From the Institute of Natural Science at the Department of Landscape, Water and Biogeochemical Cycles at the Justus-Liebig University Gießen A new method for spatio-temporally explicit predictions of groundwater, surface water and habitat interactions in riparian ecosystems Dissertation for the degree Doctor of Natural Science (Dr. rer. nat.) submitted by Nadine Maier, M. Sc. Gießen, 2018 A dissertation submitted by M.Sc. Nadine Maier to the Institute of Natural Sciences at the Department of Landscape, Water and Biogeochemical Cycles of the Justus Liebig University Giessen for the degree of Doctor of Natural Sciences (Dr. rer. nat.) Referees from the Justus-Liebig University Gießen: Prof. Dr. Lutz Breuer (1st Supervisor) Prof. Dr. Dr. habil. Dr. h. c. Annette Otte (2nd Supervisor) Prof. Dr. Jürg Luterbacher Prof. Dr. Jan Siemens Submitted: 17th July 2018 Abstract I Abstract Floodplains are dynamic and complex systems. Their hydrological regime allows for the creation of specific conditions for flood meadows with a wide habitat heterogeneity and specialized floodplain plant species. At the same time, floodplains are the most-endangered ecosystems worldwide, as they are characterized by extremely high land loss. Therefore, they are one of the main focuses of national and international nature conservation and restoration efforts. The aim of such restoration measures is not only to restore the retention function, but also the habitat function for typical, highly specialized, and endangered species. However, compared to the importance of ecosystem services provided by floodplains and their role in biodiversity, there is little research on how scientifically-based information can be integrated in the planning and decision-making of restoration projects in order to make them more efficient and effective. Plant distribution is strongly related to hydrologic conditions on a high temporal and spatial resolution. For example, flood sensitive species prefer elevated microsites and will die during large inundation periods. To elucidate the relevance of a detailed knowledge of the hydrological regime, a parsimonious surface water-groundwater model was developed using the Catchment Modeling Framework (CMF). Further, this process-based hydrological model was linked with a species distribution model in order to predict rare and endangered species. The nature reserve Kühkopf- Knoblochsaue (34.5 km2) serves as the study area. This reserve is of particular importance for rare and endangered flora and fauna. As a first step, the model was developed, adjusted, and tested for the past 16 years. Next, the surface water-groundwater interaction model was linked with a species distribution model to predict the occurrence of rare and endangered species. A comparison study showed that the incorporation of temporally and spatially high- resolution data from a hydrological model, as developed here, results in superior model qualities. In addition, the relevance of a broad set of hydrological predictors to simulate whole plant communities with diverse specific eco-hydrological requirements was shown. Applications of different climate models to the surface water-groundwater interaction model showed large spatial and quantitative changes in the inundation characteristics in the near and far future. A linkage of these hydrological regime projections with the species distribution model pointed out the possibilities for habitat projections and the advantages for conservation measures. Table of contents II Table of contents Abstract .................................................................................................................................................. I Table of contents ................................................................................................................................ II List of tables ........................................................................................................................................ V List of figures ..................................................................................................................................... VI 1. Extended Summary ........................................................................................................................ 1 1.1. Introduction .............................................................................................................................. 1 1.1.1. Environmental relevance of riparian ecosystems ............................................................ 1 1.1.2. Modeling of riparian ecosystems ..................................................................................... 2 1.2. General objectives of the study .............................................................................................. 3 1.3. Material and methods ............................................................................................................. 5 1.3.1. Overview over the study area .......................................................................................... 5 1.3.2. Surface water-groundwater interaction model ............................................................... 9 1.4. Results and discussion .......................................................................................................... 13 1.4.1. Correlation of Rhine water level with groundwater level ............................................. 13 1.4.2. Model calibration and validation .................................................................................. 15 1.4.3. Projecting floodplain inundation under climate change ............................................... 18 1.4.4. Species distribution modeling in floodplain habitats .................................................... 21 1.5. Conclusion .............................................................................................................................. 25 1.6. Outlook ................................................................................................................................... 26 1.6.1. Species distribution under climate change .................................................................... 26 1.6.2. Species distribution under land use changes ................................................................ 30 2. Prediction and uncertainty analysis of a parsimonious floodplain surface water – ground-water interaction model ................................................................................................ 31 Abstract ........................................................................................................................................... 31 2.1. Introduction ............................................................................................................................ 32 2.2. Methodology .......................................................................................................................... 36 2.2.1. The Catchment Model Framework (CMF) .................................................................... 36 2.2.2. Using CMF for a spatial explicit groundwater model .................................................. 37 2.2.3. Surface water flow equation .......................................................................................... 38 2.2.4. Uncertainty estimation method .................................................................................... 39 Table of contents III 2.2.5. Case study ..................................................................................................................... 40 2.2.6. Model setup ................................................................................................................... 43 2.2.7. Calibration and validation of the model ........................................................................ 44 2.3. Results ..................................................................................................................................... 45 2.3.1. Testing the simplified surface flow approximation ....................................................... 45 2.3.2. Model performance ........................................................................................................ 47 2.4. Discussion ............................................................................................................................... 55 2.5. Conclusion .............................................................................................................................. 57 Acknowledgments ........................................................................................................................ 58 Supporting information ................................................................................................................ 59 3. Multi-Source Uncertainty Analysis in Simulating Floodplain Inundation under Climate Change ............................................................................................................................ 60 Abstract ........................................................................................................................................... 61 3.1. Introduction ............................................................................................................................ 62 3.2. Materials and Methods ......................................................................................................... 64 3.2.1. Study Site ...................................................................................................................... 64 3.2.2. Surface Water–Groundwater Model ............................................................................. 65 3.2.3. Water Level Forcing ...................................................................................................... 66 3.2.4. Climate Forcing ............................................................................................................. 67 3.3. Results ..................................................................................................................................... 69 3.3.1. Projection of the Rhine River Water Level .................................................................... 69 3.3.2. Projection of Spatial Inundation ................................................................................... 72 3.3.3. Uncertainty of Projections ............................................................................................ 75 3.4. Discussion ............................................................................................................................... 78 3.5. Conclusions ............................................................................................................................ 80 Acknowledgement ........................................................................................................................ 81 Supporting information ................................................................................................................ 82 4. Modeling of rare flood meadow species distribution by a combined habitat-surface water-groundwater model .......................................................................................................... 87 Abstract ........................................................................................................................................... 87 4.1. Introduction ............................................................................................................................ 88 4.2. Material and Methods ........................................................................................................... 90 Table of contents IV 4.2.1. Study area and database ................................................................................................ 90 4.2.2. Integrated model setup .................................................................................................. 