Experimental and numerical techniques to predict the response of complex and ecologically sensitive fen wetlands to shifts in climate and land use changes

The Spreewald Biosphere Reserve is a floodplain landscape characterised by fen wetlands. It is situated in north-eastern Germany, about 70 km south-east of Berlin, shaped by a rather unique water distribution system. The waterway network consists of about 1,500 km of both man-made canals and ditches — circa two thirds of the network is artificial —, as well as natural rivers and streams. Located within the catchment of the middle Spree River, it lies directly downstream of the historical region of Lusatia. Since the late 1800s, the region has been shaped by lignite mining, which reached its peak in the 1980s and began to gradually wind down after 1989. The mining activity has had a profound impact in the water balance of the region, which is especially reflected on the artificially augmented discharge of the Spree River. Currently, during dry summer periods, up to 75% of the Spree River runoff originates from the groundwater extracted to dewater the mines. With the (unavoidable) decrease in the river discharge — as a consequence of the complete phase-out of the mining activity planned by 2030/2038 — superimposed with the projected impacts of climate change for the region, the Spreewald will probably face high water scarcity. Our research question centres on how to mimic such a complex and highly dynamic ecohydro(geo)logical system with limited data availability, while addressing the feasibility of developing a discharge prediction model for the Spreewald with the help of Artificial Neural Networks (ANNs). To date, no comprehensive study incorporating fully coupled surface water and groundwater modelling has been published for such an ecohydrological system. Furthermore, our proposed approach is unprecedented for the Spreewald. By opting for a machine learning-based method, we expect to overcome some of the challenges associated with mechanistic flow models, particularly in areas of such high heterogeneity, when deterministic description of the rivers system fails. To train our neural network, we first feed it with known time series datasets of groundwater and stream level, stream discharge, meteorological data, and relevant vegetation parameter, all within the same time frame and temporal resolution. Once we obtain acceptable accuracy levels, we can use it to forecast unknown discharge values from known inputs, such as precipitation and evapotranspiration. Knowing that the existing measured data are not frequent enough to develop our ANN, we measure, in a high spatial and temporal resolution and over an entire hydrological cycle, groundwater levels and stream discharge, sap-flow of wetland-specific tree species (e.g., alder), and electrical conductivity of the stream water. Additionally, for a better understanding of the surface water-groundwater exchange rates, we measure water stable isotopes in all hydrological compartments. Our main research goal is to come up with a new approach for quantifying the overlapping impacts that climate change and the re-establishment of a semi-natural runoff of the Spree River could have on the Spreewald.