Data-driven approaches for estimation of sediment discharge in rivers

Sediment discharge in rivers is among the most important water and environmental engineering issues. The present study employed the published reliable field data set and three white-box data-driven methods, including the group method of data handling (GMDH), gene expression programming (GEP), and the multivariate adaptive regression splines (MARS) approach for modeling sediment discharge in rivers (Qs). In addition, the artificial neural network (ANN) model, known as the popular and widely used black-box data-driven model, was employed for modeling sediment discharge. The performance of the proposed methods was explored by statistical measures, scatter plots, and Taylor and Violin plots. The main feature of white-box data-driven models is that they provide explicit mathematical expressions for the prediction of Qs. The outcomes of the proposed methods provide better and more competitive results than the earlier study conducted using the model tree (MT) approach. Statical measurements and graphical plots indicated that all proposed methods have similar results for the prediction of sediment discharge. However, GMDH and GEP were more accurate than the MARS and ANN models. For the overall evaluation of the proposed models, the ranking mean (RM) method was used. This method showed that the GMDH model with RM = 1.86 had better performance in estimating sediment discharge, followed by the GEP with RM = 2.29, MARS with RM = 2.86, and ANN with RM = 3. It is worth mentioning that simple, explicit mathematical expressions generated by GMDH and MARS were more straightforward calculations for estimating sediment discharge compared to GEP and ANN.