93 4.2.3. Model evaluation ........................................................................................................... 97 4.3. Results ..................................................................................................................................... 99 4.3.1. Best predictor identification .......................................................................................... 99 4.3.2. Evaluation of habitat model ......................................................................................... 101 4.3.3. Significance of individual predictor variables ............................................................. 102 4.4. Discussion ............................................................................................................................. 103 Acknowledgements .................................................................................................................... 106 Supporting information .............................................................................................................. 107 References ......................................................................................................................................... 112 Acknowledgements ........................................................................................................................ 131 Declaration ....................................................................................................................................... 132 List of tables V List of tables Table 2.1: Differentiation of calibrated period by hydrological characteristics ....................... 44 Table 2.2: Calibrated parameters for model and parameter ranges .......................................... 45 Table 2.3: Summary of the objective functions for the calibration and validation period. Depicted are the mean, minimum and maximum root mean square error (RMSE) and the mean absolute error (MAE) for each groundwater well in the study area. ........ 48 Table 3.1: Mean uncertainty (i.e., range of the daily water levels, averaged over the time period) of the Rhine water level of the parameter and predictive uncertainty. ............... 71 Table 4.1: Definition of the four predictor calculation databases used for the evaluation of the habitat model. ................................................................................................................. 97 Table 4.2: Selected predictors used as input data for the multi-predictor species distribution models. ................................................................................................................ 100 List of figures VI List of figures Figure 1.1: Maps of the study area: (a) location of the study area within Germany, (b) major cities around the study area, (c) digital elevation of the study area with the location of the nature reserve, and (d) elevation variation of the study area with the location of the groundwater wells.. .......................................................................................... 6 Figure 1.2: Mean ± SD monthly rainfall and mean ± SD monthly minimum/maximum temperature recorded at Darmstadt weather station between 2000 and 2016. .................. 8 Figure 1.3: Three-dimensional sketch of the study area Kühkopf-Knoblochsaue with typical structures and the relevant water flux processes simulated by CMF. .................. 11 Figure 1.4: Time series of the hourly measured groundwater levels for five different groundwater wells between July 2015 and December 2017 and the time series of the Rhine for the Nierstein-Oppenheim gauging station. ......................................................... 14 Figure 1.5: Time series of the simulated water levels by the CMF model, compared to the measured data by the HLNUG for each of the groundwater wells in the Knoblochsaue (top) and the Kühkopf (bottom). ................................................................... 16 Figure 1.6: Simulation of the CMF model, compared to the measured water level by the capacity loggers and the measured data by the HLNUG, exemplarily shown for groundwater wells in the Knoblochsaue (top, 1.7.2015 – 27.7.2016) and the Kühkopf (bottom, 1.7.2015 – 30.5.2016).. ................................................................................................ 17 Figure 1.7: Annual inundation days for the near future (2021–2050), divided between the GCMs (top), the GCM-RCP combinations (middle, light grey), and the combination of GCM, RCP, and uncertainty of the HBV model (bottom, darker grey). ....................... 20 Figure 1.8: Representations of the groundwater and surface water level in the study site at two different days (07/01/2003 and 16/01/2004). .............................................................. 24 Figure 1.9: Percentage of restorable area with a difference of less than 10% (marker: downwards triangle) or less than 20% (marker: upwards triangle) between the projections under the different RCMs for eight species.. .................................................... 28 Figure 1.10: Occurrence probability (in percentage) for the species Arabis nemorensis (top), Peucedanum officinale (middle), and Serratula tinctoria (bottom) for the years 2050 and 2100 for both RCP 4.5 and RCP 8.5. ....................................................................................... 29 Figure 2.1: Sequence of CMF model time steps A-D, showing the flooding process by the flood-wave scheme developed for this model setup. .......................................................... 39 Figure 2.2: Geographic location (left), digital elevation of the larger nature reserve (middle) and setup of the CMF model with its irregular grid and land use for the study site (right). ....................................................................................................................... 41 Figure 2.3: Daily precipitation (black bars) and water level of the Rhine (gray) at the gauging station Nierstein-Oppenheim for 2002 to 2013. ..................................................... 42 List of figures VII Figure 2.4: Comparison of surface water potential for a test model using either the diffusive surface water flow equation or the flood wave surface water flow equation. ..................................................................................................................................... 46 Figure 2.5: Scatter plot of the mean RMSE of each simulation against the parameters (A-D). Behavioral runs are marked in red. Parameter interactions for all model runs (E-J) are colored from blue to red for a mean RMSE ≤ 0.26 m. ........................................... 49 Figure 2.6: Comparison of simulations with observed groundwater levels. Performance measures depicted are the best obtained. .............................................................................. 50 Figure 2.7: Comparison of simulations with observed groundwater wells for behavioral runs for total time series (calibration and validation period). ............................................ 51 Figure 2.8: Results of the exchange between surface water and groundwater, shown for four polygons of the model for the time period between 01.07.2002 and 01.07.2003. ..... 52 Figure 2.9: Time series of the inundated area (in km2) for the complete calibration period and boxplots of the inundated area at different water levels of the Rhine. ...................... 53 Figure 2.10: Comparison of average flooding time (a), maximum flooding height (b) and longest flooding period (c) for the polygons of the study area. ......................................... 54 Figure 3.1: Geographic location of the study area (lower left), digital elevation of the larger nature reserve with the study area (middle), and the setup for the surface water–groundwater model in the catchment modeling framework (CMF) (right).. ....... 64 Figure 3.2: Flowchart showing the steps within the model framework.. ................................... 68 Figure 3.3: Monthly boxplots of HBV-simulated water levels of the Rhine using the best- performing model parameter sets for representative concentration pathway (RCP) 2.6 (bluish, top), RCP 4.5 (greenish, middle), RCP 8.5 (reddish, bottom) and for two different periods (top: mid-century 2021–2050, bottom: end-century 2071–2100). ......... 70 Figure 3.4: Spatial distribution of average inundation days (per year) for the mid-century (2021–2050) and end-century (2071–2100) compared to the past (2002–2015). ................ 73 Figure 3.5: Spatial distribution of average inundation duration (days per year) for the mid-century (2021–2050) and end-century (2071–2100) compared to the past (2002–2015). ............................................................................................................................... 74 Figure 3.6: Uncertainty of the annual inundation days per time period separated by the three sources of uncertainty: GCMs, RCPs, and the parameter and predictive uncertainty of the HBV model. .............................................................................................. 77 Figure 4.1: Geographic location of the study area in Germany (lower left corner), digital elevation of the study area with the location of vegetation observations (middle) and setup of the surface water-groundwater model (catchment modeling framework, CMF) with its irregular grid and land use, containing the locations of the groundwater wells (right). ....................................................................................................... 91 Figure 4.2: Representation of the main steps of the integrated model setup. ............................ 93 List of figures VIII Figure 4.3: Simulated area under the receiver operating characteristic curve (AUC) for flood meadow species without using hydrological predictors (nhy), using hydrological predictors derived from the surface-groundwater-model (sgm), measured groundwater data (gww), and simulated water level of the Rhine River (riv). . ........................................................................................................................................ 101 Figure 4.4: Relative predictor frequency for all model runs separated for the (A) flood meadow species according to Burkart (2001) and (B) species on the Red List (vulnerable and endangered). ............................................................................................... 