Understanding the variations of streamflow is critical for studying the ecology of river systems. Low flow periods can pose significant pressures to river ecosystems, including flow intermittence, drying, increasing water temperature or pollutant concentration. Process-based spatially distributed hydrological models, can be used to simulate streamflow along river networks and provide valuable insights to study river ecology. However, until now, the use of these models to simulate streamflow at a high resolution, with a focus on low flow periods, has been limited. Therefore, traditional calibration techniques need to be adapted and refined to effectively address the challenges of sustainable river system management.

In this study, we present a new approach to calibrate a process-based spatially distributed hydrological model (JAMS/J2000) to optimize the simulation of low flows in river systems. This approach combines traditional efficiency criteria (KGE, NSE, pBias, RMSE, etc.) with hydrological signatures specific to low flows (KGE(sqrt(Q)), 10th quantile, base flow index, etc.) to optimize the model at multiple gauging stations. In order to select the optimized parameter set, simulated ensembles of different parameter sets were generated using Latin Hypercube sampling and an objective function developed that combines the efficiency criteria at each gauging station. This allowed us to evaluate the performance of multiple potential parameter sets and select the one that optimizes the simulation of low flows in the river system.

This calibration method was applied in 6 mesoscale catchments in different European countries (Croatia, Spain, Finland, France, Hungary, Czech Republic), which cover different ecoregions in Europe. The study was conducted as part of the Horizon 2020 DRYvER project (Datry et al. 2021). Our results show that the integration of hydrological signatures in the objective function has a strong impact on the calibration procedure and improves model performance during low flow periods.

Datry et al. (2021) Securing Biodiversity, Functional Integrity, and Ecosystem Services in Drying River Networks (DRYvER). Research Ideas and Outcomes. https://doi.org/10.3897/rio.7.e77750

How to cite: Künne, A., Mimeau, L., Lauvernet, C., Devers, A., Branger, F., and Kralisch, S.: Optimizing low flow predictions in river systems: a multi-objective, multi-gauge calibration approach for process-based hydrological models, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14448, https://doi.org/10.5194/egusphere-egu23-14448, 2023.

Regional scale hydrological models are often constrained by a group of observation stations, typically for discharge, which each represent a lumped catchment response. While multi-station calibration greatly improves model fidelity, other sources of data and different calibration objectives are often required to improve models for other variables and increase robustness for ungauged areas. Satellite data has often been utilized as an additional source of information for multi-objective optimization. However, in many cases satellite-based data for other variables, such as soil moisture, AET, snow cover, storage change etc. has been applied as timeseries of catchment averages, thereby underutilizing the unique spatial pattern information they carry.

In a series of studies a simple alternative approach has been developed to capitalize on the benefits of combining spatial pattern information from satellite data with classical discharge and groundwater head observations for model optimization. By limiting the constraint by the satellite data to pattern information only a very limited tradeoff with other observations is achieved. Meanwhile, the approach ensures realistic spatial patterns of parameter fields and simulations leading to improved transferability to ungauged basins.

In light of equifinality, which is often encountered for regional scale models constrained by multiple discharge stations, the approach can as such also be seen as an efficient way of identifying spatially consistent and realistic solutions among a large sample of plausible parameter sets.

Here we present two cases, one across six central-European basins using a mesoscale hydrological model (mHM) and another using a national scale groundwater-surface model (MIKE-SHE).

How to cite: Stisen, S., Cüneyd Demirel, M., Soltani, M., and Koch, J.: Spatial pattern oriented optimization of regional scale hydrological models, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-7955, https://doi.org/10.5194/egusphere-egu23-7955, 2023.

In the last years, different providers (e.g. NOAA, MeteoFrance, ECMWF, DWD) produced meteorological gridded global and regional data sets. These model outputs are now reanalysis and operative forecasts distributed as open data, with different spatial and temporal resolutions. The aim of this contribution is the comparison of short-term streamflow forecasting outputs using different meteorological data sets as forcing.

