Leveraging machine learning and satellite observation for skillful flood forecasts

Urban systems are highly exposed and vulnerable to extreme rainfall and flooding. Flood impacts span across various sectors, causing disruptions in transportation network, power supply, and access to emergency services. These impacts are expected to increase with expanding urban development, aging flood control infrastructure, and intensifying rainfall events. Reliable prediction of flood hazards is crucial to inform the design of sustainable risk management strategies. This study aims to advance predictive understanding of flood hazards by leveraging recent advances in numerical weather prediction, machine learning, satellite observations and high-performance computing. We compare the predictive skill of standalone machine learning with the hybrid models built by integrating process-based hydrodynamic model outputs with machine learning algorithms. We demonstrate the ability of machine learning surrogate models to capture spatio-temporal flood dynamics with reduced computational expense. This work contributes to strengthening the scientific foundation for flood-risk prediction that is of utmost importance to enhance community resilience in the face of evolving weather and climate extremes.