Deep learning (DL) models have been developed to propagate extreme water level (EWL) dynamics across an entire open water domain throughout the evolution of extreme events, using data from a few gauge stations. These models estimate water level values for initially empty cells/nodes on a spatial grid by leveraging interactions between sparse point data and other spatial predictor features. This approach marks a significant advancement over traditional DL image recognition methods, enabling simulations comparable to physics-based models. Notably, the DL models were trained on EWL events caused by tropical cyclones in Galveston Bay and applied to accurately predict spatiotemporal EWLs caused by tropical cyclones in Chesapeake Bay.