Deep learning enables super-resolution hydrodynamic flooding process modeling under spatiotemporally varying...

Area of Science:

• Environmental Science
• Hydrology
• Artificial Intelligence

Background:

• Effective metropolitan flood emergency management requires real-time data on flood extent, severity, and duration.
• Current pluvial flood analysis methods struggle with real-time simulation of regional flooding under dynamic rainfall conditions.

Purpose of the Study:

• To develop a deep learning-enabled super-resolution hydrodynamic flood analysis method for real-time simulation of pluvial flooding.
• To provide high-resolution flow depth and velocity predictions for enhanced flood emergency management.
• To significantly reduce computational time compared to traditional methods.

Main Methods:

• A coarse-grid hydrodynamic model generates low-resolution flood map time series.
• A deep learning model converts low-resolution maps to high-resolution flood maps.
• The deep learning model is trained using limited rainfall scenarios, reducing data preparation effort.

Main Results:

• The method accurately simulates spatiotemporal variations in flood depth (RMSE 0.082 m) and velocity (RMSE 0.088 m/s).
• High correlation coefficients (0.962 for depth, 0.921 for velocity) indicate strong predictive performance.
• A 48-hour rainfall event simulation completed in under 30 seconds, 2690 times faster than direct fine-grid analysis.

Conclusions:

• The deep learning-enabled super-resolution hydrodynamic flood analysis method offers a computationally efficient tool for real-time flood risk management.
• This approach provides comprehensive, high-resolution data for improved emergency response.
• The method is applicable to complex terrains, including urban and mountainous areas.