Application of Machine Learning Approaches in Particle Tracking Model to Estimate Sediment Transport in Natural Streams

Numerous empirical equations and machine learning (ML) techniques have emerged to forecast dispersion coefficients in open channels. However, the efficacy of certain learning-based models in predicting these coefficients remains unstudied. Also, the direct application of machine learning-derived dispersion coefficients to Lagrangian sediment transport models has not been investigated. The present study utilizes data from prior research to assess the performance of ensemble ML-based models, specifically, random forest regression (RFR) and gradient boosting regression (GBR) inn estimating longitudinal and transverse dispersion in natural streams. The optimal hyper-parameters of these ensemble models were fine-tuned using grid-search cross-validation. The ML-based dispersion models were then integrated into a Lagrangian particle tracking model (PTM) to simulate suspended sediment concentration in natural streams. Suspended sediment concentration distribution maps generated from developed PTM with ML-based dispersion coefficients were compared with field data. The findings indicated that the GBR model, with a coefficient of determination (R2) of 0.95, outperformed the RFR model, which had an R2 of 0.9, in predicting longitudinal dispersion coefficients in a natural stream across both training and testing stages. However, during the testing phase, the RFR model with an R2 of 0.94 performed better than the GBR model with an R2 of 0.91 in predicting transverse dispersion. Both models consistently underestimated dispersion coefficients in both training and testing stages. Comparisons between the PTM with ensemble dispersion coefficients and empirical-based dispersion relationships revealed the superior performance of the GBR model compared to the other two methods.