Bankfull and Mean-Flow Channel Geometry Estimation Through Machine Learning Algorithms Across the CONtiguous United States (CONUS)

Abstract

Widely adopted models for estimating hydraulic geometry attributes rely on simplistic power-law equations, which can introduce inaccuracy due to their inability to capture spatial variability. This study introduces a new model for predicting channel geometry utilizing advanced tree-based Machine Learning (ML) algorithms. The research enhances the quality of the extensive HYDRoacoustic data set supporting Surface Water Oceanographic Topography (HYDRoSWOT) through a proposed preprocessing method. Observations of bankfull and mean-flow conditions at each gauge site are identified and extracted as target variables for model development. HYDRoSWOT-extracted attributes, along with other predictors from various sources, such as National Hydrography Data set Plus (NHDPlusV2.1), are used to train and validate predictive models. The models achieve average R² values of 0.85 for channel width and 0.69 for channel depth, demonstrating high accuracy in capturing spatial variability in hydraulic geometry attributes. Independent evaluations further test the models' performance in predicting reach-averaged conditions at locations outside the training and testing data sets. The results show that the proposed model significantly outperforms existing regional hydraulic geometry relations, with accuracy improvements of 30% for bankfull width and 76% for bankfull depth. The proposed model is then utilized to generate channel width and depth under bankfull and mean-flow conditions data set across approximately 2.7 million streams within NHDPlusV2.1 data set across the CONtiguous United State (CONUS). This data set is a valuable resource for water-related sciences, including hydrology, geomorphology, flood modeling, water quality assessment, and flood management.