Hybrid Extreme Gradient Boosting and Nonlinear Ensemble Models for Suspended Sediment Load Prediction in an Agricultural Catchment

In this study, four individual models namely Hammerstein-Weiner (HW), Extreme Learning Machine (ELM), Long Short-Term Memory (LSTM) and Least Square Support Vector Machine (LSSVM) were utilized for modeling the SSL of Katar catchment in Ethiopia. Then, two strategies were applied to improve the overall predictive accuracy. The first strategy involves the development of four ensemble techniques such as simple average ensemble (SE), weighted average ensemble (WE), neuro-fuzzy ensemble (NFE) and HW ensemble (HWE) using the SSL output of individual models. In the second strategy, a hybrid Extreme Gradient Boosting (XGB) model was developed to improve the performance of the base models. The accuracy of the models was evaluated using the percent bias (Pbias), Nash–Sutcliffe coefficient efficiency (NSE), mean absolute error (MAE), root mean square error (RMSE) and determination coefficient (R2). The result showed that the HW model provided the best modeling result among the individual models with NSE and Pbias values of 0.922 and -8.8%, respectively, for the testing period. Among the ensemble techniques, NFE provided the best performance by increasing the NSE values of the individual models by 4.88% to 62.52%, during the testing period. Examination of the hybrid XGB models showed that all hybrid models performed reliably, with the HW-XGB model achieving the best predictive performance (NSE = 0.989). Overall, the results of this study showed the promising potential of ensemble and hybrid XGB models for SSL modeling in an agricultural catchment.