Physics-based AI predicts wildfire path
Researchers in the US have adapted a generative AI framework to work with satellite data to predict rather than detect wildfires.
The model developed at the University of Southern California (USC) uses satellite data to track the progression of a wildfire in real time, then feeds this information into a generative AI framework that can accurately forecast the fire’s likely path, intensity and growth rate.
Multiple blazes, fueled by a dangerous combination of wind, drought and extreme heat, are currently raging across California, and there are many different techniques being developed to monitor and track wildfires. The Lake Fire, the largest wildfire in the state this year has already scorched over 38,000 acres in Santa Barbara County. Several AI-based technologies have been developed to detect, monitor and track wildfires.
The physics-informed approach for inferring the history of a wildfire from satellite measurements provides the necessary information to initialize coupled atmosphere-wildfire models from a measured wildfire state.
The fire arrival time, which is the time the fire reaches a given spatial location, acts as a representation of the history of a wildfire.
A conditional Wasserstein Generative Adversarial Network (cWGAN), trained with simple simulation data, is used to infer the fire arrival time from satellite active fire data. By carefully studying the behaviour of past wildfires, the researchers were able to track how each fire started, spread and was eventually contained. Their comprehensive analysis revealed patterns influenced by different factors such as weather, trees and brush as well as the impact of the terrain.
The cWGAN is used to produce samples of likely fire arrival times from the conditional distribution of arrival times given satellite active fire detections. Samples produced by the cWGAN are further used to assess the uncertainty of predictions.
The cWGAN was tested on four California wildfires occurring between 2020 and 2022 and predictions for fire extent are compared against high resolution airborne infrared measurements.
The predicted ignition times were then compared with reported ignition times showed an average ignition time difference of 32 minutes, suggesting that the method is highly accurate.
“This model represents an important step forward in our ability to combat wildfires,” said Bryan Shaddy, a doctoral student in the Department of Aerospace and Mechanical Engineering at the USC Viterbi School of Engineering and the study’s corresponding author. “By offering more precise and timely data, our tool strengthens the efforts of firefighters and evacuation teams battling wildfires on the front lines.”
“By studying how past fires behaved, we can create a model that anticipates how future fires might spread,” said Assad Oberai, Hughes Professor and Professor of Aerospace and Mechanical Engineering at USC Viterbi and co-author of the study.
Oberai and Shaddy were impressed that the cWGAN, initially trained on simple simulated data under ideal conditions like flat terrain and unidirectional wind, performed well in its tests on real California wildfires. They attribute this success to the fact that the cWGAN was used in conjunction with actual wildfire data from satellite imagery, rather than on its own.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
