Machine Learning Joins the Wildfire Fight in California
2018’s Camp Fire was the deadliest wildfire in California history, burning more than 153,000 acres, destroying 19,0001 buildings, and killing dozens. And while the dust settles and towns rebuild, it seems there’s no end in sight: scientists warn that wildfires will continue to worsen2 in the coming years due to climate change.
To combat future devastation, scientists across the public and private sectors are turning to data-driven technologies to better detect wildfires before they spread. Machine learning is one technology proving critical in providing actionable insights to help first responders identify wildfires early and come up with informed strategies to combat the tens of thousands of wildfires4 that occur in the U.S. — and burn increasingly more acreage5 — each year.
Monitoring Wildfires from Space
One of the most critical aspects of curbing a wildfire’s devastation is early detection. Some areas are equipped with lookout towers and surveillance cameras to aide in this, but first responders today rely primarily on calls to emergency services to identify wildfires. And when fires ignite in remote areas, as the Camp Fire did, that method falls short.
One organization working to detect wildfire ignitions in near real-time6 with the help of artificial intelligence (specifically, machine learning) is the Center for Spatial Technologies and Remote Sensing (CSTARS) at the University of California, Davis.
Alex Koltunov, a CSTARS scientist, is co-leading the effort. Koltunov and his team have developed an algorithm that conducts image analysis on data from the GOES weather satellites which scans approximately 25-square-kilometer areas in California (called pixels) about every 15 minutes. The algorithm, called GOES-Early Fire Detection7 or GOES-EFD, must be able to perceive incredibly subtle environmental abnormalities to detect wildfires early, such as temperature changes, while considering other factors that could affect temperature like fog, clouds or wind. To accomplish this, the team developed a detection model that uses machine learning to compare a pixel’s current conditions to past and current data from similar pixels under normal conditions. If the model detects a wildfire, it can send alerts to first responders. In early experiments, Koltunov has been able to detect wildfires before they were reported by conventional methods. Moreover, the GOES-178 satellite — deployed on February 12, 2019 — has pixels that are four times smaller and scans at five-minute intervals, meaning the CSTARS algorithm could help make early detection routine.
“Every minute counts,” Koltunov explains. “If we can detect a wildfire earlier to help issue an evacuation order, it can make a difference between people escaping the fire or not.”
Forecasting Wildfires Before They Start
Beyond containing existing fires, researchers are also exploring ways to predict wildfires before they even have a chance to ignite. The U.S. Forest Service and technology startup risQ9 are partnering to take on this challenge.
The government agency creates statistical models that produce thousands of hypothetical weather scenarios based on historical weather data, satellite images and current weather conditions. From that, the risQ team uses machine learning algorithms to develop a hybrid of statistical and physical models to determine the probability of wildfires in certain areas in a given year.
The company built a risk model in collaboration with the U.S. Forest Service that tracked the risk level of wildfires (see right). The map shows the 500-Year (meaning 1/500 or 0.2 percent) chance of seeing a certain flame length. That means at each location on the map, there’s a 0.2 percent probability of seeing a flame length equal to or greater than what’s seen here in any given year. This kind of map (and underlying data) can be useful for risk management and planning decisions.
Jameson Toole, co-founder and CEO of Fritz10, a Boston-based startup that helps developers build intelligent devices to solve real-world challenges, explains that once a high-risk area is identified, local fire departments could dispatch an autonomous drone with high-resolution cameras to survey the area and send insights back to the team.
Data from the forecast and the drone would help first responders come up with informed strategies on how to allocate resources and prepare for fires, months before wildfire season.
A Bright Future for Machine Learning
Looking ahead to the not-so-distant future, data scientists are optimistic about the opportunity to make high-tech wildfire detection systems a reality. And the massive amount of data generated each year (an estimated 90 percent of the data in the world in mid-2018 had been created in the two years prior11) will only help make machine learning models more accurate.
“We already have a lot of these systems in place — we have autonomous drones, we have a lot of IoT sensors,” says Toole. “We’re focused on making them more intelligent, more reliable and much easier to manage, but this is not science fiction anymore.”
(Blog) – Top Machine Learning Trends
(Infographic) – Facing a Forever Fire Season
This content is produced by WIRED Brand Lab in collaboration with Western Digital Corporation.
- Firefighters Reflect on Northern California’s Deadly Camp Fire After Blaze Fully Contained
- Yes, Climate Change is Making Wildfires Worse.
- Why California’s Wildfires are So Hard to Fight
- Wildfires in the U.S. are Getting Bigger
- Wildfires Have Gotten Bigger in Recent Years, and the Trend is Likely to Continue
- GOES-Early Fire Detection
- Fritz Labs
- How Much Data Do We Create Every Day?