The use of the above-predicted weather time series as input for the hydrological models allows the evaluation of the streamflow at the catchment closure. Hourly data sets of temperature and precipitation from the different providers were selected for this contribution and the evaluation of the short-term streamflow forecasting results on three small catchments in the Alps was carried out.

A data-driven forecasting procedure with the Support Vector Regression machine learning algorithm as a tool for hydrological modeling was implemented. For each provider, training and testing phase were performed using as forcing the weather model outputs of the same provider, in this way the simulations are consistent. The training phase takes advantage of reanalysis data sets when available, otherwise historical forecasting products are used as an alternative. Instead, the testing period is forced by lags of the temperature, precipitation and streamflow metered data for the past. The number of lag hours is defined in the training stage with a grid search approach. Moreover, the actual temperature and precipitation forecasting data sets cover the prediction lead time. The training was carried on for each day of the training period and the output of each run is the hourly short-term streamflow prediction.

The use of multiple data sources as input allows us to emphasize the differences between global and local meteorological forcing. Moreover, the simulation ensembles outputs allow the identification and quantification of uncertainty that lead to a better interpretation of the prediction. These results are promising in hydrological modeling to increase the final accuracy of the streamflow predictions and the decision-making under uncertainty.

How to cite: Dalla Torre, D., Menapace, A., Zanfei, A., and Righetti, M.: Data-driven streamflow forecasting analysis leveraging multiple meteorological providers, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11347, https://doi.org/10.5194/egusphere-egu23-11347, 2023.

Digital twins are part of ongoing digital transformation to test, monitor, and maintain physical environments virtually. The collaboration of smart measurement sensors, advanced communication networks, cloud data storage capacity, and cutting-edge computing techniques has the potential to create a digital twin of a river basin with greater physical, spatial, and temporal scalability. The digital twin is defined as a realistic virtual representation of the physical river basin that aids in improved decision-making through real-time data connectivity, association, and relationship. Because of the continuous bidirectional interactions between virtual and physical entities, the digital twin is unique to the physical river basin. The digital twin has the advantage of adapting to changing real-time river basin characteristics, resulting in increased operational efficiency, better uncertainty quantification, early warning detection, and identification of emergency management. By imagining smart river basin management via the digital twin concept, we are venturing into uncharted territory, with the goal of improving the ecological status of a river basin by balancing environmental and socioeconomic interdependence while minimizing natural resource depletion. This poster provides an overview of the concept, framework, methodology, and challenges involved in developing a digital twin of a river basin. The framework’s six dimensions are river basin, data, modeling, infrastructure, service, and connectivity. The methodology emphasizes the digital twin’s purpose identification, maturity spectrum, workflow architecture, technical core, data layers, model simulations, knowledge creation, and effective application. We discuss the key services provided by the digital twin for the river basin, as well as its future prospects for autonomous control in the physical river basin.

How to cite: Pal, D., Ala-Aho, P., Ronkanen, A.-K., Marttila, H., Lotsari, E., Korppoo, M., Silander, J., Blåfield, L., Alho, P., Uvo, C., Kämäri, M., and Kaartinen, H.: Blueprint for a digital twin of a river basin, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15057, https://doi.org/10.5194/egusphere-egu23-15057, 2023.

Bayesian non-parametric models are rarely used for predictive modeling of recreational waters. In the present study, we use a Dirichlet Process Gaussian Mixture Model (DPMM) for model-based clustering of hydrologic data collected at three river bathing sites (3 rivers, N = 256, N = 281, N = 1170). The three sites differ in their climatic conditions. Rivers 1 and 3 are continentally influenced (highly unbalanced dataset with few but severe contamination episodes); River 2 is more maritime-influenced (regular rainfall leads to balanced data set with regularly occurring pollution episodes); DP models can be used for model-based clustering, where the number of clusters does not have to be pre-defined but is inferred from the dataset itself. For each new observation x­_I, the probability of belonging to an already existing cluster as well as the probability of belonging to a new cluster is calculated. We used this property to identify unknown, i.e. high-risk situations, at the individual river sites.

