Large Scale Prediction of Channel Roughness Coefficient Using Machine Learning
Abstract
The channel roughness is one of the most critical parameters in river hydraulics and fluvial flooding. It is spatio-temporally dynamic and subject to several external controlling factors. Roughness coefficient (Manning's n) was typically assumed to be spatially constant or near constant, estimated using empirical equations, or optimized through calibration. Accurate estimation of spatially explicit roughness coefficient remains elusive due to its dependency on multiple geographic and hydraulic factors. Elucidating the spatial relationships between these controlling factors and channel roughness might lead to improved flood and hydrological predictions. In this study, we develop a novel data-informed framework to predict Manning's n in the stream networks of the Contiguous United States (CONUS) using observed data. Several open-source datasets, including a large dataset of acoustic current doppler profiler (ADCP) surveys are used to retrieve the features for the analysis. For the first time, a national-scale estimation of Manning's n for thousands of locations across the CONUS is undertaken using quasi-steady-flow assumptions and flow velocities observed by ADCP. Machine Learning (ML) based regressors are applied to predict Manning's n as a function of a set of influencing factors. The outcomes of the study show the applicability and robustness of the ML models in a large-scale deployment with a few data-scarce regions. Random Forest (RF) and Multilayer Perceptron (MLP) with spatial cross-validation showed satisfactory performance in predicting channel roughness. The Feature Importance scheme reveals the rank of features with river discharge, total drainage area, forest, and agricultural land covers and Normalized Difference Vegetation Index (NDVI) being the top five most influential features. The proposed framework reveals the hidden linkages among the myriad of influencing factors and Manning's n, a significant step towards resolving the complexities of channel roughness and ultimately improving flood prediction using a large multi-dimensional dataset in a computationally efficient manner.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H45I1491M