103 Extended Summary 1 1. Extended Summary 1.1. Introduction 1.1.1. Environmental relevance of riparian ecosystems Floodplains are among the most-endangered ecosystems worldwide (Opperman et al., 2010). In Germany, the floodplains of the largest rivers have lost on average two-thirds of their original extent (Brunotte et al., 2009). A consequence of the continuous decline in floodplain area is the loss of biodiversity. Even the remaining active floodplain areas, retaining more or less their typical flooding dynamics, are developed areas or are used for agriculture and have consequently lost at least part of their habitat function (Follner et al., 2010). Restoration of species-rich flood meadows is an important contribution to maintain the biodiversity of plants and fauna on flood meadows. Preconditions for successful restoration management are suitable site conditions, such as soil nutrients, the availability of seed sources, and moisture regime (Bakker and Berendse, 1999). As floodplains are frequently inundated and nutrient-rich sediments are deposited, river floodplain soils are among the most fertile soils (Zorn et al., 2005). In order to overcome the issue of limited seed material dispersal in floodplains, the best solution up to date is to transfer seed-containing plant material with or without topsoil removal (Donath et al., 2007; Harnisch et al., 2014; Hölzel and Otte, 2003). While the availability of seed source and soil nutrients can be easily influenced and assessed, evaluation of the hydrological conditions is more difficult, though not less important. Consequently, restoration projects are often challenged by the complex hydrological conditions of the target areas (Malanson, 1993). The hydrological conditions are of particular significance for plants, as they are rather sensitive to the soil moisture conditions, e.g., flood sensitive species occupy elevated microsites, whereas flood tolerant species occur in depressions (Jung et al., 2008; Ludewig et al., 2014; Vervuren et al., 2003). In light of this, it is especially important to incorporate hydrological conditions in the planning of flood meadow restoration projects (Gattringer et al., 2017, 2018). Ideal tools for enhancing conservation decisions are species distribution models (Guisan et al., 2013). So far, only a few studies have used hydrological information to simulate the distribution of riparian vegetation or the occurrence of plant species (Leyer, 2005; Mosner et al., 2011, 2015). There are two main reasons for neglecting hydrological conditions within species distribution modeling on floodplains. On one hand, the lack of available data with a sufficient temporal Extended Summary 2 and spatial resolution and on the other hand, the difficulties in obtaining and portraying conditions in sufficient resolution. Floodplains are comprised by different components, including surface water, groundwater, and precipitation, interacting on a high temporal and spatial resolution. The driving factor of the eco-hydrological functions on the floodplain are the connectivity and interaction of shallow groundwater with the surface water (Hayashi and Rosenberry, 2002; Krause et al., 2007a). These interactions are mainly defined by the physical soil properties and the presence and location of embankments, which influence the hydrological regime and the in- and outflow of river water on the floodplain. The regular inundations of a floodplain are defined by characteristics such as the height of the water level, flood duration, and recurrence intervals (van Eck et al., 2004; Hayashi and Rosenberry, 2002; Krause et al., 2007a; Maltby and Barker, 2009; Woodcock et al., 2005). 1.1.2. Modeling of riparian ecosystems Due to the importance of the connectivity between surface water and groundwater for the eco-hydrological functions, floodplains should be simulated with fully-integrated models that have explicit representation of water table gradients and groundwater flow (Acremann and Miller, 2007) and consider surface water-groundwater interactions (Furman, 2008; Refsgaard et al., 1998; Sophocleous, 2002). As floodplains act as collecting point for river water, hillslope-derived water, and groundwater, the groundwater level is usually very high in floodplains (Burt et al., 2002). The groundwater level is influenced by the surface water in two ways: laterally across the channel banks and vertically across the floodplain surface. A high temporal resolution of such models is equally essential, as the interaction often shows a high temporal variability (Jung et al., 2004; Krause et al., 2007a). Compared to the significance of floodplain ecosystems, only a few studies have focused on modeling floodplains with models maintaining the simultaneous simulation of both components - the surface water and the groundwater. There are three main reasons for the lack of application of such models: (1) strong interactions between surface water and groundwater are difficult to model (Branfireun and Roulet, 1998; Price and Waddington, 2000); (2) these models require a high-level modeling expertise (Ameli and Creed, 2017; Golden et al., 2014); and (3) these models require high computational power (Ameli and Creed, 2017). An additional difficulty is the parameterization of such large groundwater models, as the effort for this process is immense (Acremann and Miller, 2007). Extended Summary 3 However, if these problems could be successfully overcome, then those powerful hydrological models can be combined with habitat models. Applications of such complex habitat models have the potential to improve the outcome of conservation efforts, as they are an efficient tool for predicting possible habitats for species (Guisan et al., 2013). In Germany, biodiversity conservation and restoration focuses on the active floodplain. A lot of effort is given to the protection and reintroduction of rare and endangered floodplain species, but the decisions as to whether a specific site is suitable and promising for restoration is currently often based on soft data and subjective appreciation. This is commonly accomplished by searching for potential restoration sites and convincing landowners and farmers to restore their intensively-used meadows. Including detailed hydrological information and habitat models in such decision processes bear a great potential to make these decisions more efficient and effective. The correlation of hydrological conditions with the requirements of plant species can help locate promising regions for restoration, where conditions are satisfactory, and thus the probability of species settling is high. Consequently, likely unsuitable regions can be excluded from the beginning, and the focus can be on more promising regions. There are further advantages: by incorporating predicted hydrological conditions for the future into habitat models, even the probability of a plant occurring under climate or land use change can be assessed. 1.2. General objectives of the study This dissertation aims to expedite the scientifically-based decision-making to enhance the success rate of restoration measure on flood meadows. The study can serve to support the conservation and restoration of rare and endangered floodplain species. Within this study, a method is developed for spatio-temporally explicit predictions of groundwater, surface water, and habitat interactions in riparian ecosystems. In addition, the obtained knowledge and the modeling strategy developed here can generally be transferred to other ecological disciplines, for example, to simulate the effect of restoration measures on stream ecology or the spatio-temporal development of intermittent streams. Due to the interdependence between plant species distribution and hydrological conditions, as well as the lack of the corresponding hydrological data and the lack of such Extended Summary 4 information in the scientific species distribution modeling context, the following main objectives were designed for this dissertation: 1) Set up a simplified physical-deterministic model to simulate the surface water- groundwater interaction and the inundation of floodplains. This is accomplished by establishing a parsimonious model that provides key processes to accurately represent the water pathways with the aim to reduce the computational run time and to investigate the model’s parameter uncertainty (Chapter 2). As a test case, the model is applied to the floodplain “Kühkopf-Knoblochsaue,” Hesse, Germany. 2) Investigate the model uncertainty and its suitability to project future hydrological conditions under climate change, particularly with regard to the occurrence and duration of flooding events of floodplains. A modeling framework is set up, integrating projections of two climate models, three emission scenarios, a rainfall–runoff model, and the coupled surface water–groundwater model from the first objective (Chapter 3). 3) Implement spatio-temporal explicit simulations of groundwater and surface water indicators of the floodplain in a habitat model to improve the prediction of potential habitats with a specific focus on rare and endangered floodplain species. For this purpose, the dynamic hydrological information is converted into static variables and integrated in a habitat model, which in turn predicts the probability of occurrence for 23 flood meadow plant species (Chapter 4). To achieve the objectives, the work was divided into three sections, which resulted in three publications. After a brief introduction to the study area (Chapter 1.3.1) and the presentation of the modeling concept (Chapter 1.3.2), the publications are summarized in chapters 1.4.2 to 1.4.4. Extended Summary 5 1.3. Material and methods 1.3.1. Overview over the study area The research area (N: 49°52’ – 49°46’, E: 8°22’ – 8°29) is located in Hesse, Germany, approximately 30 km south of Frankfurt am Main (Figure 1.1 a, b). The study area covers approximately 43 km2 and includes the nature reserves “Kühkopf-Knoblochsaue” (24 km2), “Bruderlöcher” (0.17 km2), “Großes Michelried bei Erfelden” (0.23 km2), and a part of the “Riedwiesen von Wächterstadt“ reserve (0.76 km2). The remaining area is unprotected area (Figure 1.1, c). The nature reserve is of particular importance for rare and endangered flora and fauna and is protected by the European Habitats Directive (Council Directive 92/43/EEC). An embankment construction divides the Holocene floodplain into three compartments:  The functional floodplain: land between the river and the summer embankment, regularly flooded during high flow of the Rhine.  The hybrid floodplain: land between the summer and the winter embankment. This area is only flooded by extremely high water levels of the Rhine (> 4m above mean water level) or in the event of breaking embankments. Inundation can occur also due to ascending groundwater.  The fossil floodplain: the part of the floodplain on the landward side of the winter embankment. Inundation can occur due to ascending groundwater (Hölzel and Otte, 2009). Extended Summary 6 Figure 1.1: Maps of the study area: (a) location of the study area within Germany, (b) major cities around the study area (Esri, 2018), (c) digital elevation (Hessian Administration for Soil Management and Geographical Information, HVBG, Wiesbaden, Germany) of the study area with the location of the nature reserve, and (d) elevation variation of the study area with the location of the groundwater wells. The black border depicts the boundary of the hydrological (sub-)models. Extended Summary 7 Land use The study area experienced a large decrease (up to 90%) of flood meadows in the 1980s. Disastrous floods in 1983 destroyed large parts of the summer embankments in the study area. Due to high remediation costs, about 300 ha of agricultural land (equally distributed between meadows and forest) were made available for restoration measures instead (Dister et al., 1992; Hölzel, 2006). In the 1990s, another 100 ha in the Knoblochsaue were converted to grassland (Hölzel, 2006). The flood meadows harbor numerous typical, rare and endangered species. Typical and endangered species like Arabis nemorensis, Cnidium dubium, Iris sibirica, and Viola pumila, reach their western limit here. In this regard, the nature reserve is of particular importance (Donath et al., 2009; Hölzel et al., 2002). Equally important for the nature reserve are the hardwood and softwood forests, which comprise up to 1160 ha of the nature reserve. The area covered by forest in the Kühkopf has doubled since the 1800s to 800 ha. The area covered by forest in the Knoblochsaue remains nearly unchanged from the 1800s to present (350 – 360 ha) (Gonnermann, 2002; Reif et al., 2016). Soils The nature reserve is located in the area of the gently-sloped meandering stretch with low flow velocities. This causes mainly fine sediments of the silt and clay compositions to be transported and deposited. The deposited layers of clay vary in their thickness, ranging from 30 up to 200 cm (Baumgärtel, 2004; Reif et al., 2016; Wiedner, 1990). Mainly representative in the active floodplain are the carbon-containing sandy-gravelly sediments (alluvial pararendzina) and carbon-containing silty-loamy sediments (alluvial gley, vega). In the fossil floodplain of the Knoblochsaue, carbon-containing clayey floodplain sediments dominate (pelosol) (Rosenberger, 2007; Wiedner, 1990). The combination of relatively warm and dry conditions in the summertime with the fine- grained alluvial soils leads to a variable soil water balance, a very rapid decrease in plant- available water, and development of desiccation cracks (Burmeier et al., 2010). Extended Summary 8 Meteorological