Enhanced water quality prediction model using advanced hybridized resampling alternating tree-based and deep learning algorithms.

Water quality modeling in riverine systems is crucial for effective water resource management and pollution mitigation planning. However, the intricate interplay of anthropogenic activities with hydrological, climatic, and fluvial processes presents significant challenges in developing robust models for predicting water quality parameters. This study develops novel deep learning (DL) models, leveraging bidirectional-LSTM (Bi-LSTM) networks and advanced ensemble-based approaches using bootstrap aggregating (BA) combined with alternating model tree (BA_AMT), to predict key water quality parameters, including daily turbidity (TU) and dissolved oxygen (DO). The proposed hybrid models were applied to the Clackamas River, USA, and their performance was benchmarked against standalone AMT models. The dataset comprised daily records of water discharge (Q), gage height (GH), water temperature (T_w), specific conductance (SC), and pH. Model performance was evaluated under six input combination scenarios to determine optimized input configurations. Results demonstrated the superior predictive accuracy of Bi-LSTM for both TU (Root mean square error-RMSE = 0.172 mg/L, Nash-Sutcliffe efficiency-NSE = 0.985, Percent of IAS-PBIAS, 0.01% and ratio of RMSE to the standard deviation of observation (RSR)-RSR = 0.11) and DO (RMSE = 1.37 mg/L, NSE = 0.713, PBIAS, 1.90% and RSR = 0.53). Sensitivity analysis revealed that models incorporating five input parameters, including Q, GH, SC, and T_w for TU, and T_w, SC, GH, PH, and Q for DO, yielded the best predictive performance. Among these, Q and GH showed the strongest correlation with TU, while T_w, SC, and GH were most influential for DO prediction. While Bi-LSTM outperformed BA-AMT in overall accuracy, the BA-AMT model demonstrated superior capability in capturing extreme values. These findings underscore the importance of optimizing Bi-LSTM models using metaheuristic techniques to enhance predictive performance. The proposed modeling framework offers a scalable and generalizable approach for water quality forecasting and environmental management in freshwater systems, providing a valuable tool for decision-makers.