Quantifying the relative contributions of different flood generating mechanisms to floods across CONUS

Abstract

The occurrence of flooding in a catchment is attributable to various mechanisms. Several researchers have used linear models to quantify the relative contributions of different flood generating mechanisms. However, the use of linear models for this quantification may not be suitable given the complex and nonlinear processes from climate to floods. In this study, we used nonlinear machine learning models with interpretability methods to quantify the contributions of multiple flood generating mechanisms to floods across the continental United States (CONUS). Four flood generating mechanisms were considered comprehensively, which varied in terms of the influence of antecedent soil moisture condition or snowmelt on flood generation. These mechanisms involved three common mechanisms (rainfall, rainfall excess, and snowmelt/rain on snow) and one complex mechanism (effective precipitation) that accounts for snowmelt/rain on snow lost to unsaturated soil zone. Consistent with previous studies, we observed that rainfall alone (i.e., without considering other hydrometeorological processes) is seldom the dominant mechanism for flood generation. Floods in most catchments are largely attributable to rainfall excess and snowmelt. Unlike previous studies, we noted that in the midwest and northeastern CONUS, snowmelt/rain on snow upon saturated soil was the greatest contributor to local flood generation. This finding highlights that both snowmelt and soil moisture condition affect flood generation in the midwest and northeastern CONUS, instead of only snowmelt as has been reported previously. Overall, this study provides new insights into flood generating mechanisms in CONUS and demonstrates the potential of applying interpretable machine learning to examine flood generating mechanisms.