We first applied the DPMM to the available hydraulic training data for model training before conditionally updating a predefined lognormal prior for each cluster, representing the E.coli concentration in the river. For prediction, we first evaluated whether a new observation belongs to an existing cluster or whether it constitutes a new cluster. Based on this evaluation, we used either the posterior predictive distribution or the prior predictive distribution for cases where a new cluster was identified. The water quality assessment was subsequently based on the 90^th and 95^th percentiles of the individual predictive distribution. Model performance was evaluated by means of calculating four criteria: (i) the root mean squared error (RMSE), (ii) the percentage coverage of predictive intervals in relation to the test data (80%), (iii) the detection rate of confirmed contaminations (E.coli > 1800 MPN/100 mL), and (iv) the number of predicted bathing days in the test data. The ratio between training and test data was incrementally altered from 10-70%. We compared the DPMM model with four alternative data-driven algorithms: (i) an intercept-only model (zero model), (ii) a multiple linear regression based on stepwise variable selection (stepwise), (iii) a quantile random forest (QRM) and (iv) a Bayesian updating approach, where individual clusters were predetermined manually based on hydrologic characteristics instead of being inferred by the DPMM. The results show that especially for River 1 and 3, only the Bayesian models could predict over 90% of observed contaminations. Through its ability to identify unknown hydraulic situations and its combination with a prior predictive distribution, the DPMM algorithm can predict high-risk periods without the need to be trained on a dataset that includes this specific contamination information. This is achieved as it identified new hydrologic information as anomalies related to the training set. Thereby, the approach is especially suitable as a precautionary approach for recreational waters, where information-rich datasets are often missing.

How to cite: Seis, W., Rouault, P., Steffelbauer, D., Ten Veldhius, M.-C., and Medema, G.: Non-parametric Bayesian modeling for risk-based management of Bathing Water Quality, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-8832, https://doi.org/10.5194/egusphere-egu23-8832, 2023.

Phosphorus (P) inputs from anthropogenic activities are subject to riverine (hydrologic) P export, causing water quality problems in lakes and coastal systems. Nutrient budgets have been used as a quantitative means of assessing the amount of nutrients imported to and exported from a system. However, at large spatial scales, estimates of hydrologic P losses are usually not available or assumed as a fixed fraction of the budget terms. In addition, fluxes in nutrient budgets are generally not quantified at regular intervals. In this study, we estimate P losses across 150 US watersheds at an approximately 4-digit Hydrologic Unit Code (HUC 4) watershed scale from 1997-2017. To explain the spatio-temporal variability in these estimates, we develop a Bayesian model based on various anthropogenic P inputs (e.g., fertilizer, animal manure, point sources, and atmospheric deposition) and outputs (crop removal) from national inventories, climatic factors, background soil P content, and watershed characteristics. In addition, a hierarchical approach accounts for additional sources of variability across different regions. Model results help us identify hot spots of P loss, along with the primary factors contributing to these losses. Results indicate that the greatest P losses (per unit area) occur in the Mid-Atlantic and Great Lakes regions, mainly due to high anthropogenic inputs. Additionally, the Upper Colorado region is found to have the highest temporal variability in P loss, whereas the Lower Mississippi region has the lowest.

How to cite: Karimi, K. and Obenour, D.: Bayesian hierarchical modeling characterizes spatio-temporal variability in phosphorus export across the contiguous United States, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-8609, https://doi.org/10.5194/egusphere-egu23-8609, 2023.

The Red-Assiniboine River Basin (RARB) spans the Canada-USA border and discharges into Lake Winnipeg via the Red River. Recurrent harmful algal blooms caused by nutrient runoff coined Lake Winnipeg as “Canada’s sickest lake” and “most threatened in the World”. Invasive species such zebra mussels and spiny-water fleas have disrupted the aquatic ecosystem.

