Physics-Informed Data Denoising for Enhanced IoT-based Water Quality Monitoring

Eutrophication in water bodies is a critical problem, particularly in harbors with high nutrient concentrations. This excess nutrient load promotes the growth of harmful algae and bacteria, leading to oxygen depletion and hypoxic conditions, which can disrupt aquatic ecosystems and pose health risks to humans by contaminating drinking water and seafood. The integration of IoT in water quality monitoring enables real-time data collection and analysis, facilitating timely intervention. However, environmental noise, calibration drift, and transmission errors often affect sensor measurements, leading to unreliable data. Accurate water quality data is essential for effective monitoring. In this paper, we propose a Physics-Informed Data Denoising approach to enhance the reliability of water quality data. Empirical evaluations using two factual datasets of Tolo Harbor (Hong Kong) and New York Harbor demonstrate significant performance improvement over state-of-the-art water quality assessment systems. Our method reduces the root mean square error (RMSE) and improves the coefficient of determination (R²) scores for multiple water quality parameters. Our results indicate that integrating physics-based constraints into the data analysis improves prediction accuracy and sensor reliability. Additionally, we conducted an ablation study on the effect of physics-informed denoising. The study reveals that incorporating physical constraints improves stability and predictive performance across monitoring stations, emphasizing the importance of a physics-informed approach in data-driven water quality modeling.