Satellite-informed fire forecasting with artificial intelligence: a spark of innovation
By Theresa Barosh | Oct. 2024 | Originally published in SOURCE
Growing up in Minnesota, Kyle Hilburn regularly experienced extreme weather, like flooding and heat waves that inspired him to study meteorology and become an expert of using satellite observations in weather studies.
Then, he could not seem to get away from wildfires.
Hilburn, now a scientist at Colorado State University’s Cooperative Institute for Research in the Atmosphere or CIRA, lived in Santa Rosa, California, where the nearby Valley Fire of 2015 and Tubbs Fire of 2017 collectively burned over 7,000 structures and resulted in 26 deaths. Shortly after moving to Colorado, Hilburn watched the Cameron Peak Fire of 2020 destroy 469 structures.
At CIRA, Hilburn found an opportunity to make a difference in these situations by using machine learning and satellite data. CIRA represents a partnership between Colorado State University and the National Oceanic and Atmospheric Administration.
Satellite Technology and Machine Learning
Currently, use of satellite data in fire modeling is limited. Hilburn and fellow CIRA researchers forecast wildland fire behavior using satellite information through the development of tools that simplify data acquisition and processing.
“Satellites provide a wealth of information about wildfires,” said Hilburn. “This information ranges from detecting the lightning strikes that can ignite fires, to the thermal hot spots from new fires. Connecting satellite observations with wildfire models is pushing the state of the art.”
Hilburn’s use of this satellite data is not only innovative; it is also practical because satellites can catch early development of wildfires, sometimes, before someone on the ground.
“Satellites provide the most accurate and up-to-date situational information in the early hours of a new fire,” said Hilburn. “The early hours of a fire is when intervention is most likely to change the outcome.”
Satellites also provide information about fuels, weather, and topography that impact wildland fires. Fuels include any materials that burn, such as trees, grass and brush. Temperature, humidity, and precipitation impact fire spread, while lightning can start new wildfires. Topography can also impact weather, smoke, and a fire’s path. For example, fires typically spread faster uphill where fuels have already been heated by the fire. Hilburn’s fire forecasts take all these factors into account while using the underlying science.
Hilburn’s fire forecasts are done with coupled models – computer-based models of the Earth’s climate in which weather, fire, fuels, and smoke can interact in the simulations. This modeling strategy is unique compared to other fire modeling and, Hilburn says, necessary because fires create their own weather.
“Fires nowadays are not only more numerous, but they’re getting bigger and hotter. The fires produce fluxes of heat, and that heat creates buoyancy. Anything that creates buoyancy can create weather, can create storms,” said Hilburn. “Fires can create thunderstorms, known as pyrocumulonimbus, which in turn drives the fire behavior because convection creates new wind patterns. More erratic wind patterns and the outflow from the convection can push the fire in new directions.”
In the case of the 2018 Mallard Fire in Texas, the fire created a pyrocumulonimbus that created lightning, which resulted in new fires.
“We’ve demonstrated a few fires where the model has generated pyrocumulus. The Mallard Fire was one. The Creek Fire was one. The Dixie Fire was one,” said Hilburn. “So, these are all cases when our model forecast big plumes and thunderstorms. And observations showed there was actually something that formed in real life.”
Pyrocumulus clouds are formed by hot air and smoke released into the sky. When a pyrocumulus column becomes unstable, it can collapse onto itself causing fierce winds and potentially harming fire suppression teams or anyone else nearby. Pyrocumulus development prediction supports decisions about when and where to send fire hotshot crews.
Overcoming Challenges on Spatial and Temporal Scales
When Hilburn started his research in 2018, he had multiple obstacles to overcome. Hilburn was faced with low-resolution satellite observations with limited observations throughout the day. By using observations from multiple new instruments, Hilburn overcame the challenges of resolution and timing.
New technology launched on satellites in 2017 and 2018 provided much-needed information for Hilburn’s forecasts. The Geostationary Operational Environmental Satellites (GOES) satellites at a high orbit help identify when a fire started because they constantly monitor North America. The Visible Infrared Imaging Radiometer Suite (VIIRS) sensor on satellites at lower orbit has a useful, finer spatial scale. However, the VIIRS sensor orbits around the poles, so provides accurate information about where a fire is located only during the periods when the sensor is passing above said fire. This scale of detail can help identify the difference between a fire’s location just north versus south of a highway or river, information that significantly impacts the forecast of how a fire may move. By combining information from multiple different satellite technologies, Hilburn captured the information needed for his forecasts.
There were also concerns about how quickly satellite data could inform Hilburn’s fire forecast model. Would the major moments in a fire already be passed by the time the model could make a prediction? Would the model be too expensive to fund for real-time operations? By incorporating machine learning, Hilburn can process lots of data very quickly. Hilburn’s research over the last few years has demonstrated that the model forecasts can indeed operate on a time scale that is relevant for fire responders.
Getting From the Lab to Operations
The research has gone through the first stages of the NASA standard Technology Readiness Levels, which means that the model is progressing from a research project toward a product that could be used by public forecasters. These Readiness levels indicate NASA’s assessment of the maturity of technologies from concept to successful operations so that new innovations can one day be used more widely.
“We showed we can do this, we can get the satellite data in, we can get the forecast out quickly,” said Hilburn. “Next, it’s about connecting that with NOAA’s efforts and the partners who are involved.”
Hilburn is also working on a NOAA National Environmental Satellite, Data, and Information Service project with Mike Pavolonis, a research scientist at NESDIS Center for Satellite Applications Research. The software system is open source to support accessibility and innovation. In the first year of the project, they ran more than a thousand forecasts in near real-time for the 2023 season. They are currently running forecasts for the 2024 season. Hilburn said he plans to continue working to get the model into operations.
“People talk about fire, like it’s this kind of distinct entity but it’s part of the Earth system. It’s an essential part of it. We can’t neglect it anymore in our modeling,” said Hilburn. “I’m hopeful that we’ll get to operationally forecasting fires because I see that progress is being made.”