Machine Learning: The Next Firefighting Tool
By Brett Bralley Illustration by Yeab Kebede
Could artificial intelligence hold the key to cutting-edge wildfire prediction?
Editors' note: This is the second piece in a two-part series in Washington Square examining San José State University’s interdisciplinary approach to wildfire research. Read the first part, Fighting Fire With Fire, here.
Imagine knowing exactly what a wildfire will do, where it will be, hour by hour. When flames will leap across hillsides, when billowing smoke will darken skies, when neighborhoods will be engulfed.
With this knowledge, firefighters could be better prepared to defend against flames and smoke charging through landscapes; aerial water drops could be more targeted and effective. And residents could know exactly how much time they have to safely evacuate their homes. Better yet — responders could stop flames before they even become a threat.
For Adam Kochanski, assistant professor of meteorology and climate science at San José State University, this hypothetical world of fire prediction isn’t far-fetched. Kochanski believes the approach to better understanding this massive force of nature lies in technology — including machine learning.
Leveraging artificial intelligence for fire spread forecasting isn’t a novel concept in wildfire research, but “up to this point, we haven’t had enough data to build those models,” Kochanski explains.
But he sees a way to gather that data and harness it so that “we could get information about wildfires any time we want, to see the size of the fire, how it is progressing. We could estimate how fast the fire is going to spread, and how intense it will be.”
Kochanski is part of the university’s Wildfire Interdisciplinary Research Center (WIRC), a group of seven SJSU scientists from a number of disciplines, including ecology, social sciences, engineering and meteorology. Five of those researchers were hired between August 2020 and January 2021 — a move that established WIRC as the largest academic wildfire interdisciplinary research center in the country.
Momentum for wildfire research at SJSU has since surged. In August 2021, WIRC received a grant from the National Science Foundation designating it an Industry-University Cooperative Research Center (IUCRC).
This grant and designation have been a game changer, explains Craig Clements, professor of meteorology and climate science and WIRC director, because now, research can move forward at an unprecedented speed. Usually, the academic research process can take months waiting for funding approval.
But as an IUCRC, funding is immediately available, and research can start as projects are approved by partnering stakeholders, which, for San José State, include San Diego Gas & Electric Company; Pacific Gas & Electric Company; Southern California Edison; Technosylva, Inc.; Jupiter Intelligence, Inc.; State Farm Insurance; CSAA Insurance Group; and Lawrence Livermore National Laboratory, among others. Kochanski’s foray into artificial intelligence is one of several projects WIRC faculty are tackling that are funded by support from IUCRC stakeholders.
Just seven months later, nearly $1.2 million from President Biden’s federal spending plan was set aside this March for cutting-edge fire modeling and prediction technology at SJSU. That funding will support the development of four new facilities: a national wildfire data and computing hub, a remote-sensing laboratory, a wildfire dynamics laboratory and a community wildfire resilience laboratory.
“We could estimate how fast the fire is going to spread, and how intense it will be.”
— Adam Kochanski
These snapshots of San José fires simulations use fire progression reconstructed using machine learning. Images courtesy of Adam Kochanski.
Teaching machines to teach us
Kochanski has already completed an important step in teaching machines how to predict wildfire behavior: He co-developed WRF-SFIRE*, a wildfire forecast and modeling system that allows wildfire experts to predict which way smoke and flames will move based on current weather conditions. WRF-SFIRE is already a crucial resource that has helped curb the spread of wildfires across California and the globe.
However, the model utilizes data from satellite and aircraft observations of fires that are only available once or twice a day. And that isn’t nearly enough to teach a machine how to predict fire spread more accurately, Kochanski says. In fact, he says artificial intelligence networks “need thousands of data points” to do the job.
To gather them all, he’s relying on the past. He is combining historical satellite images of wildfires with the WRF-SFIRE model. Then, he’ll factor in regularly updated information about the current dryness levels of landscape vegetation, otherwise referred to by wildfire experts as “fuels.”
Scientists, including researchers and students from WIRC, gather this data on fuels by collecting them — sometimes even on the front lines of a wildfire — and measuring their moisture content by weighing them, drying them out, then weighing them again.
Jack Drucker, ’21 BS, ’24 MS Meteorology, is one of those students. The Los Angeles native joined Kochanski’s lab as an undergraduate and made it his senior project to create a database that aggregates fuel moisture-level findings from WIRC and other sources. The database is open for users to access based on their needs and has been key to Kochanski’s research, as well as to Drucker’s long-term career goals. He plans to continue wildfire research after he earns his master’s degree.
“The more I learn about wildfires, the more I realize there’s so much we don’t know,” he says. “Uncovering the secrets behind how wildfires develop is fascinating and drives my curiosity to learn more.”
