It’s fire season in California. 100,000 people have faced evacuation in the south of the state, and many more risk losing power. Last year, more than 1.5 million acres were burned across California and 93 people lost their lives. These fires are hugely destructive and monitoring them is currently a manual process. AI could change that.
The California Air National Guard (CA ANG) is one agency that does this work. For years it has relied on a team of analysts who sometimes work round the clock to interpret aerial footage of fires, so that the perimeter of the infernos can be continuously mapped. That vital information is then passed on to first responders.
Videos by VICE
But CA ANG is hoping to get AI to spot the wildfires instead, because it’s way quicker than humans.
“It’s a lot faster than the current process, which is looking at the video, comparing that to a map then drawing it onto a map,” explained Devaki Raj at CrowdAI, the firm that built the new automated system, dubbed FireNet.
“Imagine the time savings between automatically getting this done at 20 frames per second and updating these maps,” she added.
In a paper about FireNet published on the arXiv preprint server, the research team (which included CA ANG staff as well as the Department of Defense’s Joint Artificial Intelligence Center, JAIC) laid out just how laborious the manual process is. “An analyst must manually point-and-click for hours-on-end to create a fire perimeter polygon, often while working around-the-clock shifts,” the authors wrote.
Wildfires don’t just hang around. They can spread at speeds of seven to ten miles per hour in forests and even faster over grassland. That means that, in the two hours it might take an analyst to do their work, a fire could end up 20 miles or more from where it was when the video was captured. An AI in the pipeline could lead to faster response times.
The team’s tool was inspired by AI systems developed to analyse medical images, for example to mark where a tumour ends and healthy tissue begins in a medical scan. In this case, the team trained their model on thousands of frames from videos of wildfires. The resulting system can identify the edge of a fire in infrared footage of a burning forest.
Crucially, FireNet runs at a high frame rate–up to roughly 20 frames per second with 92 percent accuracy. This means that it can plot the wildfire perimeter in near-real-time, as live video from aircrews is beamed to National Guard analysts.
Data on the fire’s location is extracted from the video feed, and latitude and longitude information is collated so that accurate maps of the fire’s current position can be generated for first responders. The system is also able to automatically flag hotspots of fire that break out away from the main perimeter.
“We can tell them major off-shoots that are much more severe than what they had anticipated,” said Raj.
The system is currently being “actively tested,” the paper authors confirmed in a statement via a spokesperson for the JAIC.” They added that the results have been promising so far. “Analysts can automate the fire perimeter situational awareness as quickly as the sensor can reasonably view the fire,” they said. According to the JAIC spokesperson, “the CA NG is the end user and operator of the fire line detection aspect of this initiative.”
A rapid response system like FireNet is likely to be of great use to agencies responding to wildfires as they happen, according to wildfire expert and climate change scientist Dominique Bachelet at Oregon State University.
“We need all the tools in the toolbox to try and manage them or at least get people out of harm’s way,” she said via phone.
However, she notes that computer systems aren’t infallible: “This is going to be a life-and-death situation so they better have machines with great backup.”
Like many machine learning systems, FireNet is being designed to supplement, not replace, human efforts. Even with AI in the mix, the existing manual process will still be there as a backup, according to the JAIC spokesperson.
