Spatiotemporal Variability of Channel Roughness and its Substantial Impacts on Flood Modeling Errors

Abstract

Abstract Manning's roughness coefficient, n, is used to describe channel roughness, and is a widely sought‐after key parameter for estimating and predicting flood propagation. Due to its control of flow velocity and shear stress, n is critical for modeling timing of floods and pollutants, aquatic ecosystem health, infrastructural safety, and so on. While alternative formulations exist, open‐channel n is typically regarded as temporally constant, determined from lookup tables or calibration, and its spatiotemporal variability was never examined holistically at large scales. Here, we developed and analyzed a continental‐scale n dataset (along with alternative formulations) calculated from observed velocity, slope, and hydraulic radius in 200,000 surveys conducted over 5,000 U.S. sites. These large, diverse observations allowed training of a Random Forest (RF) model capable of predicting n (or alternative parameters) at high accuracy (Nash Sutcliffe model efficiency >0.7) in space and time. We show that predictable time variability explains a large fraction (∼35%) of n variance compared to spatial variability (50%). While exceptions abound, n is generally lower and more stable under higher streamflow conditions. Other factorial influences on n including land cover, sinuosity, and particle sizes largely agree with conventional intuition. Accounting for temporal variability in n could lead to substantially larger (45% at the median site) estimated flow velocities under high‐flow conditions or lower (44%) velocities under low‐flow conditions. Habitual exclusion of n temporal dynamics means flood peaks could arrive days before model‐predicted flood waves, and peak magnitude estimation might also be erroneous. We therefore offer a model of great practical utility.

Keywords

• channel roughness
• fluvial hydraulics
• flood modeling
• big data