A Novel Bagging Enhanced Feature Selection Method in Coupling with Bayesian Optimization for Reliable Water Quality and Quantity Modeling

Effective water resource monitoring is essential for sustainable management and public well-being. This study utilizes machine learning (ML) techniques to analyze discharge and quality parameters, namely DO, EC, and pH of water bodies. A novel feature selection approach based on Bootstrap Aggregation (Bagging) was developed, drawing inspiration from Neighborhood Component Analysis (NCA). To improve predictive performance, two ensemble learning models were developed: Bootstraped-NCA-Ensemble Machine Learning (BNCA-EML) and Bootstraped-NCA-Ensemble Bayesian Machine Learning (BNCA-EBNML). Both frameworks were built using the Ensemble Averaging method, ensuring a well-balanced combination of base models. Before integration into the ensemble structure, individual basic-models, including Gaussian Process Regression (GPR), Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Regression Trees (RT), were optimized using Bayesian hyperparameter tuning. The BNCA-EBNML model, optimized through Bayesian optimization, demonstrated superior accuracy compared to conventional models, achieving an R² of ۰.۹۹۸۷ for discharge prediction and R۲ values of ۰.۹۴, ۰.۹۶, and ۰.۹۹ for other water quality indicators. Uncertainty analysis confirmed that BNCA-EBNML was the most stable and reliable framework among the tested models. These results emphasize the effectiveness of the proposed methodology, making it a valuable and data-driven tool for researchers and decision-makers in water resource management.

کلیدواژه ها:

Water Quality ، Neighborhood Component Analysis ، Bootstrap Aggregation ، Bayesian Optimization ، Ensemble learning