A new technique developed by the National Center for Atmospheric Research (NCAR) uses artificial intelligence to efficiently update vegetation maps relied on by computer fire models to accurately predict fire behavior and spread.
In a recent study, scientists demonstrated a method using the 2020 East Troublesome Fire in Colorado, which burned land that was mistakenly characterized in fuel reserves as healthy forest. In fact, the fire, which grew explosively, burned the area, which was recently devastated by pine beetles and storms, leaving significant parts of dead and fallen material.
The research team compared fire simulations generated by a state-of-the-art fire behavior model developed at NCAR using both the standard fuel inventory for the area and the one updated with artificial intelligence (AI). Simulations that used fuels updated with artificial intelligence did a much better job of predicting the area burned in the fire, which eventually grew to more than 190,000 hectares of land on both sides of the continental divide.
“One of our main challenges in modeling fires has been to get accurate data, including fuel data,” said NCAR scientist and lead author Amy Dekastro. “In this study, we show that the combined use of machine learning and satellite imagery provides a sustainable solution.”
The research was funded by the US National Science Foundation, which is a NCAR sponsor. Modeling simulations were conducted at the NCAR-Vioming Supercomputing Center on the Cheyenne system.
Using satellites to calculate damage from the pine beetle
In order for the model to accurately simulate a forest fire, it needs detailed information about current conditions. This includes local time and terrain, as well as the characteristics of the plant matter that provides the fuel for the flame - what is actually available for combustion and in what condition. Is he dead or alive? Is it wet or dry? What kind of vegetation is that? How many? How deep is the fuel stacked on the ground?
The Gold Standard for Fuel Datasets is produced by LANDFIRE, a federal program that produces a range of geospatial datasets including information on fuel from forest fires. The process of creating these datasets on forest fire fuels is extensive and includes satellite imagery, landscape simulation, and information collected personally during the research. However, the amount of resources needed to produce them means that, in practical terms, they cannot be updated frequently, and forest disturbances - including fires, insect infestations and development - can meanwhile radically alter available fuels.
In the case of the East Troublesome Fire, which began in Grand County, Colorado, and burned east in the Rocky Mountain National Park, the latest LANDFIRE fuel dataset was released in 2016. In the past four years, pine beetles have caused widespread tree mortality in surroundings.
To update the fuel data set, the researchers turned to the Sentinel satellites, which are part of the European Space Agency’s Copernicus program. Sentinel-1 provides information on the texture of the surface, which can be used to identify the type of vegetation. (Grass has a very different texture from trees, for example.) And Sentinel-2 provides information that can be used to conclude the health of a plant based on its greenery. Scientists have entered satellite data into a machine learning model known as “random forest” that they have trained in insect and disease research by the U.S. Forest Service. The survey is conducted every year by trained staff who assess the mortality of trees from the air.
The result was that the machine learning model was able to accurately update LANDFIRE fuel data, converting most fuels categorized as “wooden litters” or “wood substrates” to “slash whitewood”, a label used for forest regions with heavy tree mortality .
“LANDFIRE data is super valuable and provides a reliable platform for upgrading,” said Dekastro. “Artificial intelligence has proven to be an effective tool for updating data in a less intensive way.
Positioned for positive impact
To test the effect that an updated fuel inventory would have on simulating forest fires, the scientists used a version of NCAR’s model for research and forecasting, known as VRF-Fire, which was specifically developed to simulate the behavior of forest fires.
When VRF-Fire was used to simulate the Eastern Problematic Fire using an unadjusted LANDFIRE fuel data set, it significantly underestimated the amount of area on which the fire would burn. When the model was restarted with a customized data set, it could predict the area burned with a much higher degree of accuracy, indicating that dead and fallen wood helped spread the fire much more than if the trees were still alive. .
For now, the machine learning model is designed to update the existing fuel map and can get the job done quickly (in minutes). But the success of the project also shows the promise of using machine learning systems to start regular production and updating fuel maps from scratch in large regions at risk of forest fires.
New research at NCAR is part of a broader trend to explore possible AI applications for fires, including efforts to use AI to more quickly assess fire perimeters. NCAR researchers also hope that machine learning could help address other persistent challenges in modeling fire behavior. For example, machine learning may be able to improve our ability to predict the properties of a fire-induced embers (how large, how hot, and how dense), as well as the likelihood that that embers could cause a fire.
“We have so much technology and so much computing power and so many resources at our fingertips to solve these problems and protect people,” said NCAR scientist Timothy Juliano, co-author of the study. “We are well positioned to make a positive impact; we just need to keep working on it.”