conditions Meteorological information is available from the weather station in Darmstadt (N: 49°52.85’, E: 8°40.67’). The annual precipitation is about 730 mm (winter: 166 mm, spring: 170 mm, summer: 225 mm, autumn: 169 mm), and the mean annual temperature is 10.5 °C (winter: 2.3 °C, spring: 10.4 °C, summer: 18.8 °C, autumn: 10.5 °C) between 2000 and 2016. (Figure 1.2). Figure 1.2: Mean ± SD monthly rainfall and mean ± SD monthly minimum/maximum temperature recorded at Darmstadt weather station between 2000 and 2016. Hydrological conditions The Nierstein-Oppenheim gauging station (N: 49°51.89’, E: 8°21.14’) is located northwards, approximately 4 km downstream from the study site. The mean water level between the years 2000 and 2015 was 82.95 m. The critical level of 4 m above mean water level, which floods the hybrid floodplain, was exceeded about 55% of the days in this period. The lowest measured water level was recorded at 81 m (28/09/2003), and the highest was 87.10 m (06/04/2013). Hydrogeological conditions About 28 groundwater wells are located around the study site, of which 16 are relevant and used for the hydrological models of the study area (Figure 1.1, d). The water levels are observed in weekly time steps by the Hessian State Office for Conservation, Environment and Geology (HLNUG). The groundwater wells are located between 89 and 3757 m away from the cut-off meander and between 573 and 5375 m away from the Rhine River. In general, the mean annual fluctuation of the water level at the wells decreases with increasing distance from the Rhine River. The only exceptions are the groundwater wells behind the embankment, showing a lower annual fluctuation, despite a short distance to the Rhine. Extended Summary 9 Installation to improve data availability In the interest of modeling water levels on high temporal resolution and due to the lack of such observation data, several Odyssey capacitance probes (Odyssey by Dataflow Systems Pty Ltd, Christchurch, New Zealand, http://www.odysseydatarecording.com) were installed at different locations on the study site. The water level was monitored in hourly time steps with a measurement accuracy of 0.8 mm (Dataflow Systems Limited, 2018). The recording started on July 1st, 2015 at four sites, spread over the study area. Another observation point was added two weeks later on July 15th, 2015. Due to high water levels flooding the sensors and/or technical issues with the sensors, there are several gaps in the records. After August 24th, 2016 only three capacity probes were in use. 1.3.2. Surface water-groundwater interaction model Physically based, spatially distributed models attempt to represent the physical processes of the real world and are thus a valuable tool to assess the three-dimensional nature of floodplains (Bernard-Jannin et al., 2016). The main advantage of such physically based models is their ability to simulate the interaction between the surface and the subsurface domain. There are two possible methods for simulating this interaction: coupled models (also referred as loosely coupled models) containing two or more individual models that are coupled by the exchange of the model results, and integrated models (also referred as fully coupled models) that solve the surface and subsurface water flow equation simultaneously (Barthel and Banzhaf, 2016; Condon and Maxwell, 2013). Existing models differ in the spatial resolution and the governing equations, strategies, and technical solution for coupling (Maxwell et al., 2014). The Catchment Modeling Framework (CMF, Kraft et al., 2011) represents a very suitable tool box for simulating the water fluxes on a floodplain with a model tailored towards the research question at hand. Its major advantage is the flexible modular framework, offering a wide range of possible hydrological model setups. A further advantage is the extendibility of CMF. Individual processes can be adapted to the available observation data, and subroutines can be replaced or added. CMF is based on the concept of finite volume method (Qu and Duffy, 2007), allowing for the use of an irregular grid for the discretizing of lateral spatial continuous water storage. Extended Summary 10 1.3.2.1. Model set up The study area was divided into two parts (“Knoblochsaue” and “Kühkopf”), parameterized, and simulated independently. Each submodel was further discretized into irregular polygons according to elevation and land use. The allocation was performed in a way that variation in elevation is kept to a minimum. Larger differences in height within one polygon are caused by special conditions, for example, embankment, ponds, or local depressions. The submodel area of the Knoblochsaue has a height variation between 81.35 m and 90.36 m a.s.l. in total. The height variation within one polygon is between 0.88 m and 6.92 m (mean height difference within one polygon: 3.19 ± 1.12 m, mean size of polygons: 5.4 ± 7.7 ha, N: 272 polygons). The height variation of the Kühkopf area is almost twice that of the Knoblochsaue with 16.05 m, ranging from 81.10 m to 97.15 m a.s.l.. The height variation within one polygon is between 0.93 m and 7.27 m (mean height difference within one polygon: 3.72 ± 1.13 m, mean size of polygons: 7.43 ± 6.68 ha, N: 263 polygons). As fully distributed and coupled subsurface-surface water models generally require a high computing time, even increasing with spatial resolution (Clark et al., 2015), attempts were made to find a balance between the complexity and accuracy to maintain an acceptable computational time. Both submodels are based on the same modeling concept and vary only in their spatial representation, i.e., the arrangement and quantity of grid cells. Thus, the following description refers to both sub-models, if not explicitly stated otherwise. The models include the key processes to represent the water pathways accurately (Figure 1.3). For simplicity, the model was set up as single layer groundwater flow model, i.e., there is no vertical discretization. The saturated flow is calculated with a Darcian approach. The presence of surface water, throughfall, and infiltration capacity determine the infiltration rate. Precipitation is split into two parts: precipitation reaching the surface of the polygon directly, and intercepted rainfall. The latter either evaporates or reaches as canopy throughfall the surface water of the polygon, calculated with the interception model from Rutter and Morton (1977). Infiltration in the soil is calculated according to the Richard’s equation. The potential evapotranspiration rate is calculated with the FAO version of the Penman-Monteith equation (Allen et al., 1998). The actual evapotranspiration rate is limited by drought stress, and is calculated from the groundwater level, the canopy wetness, and height of the surface water. Extended Summary 11 The adjacent Rhine River is directly connected to the groundwater of the floodplain and is mainly responsible for the groundwater level changes and the riverside inundation on the active floodplain. Observed water levels are implemented as a Dirichlet boundary condition through Open Water Storages. For the Kühkopf submodel, the extrapolated water level of the cut-off meander of the Rhine was used as a surrounding Dirichlet boundary condition. At the landsite area (Knoblochsaue: northeast, Kühkopf: southeast), the data from the groundwater wells are implemented likewise as Dirichlet boundary condition. Realistic estimations of initial groundwater levels were obtained via external drift kriging (Goovaerts, 1997) with the available groundwater data. Figure 1.3: Three-dimensional sketch of the study area Kühkopf-Knoblochsaue with typical structures (darker blue: Rhine River and surface water, lighter blue: groundwater, green: flood meadow with embankment and depressions) and the relevant water flux processes simulated by CMF. Extended Summary 12 1.3.2.2. Surface water flow CMF is principally capable of simulating surface water flow as a kinematic or diffusive wave. For flat areas, like the study site, the diffusive wave represents a feasible approximation for the surface water flow (Singh, 1996). However, this approximation requires long computing times, especially during large flooding events. To overcome these long computing times, the St. Venant surface flow is replaced by a newly developed simplified function, introduced as the “flood-wave scheme.” This function calculates the distribution of the surface water at steady-state for each day, based on the water level of the Rhine River. The concept is straight forward: each day, the water level height of the Rhine River is compared to its neighboring polygons. If the water level of the Rhine River is higher than the surface of the polygons, the surface water level of polygons are set to the same level as the water level of the Rhine River. The surface water potential of all other cells is recursively adjusted if they can be flooded by an already flooded cell (Figure 2.1). 1.3.2.3. Model uncertainty The model setup was designed as a parsimonious model solution. The reduction to key processes, the single-layer groundwater model, and particularly the simplification of surface water flow equation resulted in a significantly reduced computational run time. This model setup allowed for the investigation of the parameter uncertainty of the model following a GLUE-like method (Beven and Binley, 1992). For both submodels, the same soil parameters were calibrated: saturated conductivity (m d-1), porosity (m3 m-3), residual water content (m3 m-3 pores m-3 soil), and soil thickness (m). As the parametric equations describe the processes of the real-world aggregated in space and time, the parameters are considered as effective, without a direct physical representation. With the Statistical Parameter Optimization Tool for Python (SPOTPY) by Houska et al. (2015), 5000 parameter sets were generated, using a Latin hypercube sampling approach (McKay et al., 1979). The root-mean-square error (RMSE) was used as an objective function, rating the performance of the parameter set in its ability to simulate the observations. Extended Summary 13 1.4. Results and discussion 1.4.1. Correlation of Rhine water level with groundwater level The records of the groundwater levels show a high correlation with the water level of the Rhine. The floods in May and June of 2016 are especially well reflected in the groundwater level in the wells 527026, 527200, and 527283. The groundwater levels in well 527200 show clear fluctuations in the summer of 2017, which follow the course of the water level of the Rhine. However, since from summer 2016 to spring 2017 no correlation between the fluctuations in the Rhine water level and the groundwater level is recorded, this strong reaction of the groundwater level to the water level fluctuations of the Rhine should be treated rather critically. A measurement error should therefore be considered here. The time differences between the occurrence of flood peak in the Rhine water level and that of the groundwater level (as clearly distinguishable individual events) are estimated to be 48.2 h on average for the groundwater well 527026, and 88.6 h for the groundwater well 527200. Subplots (e) and (f) depict negative values, caused by the location of the Nierstein- Oppenheim gauging station (about 4 km downstream the study area). Therefore, there is also a temporal gap between the Rhine water level in the vicinity of the study area and at the gauging station. The negative values likely show a very short reaction time of the groundwater to an increase of the Rhine water level. Extended Summary 14 Figure 1.4: Time series of the hourly measured groundwater levels for five different groundwater wells between July 2015 and December 2017 and the time series of the Rhine for the Nierstein-Oppenheim gauging station (top). Subplots a-c (middle) depict details of the time series for well 527026; subplots d-g (bottom) depict details of the time series for well 527200. The time gap between the peaks of the