated by an airplane or drone, provided 3D maps of thousands of acres with as many as 500 data points per square meter. “It’s amazing,” Cardil said. “You can even see the leaves.”

AI mapped the remaining 60-70% of the state. By analyzing the land viewed with Lidar, it could understand what vegetation was present elsewhere, but only captured in lower-quality imagery. Technosylva then used a rigorous verification process to ensure the AI ​​got it right.

From there, they can process weather data and run models every day to calculate the moisture in the vegetation, Cardil said. “If the plants are drier, the ignition and spread of the fire will be easier.” Added to that is the chance of ignition.

Technosylva’s work is part of a wave of new fire modeling around the world that uses AI and approaches each project differently, with the same three factors: fuel, weather, ignition. Many aren’t yet operational, but their creators expect them to be in the next few years.

For example, the U.S. Fire Department is tasked with maintaining a fuel map for the United States. They do this in 100-foot by 100-foot squares, about the size of two basketball courts.

AI and Google Earth imagery are used to achieve even higher resolutions in a single project, down to tens of centimeters. This allows scientists to record bare soil and rocks between plants, which can serve as natural fire barriers.

“We really need very precise information about the spatial patterns of even blades of grass,” said Greg Dillon of the fire department.

“And the more data you get, the more you need machine learning and AI-like classifications to make sense of it.”

The third factor, ignition, poses other problems, for obvious reasons. “One of the hardest things to predict is lightning-caused fires,” said Piyush Jain, a Canadian government scientist.

A wide range of scientists are experimenting with using AI to predict lightning, including those from several US federal agencies – NASA, the US National Oceanic and Atmospheric Administration (NOAA), and the US Fire Administration – as well as academics and private companies around the world.

In many of their projects, machines crunch years, sometimes even decades, of weather data to find clues about the causes associated with lightning strikes.

For example, one NOAA AI-based model predicts lightning across the entire U.S. for the next hour. It’s building on that with a new project to predict lightning that poses a risk to wildfires, specifically “dry” lightning without rain.

A firefighting model aims to predict fire-causing lightning strikes for every 20-square-kilometer area in the U.S. one week in advance, using a statistical model based on 25 years of hourly lightning data — a massive data-crunching effort made possible, again, by AI.

Yet one of the world’s most ambitious fire forecasting projects abandoned the idea of ​​focusing on a single risk factor and instead zoomed out to the entire planet, testing the limits of AI. Scientists from the European Centre for Medium-Range Weather Forecasts (ECMWF) set out about a week in advance to predict wildfires around the world.

Working at this scale, they avoided mapping vegetation on the ground, as it is not consistent across the globe. Instead, they took basic land classification maps, which show, for example, whether an area is evergreen or savannah-like, and fed their AI program with satellite data that Co₂ levels in the air.

“That will tell us … how active the vegetation is,” said Joe McNorton, a researcher at ECMWF. In other words, it measures carbon sinks and infers how much fuel is available to burn in a given area.

Their program also feeds global satellite weather data, and the AI ​​was trained by looking for huge fires that could be spotted by satellite. Now it produces one global wildfire forecast per day, for the next 10 days, with a resolution of nine kilometers around the planet. Surprisingly, even for those who created it, it seems to work for this largest category of fires. It was able to predict last year’s Canadian wildfires about 10 days in advance, McNorton said.

In California, SeLegue said he still doesn’t know all the details of the role AI is playing, but he doesn’t need to: “It’s baked in,” he shrugged. But what’s more important, he said, was clear: “It improved accuracy.”