Beyond conventional modeling: A cutting-edge hybrid IAER-AMT decision-tree-based algorithm for high-resolution river turbidity prediction

Water Turbidity (TU) is a widely used indicator of water quality. Given the time-consuming nature of direct TU measurement, developing an accurate predictive model is imperative. In this study, the alternating model tree (AMT) and its ensemble version through iterative absolute error regression (IAER), bootstrap aggregating (BA), weighted instance handler wrapper (WIHW), and random subspace (RS) were used to predict TU at Clackamas River, USA. Daily time-series data of the physicochemical water quality variables from 2006 to 2023, including water temperature (Tw), specific conductance (SC), dissolved oxygen (DO), pH, as well as physical river parameters, including daily water discharge (Q) and water stage (WS), were used as potential input variables to predict TU. The manual approach, principal component analysis (PCA), and correlation-based feature selection subset evaluation (CfsSubsetEval) techniques were compared under different input scenarios. Finally, the performance of the models was evaluated using various statistical metrics, including the coefficient of determination (R2), root mean squared error (RMSE), percentage of bias (PBIAS), Nash-Sutcliffe efficiency (NSE), and root mean standard deviation ratio (RSR). WS had the highest impact on TU prediction, whereas Tw was less correlated. In addition, the input scenario that included all variables led to the highest model performance. Based on the testing dataset, the novel IAER-AMT hybrid algorithm outperformed others, achieving an RMSE of 1.20 Formazin Nephelometric Units (FNU), an NSE of 0.72, a PBIAS of 3.17 %, and an RSR of 0.53 followed by BA-AMT (RMSE = 1.30 FNU, NSE = 0.67, PBIAS = −9.73%, and RSR = 0.57), WIHW-AMT (RMSE = 1.34 FNU, NSE = 0.65, PBIAS = −0.35%, RSR = 0.58), RS-AMT (RMSE = 1.37 FNU, NSE = 0.64, PBIAS = −20.95%, and RSR = 0.60), and AMT (RMSE = 1.38 FNU, NSE = 0.63, PBIAS = −26.81%, and RSR = 0.60). © 2025 Elsevier B.V., All rights reserved.