Predicting flow velocity in a vegetative alluvial channel using standalone and hybrid machine learning techniques

Abstract

The presence of vegetation in the water bodies has a profound effect on the flow velocity in an open channel due to the resistance offered by it. In rivers, estuaries, and coastal locations, vegetation significantly impacts the local hydrodynamics, which in turn affects various morphodynamic and biophysical processes. In this context, an accurate prediction of flow velocity in a vegetative alluvial channel is paramount. Several empirical and data-driven methodologies have been proposed as viable solutions in the literature to predict the flow velocity in a vegetative alluvial channel. Empirical equations cannot always be trusted to be accurate, but they have the advantage of being simple and physically appealing. Since machine learning (ML) techniques can capture complicated non-linear correlations, they are frequently employed to map natural processes. In this work, we investigate the performance of multiple standalone and hybrid Machine Learning (ML) techniques for predicting flow velocity (U) in a vegetative alluvial channel. An array of datasets available in the literature, comprising wide ranges of the number of cylinders per unit horizontal area (m), flow depth (h), channel slope (i), height of the vegetation (k), diameter of cylindrical vegetation (D), and non-dimensional drag coefficient (Cd) have been utilized in this study. For standalone methods, we made use of the M5Prime and Random Tree (RT) methods, while for hybrid ML method approaches, we made use of the Additive Regressor (AR) and Bagging (BA) methods. In the present study, six ML methods, viz., M5P, AR-M5P, BA-M5P, RT, BA-RT, and AR-RT, have been explored and their performance has also been analyzed. Among the proposed methods, AR-M5P provides the highest prediction (R2 = 0.954, CC = 0.977, NSE = 0.954, MAE = 0.042, MSE = 0.003, and Pbias = 1.466), followed by BA-M5P, BA-RT, M5P, RT, and AR-RT for the prediction of flow velocity in a vegetative alluvial channel. We have also performed the sensitivity analysis and found that the height of vegetation is the most sensitive variable in flow velocity prediction.