SPAtially Referenced Regression On Watershed Attributes (SPARROW) is a watershed model that follows a stream network and relates water quality conditions to nutrient sources, landscape delivery factors, nutrient transport, and losses in streams and reservoirs/lakes. We ported components of the first binational SPARROW model (deterministic calibrations) for RARB to a Bayesian framework to account for the main sources of uncertainty: errors resulting from loading estimates, parametric uncertainty (incorporated with prior distributions based on literature values), and model structural error. Open-source languages were used: Python, R, and WinBUGS. The limits of the computational capacity of WinBUGS were tested, with a total of over 70,000 catchments across the RARB.

We identified hot spots in Canada and USA at a sub-watershed scale, where wastewater treatment plants and agricultural inputs were the main contributors of Total Phosphorus (TP) to Lake Winnipeg. Within those hot spots, we looked at hot spots at the catchment scale, where wastewater treatment plants were contributing to nearly 100% of the TP loading.

We will present various scenarios testing hypotheses on different TP sources, such as the inclusion of urban areas, the subdivision of selected source variables, and the variation of wastewater treatment plant loadings.

How to cite: Ouellet, F., Richards, A., Neumann, A., Benoy, G., and Arhonditsis, G.: Applying a Bayesian Framework to Track Binational (Canada-USA) Loads and Sources, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10477, https://doi.org/10.5194/egusphere-egu23-10477, 2023.

Phosphorus (P) pollution from agriculture is a major pressure on maintaining and improving water quality worldwide. In Ireland, the Agricultural Catchments Programme (ACP) was created to evaluate the Good Agricultural Practice measures implemented under the EU Nitrates Directive. Considerable monitoring and research have been done into the drivers of, and controls on, nutrient loss in these catchments.

Managing P pollution in agricultural catchments requires informed decisions about the pollution risks using catchment-scale understanding which, in turn, requires a systemic modelling and assessment approach. Bayesian Belief Networks (BBNs) support system-level thinking as they can represent complex systems (such as rivers and catchments) and integrate disparate information sources while representing uncertainty. In a previous study, a BBN was developed using the ‘source-mobilisation-transport-continuum’ and parameterized for a 12 km² agricultural catchment with flashy hydrology on poorly drained soils and grassland as the dominant land use. Seven years of hourly turbidity and discharge measurements at catchment outlet, and mapped soil P content were used to inform parameterization. Literature data and expert opinion were included to complement the dataset when information on point-source pollution (farmyard and septic tank nutrient losses) was lacking.

In the current study, the BBN is developed further and is parameterized using a monthly time-step for three additional diverse ACP catchments: two arable land-dominated catchments with contrasting hydrology (well-drained vs poorly drained) and a well-drained grassland catchment to test model transferability. In a step forward from the previous model, we quantify P losses from a sewage treatment plant in one of the arable catchments, and we consider biota in-stream P removal as an additional process. Lastly, the observed TRP concentrations were bootstrapped to obtain monthly TRP distributions which are compared to predictions from the BBN to validate the model.

Results showed that the model performs well for the target catchment but applying it to other catchments is key to assessing its generalizability and utility. Here, preliminary results explore whether the BBN can capture the differences in P loss risk between catchments and the reasons for this.

In addition, testing the model transferability to other catchments is important to (a) inform on the differences in P loss between catchments, and (b) inform model testing in data-sparse catchments. Future research will be focussed on integrating climate change scenarios in the model to inform the targeting of mitigation measures under future change, foster discussion with stakeholders, and provide support to decision-makers.

How to cite: Negri, C., Schurch, N., Wade, A. J., Mellander, P.-E., and Glendell, M.: Testing the transferability of a Bayesian Belief Network to diverse agricultural catchments using high-resolution hydrology and land management data sets, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2747, https://doi.org/10.5194/egusphere-egu23-2747, 2023.