Advanced machine learning models development for suspended sediment prediction: comparative analysis study

Abstract

Accurate estimation of suspended sediment (SS) is very essential for planning and management of hydraulic structures. The study investigates the accuracy of four machine learning methods, dynamic evolving neural-fuzzy inference systems (DENFIS), fuzzy c-means based adaptive neuro fuzzy system (ANFIS-FCM), multivariate adaptive regression spline (MARS) and M5 model tree (M5Tree), in estimating suspended sediments. Several input scenarios including streamflow (Q) and sediment (S) data obtained from Ain Hamara Station in Wadi Abd basin, Algeria were constructed to find the most effective one. The research results indicate that the DENFIS model with current streamflow (Qt) and 1 previous sediment (St-1) values performs superior to the other alternatives in SS estimation; it increases the efficiency of the best ANFIS-FCM, MARS and M5Tree by 1.6%, 15.7% and 9.6% with respect to RMSE (root mean square error), respectively. Variation of Q and S data on models’ estimation ability was also investigated and it was found that the variation input considerably increase the prediction ability of MARS method; increments in RMSE and MAE (mean absolute error) are by 10.8 and 4.9% and decrement in NSE (Nash-Sutcliffe efficiency) is by 12.9%.

Keywords:

• Mapping streamflow-sediment relationship
• DENFIS
• ANFIS-FCM
• MARS
• M5 model tree

Acknowledgement

The data (streamflow and suspended sediment) used in this study were obtained from the National Agency of Hydraulic Resources (ANRH).

Disclosure statement

The authors have no conflict of interest to declare.

Data availability

Data is available upon request from the corresponding author.