A Stacked Machine Learning Approach for HighResolution Urban Water Quality Monitoring 

Water quality is a critical parameter for human health, agriculture, and aquatic ecosystems. Accurate and timely prediction of the water quality index (WQI) enables effective water resource management and proactive measures for contamination control. This study proposes a hybrid deep learning framework that combines multiple approaches for improved WQI prediction. The framework includes a hybrid stacking ensemble of Random Forest (RF), Gradient Boosting (GB), XGBoost (XGB), and Multilayer Perceptron (MLP), as well as deep learning models such as MLP + Single-Layer Perceptron (SLP) + CNN and MLP + SLP + Deep CNN. The hybrid deep learning models leverage the complementary strengths of each architecture: MLP captures nonlinear interactions among features, SLP ensures generalization, and CNNs extract local correlations and patterns across sequential or structured features. A dataset containing physicochemical attributes such as water speed, direction, chlorophyll, temperature, dissolved oxygen, pH, salinity, and turbidity is utilized. Models are trained and validated on real-time measurements and evaluated through standard metrics including Mean Squared Error (MSE), Mean Absolute Error (MAE), and R² score. Results indicate that the hybrid approaches outperform individual models, achieving superior accuracy and reliability in WQI prediction. The framework also provides interpretability through feature importance analysis and generates actionable suggestions based on predicted water quality categories, ranging from excellent to poor. These methods demonstrate significant potential for environmental monitoring and decision support in water quality management. Future work may integrate satellite and IoT sensor data for enhanced predictive capabilities.

Keywords: Water Quality Index, Hybrid Stacking, Deep Learning, MLP, CNN, SLP, RF, GB, XGB, Prediction, Feature Importance, Environmental Monitoring