Nation-wide estimation of groundwater redox conditions and nitrate concentrations through machine learning

Abstract

The protection of water resources and development of mitigation strategies require large-scale information on water pollution such as nitrate. Machine learning techniques like random forest (RF) have proven their worth for estimating groundwater quality based on spatial environmental predictors. We investigate the potential of RF and quantile random forest (QRF) to estimate redox conditions and nitrate concentration in groundwater (1 km × 1 km resolution) using the European Water Framework Directive groundwater monitoring network as well as spatial environmental information available throughout Germany. The RF model for nitrate achieves a good predictive performance with an R 2 of 0.52. Dominant predictors are the redox conditions in the groundwater body, hydrogeological units and the percentage of arable land. An uncertainty assessment using QRF shows rather large uncertainties with a mean prediction interval (MPI) of 53.0 mg l −1 . This study represents the first nation-wide data-driven assessment of the spatial distribution of groundwater nitrate concentrations for Germany.

Spatial Distribution of Integrated Nitrate Reduction across the Unsaturated Zone and the Groundwater Body in Germany

“…The remaining share represents the given groundwater recharge (Q gw ) [38] and the estimated NO 3 -N input load transported across the unsaturated zone to the groundwater by Q gw (NO 3 -N_load input_gw ). For the groundwater body, a groundwater nitrate concentration map (c(NO 3 ) gw ) is taken from a previous study [26]. Given the groundwater recharge Q gw , the NO 3 -N load in the groundwater (NO 3 -N_load gw ) can be calculated.…”

“…The spatial distribution of the five land-use types was given by the land cover model [39] and their area proportions were assigned to the grid cells. An area-weighted and land-use-specific hydrospheric N surplus was calculated according to Knoll et al [26] and is considered to be equivalent to the NO 3 -N input load into the unsaturated zone (NO 3 −N_load input_uz ):…”

“…In a study by Knoll et al [26], a 1 km × 1 km grid map of the groundwater nitrate concentrations in Germany was estimated using a data-driven approach based on the measured concentrations from the period 2009 to 2018. Applying the "random forest" machine learning technique, the groundwater nitrate concentration map was modelled using several spatial predictors, such as the land use, hydrogeology and redox conditions throughout Germany.…”

“…Such an approach is appropriate because data for the validation of the reduction processes in the unsaturated zone are very limited. Since both the initial value (NO 3 -N input load) and the resulting value (groundwater NO 3 -N load) are available from previous studies [26,36], all possible natural degradation processes are taken into account when determining the nitrate reduction. The integrated nitrate reduction, the reduction of the NO 3 -N loads, is quantified as the ratio of the groundwater NO 3 -N load to the NO 3 -N input load to the groundwater:…”

“…Both processes have been identified in several studies on nitrate degradation in groundwater [21][22][23][24]. Other studies have found that the nitrate concentration in groundwater strongly depends on the hydrogeological conditions and the prevailing redox conditions [25,26]. Due to the complexity of the environmental conditions that favour degradation processes and the difficulty of measuring these in the field, models are useful tools to estimate the magnitude of nitrate reduction at different scales [27].…”