Large scale prediction of groundwater nitrate concentrations from spatial data using machine learning.

Reducing nitrogen inputs, in particular nitrate, to groundwater is becoming increasingly important to fulfil requirements of the European Water Framework Directive. When developing management plans for mitigation measures at larger scales, complex hydro-biogeochemical models reach their limits due to data availability and spatial discretization. To circumvent this problem, the spatial distribution of nitrate concentration in groundwater is estimated using a parsimonious GIS-based statistical approach. Point nitrate concentrations and spatial environmental data as predictors are used to train statistical models. In order to compile the spatial predictors with the respective monitoring sites, different designs of contributing areas (buffer zones) and their effects on the performance of different statistical models are investigated. Multiple Linear Regression (MLR), Classification and Regression Trees (CART), Random Forest (RF) and Boosted Regression Trees (BRT) are compared in terms of the predictive performance of each model according to various objective functions. We determine the most influential spatial predictors used in the respective models. After training the models with a subset of the data, we then predict the spatial nitrate distribution in groundwater for the entire federal state of Hesse, Germany on a 1 × 1 km grid by only the spatial environmental data. The Random Forest model outperforms the other models (R = 0.54), relying on hydrogeological units, the percentage of arable land and the nitrogen balance as the three most influencing predictors based on a 1000 m circular contributing area. The use of exclusively spatial available predictors is a big step forward in the prediction of nitrate in groundwater on regional scale.