NSF Funding Fuels Wildland Fire AI Research at UMD

The approximate doubling of wildfire activity across the United States over two decades has been easy enough to track, but forecasting where and when the blazes will spread to threaten communities, wildlife and natural resources is far less straightforward.

University of Maryland computer science Professor Heng Huang is leading a project to close that gap using the power of artificial intelligence. Huang was recently awarded $1.86 million from the National Science Foundation to advance AI-driven methods for detecting and predicting wildland fires. 

The study, conducted in collaboration with American University researchers, aims to address the extreme complexity of fire dynamics, which has limited the ability of traditional computer modeling approaches to forecast events.

Huang and his team will develop large-scale AI and machine learning algorithms capable of digesting the diverse sources of geoscientific data, including satellite observations, atmospheric records, fuel information and historical fire data. The goal is a framework that boosts detection accuracy and forecasting capabilities while remaining usable in real-world conditions, where data is often incomplete.

“I am excited for this project to develop and apply advanced AI techniques to address the challenging wildland fire prediction and prevention problems, which could make a large impact on human communities, wildlife and the environment,” said Huang, the Brendan Iribe Endowed Professor in Computer Science.

Data will be collected from geostationary and low-Earth orbit satellites, as the system will also include reanalysis of atmospheric trends and ground-level fire and surface characteristics. To integrate these diverse sources, the framework will feature an interpretable multimodal transformer—a type of AI algorithm that can utilize different data types to create outputs understandable to users. It will also include a time-series deep learning model to forecast fire events and a federated learning platform, allowing training of the model by different research groups, to support collaboration.

By combining multimodal and longitudinal datasets, researchers seek to capture not only current fire activity but also patterns that could inform predictions of future events. This approach, Huang said, is designed to leverage both large-scale computational power and innovative algorithm design.

Another aspect of the project is the development of an open-source, integrated dataset that can be shared with the broader research community. By releasing these resources, the team intends to assist other researchers in benchmarking their methods and accelerate advancements in wildland fire science.

Partnership with federal, state and local agencies will play a central role, Huang said. Organizations such as NASA, the U.S. Forest Service and the National Park Service will help ensure that the tools align with operational needs for fire tracking and management. The collaboration aims to translate academic research into systems that can be used on the ground during fire seasons.

“By …  accurately forecasting wildland fire events and implementing preparedness and mitigation strategies in advance,” Huang said, “societies can better shield themselves against the devastating impacts of such disasters.”