Abderrahim Jardani

River discharge is an integrative hydrological variable that is essential for water resource management, flood risk assessment, and ecosystem monitoring. However, long and continuous discharge records are often incomplete or unavailable. This study develops an explainable and uncertainty-aware deep-learning framework for daily discharge simulation. It is applied to three contrasting gauging stations of the Seine system: the Seine at Vernon, the Eure at Louviers and the Risle at Pont-Authou, representing distinct hydrological contexts, corresponding to a large integrated basin, a chalk-influenced basin and a karst-affected basin, respectively. The objective is to investigate how basin hydrological functioning shapes model interpretability, the relevance of CNN-based spatial feature extraction, and the reliability of predictive uncertainty. SHAP analysis shows basin-dependent controls: precipitation- and runoff-related variables dominate the Seine simulation, whereas radiation-related variables are the main predictors for the Eure and Risle, consistent with stronger subsurface storage and delayed hydrological responses. Model evaluation using statistical metrics, hydrological indicators and selected high-flow events shows that enriched LSTM inputs perform best for the Seine, while CNN-derived spatial representations improve simulations for the Eure and Risle. Across independent test periods, the best configurations achieve high daily discharge accuracy, with NSE values generally above 0.9. Uncertainty analysis indicates that standard deep ensembles underestimate predictive uncertainty, whereas probabilistic and hybrid approaches improve interval reliability. The extended ensemble with input perturbation consistently reaches the nominal 95% coverage target, although with wider prediction intervals.

In karstic environments, water supply wells are vulnerable to rapid sediment transfer during intense rainfall events, often generating turbidity peaks that disrupt water-treatment operations. In Normandy (France), the high density of sinkholes and the complexity of transport processes in karsts complicate the identification and prioritization of sinkholes requiring mitigation to reduce sediment fluxes at water supply wells. This study aims to quantify the time-lagged impact of each sinkhole on turbidity peaks at a supply well using a cascade modeling approach that couples numerical surface erosion–runoff simulations with deep learning models representing hydrosedimentary responses through the karst network. Surface erosion–runoff was simulated using WaterSed. Hydroclimatic time series and WaterSed model outputs were used as inputs for our deep learning models. Several deep learning architectures were compared and optimized across multiple rounds to identify a best-performing model, which was then interpreted using interpretability methods. Interpretability analyses show that turbidity is primarily controlled by seasonal conditions and short-term rainfall accumulation, while multiple sinkholes contribute jointly to short time lags. Temporal attributions reveal rapid karst response followed by attenuation, consistent with reactive karst behavior. The contribution of each sinkhole to turbidity peaks allows us to identify the most important sinkholes requiring mitigation by stakeholders.

In a context where anticipating future trends and long‐term variations in water resources is crucial, improving our knowledge about most types of aquifer responses to climate variability and change is necessary. Aquifers with variability dominated by seasonal (marked annual cycle) or low‐frequency variations (interannual to decadal variations driven by large‐scale climate dynamics) may encounter different sensitivities to climate change. We investigated this hypothesis by generating groundwater level projections using deep learning models for annual, inertial (low‐frequency dominated) or mixed annual/low‐frequency aquifer types in northern France from 16 CMIP6 climate model inputs in an ensemble approach. Generated projections were then analyzed for trends and changes in variability. Generally, groundwater levels tended to decrease for all types and scenarios across 2030–2100 without any significant differences between emission scenarios. However, when comparing future projections to historical data, groundwater levels appeared slightly higher in the near future (2030–2050), with decreasing intensities in later periods. The variability of projections showed slightly increasing variability for annual types for all scenarios but decreasing variability for mixed and inertial types. As the severity of the scenario increased, more mixed and inertial‐type stations appeared to be affected by decreasing variability. Focusing on low‐frequency confirmed this observation: while a significant amount of stations showed increasing variability for the less severe SSP2‐4.5 scenario, low‐frequency variability eventually showed slight yet statistically significant decreasing trends as the severity of the scenario increased. For the most severe scenario, almost all stations were affected by decreasing low‐frequency variability.

In a context where anticipating future trends and long-term variations in water resources is crucial, improving our knowledge about most types of aquifer responses to climate variability and change is necessary. Aquifers with variability dominated by seasonal (marked annual cycle) or low-frequency variations (interannual to decadal variations driven by large-scale climate dynamics) may encounter different sensitivities to climate change. We investigated this hypothesis by generating groundwater level projections using deep learning models for annual, inertial (low-frequency dominated) or mixed annual/low-frequency aquifer types in northern France from 16 CMIP6 climate model inputs in an ensemble approach. Generated projections were then analysed for trends and changes in variability. Generally, groundwater levels tended to decrease for all types and scenarios across the 2030-2100. The variability of projections showed slightly increasing variability for annual types for all scenarios but decreasing variability for mixed and inertial types. As the severity of the scenario increased, more mixed and inertial-type stations appeared to be affected by decreasing variability. Focusing on low-frequency confirmed this observation: while a significant amount of stations showed increasing variability for the less severe SSP 2-4.5 scenario, low-frequency variability eventually showed slight yet statistically significant decreasing trends as the severity of the scenario increased. For the most severe scenario, almost all stations were affected by decreasing low-frequency variability. Finally, groundwater levels seemed, in most instances, slightly higher in the future than in the historical period, without any significant differences between emission scenarios.

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