Enhancing Water Quality Monitoring: An Integrated Machine Learning Framework with Physical Constraints for Imputation and Time Series Downscaling

Extreme climate change can lead to a drastic deterioration in water quality. However, researchers often struggle to find long-term water quality monitoring data, especially at a daily scale, which hinders the understanding of the response relationship between extreme climate, hydrology, and water quality. This study proposes an integrated machine learning framework with physical constraints from various environmental domains such as meteorology and water quantity that can effectively impute a high percentage of missing data and downscale time series data of water quality, producing satisfactory results. Over 78% of the physical water quality variables exhibit NSE (Nash-Sutcliffe efficiency coefficient) values greater than 0.45, and more than 66% of the chemical water quality variables achieve NSE values greater than 0.35. The results of this work demonstrate the effectiveness of the proposed framework as a data augmentation and temporal interpolation tool to enhance the adequacy of water quality monitoring and explore the mechanisms underlying the impact of extreme climate on water quality.