Water quality (WQ) monitoring in the surface water resources is a crucial concerning which has a direct impact on the human health and ecosystem equilibrium. In recent years, accurate simulation of river water quality indicators with data mining techniques, among the numerous available variables, is still a challenge for researchers. In this study, two smart dual-preprocessing hybridized with Elman recurrent neural network (ERNN) is developed for accurate simulation of two surface WQ indices including electrical conductivity (EC) and bicarbonate (HCO3-) in two northern zones of Karun River, the longest River of Iran. Here, nine input features including sodium adsorption ratio (SAR), Mg+2, Ca+2, sum of the anions (Sum.A), SO4-2, Cl-, pH, discharge (Q), Na+ are employed to simulate the EC and HCO3-. Two Boruta-data filtering strategies including Boruta-GXBoost (BXGB) and Boruta-Extra tree (BET) were utilized to extract the most important WQ predictors among available features and coupled with the best subset regression (BSR) scheme to optimize the candidate input combinations of both targets. Afterward, three superior scanrios (C1, C2, and C3) for each target were considered to feed the machine learning (ML) models. The aforementioned dual pre-processing was hybridized with ERNN and the comparative advanced ML approaches comprised of the long short term memory (LSTM), Kernel ridge repression (KRR), and Elastic net regression (ELNET). Eight hybrid paradigms (BXGB-ERNN, BET-ERNN, BXGB-KRR, BET-KRR, BXGB-ELNET, BET-ELNET, BXGB-LSTM, and BET-LSTM) were evaluated using statistical measures such as correlation coefficient (R), root mean square error (RMSE), and King-Gupta efficiency (KGE). The outcomes demonstrated that the BET-ERNN-C1 regarding (R = 0.9847, RMSE = 0.0793 mEq/L, and KGE = 0.9782) and (R = 0.9543, RMSE = 51.0260 , and KGE = 0.9406) outperformed the other hybrid models for HCO3- and EC simulation followed by the BXGB-ERNN-C3, and BXGB-KRR approaches.
