Deep Learning Approaches for Numerical Modeling and Historical Reconstruction of Water Quality Parameters in Lower Seine

Water quality monitoring is essential for managing water resources and ensuring human environmental health. However, obtaining reliable data can be challenging costly, especially in complex systems such as estuaries. To address this problem, we propose a novel deep learning-based approach that uses limited available to accurately estimate reconstruct critical variables, electrical conductivity, dissolved oxygen, turbidity. Our included two tasks, numerical modeling historical reconstruction, was applied the Seine River Normandy region of France at four stations. In first task, evaluated learning approaches (GRU, BiLSTM, BiLSTM-Attention, CNN-BiLSTM-Attention) numerically simulate each variable station under different input selection scenarios. We found incorporating with level collected various stations into improved accuracy simulation. Combining levels from multiple reliably reproduced near sea where tidal fluctuations control saltwater intrusion area. While model had its strengths, CNN-BiLSTM-Attention performed best tasks dissimilar trends, GRU outperformed other models simple similar input-target trends. The second task involved automatically searching optimal configurations completing missing sequential order using results. conductivity were filled before oxygen data, which turn more than turbidity reconstructed 15 years only six half data. Overall, research demonstrates potential their limitations discusses improve reconstruction.