Physics-informed Machine Learning for Compound Flood Mapping

Objective:

This project aims to revolutionize regional compound flood mapping by integrating physics-informed machine learning with remotely sensed data to deliver fast, accurate, and scalable flood impact assessments in ungauged coastal regions.

Approach:

This project adopts a hybrid modeling framework that combines physics-based hydrodynamic simulation with advanced machine learning (ML) and data fusion techniques to assess compound flood hazards in freshwater-influenced coastal systems. We use the D-FLOW FM hydrodynamic model, thoroughly calibrated with historical storm and tide data, to simulate complex flood dynamics driven by concurrent oceanic and hydrologic inputs. These simulations serve as training data for ML models capable of learning flood patterns across varied scenarios. The ML models are designed to be physics-informed, ensuring that their outputs align with known flood processes rather than relying on black-box predictions. We have also built a cloud-based mapping platform on Google Earth Engine that integrates remotely sensed observations with ML outputs and precomputed simulation results. This approach offers a scalable solution for large-area flood mapping and rapid post-storm assessments. The combined methodology improves National Water Model validation, enhances exposure mapping in data-scarce regions, and enables faster reparametrization of models after major flood events.

Impact:

Abstract:

This project addresses a critical gap in our ability to rapidly assess and communicate compound flood risks across U.S. Gulf Coast communities, especially in data-scarce regions. We are developing a hybrid flood mapping framework that combines hydrodynamic modeling with physics-informed machine learning algorithms and remotely sensed observations to improve the accuracy and scalability of compound flood impact assessments. The challenge we address is the lack of regional compound flood mapping tools that are both physically sound and operationally feasible—particularly for areas without dense monitoring networks. Our products include:

- A suite of ML models trained on physically simulated flood scenarios for faster, post-storm hazard mapping.
- A cloud-based mapping tool (Google Earth Engine) for regional compound flood prediction using integrated simulation and satellite data.
- Synthetic rapid-intensification hurricane scenarios for robust, climate-aware risk analysis.
- Tools for recalibrating flood forecasting systems like the National Water Model, enabling smarter emergency response.

Today’s flood mapping tools are often siloed, data-heavy, and slow. Our approach enhances coverage, reduces computational costs, and speeds up post-storm decision-making. Ultimately, our framework helps NOAA and emergency managers validate flood forecasts, evaluate flood exposure as soon as remotely sensed data become available, and improve communication with vulnerable communities—paving the way for a more adaptive, and responsive flood management system.