Authors: Suma Bhanu Battula – University of Alabama
Title: Forecast-Horizon Based Probabilistic Approach for Communicating Uncertainty in Operational Flood Inundation Mapping
Presentation Type:
Abstract: North Carolina experienced severe flooding triggered by 20 to 30 inches of heavy rainfall over a three day period in September 2024, an event with an estimated 1000 year recurrence interval associated with Tropical Cyclone Helene. The French Broad River at Asheville and the Swannanoa River at Biltmore in North Carolina reached peak stage of 23.1 ft and 27.33 ft, respectively. At least 63 river gauges exceeded their recorded peak flood level during Helene, from northwestern South Carolina into southwestern North Carolina, northeastern Tennessee and southwestern Virginia. Nearly 30% of the 63 gauges have missing observational data during the event between 23 to 28 September.
This presentation will propose a novel data-driven (AI) method to not only fill in the missing streamflow records, but also to provide spatially complete forecasts throughout the region. Generating fine resolution ensemble forecasts of streamflow at sub-daily timescales with process-based models is computationally expensive, is highly susceptible to biased forcing inputs and has operational challenges when assimilating observations. Here, we use streamflow data from all gauges over the Southeast and apply kriging to generate spatially and temporally continuous streamflow predictions over the storm region. To forecast peak flows, the interpolated streamflow records were postprocessed with a neural network. We apply deep learning and observation perturbations using Mixture density functions to generate streamflow ensemble forecasts. We evaluate the performance of this ensemble using Continuous Ranked Probability Skill Score (CRPSS) to forecast peak flows. Our approach provides a framework to characterize streamflow forecast uncertainty for peak flows in ungauged basins.