Nitrate pollution of groundwater has been a well-known problem for decades, but today the debate about causes and possible solutions is very controversial. The EU Water Framework Directive has been implemented to restore the good status of water systems in Europe, which requires, among other things, a reduction of nitrate inputs to groundwater. In order to achieve this, harmonized and transparent approaches for evaluating the quality of groundwater bodies and clearly defined regulations for the implementation of mitigation measures need to be established. Important aspects to be taken into account when developing action plans are the estimation of nitrogen inputs to the subsurface, the spatial distribution of groundwater nitrate concentrations and the natural nitrate reduction capacity of the aquifer. For groundwater quality modelling on a country-wide scale, data-driven approaches, in particular machine learning techniques, have proven their worth. In this dissertation, an approach for the large-scale regionalisation of groundwater nitrate concentrations depending on spatial environmental parameters is developed. In a first step, several approaches for linking point information from monitoring sites with the spatial data from maps are investigated using the example of the federal state of Hesse, Germany. Four machine learning techniques based on different statistical model types are compared regarding their predictive performance. It can be shown that a 1,000 m circular buffer can describe the contribution area of a monitoring site in a simplified way and can be used for compiling the factors influencing the groundwater nitrate concentration. The random forest model outperforms classical multiple linear regression, simple classification and regression trees and boosted regression trees. In a second step, the approach will then be transferred to the entire federal republic of Germany. Based on the Water Framework Directive groundwater monitoring network, a random forest model is trained and the nitrate concentrations in groundwater is estimated for a 1 km x 1 km grid map. Good model predictive performance can be achieved with an R² of 0.52 where the redox conditions, the hydrogeological units and the percentage of arable land are identified as the most influential predictors for the estimation of groundwater nitrate concentration. By using quantile random forest for an uncertainty analysis, with a mean prediction interval of 53 mg NO3-/l large uncertainties are determined. Finally, the third part of this dissertation focuses on the estimation of nitrate reduction in groundwater. The estimated spatial distribution of groundwater nitrate concentrations together with data on nitrogen surplus are used in a simplified conceptual approach to estimate the integrated nitrate reduction across the unsaturated zone and the groundwater body. The determined nitrate reduction rates are on average 57% and strongly vary from no reduction up to a degradation of 100% with predominantly high reduction rates in northern Germany and lower reduction rates in the central and southern part of Germany. Nitrate reduction capacity is highly dependent on hydrogeological conditions, with reduction rates in porous aquifers and under anaerobic conditions, being significantly higher than in fractured consolidated aquifers and under aerobic conditions. With the overall approach presented here, spatial predictions can be made based on freely available geodata, which makes it an important contribution to the large-scale assessment of groundwater quality and can be used in the planning of mitigation measures.
Verknüpfung zu Publikationen oder weiteren Datensätzen