“The more I learn about wildfires, the more I realize there’s so much we don’t know. Uncovering the secrets behind how wildfires develop is fascinating and drives my curiosity to learn more.”
— Jack Drucker
A clearer future
If Kochanski is successful in his endeavor, the effects will be profound.
“First, we can have a better forecast of what a fire is going to do, where a fire is going to be 24, 48 hours from now,” he notes.
But that’s just the beginning: He says that machine learning can give us an idea of what wildfire activity will look like decades down the road. That snapshot into the future could inform many decisions made around wildfire prevention — from where we choose to build homes, to whether utility companies should consider placing infrastructure underground, to if and how much governments should invest in preventing wildfires.
One such area the government could invest more in is better forest management, Kochanski notes — an idea shared by other WIRC researchers, including Kate Wilkin, assistant professor of fire ecology.
In general, Wilkin says “we have chosen a hands-off land management approach to many of our forests and woodlands, and what we’ve ended up with are really, really dense forests — and in some cases, woodlands that have become forests — because we have not been allowing a natural process to occur, which is fire.”
Wildfires are an important, natural occurrence, and whenever we extinguish smaller, harmless fires, we pave the way for the massive, catastrophic ones we experience every year, she explains.
Prescribed fires — fires that burn under controlled conditions — are a viable solution. In fact, California Governor Gavin Newsom’s Wildfire and Forest Resilience Task Force issued a strategic plan in March for expanding use of prescribed fire and other beneficial burning tactics to help mitigate the spread of wildfires.
“Whether it be for forest health or biodiversity, there is a lot of evidence that it is good when we allow these fires to burn,” adds Wilkin.
In federal funding awarded by President Biden's March 2022 infrastructure bill to support fire modeling and prediction technology
Wildlife Interdisciplinary Research Center faculty members focused on cutting-edge research
Acres burned in California wildfires this year according to CalFire, as of August 2022
Professor Adam Kochanski (middle) at the data center with Jeremy Benik, '23 MS Meteorology, and Kathleen Clough, '23 MS Meteorology. Photos: Robert C. Bain
“I’m interested in doing work that directly and positively impacts mountain communities."
— Kathleen Clough
Next generation’s fire fighters
Growing up in Colorado, Kathleen Clough, ’23 MS Meteorology, has seen the damage wildfires and drought can have on communities.
Clough had no idea wildfire research was a field she could even pursue, though she knew she was interested in studying the effects of climate change. Then, at a conference during her senior year of college, she learned about Kochanski’s work. When she found out that he was at San José State University, “I knew that’s where I had to go.”
Now, Clough is analyzing the performance of WRF-SFIRE in order to help improve the prediction model’s accuracy over time. She plans to pursue a PhD in wildfire research, then hopes to work for NASA or Lawrence Livermore National Laboratory “to be directly involved in the implementation and improvement of WRF-SFIRE,” she says. “I’m interested in doing work that directly and positively impacts mountain communities.”
SJSU is also working to make wildfire science education available to more students. In fall 2021, the College of Science introduced a wildfire science minor program. While specifically designed for careers related to environmental- and climate change-related fields, Kochanski emphasizes that it can complement a variety of fields of study.
“We need more students, more people who will know how to use this technology or work in utility companies, insurance companies or other places,” Kochanski says.
Those students studying wildfire science today might be part of creating a future prediction model that could update residents hour by hour on smoke levels in the air, Kochanski imagines. They could be part of the government’s future efforts to restore our forests to better health. Or they could be leaders in their communities, educating residents how to better safeguard their homes against wildfires.
“There is big potential for careers outside of [wildfire] science itself,” he asserts. “The fire problem is not going away.”
*WRF-SFIRE is a model that combines a weather forecast model — Weather Research Forecasting System (WRF) — with the fire-spread model known as SFIRE.
Little Fires Everywhere
While Kochanski is using machine learning to understand how wildfires spread, Ali Tohidi, assistant professor of mechanical engineering and wildfire dynamics, is investigating a crucial piece of that puzzle: How do small embers shed by wildfires contribute to their propagation?
The impact these embers can have in starting another little fire, a process known as “spotfire,” depends on multiple factors, in addition to current weather conditions: how they move through the air, how much energy they transfer when they land on a certain fuel, and what role the fuel plays. Sometimes, those spot fires grow so much that they become their own separate fires.
“There are a lot of variables,” Tohidi says. “So we don’t have a good understanding of exactly what happens in this process.”
His current research project looks into creating a physics-based model for this phenomenon and incorporating it into WRF-SFIRE that, in turn, could make the prediction model much more precise.
Washington Square: San José State University's Magazine © 2022. All Rights Reserved | Land Acknowledgement