Implementing Machine Learning to Support National-Level Flood Impact Forecasting for the U.S.

The National Water Center’s Flood Hazard Outlook (FHO) is a daily national product available to the public, highlighting potential inland flood impacts across the United States and U.S. territories. Development of the FHO currently relies heavily on forecasters’ critical assessment of over 30 variables to classify and select hydrologic unit codes (HUCs) based on the severity and timing of potential impacts. This method involves inherent subjectivity in the selection of HUCs that can be affected by different factors, such as how forecasters weigh different variables, local knowledge, biases, experience, and even forecaster fatigue. An objective and quantitative approach to defining and characterizing potential flood impacts across the U.S. is needed to provide a baseline classification or “First-Look” FHO. A First-Look FHO increases the efficiency of FHO production by creating a starting point for the forecaster while also bringing the attention of a forecaster to areas that may be less obvious. It would highlight areas of concern by considering key variables that forecasters use, such as population density, land use, terrain, soil type and moisture, quantitative precipitation forecasts (QPF), and others.