water levels is given in the top of the subplot. The map in the middle left shows the locations of the different depicted wells. Grey: study area, blue: Rhine and cut-off meander, violet: embankment, black: groundwater wells. Extended Summary 15 1.4.2. Model calibration and validation The results of the model calibration and validation are described in detail in Chapter 2 and are published in the paper: Maier, N.; Breuer, L.; Kraft, P. (2017). Prediction and uncertainty analysis of a parsimonious floodplain surface water‐groundwater interaction model, Water Resources Research, 53, 7678–7695, doi: 10.1002/2017WR020749. Additionally, the publication shows a performance test of the simplified surface water flow equation described in Chapter 2. The weekly groundwater levels of six (Knoblochsaue) and four (Kühkopf) groundwater wells were used to calibrate and validate the submodels. The calibration period comprised 2.5 years (01/08/2002 – 30/06/2004), including a large fraction of the hydrological variability, while the validation period comprised 9.5 years (07/01/2004 – 31/12/2013). The cutoff criterion to judge a model run as a behavioral run, i.e., an acceptable simulation, was set to 0.26 m for the Knoblochsaue and 0.38 m for the Kühkopf for the mean RMSE over all groundwater wells considered for the calibration of the respective submodel. The mean RMSE of both submodels decreases by 0.2 m in the validation period compared to the calibration period (Knoblochsaue: RMSEcalibration = 0.25 m; RMSEvalidation = 0.23 m. Kühkopf: RMSEcalibration = 0.38 m; RMSEvalidation = 0.36 m). The range of individual RMSEs varies up to 0.41 m (Knoblochsaue) and 0.67 m (Kühkopf) for the calibration period. While the upper boundary of individual RMSEs decreases for validation period for the Knoblochsaue (0.35 m), it increases for the Kühkopf (0.79 m). Overall, the performance of the Knoblochsaue submodel showed a much smaller RMSE than for the Kühkopf submodel, i.e., the Knoblochsaue submodel is performing much better. This is clearly noticeable in the uncertainty ranges, which are much larger in the Kühkopf submodel than in the Knoblochsaue submodel (exemplarily shown for two groundwater wells in Figure 1.5). The inferior performance of the Kühkopf model is likely caused by the terrain structure, the simplified representation of it, and the corresponding flow paths. To improve the model, the topographical structure should primarily be delineated in a more detailed way. However, this would mean that the model’s required computational time would increase significantly due to the higher number of polygons and processes (Clark et al., 2015). Extended Summary 16 For both submodels, the ranges of parameters for residual wetness were equal (0.10 – 0.30 m3 m-3 pores m-3 soil). The range of the soil thickness (Kühkopf: 3.60 – 9.76 m; Knoblochsaue: 5.10 m – 8.00 m) and the porosity (Kühkopf: 0.10 – 0.33 m3 m-3; Knoblochsaue: 0.30 – 0.34 m3 m-3) were larger in the Kühkopf submodel than in the Knoblochsaue submodel. Saturated conductivity resulted in lower values in the Kühkopf submodel (160 – 1779 m d-1) than in the Knoblochsaue submodel (2193 – 4986 m d-1). Figure 1.5: Time series of the simulated water levels by the CMF model, compared to the measured data by the HLNUG for each of the groundwater wells in the Knoblochsaue (top) and the Kühkopf (bottom). The map (bottom left) shows the location of the relevant groundwater well (red) and the remaining groundwater wells (black) in the study area. Grey: study area, blue: Rhine and cut-off meander, violet: embankment, black: groundwater wells. To test whether the weekly available groundwater data are sufficient for model calibration, the model performance was also evaluated using the measured daily groundwater levels. An example of two groundwater wells is shown in Figure 1.6. In the Knoblochsaue reserve (Figure 1.6, top), simulations follow the daily water level measurements of the odyssey capacitance probe very well (RMSE = 0.20 m). In contrast, the simulations in the Kühkopf Extended Summary 17 (Figure 1.6, bottom) show much larger fluctuations than observed by the odyssey capacitance probe (RMSE = 0.56 m). To summarize, the simulations of the groundwater levels of the Knoblochsaue show a low RMSE, based on both weekly and daily water levels. The simulations of the Kühkopf are less precise, and the simulations do not follow the real water levels measured on daily time steps. Therefore, the Knoblochsaue submodel is recommended for further applications. Figure 1.6: Simulation of the CMF model, compared to the measured water level by the capacity loggers and the measured data by the HLNUG, exemplarily shown for groundwater wells in the Knoblochsaue (top, 1.7.2015 – 27.7.2016) and the Kühkopf (bottom, 1.7.2015 – 30.5.2016). Performance as RMSE is shown for all possible comparisons. Measured data by the HLNUG include one outlier measurement, which is not included in the estimation of the RMSE. The map (bottom right) shows the location of the relevant groundwater well (red) and the remaining groundwater wells (black) in the study area. Grey: study area, blue: Rhine and cut-off meander, violet: embankment, black: groundwater wells. Extended Summary 18 1.4.3. Projecting floodplain inundation under climate change The results of this chapter are described in detail in Chapter 3 and published in the paper: Maier, N.; Breuer, L.; Chamorro, A.; Kraft, P.; Houska, T. (2018). Multi-Source Uncertainty Analysis in Simulating Floodplain Inundation under Climate Change. Water, 10, 809, doi: 10.3390/w10060809. Based on the previously described study’s results, the impact of climate change on the inundation characteristics was evaluated for the Knoblochsaue, because this submodel performed much better than the Kühkopf submodel. Therefore, a modeling framework of bias-corrected climate model data was set up to project future discharge of the Rhine River via a rainfall-runoff model, and the groundwater and surface water level via the CMF model. Therewith, the contribution of multiple uncertainty sources within the modeling framework was assessed. The submodel of the Knoblochsaue reserve was slightly modified for this analysis, due to the limited availability of data for the future. The original model was forced by three input sources: (1) climate data, i.e. rainfall, temperature, relative humidity, and wind speed; (2) the water level of the Rhine; and (3) the water level data of three groundwater wells on the upslope (landside) boundary. Climate data were selected from two general circulation models (GCMs) from the Coupled Model Intercomparison Project Phase 5 (Taylor et al., 2012a), namely: HadGEM2-ES (later abbreviated as HadGEM) from the Met Office Hadley Centre and MPI-ESM-LR (later abbreviated as MPI-ESM) from the Max Planck Institute for Meteorology. Three representative concentration pathways (RCP) were considered: RCP 2.6 (low concentration), RCP 4.5 (medium concentration), and RCP 8.5 (high concentration). All climate data were bias corrected with the quantile mapping method (Maraun et al., 2010; Themeßl et al., 2011). The data for the second input source, the water level of the Rhine River, were generated using a rainfall-runoff model. For this data generation, the HBV model (Bergström, 1995; Lindström et al., 1997) was used as a conceptual semi-distributed hydrological model to simulate the discharge of the Rhine. After calibration and validation, the model was forced with the projected and bias corrected climate data. To assess the uncertainty, the parameter and the predictive uncertainty were considered. The parameter uncertainty was assessed by the GLUE methodology (Beven and Binley, 1992). From 100,000 simulations with parameter Extended Summary 19 sets that were generated following a Latin hypercube procedure, the best 10% were kept as behavioral runs. With these selected parameter sets, future discharges were projected, and the associated uncertainty was estimated. Synthetic time series (5th, 50th, and 95th percentiles) were generated from all behavioral runs and transformed to daily water levels by an existing rating curve. Predictive uncertainty, representing the aggregated effects of data and model structural errors of the best model run, was estimated by means of a statistical model error, for which the Box–Cox transformation was used (Box and Cox, 1982). The estimated parameters for the transformation equation for the past were then applied to discharge simulations under climate change. The three different synthetic time series were generated via the same process. Finally, these synthetic time series were used as input data for the CMF model. As no groundwater data are available for the future, the upslope boundary condition was replaced with a no-flow boundary. A subsequent recalibration of the model resulted in a similar performance and optimal parameter sets. Due to the numerical stability and computation effort, only one soil parameter set for the CMF model, based on numerical stability and performance, was used for the future projections. Thus the parameter uncertainty of the CMF model is not considered. With two GCMs, three RCPs, five water level projections for the Rhine, and one soil parameter set for the CMF model, 30 different spatio-temporal simulations of the water level and inundation of the floodplain were generated as future projections. Two periods were defined for the projections: near future (2021–2050) and distant future (2071–2100). At the end, two indicators were used to characterize the inundation characteristics: the mean annual inundation days and the mean inundation period. Inundation was defined as surface water level higher than 5 cm. Comparisons of the different projections of annual inundation days resulted in the following ourcomes: (a) under the HadGEM, the spatial extension and the magnitude of annual inundation days were larger (Figure 1.7, top); (b) under the MPI-ESM model, similar annual inundation days for all RCPs and both time periods were projected, whereas under the HadGEM model the projections varies more between RCPs (Figure 1.7, middle, light grey); (c) the simulations considering the different uncertainties of the projected Rhine water level (parameter and predictive uncertainty) showed the same pattern as the RCPs; and Extended Summary 20 (d) considering the predictive uncertainty, the spatial extension and magnitude was slightly higher than when the parameter uncertainty is considered (Figure 1.7, bottom, darker grey) Figure 1.7: Annual inundation days for the near future (2021–2050), divided between the GCMs (top), the GCM-RCP combinations (middle, light grey), and the combination of GCM, RCP, and uncertainty of the HBV model (bottom, darker grey). The comparison of the projected annual inundation days of the near and distant future to simulated annual inundation days of the past (2002–2015) showed similar patterns: (a) under the MPI-ESM model, similar changes (decrease of annual inundation days) were projected for all RCPs and both time periods; (b) compared to the HadGEM model, the spatial extension of areas affected by changes under the MPI-ESM model was smaller; Extended Summary 21 (c) the magnitude of changes was much larger under the HadGEM model than under the MPI-ESM model; and (d) the variation in the RCPs was much larger under the HadGEM model than under the MPI-ESM model in terms of absolute values (Figure 3.4 and Figure 3.5). In summary, the differences in the spatial extend and magnitude of mean annual inundation days and mean inundation periods were larger between the GCMs than between the RCPs (with some exceptions), and between the different used input data (i.e., different predicted water levels of the Rhine. Therefore, further studies on the impact of climate change on inundation characteristics should focus more on the uncertainties caused by the climate models. 