Every morning, California’s top firefighters get a forecast of the day in wildfire terms – when the wind will shift, how dry the ground is and a host of other ingredients that can start or spread a fire.
Lately, the routine has an extra step: checking a machine’s opinion.
“If we know that there’s going to be heightened fire activity or heightened fire weather in a certain area, we can use [a programme with AI] to validate that yes, this area’s purple,” meaning the highest alert level, explained chief Phillip SeLegue.
If the AI gives that nod, “every response that we’re going to initiate, we’re going to augment that response with additional resources”.
How drastically AI can change firefighting is already clear in wildfire responses. In early July, SeLegue was fighting a blaze in Los Padres National Forest, north of Los Angeles.
A few years ago, if a 911 call came in reporting that a fire had started or spread, an analyst would hustle to predict its path, “to go in and pull all those different attributes. What is the fuel? What is the weather?” said SeLegue, listing half a dozen different factors. This took “a couple hours to several hours,” depending on the person doing it.
Now, that entire process is automated. Any 911 fire call, once dispatched, also generates an AI prediction in “about 18 to 20 seconds,” said SeLegue, with unlimited new reports created on request.
“We used them probably … 12 to 14 times just this morning,” he said one day from the Los Padres fire.
After the cataclysmic wildfire seasons of the past few years, the pressure is on to get a step ahead of fires. Predicting their spread is increasingly important, but so is an even harder task: predicting, days in advance, when a big blaze is likely to break out.
You need to know, first, how much burnable vegetation there is across vast swaths of land, and second, how dried out it is. The third factor, and the hardest to predict, is a spark of ignition, which can be something human-caused like a cigarette butt or a lightning strike. What’s more, all that data needs to be recrunched anew every day as the weather forecast changes.
Humans can do this. But AI seems to be able to do it better, digesting mammoth datasets to predict wildfires fairly accurately as much as a week or even 10 days before they start.
“We cannot run the typical algorithms to do this type of analysis. The amount of data is huge … you need more power,” said Adrián Cardil.
Cardil is a scientist at Technosylva, which provides California with the AI-informed programme that SeLegue has been using, as well as similar versions in several other American states, in addition to Chile, Spain, the Netherlands and other places.
His team first had to get an accurate portrait of California’s shrubland and timber forests. Lidar, a very high-resolution sensor operated by plane or drone, provided 3D maps of thousands of acres with as many as 500 data points a square metre. “It’s amazing,” said Cardil. “You can even see the leaves.”
AI mapped the other 60-70% of the state. Analysing the Lidar-viewed land, it was able to understand what vegetation was present elsewhere but only captured in lower-quality images. Technosylva used a rigorous verification process afterwards to ensure the AI was getting it right.
From there, they are able to incorporate weather data, running models every day to calculate the vegetation’s moisture, said Cardil. “If the plants are drier, the ignition and the fire spread are going to be easier.” Added to that is the chance of ignition.
Technosylva’s work is part of a wave of new fire modelling around the world drawing on AI and grappling each project in a different way, with the same three factors: fuel, weather, ignition. Many are not yet operational, but their creators expect them to be in the next couple of years.
The US fire service, for example, is tasked with maintaining a fuel map for the United States, which it does in increments of 30 by 30 metres, about the size of two basketball courts.
It uses AI and Google Earth images to go even higher-res in one project, down to tens of centimetres, allowing scientists to register bare ground and rocks between plants, which can serve as natural fire-breakers.
“We need really fine-resolution information about the spatial patterning of even just grass clumps,” said Greg Dillon of the Fire Service.
“And the more data you get, the more you need machine learning and AI-type classifiers to make sense of it.”
The third factor, ignition, presents other problems, for obvious reasons. “One of the hardest things to predict is lightning-caused fires,” said Piyush Jain, a Canadian government scientist.
A whole host of scientists are experimenting with using AI to predict lightning, including those at several American federal agencies – Nasa, the US National Oceanic and Atmospheric Administration (NOAA), and the US fire service – as well as academics and private companies around the world.
Many of their projects ask machines to process years, sometimes decades, of weather records in order to find clues about what tends to accompany lightning strikes.
One NOAA AI-based model, for example, predicts lightning across the entire US for the following hour. It is building on this with a new project meant to predict lightning that poses a wildfire risk, especially “dry” lightning unaccompanied by rain.
A fire service model aims to predict fire-starting lightning for each 20-square-kilometre patch of the US a full week ahead of time, using a statistical model based on 25 years of hourly lightning data – a massive data-processing effort that, again, was made possible by AI.
Still, one of the world’s most ambitious fire prediction projects abandoned the idea of zeroing in on one risk factor, and instead, zoomed out to the whole planet, testing the limits of AI in the process. Scientists from the European Centre for Medium-Range Weather Forecasts (ECMWF) set out to predict wildfires anywhere on Earth about a week ahead of time.
Working on this scale, they eschewed on-the-ground vegetation mapping, since it is not consistent around the world. Instead, they took basic land classification maps, showing if an area is evergreen or savannah, for example, and fed their AI programme with satellite data that measures Co2 levels in the air.
“That will tell us … how active 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 programme also feeds in global satellite weather data, and the AI was trained by looking for huge fires that could be spotted by a satellite. Now, it produces one global wildfire prediction a day, for the next 10 days, with a nine km resolution around the planet. Surprisingly even to those who made it, it seems to work for this biggest category of fire. It was able to predict the Canadian wildfires last year about 10 days ahead, said McNorton.
In California, SeLegue said he still doesn’t know all the details of the part AI plays, but he doesn’t need to: “It’s baked into it,” he shrugged. But what’s more important was obvious, he said: “It improved the accuracy.”
