HS8.2.11

Estimating the effects of climate change on water resources is an important topic for researchers in hydrological sciences. Climate model outputs can be used as inputs for calibrated numerical groundwater models to predict the effect of different climate scenarios on groundwater levels. Unfortunately, spatially explicit numerical models are spatially constrained, data-hungry, and difficult to set-up and to calibrate. Furthermore, the involved model run-times make proper uncertainty analysis of the model predictions computationally expensive.
Machine-learning (ML) tools have become recognized as a powerful alternative for numerical models for different applications in hydrological science, but come with their own challenges. They have been successfully used for the prediction of groundwater levels in single bores based on historical data, but their application for estimating groundwater levels for climate scenarios is still a matter of active research.
To identify the potential of ML techniques for this application, two different ML algorithms have been applied to predict groundwater levels for several climate scenarios at a number of groundwater wells with long-term historical data time series. The two ML algorithms are (i) multi-layer perceptrons (MLP) with a closed feedback loop and (ii) long short-term memory (LSTM) networks. For each observation well, several thousand versions of these ML models with differing setups are trained to historical, monthly data time series up to 2015 and then applied for the estimation of monthly groundwater levels. These estimations are computed by using climate model outputs for different scenarios, as well as artificially generated scenarios, as drivers for the ML models to test their sensitivity and plausibility to those input data series. The high quantity of model versions for each bore are utilized to generate mean groundwater level estimates and accompanying uncertainty bands via Bayesian Model Averaging (BMA).
Both ML techniques are able to match the historical data time series for the different bores with small uncertainty, but differ in their ability to predict long-term groundwater levels from climate change and artificial scenarios. The long-term simulations of the MLP models show believable trends and appropriate uncertainty bands, while the LSTM networks seem to underestimate the uncertainty of their future predictions.

How to cite: Gosses, M. and Wöhling, T.: Long-term prediction of groundwater levels for climate scenarios with machine-learning tools, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4809, https://doi.org/10.5194/egusphere-egu22-4809, 2022.

The summers of 2017 to 2020 were characterized by exceptional dry spells throughout Europe. Climate models show that such periods of drought could occur more frequently and become even more extreme in the future. The recent periods of intense droughts lead to significant ecological, economic and even societal damages in Flanders (Belgium). During these summers, receding groundwater levels were observed throughout Flanders that reached historical low levels. To monitor low ground water levels and to support a proactive drought management, the Flemish government developed an operational ground water indicator. This indicator gives an overview of the current phreatic ground water levels combined with a prediction for the next month for a selected number of phreatic wells. To increase the spatial resolution of the indicator, we developed a novel data driven regional ground water model for phreatic aquifers.

The ML model combines a gradient boosting decision tree model (CatBoost) with a Long Short Term Memory (LSTM) network. CatBoost is used to model the average ground water depth at each location. This value is passed to the LSTM network that predicts the temporal evolution of the groundwater at each location around its average. The training dataset for the CatBoost model contains the average groundwater depth of 5.673 wells spread across Flanders and a large set of explanatory variables related to soil texture, distance to a drainage, geology, topography, meteorology and land use. The model performance is evaluated using cross-validation which showed the model generalizes well with a mean absolute error of 69cm. The most important explanatory variables for the model are the thickness of the phreatic aquifer, the vertical distance to closest drain, the topographic index and the precipitation surplus.

The training dataset for the LSTM model contains 408 wells that have sufficiently long and reliable observations for training. The input data to the LSTM consists of rainfall and evapotranspiration up to 10 years prior to each observation, combined with the same explanatory variables as the CatBoost model. A single regional LSTM model is trained on all 408 wells simultaneously. The resulting model is accurate with a median RMSE of 20cm for the validation data, outperforming the currently used SWAP models [1]. The ML model is however less performant in simulating the higher ground water depths during summer and shows a consistent bias towards lower ground water depths during long dry spells.

[1] Kroes, J.G., J.C. van Dam, R.P. Bartholomeus, P. Groenendijk, M. Heinen, R.F.A. Hendriks, H.M. Mulder, I. Supit, P.E.V. van Walsum, 2017. SWAP version 4; Theory description and user manual. Wageningen, Wageningen Environmental Research, Report 2780

How to cite: Wolfs, V., Franken, T., Gullentops, C., Lermytte, J., and Corluy, J.: A regional data driven model for simulating phreatic ground water levels in Flanders, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6263, https://doi.org/10.5194/egusphere-egu22-6263, 2022.