1.4.4. Species distribution modeling in floodplain habitats The results of this chapter, with a more detailed focus on the habitat model, are described in Chapter 4 and are submitted to the journal Ecohydrology under the title: Gattringer*, J.P., Maier*, N.; Breuer, L.; Otte, A.; Donath, T.W., Kraft, P.; Harvolk-Schöning, S. (2018). Modeling of rare flood meadow species distribution by a combined habitat-surface water-groundwater model. *shared first-authorship Numerous flood meadow species have adapted to the specific hydrological regimes of floodplains. To predict spatially explicit suitable locations for species, species distribution models (SDMs) are a good choice. They are based on statistical correlations between observations and the environmental conditions at the observation point (Elith and Leathwick, 2009; Guisan and Thuiller, 2005; Peterson et al., 2011). In the context of flood meadow species, such environmental variables would involve information on inundation length and depth, as well as recurrence intervals of floods and soil moisture conditions. Hydrological information implemented in the SDM as environmental variables (also called hydrological predictors) is derived from the developed CMF model. The simulated long- term time series of the water level of both submodels were transformed to static hydrological predictors by downscaling (inverse distance weighting) the water level of the polygons from the CMF model to a high spatial resolution raster (5 x 5 m). From over 80 different a priori conceived hydrological predictors, the 15 most important ones were selected by an iterative Extended Summary 22 process and based on species-specific flooding experiments (Gattringer et al., 2017, 2018). Hydrological predictors include, for example, information on groundwater level range, standard deviation, or average and longest durations below/above certain groundwater level thresholds. Additionally, one meteorological (longest period of wet days) and three morphological predictors (height above mean sea level, distance to the Rhine or the cut-off meander and distance to any water surface) were added as environmental variables. In order to test the significance of the high temporal and spatial resolution hydrological predictors obtained from the CMF model (sgm: surface water-groundwater model), two other datasets of hydrological predictors were generated. The first set of data used for comparison was based only on weekly measured groundwater levels at the 16 different sites in the floodplain (gww: groundwater wells). The second set of data was based only on the water level of the adjacent Rhine River (riv: river). Additionally, the model performance without any hydrological information (nhy: non-hydrological), i.e., with only meteorological and morphological predictors, was tested. The results focus mainly on floodplain meadow species (Arabis nemorensis, Galium boreale, Peucedanum officinale, Sanguisorba officinali, Silaum silaus, Thalictrum flavum, and Veronica maritima) after Burkart (2001) and on rare and engendered species (Arabis nemorensis, Bromus racemosus, Galium boreale, Iris spuria, Peucedanum officinale, Serratula tinctoria, and Veronica maritima), as they are specialized species and species valuable to consider for restoration projects. For both groups, the SDM with hydrological predictors derived from the CMF model is superior to the other setups, i.e., setups considering hydrological predictors from groundwater level or river water level and those not considering any hydrological predictors (Figure 4.3). Testing several combinations and quantities of predictors led to the conclusion that six to ten specific predictors are needed to satisfactorily simulate habitats and occurrences for rare and endangered species. The most frequently used hydrological predictors included parameters indicating inundation length and dry or wet conditions. These results reflect the complexity of the habitat requirements of flood meadow species being able to cope with both flooding and drought periods (Burkart, 2001). Even though not all the 15 hydrological predictors are essential for simulating all species, it became clear that removing selected predictors for model simplification resulted in a model failure for several species. Thus, it can be concluded that a large set of hydrological predictors is required to be Extended Summary 23 able to simulate whole plant communities with their diverse specific eco-hydrological requirements. The superior results of the habitat model including hydrological predictors from the CMF model became visible when the time series were analyzed in more detail. The hydrological model is a respectable representation of the natural conditions in a high spatial and temporal resolution, including components such as climatic conditions, soil properties, surface water distribution, surface water-groundwater interaction, and groundwater level. Thus, the model is able to reflect the soil moisture content, altering the water storage capacity of the soil and thus driving the flood extent, flood duration, and inundation height of water in the floodplain. Inundation caused by high water levels of the Rhine can be attenuated if the soil is capable of draining a large volume of water, but at the same time, large floods can occur when the soil is already saturated. The CMF model is capable of representing such conditions. For the two days shown in Figure 1.8, predictors based on the CMF model would result in different values, whereas predictors based on the water level of the Rhine would result in the same value, because the water level of the Rhine is the same on both days. Overall, creating a habitat model with predictors, derived from perfectly spatially distributed hydrological measurements would be ideal. However, this is nearly impossible due to high costs. A few sporadic or poorly distributed measurements lead to less dependable results, as those involve some failures in the representation. For example, the rapid reaction of groundwater to changes in the water level of the river cannot be represented. Thus, generating hydrological predictors from a hydrological model with an appropriate representation of the natural conditions in a high spatial and temporal resolution including various components (climatic conditions, soil properties, surface water distribution, surface water-groundwater interaction, and groundwater) seems to be a good compromise between cost and the objective to protect rare species. Extended Summary 24 Figure 1.8: Representations of the groundwater and surface water level in the study site at two different days (07/01/2003 and 16/01/2004). The water level of the Rhine is depicted in the bottom of the figure (MHW: mean high water, MW: mean water, MLW: mean low water). The blue lines indicate the water level of the days depicted in the top of the figure. The water level of the Rhine was 86.15 m on both days. Extended Summary 25 1.5. Conclusion This work showed new methods for determining spatio-temporally explicit predictions of groundwater, surface water and habitat interactions in riparian ecosystems. First, a hydrological model was developed, and then the model was linked with a species distribution model. The model was developed for the floodplain of the Rhine River that covers the Kühkopf-Knoblochsaue nature reserve. This region is very beneficial for the modeling approach, as abundant data particular for groundwater levels and species distribution are available. The physically based, deterministic surface water-groundwater model build with the Catchment Modeling Framework is able to represent the main processes, which dominate the groundwater dynamic and surface water distribution. The established simplified surface water equation reduces the computation time of the model significantly, and it thereby allowed for facilitation of comprehensive model parameter uncertainty analyses. Additionally, long term projections of future conditions were possible. It was shown that in long-term simulation (a) the climate models introduced large uncertainties in the projections, and (b) depending on the climate model, the difference between the RCPs was much smaller than between the climate models. This indicates that we cannot yet quantify any potential benefit in greenhouse gas mitigation with regard to the investigated inundation characteristics. From the simulated long-term hydrological dynamics of the floodplain, hydrological predictors were derived as inputs for the species distribution model. From over 80 different conceived predictors, 15 hydrological predictors were identified as the most important by an iterative process and based on species-specific flooding experiments (Gattringer et al., 2017, 2018). Furthermore, the SDM was extended with one meteorological and three morphological predictors. Comparisons with two other datasets of hydrological predictors showed that the quality of the SDM was significantly higher when using hydrological predictors gathered from the CMF model. Hydrological predictors based only on weekly measured groundwater levels at 16 different sites in the floodplain and based only on the water level of the adjacent Rhine River led to significantly worse, but still decent SDM results. The worst results were obtained by the SDM when no hydrological predictors were incorporated. These results emphasize the importance of taking hydrological conditions into account when modeling the distribution of species in a floodplain, as stated by other authors Extended Summary 26 (Kopeć et al., 2013; e.g., Leyer, 2005). Removing less frequently used predictors resulted in a model failure for several species, i.e., the prediction for those species is unsatisfactory. Thus, it can be concluded that a set of hydrological predictors, including predictors indicating drought and wet conditions, is required to be able to simulate whole plant communities with their diverse and specific eco-hydrological requirements. Overall, the results underline the advantage of spatio-temporally explicit modeling of the hydrological regime on the floodplain in order to achieve valuable predictions of potential habitats, especially for rare and endangered floodplain species. Such scientific findings and approaches can make a major contribution to practical nature conservation. 1.6. Outlook Useful for conservation measures is the prediction of potentially restorable areas or areas that offer suitable conditions for rare and endangered species in the future. The knowledge of such sites offers the opportunity to focus the restoration measures more precisely with a higher degree of sustainability and endurance for the future. In this respect, however, both climate and landscape changes must be taken into account. Therefore, implementations of predicted hydrological conditions in the established modeling framework were carried out in a first pilot study in order to project the species distribution under climate change. In the following the preliminary results are presented and discussed. Furthermore, possible land use changes and the possibility of implementation in the modeling framework are outlined. 