Regionalisation methods for the designation of areas with groundwater nitrate pollution in Germany

The designation of nitrate-polluted areas for groundwater protection based on national directives implementing the EU Nitrates Directive (91/676/EEC) in Germany requires the use of geostatistical or deterministic regionalization methods. The objective of this study was to assess the applicability and propose a suitable methodology for a possible national-scale area designation based on an in-depth problem analysis and empirical as well as theoretical model assessments, which identified shortcomings, uncertainties and possible biases in the methods used until now.

Suitable methods must not only be able to identify exceedance regions – as opposed to simply regionalizing nitrate concentration; they also need to take into account spatial heterogeneity and adequately represent distributional characteristics across a variety of hydrogeological settings. This requires the incorporation of ancillary information in the form of quantitative and categorical spatial predictors representing hydrogeological and general environmental conditions, but not emissions estimates in the present regulatory setting.

Kriging with external drift, geostatistical regression-kriging methods and conditional geostatistical simulations offer an established methodological toolbox that fulfils these requirements and enables transparent decision-making. These models consist of linear, potentially nonlinear spatial trend components and geostatistical interpolation components, which can be further differentiated based on hydrogeological regions to account for heterogeneity. An unbiased estimate of the total exceedance area with nitrate levels >50 mg/l can be obtained from these models and accounted for in the decision-making process. An empirical comparison highlights possible biases in the size of exceedance areas obtained with traditional approaches that ignore local prediction uncertainty and focus on spatial prediction of nitrate concentration. Potentials and challenges of combining geostatistical techniques with nonlinear machine-learning models in a regulatory context are discussed.