This paper presents the results of the 2022 Groundwater Time Series Modelling Challenge, where 15 teams from different institutes applied various data-driven models to simulate hydraulic-head time series at four monitoring wells. Three of the wells were located in Europe and one was located in the USA in different hydrogeological settings in temperate, continental, or subarctic climates. Participants were provided with approximately 15 years of measured heads at (almost) regular time intervals and daily measurements of weather data starting some 10 years prior to the first head measurements and extending around 5 years after the last head measurement. The participants were asked to simulate the measured heads (the calibration period), to provide a prediction for around 5 years after the last measurement (the validation period for which weather data were provided but not head measurements), and to include an uncertainty estimate. Three different groups of models were identified among the submissions: lumped-parameter models (three teams), machine learning models (four teams), and deep learning models (eight teams). Lumped-parameter models apply relatively simple response functions with few parameters, while the artificial intelligence models used models of varying complexity, generally with more parameters and more input, including input engineered from the provided data (e.g. multi-day averages). The models were evaluated on their performance in simulating the heads in the calibration period and in predicting the heads in the validation period. Different metrics were used to assess performance, including metrics for average relative fit, average absolute fit, fit of extreme (high or low) heads, and the coverage of the uncertainty interval. For all wells, reasonable performance was obtained by at least one team from each of the three groups. However, the performance was not consistent across submissions within each group, which implies that the application of each method to individual sites requires significant effort and experience. In particular, estimates of the uncertainty interval varied widely between teams, although some teams submitted confidence intervals rather than prediction intervals. There was not one team, let alone one method, that performed best for all wells and all performance metrics. Four of the main takeaways from the model comparison are as follows: (1) lumped-parameter models generally performed as well as artificial intelligence models, which means they capture the fundamental behaviour of the system with only a few parameters. (2) Artificial intelligence models were able to simulate extremes beyond the observed conditions, which is contrary to some persistent beliefs about these methods. (3) No overfitting was observed in any of the models, including in the models with many parameters, as performance in the validation period was generally only a bit lower than in the calibration period, which is evidence of appropriate application of the different models. (4) The presented simulations are the combined results of the applied method and the choices made by the modeller(s), which was especially visible in the performance range of the deep learning methods; underperformance does not necessarily reflect deficiencies of any of the models. In conclusion, the challenge was a successful initiative to compare different models and learn from each other. Future challenges are needed to investigate, for example, the performance of models in more variable climatic settings to simulate head series with significant gaps or to estimate the effect of drought periods.

Continuer la lecture Partager

Carbon-14 (14C) has a natural origin but is also anthropogenically released from civil nuclear facilities. Due to its long decay period (half-life: 5700 ± 30 years), it is a persistent radionuclide in the environment. In rivers, the complex speciation of carbon makes the fate of industrial 14C difficult to track. This study reports a first overview of artificial 14C cycling in a nuclearized river. A one-year sampling campaign was conducted on the French nuclearized Rhône River and two of its non-nuclearized tributaries (Durance and Ardèche rivers). Isotopic (δ13C, Δ14C) and carbon concentrations analyses were performed on the particulate organic carbon (POC), dissolved organic carbon (DOC) and dissolved inorganic carbon (DIC). Chlorophyll-a (Chl-a) and tritium analyses were performed to assess the dynamic of aquatic organic matter and the nuclear industry contribution, respectively. Comparisons of Δ14C data obtained from the Rhône River with those from the tributaries highlight significant industrial radiocarbon labelling in all carbon forms, with medians of 142, 130 and 42 ‰ for POC, DOC and DIC, that are 2–3 times higher than those of the tributaries. The high values of Chl-a/POC ratios with Δ14C-enriched POC suggest a biological uptake of artificial Δ14C in DIC by aquatic photosynthesis. The relationship of Δ14C-DIC with tritium activity indicates a response to recent releases and enables the contribution of nuclear power plants to be estimated at a median of 26 %. Sampling at the Rhône's mouth would reinforce our understanding of the fate of riverine 14C when entering the marine environment.

Continuer la lecture Partager

Groundwater level (GWL) can vary over a wide range of timescales. Previous studies highlighted that low- frequency variability (interannual (2–8 years) to decadal (>10 years)) originating from large-scale climate variability, represents a significant part of GWL variance. It remains an open question, however, how GWL, including extremes, may respond to changes in large-scale climate forcing, affecting precipitation variability. Focusing on the Seine River basin, this study therefore aims to assess how GWLs respond to changes in interannual to decadal climate variability. We implemented an empirical numerical approach, which enables an assessment of the GWL sensitivity to changes in precipitation variability over a range of timescales (up to decadal), using the Seine hydrosystem as a case study. The approach consists in: i) identifying and modifying the spectral content of precipitation in the low- frequency range; ii) using these perturbed precipitation fields as input in the physically-based hydrological/ hydrogeological CaWaQS model for the Seine River basin to simulate the corresponding GWL response; iii) comparing the spectral content, mean, variance and extremes of perturbed GWLs with reference (i.e. unperturbed) GWLs. Two interannual (2–4 yr and 5–8 yr) and one decadal (15 yr) timescales were modified individually by either increasing or decreasing their amplitude by 50 %. This led to six scenarios of perturbed low- frequency precipitation variability, which were subsequently used as CaWaQS inputs to assess the GWL response. Results indicated increased (decreased) GWL up to 5 m when low-frequency precipitation variability increased (decreased) by 50 %. This led to an increased occurrence of groundwater floods (droughts) with increased severity and decreased occurrence of groundwater droughts (floods) with decreased severity, respectively. These results indicate: i) how using biased climate data, in terms of low-frequency variability, leads to large deviations in the GWL simulation, ii) to what extent potential changes in low-frequency climate variability may affect future GWL, and particularly drought and flood occurrence and severity.