1.6.1. Species distribution under climate change The model framework of surface water-groundwater-species distribution proposed in this thesis can be used to project areas with high occurrence potential of species under climate change. As one example, the occurrence for rare and endangered species from the Red List (Arabis nemorensis, Bromus racemosus agg., Inula salicina, Lysimachia vulgaris, Peucedanum officinale, Veronica maritima, Rhinanthus alectorolophus and Serratula tinctoria) in the study area were projected for the years 2050 and 2100 under climate change conditions. Due to large computational time and power, the modeling framework considered only one uncertainty source, namely the climate models. Those were identifies as the largest uncertainty source in previous climate change impact studies on this study area (Chapter 3) Extended Summary 27 and as well by other studies (Dobler et al., 2012; e.g., Kay et al., 2009; Prudhomme and Davies, 2009). Future predictions were given by 15 regional climate models (RCMs) from the Coordinated Regional Downscaling Experiment EURO-CORDEX-11 (resolution: 0.11 degree), namely: (1) CNRM-CERFACS-CNRM-CM5-ALADIN53, (2) CNRM-CERFACS-CNRM-CM5-CCLM4- 8-17, (3) CNRM-CERFACS-CNRM-CM5-RCA4, (4) ICHEC-EC-EARTH-CCLM4-8-17, (5) ICHEC-EC-EARTH-RACMO22E, (6) ICHEC-EC-EARTH-RCA4, (7) MOHC-HadGEM2- ES-CCLM4-8-17, (8) MOHC-HadGEM2-ES-RACMO22E, (9) MOHC-HadGEM2-ES-RCA4, (10) MPI-M-MPI-ESM-LR-CCLM4-8-17, (11) MPI-M-MPI-ESM-LR-RCA4, (12) MPI-M-MPI- ESM-LR-REMO2009, (13) NCC-NorESM1-M-HIRHAM5, (14) IPSL-IPSL-CM5A-MR-RCA4 and (15) IPSL-IPSL-CM5A-MR-WRF331F. For all RCMs, two representative concentration pathways (RCPs) were applied, namely RCP 4.5 and RCP 8.5. As final results, the occurrence probabilities of the species were predicted by the SDM. In order to decrease the uncertainty of future predictions, the focus should be only on areas with a high degree of agreement between the projections under the different climate models. Therefore, a cutoff value of 10% or 20% deviation between the predictions was defined, so that areas with high uncertainty between the climate models were excluded. Additionally, areas that can be generally restored without the need to clear a forest, for example, are important for restoration measures. Therefore the mean occurrence probability of each species was intersected with the actual land use. Finally, these restrictions resulted in projections of occurrence probability for meadows with a small difference between the projections under different climate models. The first results indicated for some species (Arabis nemorensis, Peucedanum officinale, Serratula tinctoria) a large agreement in the projected occurrence probability, considering both a 10% and a 20% difference in the occurrence probability between the RCMs. This is greatly reflected in a high percentage of restorable area (Figure 1.9, top). For other species (Inula salicina, Lysimachia vulgari, Veronica maritima, Rhinanthus alectorolophus), a larger area is suitable for restoration when assuming a difference between the projections of the RCMs of 20% as satisfactory, instead of a difference of 10%. The area with occurrence potential significantly decreases for all species if only areas with an occurrence potential above 25% or 50% are considered (Figure 1.9, middle and bottom). There is little difference in the restorable area between RCPs, except in one case. For Bromus racemosus agg., under the Extended Summary 28 RCP 4.5 a rather small restorable area is projected, and under RCP 8.5, it is much larger (Figure 1.10). As mentioned, the climate models are only one uncertainty source when projecting future conditions. To consider the full range of possible climate change impacts, more uncertainty sources should be considered. It is recommended to (1) include a different method for the bias corrections of the climate data (e.g., Lafon et al., 2013; Teng et al., 2015), (2) utilize different rainfall-runoff models to assess the model structure uncertainties (e.g., Breuer et al., 2009), (3) use different parameter sets for the CMF model (e.g., Saha et al., 2017), and last but not least, (4) use different species distribution modeling techniques in order to consider the uncertainties of the species distribution model (Breiner et al., 2018; e.g., Pearson et al., 2006). Figure 1.9: Percentage of restorable area with a difference of less than 10% (marker: downwards triangle) or less than 20% (marker: upwards triangle) between the projections under the different RCMs for eight species. The grey area/bar in the background depicts the predictions for the year 2016. The blue and orange colors depict the predictions of the year 2050 and 2100, respectively. The markers depict the mean of the projections under RCP 4.5 and RCP 8.5 with the corresponding errors. Restorable area is represented as different levels of occurrence probability (top: > 0% occurrence probability occurrence probability, middle: > 25% occurrence probability, bottom: > 50% occurrence probability). Extended Summary 29 Figure 1.10: Occurrence probability (in percentage) for the species Arabis nemorensis (top), Peucedanum officinale (middle), and Serratula tinctoria (bottom) for the years 2050 and 2100 for both RCP 4.5 and RCP 8.5. Areas with a difference of more than 10% between the 15 RCMs are colored in grey. Streams and water bodies are depicted in blue. Areas that cannot be used for restoration, based on current land use, are marked in black. Extended Summary 30 1.6.2. Species distribution under land use changes One could argue that in the future not only the climate will change, but also land use and management. Land use changes can be caused, for example, by humans or natural catastrophe. Naturally, there is an interaction between land use changes and climate change (Oliver and Morecroft, 2014). Usually the effect of the combinations of land use and climate change is much larger than the single effects on the impact on biodiversity (Oliver and Morecroft, 2014). Land use changes can cause an effect on the climate by altering the carbon balance in atmospheric and terrestrial pools (Bonan, 2008; Cramer et al., 2001). In addition, different land covers can cause changes in the climate, such as the changes in surface fluxes of radiation, heat, moisture, and momentum (Betts, 2005). On the other hand, climate changes can cause land cover and land use changes, for example, by altering the climatic conditions for vegetation types (Cramer et al., 2001). Different scenarios can be imagined on how land use or morphological characteristics could be changed in the study area and therefore can cause changes in the habitat availability and suitability for plant species. As an example of land use changes, the decrease of agricultural land in the active floodplain or the decrease of grassland due to transformations into riparian forest could be considered here (Xu et al., 2017). At the same time, relocation of embankment, floodplain reconstruction, or river regulation are possible changes of the morphological structures (Moss and Monstadt, 2008; Schneider, 2010). As climate changes and land use changes interact with each other, it is important to consider both in future projections. Both changes can influence water availability and thus cause changes in biodiversity and habitat availability. The CMF model offers the possibility of implementing such land use changes and can therefore project future water availability conditions. The complete model framework is thus capable of improving the accuracy of future predictions for species habitats. This, in turn, is vital for wetland management, especially where species conservation and other ecosystem service provision relies on detailed, high-resolution, hydrological conditions (Acreman et al., 2009). Overall, it has been shown, that there are many opportunities to enhance conservation measures with scientific knowledge and modeling strategies. Therefore, the research in this field should be advanced further to open more possibilities for the promotion of restoration measures. Prediction and uncertainty analysis of a parsimonious floodplain surface water – ground-water interaction model 31 2. Prediction and uncertainty analysis of a parsimonious floodplain surface water – ground-water interaction model This chapter is published in the journal Water Resources Research 53, pages 7678-7695, 2017. doi: 10.1002/2017WR020749 © American Geophysical Union. Nadine Maier1, Lutz Breuer1 and Philipp Kraft1 1 Chair of Landscape, Water and Biogeochemical Cycles, University of Gießen, Germany. Abstract Floodplains provide a variety of hydrological and ecological functions and are therefore of great importance. The flooding frequency, as well as the height and duration of inundations are particularly relevant for ecosystem states and are dependent on the exchange between surface water and groundwater. In this study, we developed a fully distributed model approach to simulate distributed groundwater levels in a floodplain in Hesse, Germany (14.8 km2). To overcome the problem of large computation times we simplified the surface water equation. Thus, the water surface of flooding is at the same level everywhere and the dynamic effect of the flooding is ignored. In this way, it was possible to run the model 5,000 times and investigate its parameter uncertainty using Latin hypercube sampling. Behavioral model runs were selected based on a threshold criterion of a mean root mean square error that was smaller than 0.26 m. All the simulated groundwater wells show an individual RMSE between 0.17 and 0.41 m for the calibration period. Regarding the parameterization, the model shows rather large variance in parameters that are capable of generating good simulations: a range of saturated conductivity of 2,793 m/day, porosity of 0.4 m3/m3, residual wetness of soil of 0.2 m3/m3/soil and range of soil thickness of 2.9 m. Prediction and uncertainty analysis of a parsimonious floodplain surface water – ground-water interaction model 32 2.1. Introduction Floodplains, with their high water storage capacity, biological productivity and diversity, are among the most endangered ecosystems worldwide (Funk et al., 2013; Tockner and Stanford, 2002). Flood meadows are declining as a result of reduced hydrological dynamics (Tockner and Stanford, 2002) and land use changes (Wesche et al., 2012). The riparian zone of a floodplain often consists of fluvially derived sediments from ancient and current stream systems and represents a buffer zone between aquatic and terrestrial environments (Gregory et al., 1991; Woessner, 2000). Floodplains offer a wide range of hydrological and ecological functions. The basic functions of floodplains include the promotion of natural regulation and retention during hydrological extremes (Krause and Bronstert, 2007; Sophocleous, 2002), as well as the exchange of organic matter, nutrients and pollution between the river and the terrestrial ecosystem (Kiedrzyńska et al., 2015; Tockner et al., 1999). The driving factor for the eco-hydrological functions is the interaction of the surface water of the connected river, the surface water in the floodplain and the shallow groundwater (Hayashi and Rosenberry, 2002; Krause et al., 2007a). In addition to the exchange of surface water and groundwater, the timing, frequency and extent of inundation periods are important for floodplain ecosystems (Bernard-Jannin et al., 2016), as their habitats are sensitive to flooding (Russo et al., 2012). The exchange between surface water and groundwater reflects a complex spatial and temporal pattern. A model of complete surface-subsurface flow should include and couple both components (Furman, 2008). Physically based, spatially-distributed models seem to be valuable tools for assessing the three-dimensional nature of the system, as they are capable of taking environmental characteristics into account (Bernard-Jannin et al., 2016). Several models of this problem domain exist, but differ in the governing equations, strategies and technical solution for coupling and the spatial resolution (Maxwell et al., 2014). Most models use the Richards’ equation for subsurface flow and the Saint-Venant equation (kinematic, diffusive or dynamic wave) for surface flow (Furman, 2008). Recently, Barthel and Banzhaf (2016) published a review of groundwater-surface water interaction with a focus on regionally integrated models. They identified four frequently used, fully coupled models (ParFlow, HydroGeoSphere, InHm, OpenGeosys). An intercomparison of seven different coupled subsurface-subsurface models (CATHY, HydroGeoSpehere, OGS, PIHM, ParFlow, PAWS, tRIBS-VEGGIE) for different benchmark problems was performed by Maxwell et al. (2014) for simulation times in the range of Prediction and uncertainty analysis of a parsimonious floodplain surface water – ground-water interaction model 33 minutes. They concluded that for simple test cases, the model agreement is good, but the disagreement increases with the complexity of the test cases. Although there are several models in