Explainable hybrid ensemble regression tree algorithms and ElasticNet regression for river suspended sediment load estimation

Estimating the suspended sediment load (SSL) and understanding the impacted factors can aid in watershed management. This study evaluates the potential of ElasticNet linear regression (ENLR), random forest (RF), AdaBoost, CatBoost, and extreme gradient boosting (XGB) for SSL estimation at Ziway catchment (Ethiopia). The Shapley additive explanation method was used to determine the contribution of different driving factors for river SSL generation. Then, hybrid models, namely ENLRXGB, ENLRRF, ENLRAdaBoost, and ENLRCatBoost, were proposed by coupling the ENLR model with ensemble tree models to capture both the linear and non-linear patterns of SSL. The finding revealed that XGB outperformed other individual models. In addition, combining linear and non-linear models produced better performance, i.e. ENLRXGB demonstrated the highest accuracy Nash-Sutcliffe efficiency (NSE = 0.967), improving the base model’s NSE by 3.755–21.33%. Overall, the results demonstrate the efficiency of hybrid models for estimating river SSL in this data-limited agricultural catchment.