Research uses AI to track wildfire spread

In his seven years as a wildland firefighter in North Idaho, Phinehas Lampman recognized the importance of predicting fire behavior in real time.

Wind conditions, topography, the intensity of the fire and the amount of available fuels play significant roles in how far and how fast a wildfire is expected to travel.

For his doctorate work at U of I, Lampman developed an AI system that used drone-based infrared imagery to predict how fast a wildfire is moving.

Knowing where a roaring, snapping wildfire is headed and how long before its caldron of flames arrives at a rural residence, a highway or a community can determine what firefighting resources should be employed to intervene, and the best location to make a stand.

“Most incident commanders in the field rely on their vast experience to assess and ultimately predict wildfire movement,” said Lampman who worked as a firefighter for both the Clearwater-Potlatch Timber Protective Association and the United States Forest Service on engine crews and a hotshot crew in North Idaho.

Incident commanders largely rely on maps, weather reports, on-the-ground observations and aerial observations from planes and helicopters to make firefighting decisions, he said.

“Sometimes those aerial assets are grounded, or the visibility is really poor,” Lampman said. “But drones can usually still fly and gather information.”

As a graduate student in Professor Leda Kobziar’s fire science lab, Lampman spent years refining the use of drones to collect smoke samples for research projects including Kobziar’s pioneering research that tracks the movement of microbes in wildfire smoke.

As part of his dissertation, Lampman equipped drones with thermal infrared (TIR) cameras that sense heat and analyzed the resulting imagery to develop machine learning models to quantify and predict fire behavior. His system quantified rate of spread (ROS) of a moving fire, fire line intensity – the rate of heat released along the fire front line– and the radiative power of the fire, or the rate of radiant energy produced over an entire fire.

Those three metrics are essential to determine fire behavior, he said.

Lampman trained two machine learning models including a Random Forest (RF) model, which is an algorithm that combines hundreds of individual trees with large amounts of environmental and meteorological data to predict fire behavior. He also trained an Artificial Neural Network (ANN) — a machine learning model used to process complex data and predict fire behavior by identifying hidden patterns to forecast wildfire spread rates.

“Using repeat passes with drones that collected TIR imagery, we derived high-resolution metrics and trained an artificial neural network (ANN) and random forest (RF) models to predict rate of spread with low error,” Lampman said.

His research found that both ANN and RF models performed well, but RF performed better, with less model training data and stronger model performance.

The research was published in the International Journal of Wildland Fire.

Traditional fire behavior models have been reliable, Kobziar said, but using machine learning based on observations could be a better predictor of where fire will move and how hot it will be when it gets from point A to point B.

“Conceptually, we have a good understanding of how fire should behave based on empirical data and physics-based computing,” Kobziar said. “In this case we’re using real-time data from an active fire to make more precise predictions.”

Fire behavior has changed over time given new field and weather conditions that make behavior harder to predict using old models, she said.

“This is an important application of machine learning using AI techniques,” Kobziar said. “We had no intention initially to use AI to make these predictions, but Phinehas learned to create these models, and it became obvious that machine learning could be a good way to approach these questions.”