Scientists use AI to predict wildfires' next move — USC News

Applications of AI


USC researchers have developed a new way to accurately predict the spread of wildfires. Combining satellite imagery and artificial intelligence, the model could lead to a breakthrough in wildfire management and emergency response.

Challenge: Globe logo
Learn more about USC's Assignment: Earth initiative.

As detailed in early findings published in Artificial Intelligence for the Earth Systems, the USC model uses satellite data to track the progress of wildfires in real time and feeds that information into advanced computer algorithms to accurately predict the fires' expected path, intensity and growth rate.

The study comes as California and much of the Western U.S. continues to experience an increasingly severe wildfire season, with multiple fires raging across the state, fueled by a dangerous combination of wind, drought and extreme heat. Among them is the Lake Fire, the state's largest wildfire this year, which has already burned more than 38,000 acres in Santa Barbara County.

“This model represents an important step forward in our ability to fight wildfires,” said Brian Shady, a doctoral student in the Department of Aerospace and Mechanical Engineering at USC's Viterbi School of Engineering and corresponding author of the study. “By providing more accurate and timely data, our tool will enhance the efforts of firefighters and evacuation teams on the front lines battling wildfires.”

Reverse Engineering Wildfire Behavior with AI

The researchers began by collecting data on past wildfires from high-resolution satellite imagery. By carefully studying past wildfire trends, the researchers were able to track how each fire started, spread, and was ultimately extinguished. Their comprehensive analysis revealed patterns that were influenced by a variety of factors, including weather, fuels (trees, brush, etc.), and topography.

They then trained a generative AI-powered computer model called a conditional Wasserstein generative adversarial network (cWGAN) to simulate how these factors affect the progression of a wildfire over time, and taught the model to recognize patterns in satellite imagery that matched the way wildfires spread.

We then tested the cWGAN model on actual wildfires that occurred in California between 2020 and 2022 to see how accurately it could predict the spread of fires.

“By studying how fires have behaved in the past, we can create models that predict how future fires will spread,” said Asad Oberai, Hughes Professor and professor of aerospace and mechanical engineering at USC Viterbi and co-author of the study.

Using AI to predict wildfires: A great model

Oberei and Shadi were impressed that cWGAN, which was initially trained on simple simulated data under ideal conditions like flat terrain and unidirectional winds, performed well when tested on real California wildfires. They attribute this success to cWGAN being used not in isolation, but in combination with real wildfire data from satellite imagery.

Oberlay, whose research centers on developing computer models to understand the physics that underlie a wide range of phenomena, has modeled everything from turbulence over an airplane wing to infectious diseases to the way cells grow and interact with their surroundings in a tumor. Oberlay says wildfires are some of the most difficult things he has modeled.

“Wildfires involve complex processes: fuels like grasses, shrubs and trees burn, causing complex chemical reactions, generating heat and wind currents. Factors like topography and weather also affect fire behavior: fires don't spread very much in wet conditions, but can spread rapidly in dry conditions,” he says. “These are very complex, chaotic and nonlinear processes. To model this accurately you need to take into account all these different factors. That requires advanced computing.”


About the Study: Additional co-authors include Valentina Calaza, an undergraduate student in the Department of Aerospace and Mechanical Engineering at USC Viterbi, Deep Ray from the University of Maryland, College Park (formerly a postdoctoral researcher at USC Viterbi), Angel Farguell and Adam Kochanski from San Jose State University, Jan Mandel from the University of Colorado, Denver, James Haley and Kyle Hilburn from Colorado State University, Fort Collins, and Derek Mallia from the University of Utah.

The research was funded by the Army Research Office, NASA, and the Viterbi CURVE program.



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *