Deep Learning-Driven Hyperspectral Data Fusion for Real-Time Water Quality Monitoring

Water quality monitoring is essential for addressing water contamination and ensuring public safety. Particularly, managing nitrate levels has become a major concern due to their direct impact on eutrophication. Despite the high accuracy of conventional analysis methods, their practical application is often limited by high costs, labor-intensive processes, and a lack of real-time monitoring capabilities. This study presents a novel framework for real-time water quality monitoring by integrating hyperspectral and multi-sensor data through deep learning-based data fusion. The multi-sensor data includes pH, electrical conductivity (EC), dissolved oxygen (DO), and oxidation-reduction potential (ORP). A transformer-based deep learning model was applied to predict water quality concentrations by capturing correlations within time-series hyperspectral absorbance and multi-sensor data. Furthermore, transfer learning was employed to improve the performance in target domains by transferring the information contained in a pre-trained model. The data-fusion transformer model predicted water quality concentrations with high accuracy, achieving a coefficient of determination (R²) exceeding 0.99 in both deionized and tap water conditions. Specifically, the integration of multi-sensor data improved model robustness and performance compared to applying spectral data alone. This research also demonstrated that transfer learning effectively supported the model in adapting to varying flow conditions. The proposed deep learning-based data-fusion framework provides a reliable solution for real-time water quality monitoring, with aims to extend the model application to predict multiple water parameters simultaneously.