A physically informed domain-independent data-driven inundation forecast model

Area of Science:

• Hydrology and Water Resources Engineering
• Geospatial Analysis
• Artificial Intelligence in Environmental Science

Background:

• Operational flood forecasting requires accurate spatial and temporal water depth data, crucial for public safety during pluvial flood events.
• Traditional physically-based models are too slow for real-time predictions, while existing data-driven models lack domain independence.
• Current data-driven approaches often require downsampling for larger catchments, limiting their applicability.

Purpose of the Study:

• To develop a physically informed data-driven forecast system capable of real-time spatial and temporal water depth inundation predictions in previously unseen areas.
• To overcome the limitations of traditional and existing data-driven flood forecasting models.
• To enhance the domain independence and applicability of data-driven flood models.

Main Methods:

• A Convolutional Neural Network (CNN) employing an image-to-image translation process was developed, trained on catchment characteristics from Baiersdorf, Germany.
• A spatiotemporal prediction framework was implemented, featuring 10-minute time-stepping and domain-independent forecasting tested across 23 unknown areas.
• A physically informed loss function was integrated, incorporating a 2D continuity equation and kinematic wave formulation to estimate velocity and enforce physical constraints.

Main Results:

• The model achieved a Critical Success Index (CSI) of approximately 74% and a mean Root Mean Squared Error (RMSE) of 0.045 m on unknown areas.
• The physically informed loss function demonstrated superior performance compared to a standard data-driven loss function, reducing RMSE by about 25%.
• The framework successfully provided forecasts with 10-minute temporal resolution, eliminating the need for downsampling.

Conclusions:

• The proposed physically informed data-driven system offers a viable solution for operational, real-time pluvial flood inundation mapping in diverse geographical locations.
• Integrating physical principles into data-driven models significantly enhances prediction accuracy and reliability for flood forecasting.
• The developed framework demonstrates the potential for accurate and efficient flood prediction systems applicable to a wide range of catchments.