Accelerating compound flood risk assessments through active learning: a case study of Charleston County (USA)

This study introduces a novel framework that uses active learning to enhance the efficiency and accuracy of compound flood risk assessments. Focused on Charleston County, South Carolina, it addresses limitations in traditional stochastic modeling that arise due to the computational burden of simulating numerous flood scenarios involving multiple drivers like storm surge, tides, and precipitation. By employing a Treed Gaussian Process (TGP) model, the framework selects the most informative flood events to simulate, reducing redundancy and allowing more complex variables (such as driver duration and time lags) to be considered without overwhelming computational costs.

The active learning approach reduced the number of required simulations by a factor of four and cut the root mean square error in damage predictions by a factor of eight compared to traditional methods. The framework allowed the inclusion of additional variables that are typically simplified or ignored, resulting in significantly more accurate flood risk estimates. For example, excluding certain variables led to an 11.6% underestimation (USD 25.47 million) of expected annual damages. The approach is scalable across different geographic areas and outputs, although computational demands rise with model complexity. The findings suggest that this method can support more robust flood resilience planning by providing better-informed risk assessments.