Hindcasting Maximum Water Depths in Coastal Watersheds: The Importance of Incorporating Off‐Channel Data and Their Uncertainties in Machine Learning Models

Abstract In the absence of adequate observations on the off‐channel areas, flood models are typically trained and validated against stream water depths. This approach can be efficient for physics‐based models, which incorporate the underlying physical processes, but the efficiency for data‐driven models like machine learning (ML) algorithms is unclear. The existing off‐channel observations like high‐water marks (HWMs) are also subject to uncertainty. This paper addressed three research questions: (a) how useful are ML models, trained with stream gauges, for hindcasting water depths in the off‐channel areas? (b) how does incorporating the uncertainty of HWMs improve the model performance? and (c) does the uncertainty incorporation improve the model transferability to other watersheds and events? To answer these questions, we evaluated the performance of ML models across three large coastal watersheds in the US during three hurricanes—Michael, Ida and Ian. The model was developed under three scenarios, which differed in terms of the flood observational data (stream gauges and HWMs) used for their training and validation. A loss function was proposed to incorporate the uncertainty of observations. We found that ML models trained solely by stream gauges performed well only for stream hindcasts. Satisfactory hindcasts on off‐channel areas were obtained by incorporating the HWMs' uncertainty via the loss function. This uncertainty incorporation reduced the model bias and resulted in the best transferability to other coastal watersheds and flood events. Our study provides insights about developing transferable ML models for hindcasting water depths on streams and off‐channel areas in coastal watersheds during extreme events.