Authors: Riley McDermott, Sagy Cohen, Reihaneh Zarrabi – University of Alabama; Mohammad Erfani – University of South Carolina
Title: Estimation of River Channel Shape Using Regression and Machine Learning Approaches
Abstract: Emerging datasets of riverine observations at large scale and fidelity are enabling the development of more robust estimates of river hydraulic characteristics. More specifically, the availability of a large dataset containing Acoustic Doppler Current Profiler (ADCP) measurements, the gold standard of underwater channel attribute measurements, offers pathways for developing new models of channel geometry and shape. Previous methods of modeling channel geometry were limited by dataset size and quality of channel dimension measurements. The new datasets, on the other hand, can be challenging to use due to their abundance of observations, unfiltered errors, and inclusion of observations during over-bank (flooding) conditions, which do not reflect channel dimensions. Cleaning the data in order to achieve robust channel geometry estimations is challenging. This project aims to develop a first-order channel shape estimation model based on these large CONUS-scale datasets. First, we develop a methodology for identifying and removing overbank observations in order to ensure that only within-channel measurements are used. We found that a derivative of a fitted polynomial regression on observed channel width and depth dimensions and discharge works well in identifying and removing near-bankfull observation points. A data-driven model of channel shape will be developed using the derived dataset. Multiple methods of reconstructing channel cross-sectional shape from observational data will be compared to observed channel cross-sections to identify the most accurate approach, which will be used in the development of continental channel shape models. These models aim to improve the quantification of channel geometric parameters with the goal of advancing the field of hydrologic modeling and improving the generation of flood inundation mapping (FIM).