Sensor-driven machine learning framework for river water quality assessment and classification

ABSTRACT

River water quality monitoring is essential for protecting public health, ecological sustainability, and effective envi ronmental governance. However, conventional laboratory-based monitoring is expensive and fail to determine real-time spatiot emporal variations. Even though machine learning (ML) and Internet of Things (IoT) automate monitoring of various services, these approaches struggle from low interpretability and lack of uncertainty awareness. To address these challenges, this work proposes an IoT-enabled Spatio-Temporal Adaptive Interpretable and Risk-Aware Quantile Learning Network (STAIR-QXNet) for robust river water quality assessment and classification. It involves the use of Seasonal-Trend Decomposition using Locally Estimated Scatterplot Smoothing (STL) method for temporal decomposition, and an Information SHapley Additive exPlanations -Boruta Fusion (ISBF) approach for stable features selection. Then, a Temporal Boosted Convolutional Learning Framework (TB-CLF) based on Light Gradient Boosting Machine (LightGBM) and Temporal Convolutional Networks (TCN) is integrated to learn and capture non-linear feature interactions. Further, quantile regression enables uncertainty estimation for predictions, and SHAP offers interpretability for the model. Meanwhile, experimental results demonstrate approximately 99% classification accuracy, low training and testing loss (~0.2), and an AUC of 0.986, with narrow prediction intervals. Thereby, delivering precise a nd meaningful environmental decision-making for proactive river water quality monitoring.