FloodCastBench: A Large-Scale Dataset and Foundation Models for Flood Modeling and Forecasting

Area of Science:

• Environmental Science
• Hydrology
• Data Science

Background:

• Effective flood forecasting is vital for disaster management and public safety.
• Current flood datasets primarily describe static flood events, lacking dynamic process data essential for advanced machine learning (ML) models.
• There is a need for comprehensive, dynamic datasets to advance ML-based flood prediction.

Purpose of the Study:

• To introduce FloodCastBench, a novel dataset specifically designed for ML-based flood modeling and forecasting.
• To provide high-resolution spatiotemporal dynamic flood process data derived from hydrodynamic modeling.
• To establish a benchmark for neural flood forecasting using the FloodCastBench dataset.

Main Methods:

• Data acquisition involved preparing input data (topography, land use, rainfall) and collecting flood measurements (SAR-based maps, surveyed outlines).
• A finite difference numerical solution was employed to generate high-resolution (30-m spatial, 300-s temporal) dynamic flood processes.
• Flood measurement data were utilized for hydrodynamic model calibration and validation of inundation maps.

Main Results:

• FloodCastBench offers both low-fidelity and high-fidelity flood forecasting datasets suitable for ML applications.
• The dataset covers four major flood events: Pakistan 2022, UK 2015, Australia 2022, and Mozambique 2019.
• A benchmark of foundational models demonstrated the dataset's effectiveness for spatiotemporal, cross-regional, and downscaled flood forecasting.

Conclusions:

• FloodCastBench significantly enhances the capability for ML-driven flood forecasting by providing dynamic, high-resolution data.
• The dataset and benchmark models validate the potential of ML for improved flood prediction across diverse scenarios.
• This resource facilitates advancements in developing more accurate and reliable flood forecasting systems.