177 - Generalizability of Neural Network Models for Hindcasting Maximum River Flood Depths in Coastal Watersheds

Hindcasting flood characteristics can help us better predict future events. Machine learning (ML) algorithms are efficient tools for predicting flood characteristics, but their generalizability (out-of-sample) has raised concerns. This study developed a neural network model for hindcasting maximum river flood depths during Hurricane Ida in Lower Hudson Watershed in Northeast US. It also assessed the model generalizers for hindcasting maximum river flood depths other hurricane events - Isaias, Sandy, and Irene - without retraining. Key features representing physical processes, including topography, soil moisture, hydrodynamics, land surface, hydrology, and meteorology, were incorporated in the model. The model considered spatial variability of these features and processes by using information from contributing watersheds of each river location. We evaluated the neural network model performance via R2, mean absolute error (MAE), normalized root mean square error (NRMSE), and the ratio of estimated to observed maximum flood depth (FQ). Our case study results demonstrated the model ability to hindcast maximum river flood depths during Hurricane Ida (R2: 0.92, MAE: 0.66 m, NRMSE: 29%, FQ: 138%). The pre-trained model successfully transferred to other major flood events (R2 > 0.71, MAE < 1.69 m, NRMSE < 109%, FQ < 366%). This study highlighted the ML models' utility for hindcasting maximum river flood depths and the generalizability of these models across major events when informed by appropriate physical processes and spatial variability of pertinent features in coastal watersheds.