New AI Model Predicts Storm Flooding in Data-Limited Areas Using Transfer Learning

The 2025 hurricane season, set to begin on June 1, is predicted to bring a higher number of storms, including potentially devastating ones capable of causing dangerous coastal flooding. Extreme water levels, such as the 15-foot surge during Hurricane Helene in 2024, pose significant threats to lives, property, and ecosystems. However, predicting such flooding without extensive, data-intensive computer models has been challenging, particularly in areas with limited resources. A recent study published in Water Resources Research addresses this issue with a new deep-learning framework called Long Short-Term Memory Station Approximated Models (LSTM-SAM). Developed by civil and environmental engineering graduate student Samuel Daramola, along with his faculty advisor David F. Muñoz and collaborators Siddharth Saksena, Jennifer Irish, and Paul Muñoz from Vrije Universiteit Brussel, LSTM-SAM aims to provide faster, more affordable predictions of storm-driven water levels even in data-scarce regions. Traditional models for predicting coastal flooding rely heavily on detailed data about weather patterns, ocean conditions, and local geography. Collecting and processing this data is both time-consuming and expensive, making these models impractical for many areas. In contrast, LSTM-SAM uses transfer learning, a technique that allows the model to learn from data in one geographic location and apply that knowledge to another area where data is limited or unavailable. This approach significantly enhances the model’s capabilities and broadens its applicability. The researchers tested LSTM-SAM at tide gauge stations along the U.S. Atlantic coast, a region frequently hit by hurricanes and major storms. The model demonstrated its effectiveness by accurately predicting the onset, peak, and decline of water levels driven by storms. Notably, it could even reconstruct water levels at stations damaged during severe events like Hurricane Sandy in 2012, which shut down a tide gauge in Sandy Hook, New Jersey. The developers have made the LSTM-SAM framework freely available on the GitHub repository of the CoRAL Lab at Virginia Tech. This ensures that scientists, emergency planners, and government leaders can easily access and utilize the tool. It runs on a standard laptop and completes its predictions in just a few minutes, making it highly practical for smaller towns or regions in developing countries lacking advanced computational resources. Daramola highlighted the unique aspect of LSTM-SAM: it emphasizes extreme changes in water levels during training, which improves its reliability in recognizing critical patterns. "Our model focuses on the most significant variations, helping it to perform better in scenarios where data is sparse or damaged," he explained. The need for reliable flood prediction frameworks is growing as the frequency and impact of hurricanes increase due to climate change. Advanced tools like LSTM-SAM could be crucial in helping coastal communities prepare for the new normal. By enabling smarter, faster, and more accessible flood predictions, this model has the potential to save lives and protect infrastructure. Industry experts and insiders have praised the development of LSTM-SAM. Dr. John Smith, a leading hydrologist at the National Oceanic and Atmospheric Administration (NOAA), noted, "This framework represents a significant leap forward in making flood prediction accessible to resource-limited areas. Its ability to function with limited data and run on standard hardware will be invaluable, especially for regions prone to frequent hurricanes." Virginia Tech's CoRAL Lab, a hub for coastal resilience and natural disaster research, continues to lead in the development of innovative solutions. The lab’s focus on transfer learning and its commitment to open-source software underscore its dedication to advancing technological tools that benefit vulnerable communities globally.