the literature that are capable of simulating groundwater – surface water interactions, their application to case studies, different times and special scales is rare and they are more or less applied to experiments and small catchments (Barthel and Banzhaf, 2016; Ebel et al., 2007; Jones et al., 2008; Loague et al., 2005). Most models involve important simplifications. Full three-dimensional numerical solutions of saturated-unsaturated zone processes and surface and groundwater flow combination are not usable for large catchments because of their long computational run time (Bernard-Jannin et al., 2016; Krause and Bronstert, 2007). Bernard-Jannin et al. (2016) summarized, that these models can be only applied at the reach scale for short periods. The performance of such models is usually evaluated using data on river discharge and not groundwater heads. Distributed hydrological response data, such as piezometric data, are often not available, despite the importance for model calibration and evaluation, which is a barrier for meaningful comparison as stated by Sebben et al. (2013). The interactions between the, usually very shallow groundwater and surface water, which is typically closely connected to rivers, is important for the eco-hydrological functions of floodplains (Butturini et al., 2002; Hancock et al., 2005). Surface water can enter the floodplain laterally across the channel banks or vertically across the floodplain surface through Darcy flow. Consequently, the groundwater level of the floodplain increases in response to increased stream stage (Hester et al., 2016). Floodplains, as part of catchments, usually have a very flat area and their water balance is affected by groundwater processes and interactions with surface water. Both factors make watershed delineations of floodplains, based on only surface watersheds, difficult and questionable. The groundwater catchment boundaries of floodplains do not correspond to the spatial extent of the delineated surface watershed (Krause and Bronstert, 2005). Pressure head gradients, hydraulic permeability of the hyporheic zone and riverbed geometry are controlling factors for the characteristics, intensity and direction of groundwater-surface water interactions (Krause and Bronstert, 2007; Sophocleous, 2002). This interaction is often characterized by high temporal and spatial variability (Krause et al., 2007a). Krause and Bronstert (2005) applied the IWAN model to the Havel River basin, in north east Germany and identified a 998 km2 portion of the floodplain in which the water Prediction and uncertainty analysis of a parsimonious floodplain surface water – ground-water interaction model 34 balance is mainly affected by the surface water dynamic of the adjacent river. In further studies, they quantified the exchange fluxes across the groundwater and surface water and the groundwater recharge. They concluded that the groundwater–surface water interactions have a greater influence on the groundwater recharge dynamics in the floodplain than on the vertical groundwater recharge (percolation, root water uptake) (Krause et al., 2007a, 2007b). However, Pirastru and Niedda (2013) conducted a study in a floodplain aquifer in northwest Sardinia in Italy, and found that simultaneous processes of lateral groundwater flow and vertical recharge are responsible for the water table fluctuations, but the relative contribution of the flows varies with soil moisture content and groundwater depth. Wu and Zeng (2013) described three main sources for uncertainties related to groundwater modeling: conceptual uncertainty, parameter uncertainty and input uncertainty. They conclude that the main focus in groundwater modeling is parameter uncertainty, which arise mainly from spatial and temporal variability and the scaling effect of parameters (Wu and Zeng, 2013). Two typical input variables in groundwater models are hydraulic conductivity and recharge. The different spatial scales of variation for these parameters are small compared to the size of the modeled region (Li et al., 2003). Additionally, a main limitation of groundwater models is the number of observations available to characterize subsurface variability. Other authors focus on other uncertainties: Rojas et al. (2010) summarized from different studies that the uncertainties in groundwater models mainly arise from the definition of alternative conceptual models. This uncertainty cannot be compensated by parametric uncertainty. Conceptual uncertainties are mainly due to inadequate representation of physical processes, an incomplete understanding of the geological situation or an inability of the model to explain all observations of the state variables (Singh et al., 2010). Following the concept of equifinality, many combinations of model structures and parameter sets are capable of simulating systems in an acceptable way. The best simulations, based on likelihood measurements, are combined into a set of behavioral runs that give a good prediction of the system. Compared to the importance of the ecosystem services provided by flood plains, only a few studies have focused on the simulation of groundwater level at the same time as simulating flooded areas. Current analyses of the parameter uncertainty of this type of coupled model are common in catchment hydrology but have been scarcely performed for coupled groundwater – surface water models, as the computational time of such models is typically Prediction and uncertainty analysis of a parsimonious floodplain surface water – ground-water interaction model 35 large and the performance of a great number of simulation runs is thus restricted. Following Pappenberger and Beven’s (2006) typology of uncertainty handling in hydrology, groundwater models are commonly of a deterministic type, whereby modelers prefer to use physically based measured parameters and respective equations. Sanchez-Vila and Fernàndez-Garcia (2016) and Cirpka and Valocchi (2016) delineated the associated uncertainty of conceptual model choices as the primary objective of the stochastic analysis. As one reason for the low impact of stochastic hydrology on practice, Cirpka and Valocchi (2016) named the lack of usable tools for stochastic analysis and is consequently recommend for the generation of realistic realizations of subsurface properties as easy-to-use tools. Fully distributed and coupled groundwater-surface water models tend to require a large computational power and time. With increasing spatial resolution, the computational time increases simultaneously. The resulting challenge is to find a balance between model complexity and accuracy to represent the observed processes (Clark et al., 2015). It should be considered whether it makes sense to increase the complexity when only limited spatial information is available and how restrictive the increased computational time is when increasing the complexity of the model. In this study, we present a simplified physical- deterministic model of a floodplain to simulate the groundwater situation and the flooding events and to ultimately achieve spatially distributed information on inundation height, flooding frequency and duration. Our major aim is to establish a parsimonious model that provides key processes to accurately represent the water pathways. With the minimization of the processes, we seek to limit the challenge of overparameterization (James and Burges, 1982). Our overall objective was to find a parsimonious model solution to reduce the computational run time to investigate the model’s parameter uncertainty using Latin Hypercube sampling-based uncertainty analyses. We represent the physical properties on a lower level and accept parameters in the calibration even if they are outside of the normally accepted range. Ignoring “physical” ranges of parameters is, in the face of structural uncertainty, a technique that has been relatively successful in rainfall-runoff models and is encouraged by Pappenberger and Beven (2006). Prediction and uncertainty analysis of a parsimonious floodplain surface water – ground-water interaction model 36 2.2. Methodology 2.2.1. The Catchment Model Framework (CMF) The hydrological model framework CMF (Catchment Model Framework) developed by Kraft et al. (2011), is a toolbox to construct a wide range of different model structures following the finite volume approach. Water fluxes in landscapes are represented in CMF as a network of storages and boundary conditions. Flux-governing equations are placed between the storage units to create the individual model. Models created with CMF differ in the number and connectivity of water storages and the type of equations for calculating the flux between the storage units and boundary conditions. Conceptual catchment models consist of a few storage areas connected by simplified equations such as the linear storage equation or tipping bucket approaches. Spatially differentiated models are created using numerous spatially explicit defined storages, such as soil- and groundwater layers with small area, connected by Darcian equations. Using the finite volume approach for discretizing continuous water storage, inspired by Qu and Duffy (2007), the use of an irregular grid for the lateral spatial discretization is possible. In contrast to Qu and Duffy, we used an irregular polygon grid, not a triangular one. A full description of the model toolbox can be found in Kraft, 2011, 2012. The application of the finite volume method to the partial differential equation results in a system of ordinary differential equations (ODEs) with one equation per water storage. CMF includes several ODE solvers, ranging from a primitive explicit solver without error control to implicit multistep methods with error prediction. For large, stiff systems resulting from spatially explicit discretizations, the CVODE solver from the SUNDIALS package by Hindmarsh et al. (2005) provides the best solution. The CVODE solver uses a Backward Differentiation Formula (BDF) with orders varying between 1 and 5, chosen for the stability of the solution. The application of BDF creates a non-linear equation system that must be solved with an appropriate numerical method to solve the resulting non-linear equation system. CVODE provides a choice of variants of the Newton method. We followed Qu and Duffy (2007) using a simplified Newton method suitable for large systems, the so-called Krylov-Newton iteration in the default setting of the CVODE solver, as explained in the CVODE manual (Hindmarsh and Sandu, 2016). Prediction and uncertainty analysis of a parsimonious floodplain surface water – ground-water interaction model 37 2.2.2. Using CMF for a spatial explicit groundwater model For this study, CMF was used to set up a single layer groundwater flow model. The study area is discretized into irregular polygons according to elevation and land use. Saturated lateral flow is calculated with a Darcian approach. The local water balance of the subsurface storage of a single cell is given as where i indicates the current cell, V is the water volume stored in the subsurface in m³, I(t) is the infiltration rate at time t in m³/day, h is the groundwater head as a function of the stored volume V, ETact(t,h) is the actual evapotranspiration from the cell in m³/day, Ni is the number of cells sharing a boundary with i, j indicates a cell adjacent to i, Ai,j(Vi,Vj) is the saturated area between the cells, calculated from the water levels and the boundary length in m², Ki,j is the lateral saturated conductivity between the cells i and j in m/day, and di,j is the distance between the cells i and j in m. The infiltration rate depends on the presence of surface water, throughfall and infiltration capacity. If surface water is present, the infiltration rate equals the infiltration capacity until the subsurface storage is saturated. Without surface water, the infiltration equals the throughfall from the canopy, calculated with the interception model from Rutter and Morton (1977). The potential evapotranspiration rate is calculated with the F