Li Develops Revolutionary Flood Hazard Forecasting System
As Hurricane Helene was bearing down on the East Tennessee area in September of 2024, Haochen Li was receiving updates in a group chat from one of his students. The flooding was decimating the student's hometown of Erwin, Tennessee. The destruction spread far and wide, overwhelming areas well outside of the predicted flood zones.
Li, an assistant professor in the Department of Civil and Environmental Engineering, wondered if he could make a predictive tool that would better assess the likelihood of flood damage.
"I feel like it's almost become sort of personal. It wasn't in my area of interest, but I wanted to look at the problem and see what is currently available," Li said. "I realized there's sort of a research gap, and there hasn't been any tool to map the flood risk for mountainous terrain accurately and efficiently."
In collaboration with CEE Associate Professor Jon Hathaway, Biosystems Engineering and Soil Science Assistant Professor Emine Fidan, and Geography and Sustainability Associate Professor Kelsey Ellis, Li launched a project that seeks to develop a deep learning (DL) empowered real-time high-resolution flood hazard forecasting system for Southern Appalachians. The system would enable stakeholders to better assess risks, plan emergency responses, develop infrastructure, and implement mitigation strategies.
Li's flood mapping work has received support from the University of Tennessee Office of Research, Innovation, and Economic Development's (ORIED's) AI Tennessee Initiative. The Graphics Processing Units (GPUs) available through AI Initiative and Google Tensor Processing Units (TPUs) were instrumental in Li's work.
"It really gives us a higher-resolution predictive tool that we'll be able to use to tell you whether your house will get flooded or not," Li said. "That is something very important from a safety and insurance perspective."
Advanced weather technology
Flood risk in the US is expected to increase by 26.4% by 2025, resulting in annual losses of approximately $32.1 billion, according to a study published in Nature. Flooding for most home insurance is based on FEMA maps that are developed though geographical and historical data, and using hydraulic and hydrologic analysis.
"They are not always accurate for a lot of scientific reasons," Li said. "They can be good for a large portion, such as major river corridors, but there's a lot of event-specific situations that a FEMA map cannot really characterize correctly. This is particularly true for upland streams."
Using computational resources and compilers that weren't available in the past, Li was able to develop an advanced hydrodynamic solver that can model the entire watershed (19,980 km2/7,714 square miles) with 202 million degree-of-freedom on a single modern accelerator (GPU/TPU).
"Right now, for the East Tennessee area, we'd be able to perform flooding simulation to a five-meter resolution," Li said. "We can make a more accurate prediction at a very fine resolution, higher accuracy, and be able to do this for forecasting with limited computational resource."
Expanding the zone
Li's goal is to use the advanced forecasting system to provide a flood risk model for the entire Southeast region to help educate homeowners, assist government agencies in disaster response, and inform urban planners.
Li intends to publish a paper about his research and host a platform through UT web service that will provide flood risk for the southern Appalachians.
Being able to more accurately predict flood risks when natural disasters strike would help save lives, homes, businesses, farmland, and many other valuable possessions in the future. After assessing the damage from Hurricane Helene, Li realizes how critical the need has become.
"There are so many scientific problems that I came across while doing this. I realized why people couldn't do this before, and now we can do it," he said. "It really comes down to needing to leverage modern accelerators (GPU/TPU) and specialized compilers. For Tennessee and for mountain areas, you need to have very high resolution to resolve the complex terrain. It was a level of computational power that was not possible before. Now it's possible."
Provided by University of Tennessee at Knoxville
