Assistant Research Professor Florida State University
Flood models should be calibrated and validated against historical observations. However, such observations—high-water marks (HWMs)—are typically very limited and spatially sparse. Hindcasting historical flood events helps improve future flood simulations and forecasts. Hurricane Ida, a Category 4 hurricane in 2021, was a destructive event that made landfall, and broke several rainfall records. While it centered in New Orleans, it affected several areas along a thousand-mile path. Although different physically based models have been used for urban flood simulations, they are very complex, require huge amount of data, and the runtime is very long. In this study, we used spatially distributed machine learning algorithms alongside meteorologic (precipitation), hydrodynamic (tide level) and morphologic (height above the nearest drainage [HAND]) features to derive the maximum flood depth during this hurricane that affected inland and coastal urban areas (New Orleans, New York, and Philadelphia) with different flood mechanisms. Our objective was to assess the performance of computationally efficient data-driven models for hindcasting historical floods. We further examined the transferability of our hindcasting model across geographic extent. We validated our approach against HWMs collected during the event. Our framework provides a computationally efficient framework to hindcast major floods in inland and coastal areas.
