海角社区

BATON ROUGE - With approximately 40 percent of the world鈥檚 population living within 100 kilometers, or 62 miles of a coastline, the need for quick, accurate flood prediction to protect residents from coastal hazards is growing.

Nowhere is this need more present than in Louisiana, where the coastal communities face significant challenges from flooding.

A new technique using machine learning, pioneered by 海角社区 scientists, may provide a potential solution. It is called the Prediction-to-Map, or P2M framework, and was the in the journal npj Natural Hazards.

The P2M framework is noteworthy for its speed and accuracy, said Z. George Xue, a professor in the Department of Oceanography & Coastal Sciences with a joint appointment with Center for Computation and Technology, and one of the project鈥檚 creators.

Also included in development of the framework were DOCS Professor Kevin Xu, Associate Professor Matt Hiatt, Jill Trepanier of the 海角社区 Geography & Anthropology, Post-doctoral research assistant Daoyang Bao and Courtney Harris of the Virginia Institute of Marine Science.

P2M produces accurate flooding predictions more than 100,000 times faster than the sophisticated numerical models that Xue鈥檚 group recently developed.  鈥淭he P2M AI model can be carried out on a laptop and finish a 72-hour simulation in 4 seconds,鈥� Xue said. By comparison, 鈥渢he numerical model has to use 500 cores [on a supercomputer] for 15 minutes for a similar forecast.鈥�

The framework creates its flood predictions by drawing on the strengths of both numerical modeling and artificial intelligence, creating a hybrid of the two. The technique involves training AI mapping tools on information from a process based numerical model combined with observational data from a specific area, to create rapid, accurate flooding predictions, for up to a six hour timeframe. Xue said he and Bao are now working together to further increase the prediction time window.

"Developing this model has been a rewarding process. I鈥檓 excited to see how novel AI/ML techniques can be combined with traditional numerical models to deliver faster and more accurate flood forecasts," Bao said. "I believe the biggest impact of this research lies in its potential to go beyond academia. It could serve as an early warning tool, potentially providing real-world support to help save lives and reduce economic losses during extreme flood events."

The team ran the model to re-create the events of Hurricane Nicholas, and compared the results to those of a more energy- and time- intensive numerical model, and found them to be very close. In head-to-head testing, P2M edged out the numerical model, posting an R-squared score 0.06 higher and cutting RMSE by 0.1鈥塵 鈥� a measurable boost in prediction accuracy.

The P2M framework also proves effective through its use of data generated by Xue鈥檚 dynamic two-way coupled hydrological-ocean model, the paper noted.  This modeling technique can be used to predict compound flooding 鈥� ie, the type of event that can happen during a hurricane, when waters from a storm surge meet inland flooding 鈥� by taking into account the complete hydrology of an area, including both rivers and coastal waters.

While the framework is currently designed primarily to be used in short-term forecasts, the paper also notes it may have applications for other types of environmental predictions in coastal and estuarine management.

Xue emphasized that although machine learning plays a crucial role in this flood prediction framework, there is still a strong need for robust numerical models, as well as the observational data provided by federal research infrastructure, such as buoys, gauges and weather models.