Groundwater is the most important freshwater source considering its abundance, demand, quality, and accessibility. However, estimation of groundwater storage variations is notoriously difficult due to the lack of information on physical aquifer properties, the connectivity among other water storages, the scarcity of observations at regional scales and high uncertainty associated with the observations if available. In the current study, we tried to improve the simulations of groundwater storage variations of a global hydrological model with a simple bucket-type groundwater module – the WaterGAP Global Hydrological Model (WGHM) – in four French river basins. Sensitive model parameters were calibrated against in-situ streamflow observations and GRACE-based observations of total water storage anomalies in a multi-criterial calibration framework. A regional data set of in-situ observations of groundwater levels and deduced basin-average groundwater storage variations (see Hsu et al., this meeting) was used for validation of the calibration results. In a second setup, the in-situ groundwater observations were introduced as additional observations into the calibration framework.  We found significant improvement in the simulated groundwater dynamics in terms of their seasonal signals and amplitudes after calibration. Overestimated negative groundwater trends in a few river basins caused by overestimated human groundwater use in the original model could also be corrected by the calibration approach. In an additional calibration experiment, specific yield was introduced into WGHM as a new calibration parameter and its calibrated values were compared to those obtained from regionalizing the hydrogeological information.

How to cite: Hasan, H. M. M., Hsu, K.-H., Eicker, A., Longuevergne, L., and Güntner, A.: Improved groundwater simulations by multi-criteria calibration of a global hydrological model in the river basins of France, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7384, https://doi.org/10.5194/egusphere-egu22-7384, 2022.

Alluvial aquifers and springs play a vital role in the water supply of Austria as the main source of drinking water. One such example is the Grazer Feld Aquifer, located in an urban and semi-urban setting in southeast Austria. Urbanisation imposes stresses on aquifers leading to quantitative and qualitative groundwater problems. These problems are commonly addressed by numerical groundwater models. In these models, natural and anthropogenic processes need to be carefully represented. Calibration to groundwater level data disturbed by human activities will fail or produce erroneous parameter estimates if the disturbance is not adequately considered by the model. As a consequence, such model likely result in erroneous predictions and assessments. To account for relevant drivers in the system and examine groundwater level data for the calibration of a numerical groundwater model, we test the application of the time series analysis as an additional and preliminary step in a general numerical groundwater modeling framework. The results of time series models (TSM) contribute to the understanding of spatiotemporal aquifer dynamics and main driving forces as advocated by Bakker and Schaars (2019).

The objective of this study is twofold. First, time series models were set up and calibrated for each monitoring well in the aquifer, using the stresses identified in the initial hydrogeological assessment (precipitation, evapotranspiration, and river levels). Second, we create a calibration data set that flags groundwater level observations that were caused by human temporal activities (e.g., pumping, irrigation, and/or dam construction). This is achieved by constructing TSMs and conducting a visual and spatial investigation on model results. The process differentiates good fit models from no-good fit models. Then, models, not delivering a good fit, are checked for missing driving forces by engaging local stakeholders. Once the process is characterized, the period with unexplained groundwater level change is marked. The groundwater level fluctuations of 115 out of 149 observation wells are found to be reasonably simulated by considering recharge from precipitation and, if applicable, river stages as driving forces. For 34 observation wells, however, the models perform less accurately, suggesting other factors, such as construction activities and temporary groundwater abstraction, influencing groundwater level fluctuations during the part or the entire simulation period. Estimated recharge is found to be higher in urban and semi-urban areas compared to agricultural fields. The results from this study will be used in the future development of a numerical groundwater model for the entire aquifer.

How to cite: Kokimova, A., Collenteur, R., and Birk, S.: Data-driven time series modeling to support groundwater model development for the Grazer Feld Aquifer, Austria, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7734, https://doi.org/10.5194/egusphere-egu22-7734, 2022.

Recent years have seen several major groundwater drought instances in Europe, increasing the interest in their research and future predictions. Modeling methods can be used to explore groundwater drought risk in near and medium term future using climate projection data sets as an input. Therefore an optimal modeling approach has to be chosen according to the available data and modeling capacity of the historic series. 

We  explore two approaches for modeling groundwater level time series: Transfer function-noise models with Impulse response functions (TFN-IRF) and machine learning (ML). In both approaches, daily meteorological variables are used as an input and models are calibrated against historical groundwater level observations. 

TFN-IRF input parameters are daily precipitation and potential evapotranspiration. Other time series data, such as groundwater abstraction, can be added to potentially increase the fit. Machine learning models can, in addition to the aforementioned, benefit from a wider variety of site parameters, including derived parameters (e.g., meteorological indices), however at the expense of increased data collection effort. In addition, the obscure input data interpretation in ML methods can erode the trust in these models. In this study, only the most basic meteorological parameters - precipitation, temperature - and derived parameters such as potential evapotranspiration were used. 

We draw from an extensive groundwater level monitoring database of more than one thousand monitoring wells from three Baltic countries. The study provides an insight into differences between the two modelling approaches, keeping in mind the limitations of future projection data.

This research is funded by the Latvian Council of Science, project “Spatial and temporal prediction of groundwater drought with mixed models for multilayer sedimentary basin under climate change”, project No. lzp-2019/1-0165.