Continuer la lecture Partager

Environmental imprint of inorganic fertilizer uses was assessed over the last hundred years at the downstream part of large French rivers (Loire, Moselle, Rhine, Rhone, Meuse and Seine rivers) based on Potassium-40 (40K) activity concentration data sets acquired from soil monitoring (1980–2022) and from sediment coes collected from 2020 to 2022 to reconstruct the temporal trajectories of 40K activity concentrations since the beginning of the last century. Cultivated soils were significantly enriched in 40K compared to non-cultivated ones in the 1980s and 1990s when they turned back to the contents of non-cultivated soils during the following decades. In riverine sediments, all the rivers displayed close 40K temporal trajectories with peaking 40K contents in fine grain size sediments in the 1980s. Maximum 40K enrichment factors from this period were related to the proportion of agricultural areas in the river catchment. In the Loire and Moselle rivers, some high 40K contents were associated with sandy sedimentary strata deposited by flood events before the end of the 1950s due to the presence of potassium enriched minerals. The comparison of 40K activity concentration in sediments with potassic fertilizer delivery in France highlighted very similar temporal trajectories giving evidence that the uses of potassic fertilizers imprint the riverine sediments of most French large rivers. Finally, the environmental resilience face to this anthropic pressure was fast as 40K levels decreased immediately after the decreases of the delivery in most of cases.

Continuer la lecture Partager

Depuis 1985, la rade de Cherbourg fait l’objet de suivis environnementaux qui intègrent aussi bien la masse d’eau que le biota et les sédiments marins. Ces suivis sont intégrés dans le cadre du programme CHERAD, programme de monitoring de la rade. L’étroite collaboration mise en place entre l'équipe de recherche du CNAM-Intechmer et les responsables d’un élevage de salmonidés dans la rade a amené à mettre en place une série d’actions plus spécifiques rassemblées dans le projet SALMOCOT. SALMOCOT III a été mis en place dès 2023 a l’occasion d’une restructuration de l’élevage de saumons (Salmo salar). Dans ce rapport, une analyse des communautés benthiques (macrofaune benthique et analyse de la nature des sédiments), sous les cages situées en grande rade, a été menée, afin d’évaluer l’état des communautés suite à l’arrêt de l’élevage durant deux ans et avant la remise en eau d’une nouvelle population de poissons. Le 2 février 2024, au niveau du site d’implantation des cages à saumon, quatre stations dans la grande rade ont été étudiées. L’année 2024 se caractérise par une richesse taxonomique moyenne, des abondances et biomasses très faibles et un diagnostic écologique très bon à moyen, selon les stations. Des différences sédimentaires sont également constatées. Les différentes analyses montrent pour 2024, des compositions faunistiques et en abondances particulières pour chaque station échantillonnée. Les comparaisons 2014-2024 (cages en exploitation en 2014) montrent deux situations très tranchées entre 2014, avec une faune riche et caractéristique d’une zone enrichie en matière organique, en comparaison de 2024 où une faune appauvrie sans indication d’espèces témoins d’enrichissement en matière organique est observée. Ceci peut être mis en lien avec l’arrêt de l’élevage depuis 2 ans (cages vides), qui contribuait de manière directe à l’apport en matière organique localement, via les restes de nourriture et les fèces de poissons. L’apport en matière organique n’étant plus présent suite à l’arrêt provisoire de l’élevage, une faune composée d’espèces non-indicatrices d’enrichissement en matière organique s’est installée. Cette faune est peu abondante et peu diversifiée, probablement en lien avec la présence d’un substrat composé d’une grande quantité de débris coquilliers et de coquilles entières d’huîtres et de Saint-Jacques, permettant difficilement l’installation d’endofaune.

Continuer la lecture Partager

Accurate groundwater level (GWL) simulations facilitate reconstructions and projections for analysing historical and future groundwater trends and variability at the decadal scale. In this thesis, we investigate the use of deep learning (DL) approaches for GWL simulations, reconstructions, and projections, with a focus on capturing low-frequency variability and leveraging climate reanalysis and GCM model outputs. A wavelet-assisted DL framework was developed, using the Maximal Overlap Discrete Wavelet Transform (MODWT) as a pre-processing step to decompose input signals. We specifically evaluated advanced DL models, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM), for single-station and multi-station approaches. The single station approach results indicated that MODWT-assisted GRU models allowed for extracting low-frequency information and significantly outperformed standalone models in simulating GWLs, particularly for inertial-type GWL. The Shapley Additive Explanations (SHAP) technique was used to interpret model outputs and highlight important input features. For long-term GWL reconstructions, DL models were trained on ERA5 and ERA20C climate reanalysis datasets, enabling reconstructions up to 1900 and 1940, respectively. These DL-based models were able to capture multi-decadal variability in all reconstructed GWLs. Several multi-station training approaches and clustering were used for large-scale GWL simulations, incorporating dynamic climatic variables and static aquifer characteristics. Models specifically trained on different GWL types, clustered by spectral properties, performed significantly better than those trained on the whole dataset. Finally, A multi-station GRU model trained for each GWL type with boundary-corrected MODWT (BC-MODWT) pre-processing was used to generate projections until 2100. Future changes show decreasing trends in groundwater levels and variability, intensifying from SSP2-4.5 to SSP5-8.5, despite projected groundwater levels being higher on average compared to the historical period in all scenarios. We explain this seemingly counter-intuitive result by the fact that projected levels are systematically much higher at the beginning of the future period (up to ~2050) compared to the historical period. Finally, our results indicate that the variability of annual-type aquifers has increased for all emission scenarios.

Continuer la lecture Partager