AI-Driven Temporal Super-Resolution for Flooding Prediction in Norfolk, Virginia (Papers Track)
Chetan Kumar (Old Dominion University); Diana McSpadden (Thomas Jefferson National Accelerator Facility); Malachi Schram (Thomas Jefferson National Accelerator Facility); Heather Richter (Old Dominion University); Yidi Wang (University of Virginia); Binata Roy (University of Virginia); Jonathan Goodall (University of Virginia)
Abstract
Accurate and near real-time water depth predictions are essential for supporting transportation and emergency management decisions during flood events. However, traditional physics-based hydrodynamic simulations are computationally expensive and time-consuming, limiting their practicality for real-time response. To address this challenge, we employ the Fourier Neural Operator (FNO) to learn complex spatiotemporal patterns of urban flooding and enable temporal super resolution. Our approach leverages coarse-resolution water depth and rainfall data to predict high-frequency 15-minute resolution water depths. We experiment with varying the temporal resolution during training, from 15 minutes to 1 hour, while always generating predictions at a finer 15-minute temporal resolution during testing. The method is applied to five flooding events between 2017 and 2022 in Norfolk, Virginia, USA. Across different training scenarios, our model achieves an R-squared value higher than 0.79 on test data. These results demonstrate the effectiveness of FNO-based temporal super resolution for accurate and timely water depth predictions.