How to cite: Jemeļjanova, M., Bikše, J., and Kalvāns, A.: Comparison of Impulse response function and Machine learning models for use in groundwater level short to medium term future projections in the Baltic states, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7999, https://doi.org/10.5194/egusphere-egu22-7999, 2022.

Time series analysis with response functions is a versatile approach to analyze measured head series in observation wells. In such an analysis, the response function of the groundwater does not change with time. This approach works well when the groundwater recharge is a linear function of the measured rainfall and evaporation, as was shown through the analysis of synthetic time series generated with a saturated groundwater model where the groundwater recharge is applied directly to the saturated zone. In this research, the method was evaluated for situations where the groundwater recharge cannot be approximated well as a linear function of the measured rainfall and evaporation. Synthetic time series were generated with a two-dimensional saturated/unsaturated zone model (Hydrus2D) and analyzed with response functions. Performance of the time series model was improved through inclusion of a new root zone model consisting of a single reservoir. Reservoir inflow is measured rainfall. Reservoir outflow is evaporation and groundwater recharge, where the evaporation is a function of the amount of water stored in the root zone and the recharge is a function of both the amount of water stored in the root zone and the groundwater table. The new root zone model is a promising tool for the analysis of head series in areas with thick unsaturated zones and/or high potential evaporative fluxes.

How to cite: Bakker, M., Vonk, M., Collenteur, R., and Schaars, F.: Time series analysis of synthetic time series generated with a saturated/unsaturated zone model, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9995, https://doi.org/10.5194/egusphere-egu22-9995, 2022.

Distinguishing between natural and anthropogenic impacts on groundwater systems at regional scale is not straightforward using current data-driven and traditional numerical groundwater models. This limits their benefit for groundwater level predictions and thus future-oriented groundwater resource management. We propose an approach to leverage the large amount of information and variability in the characteristics of groundwater hydrographs and environmental factors to obtain generalized insights into the influences of natural and anthropogenic factors on specific patterns in groundwater hydrographs using data-driven regionalization of groundwater level dynamics. In our approach, we focus on coastal regions that are often under water stress due to the water demands of growing coastal populations building on a data set containing several thousand wells in Europe, North America, South Africa, and Australia.

The approach comprises construction and comparison of multiple unsupervised machine learning cluster models based on a) groundwater level dynamics information, aggregated into groundwater hydrograph features, and b) selected environmental drivers that potentially influence natural groundwater recharge and discharge processes. Environmental descriptors were extracted at well locations from available global map products. We discuss the extent to which our selection of features can express the range of dynamics in representative groundwater hydrographs of clustered basins. Furthermore, we compare the similarity of anthropogenic factors within and between clusters in order to test our hypothesis that hydrograph patterns differ in response to natural processes but irrespective of anthropogenic influences. This would contribute to our understanding of natural processes in coastal groundwater systems.

How to cite: Nolte, A., Haaf, E., Heudorfer, B., Bender, S., and Hartmann, J.: Analysis of hydrogeological behavior of coastal aquifers based on clusters of groundwater hydrograph features and environmental drivers, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7893, https://doi.org/10.5194/egusphere-egu22-7893, 2022.

Regular and gapless observations are necessary to perform a range of statistical analysis on the parameter of interest. Groundwater level (GWL) hydrographs  are often recorded at irregular frequencies and have time periods without any observations. As a result, groundwater level hydrographs have missing values. Typically groundwater hydrographs are removed from further analysis if large gaps are present, while each groundwater observation point is valuable and methods exist that can impute (fill in) the missing observations. 

This study aims to evaluate performance of machine learning methods to prepare gapless daily groundwater level hydrographs and to assess the imputation error according to various approaches. Filled groundwater level hydrographs will further be used  to identify typical groundwater level patterns in the Baltic region.

The performance of two machine learning imputation methods - missForest and missMDA - along with conventional approaches (linear interpolation, mean imputation) - were tested. A subset of the GWL observation data from Lithuania, Latvia and Estonia were used for the time period from 2011 to mid-2019 comprising 283 groundwater monitoring wells. Cluster analysis of the temporal distribution of actual missing values in the GWL time series provided 13 different gap patterns. Next a corresponding number of artificially generated gap distribution scenarios were defined. The performance of various gap-filing approaches were then evaluated by imputing each artificially generated gap pattern in each hydrograph. Results indicated that imputation performance varies among different clusters of missing value patterns, while generally the best performance was achieved by the missForest algorithm.

This research is funded by the Latvian Council of Science, project “Spatial and temporal prediction of groundwater drought with mixed models for multilayer sedimentary basin under climate change”, project No. lzp-2019/1-0165.

How to cite: Bikše, J., Kalvāns, A., Retike, I., and Jemeļjanova, M.: Performance analysis of missing data imputation methods for daily groundwater hydrographs using typical gap patterns, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7920, https://doi.org/10.5194/egusphere-egu22-7920, 2022.