Six days after lightning strikes ignited more than 30 wildfires in a stretch of Quebec and Ontario experiencing near-record dryness and heat in June 2023, the smoke and haze settled over Washington like a shroud.
As the National Weather Service and the EPA issued urgent warnings to stay indoors, air quality and visibility continued to plummet—even enveloping the Washington Monument in an ominous, orange pall.
And yet, in that moment, computer science professor Leah Ding saw with piercing clarity the urgent, publicly vital application of her expertise in artificial intelligence-driven satellite data analysis.
Spurred to action—and backed by a nearly $870,000 grant from the National Science Foundation—Ding is leading a three-year effort to push wildfire forecasting into a new era, just as extreme wildfire activity has more than doubled worldwide, according to NASA.
Ding and her team are building an AI system that can detect, forecast, and assess the risk of blazes in real time and with unprecedented precision by developing advanced machine learning frameworks capable of integrating multiple geoscientific datasets, ranging from satellite imagery and land-surface information to historical fire records and reanalysis.
Ding’s new AI framework and machine learning algorithms will be integrated into existing alert systems, ultimately giving firefighters more time to respond, communities more time to prepare, and leaders the information they need to save lives and protect property.
Current wildfire models fall short, Ding explains, because available satellite data forces a difficult choice: blurriness or blind spots.
Geostationary Earth orbit (GEO) satellites offer wide-angle coverage every 10 to 15 minutes but produce blurry images, often missing small, fast-moving fires. Low Earth orbit (LEO) satellites capture sharper detail, but their infrequent passes—only twice daily—leave crucial, long gaps in coverage where fires can spread rapidly.
Additionally, fire data comes from many sources—different instruments, weather models, and more—each with its own timing, resolution, and accuracy, making integration and reliable forecasting even more difficult.
Ding’s advanced AI framework directly addresses these challenges. It includes two core components: a machine learning model designed to handle incomplete data—allowing it to merge the frequency of GEO satellites with the detail of LEO satellites for reliable prediction—and a federated learning model that combines insights from many datasets for collaborative analysis while keeping the underlying data private.
To execute this vision, Ding has assembled a specialized team of wildfire scientists, remote sensing experts, and AI researchers.
“Each brings a piece of the puzzle: fire behavior expertise, satellite data knowledge, and advanced machine learning methods,” she explains. This multidisciplinary effort is strategically aligned with federal, state, and local operational partners, including NASA, the US Forest Service, and the National Park Service, ensuring that research meets real-world operational needs.
Ding’s project is designed for broad societal impact, beginning with the creation of an open-source integrated dataset. This gives scientists and students a standardized benchmark to develop and test new AI models without starting from scratch.
Currently, most wildfire studies use fragmented data, making it difficult to compare models or reproduce results. “By bringing together satellite observations, environmental variables, and fire records in one place, we’re lowering that barrier, so researchers can innovate faster in the field of AI for wildfire science,” Ding says.
For the AU students working alongside Ding, the project provides a direct line to the front lines of the climate crisis. They are gaining hands-on experience preprocessing historical data, building machine learning models, analyzing data, and testing tools.
“I want my student researchers to come away capable and thoughtful,” Ding says, “ready to participate in shaping an AI-powered future—not just be shaped by it.”