Spatial prediction of the groundwater nitrate reduction front across Flanders with borehole lithology and geochemistry data

Nitrate contamination of water bodies driven by agricultural activities remains a widespread environmental concern. In groundwater-fed surface waters, nitrate concentrations are governed by the transport of N-excess through groundwater systems. Denitrification plays a central role in mitigating this transport by reducing nitrate to inert N₂ gas in anoxic zones. The anoxic zone in an aquifer is separated from the oxic zone by a sharp boundary, i.e., the nitrate reduction front (NRF). This NRF can be retrieved from sediment colour changes, shifting from oxidised hues (yellow, brown, or red) to reduced colours (grey, green, or black), provided there are sufficient Fe-bearing minerals and redox colours are not masked by high clay or organic matter contents. Hydrochemical data from multilevel observation wells provide a more integrated signal of redox conditions, yet precise NRF delineation remains challenging since usually, groundwater sampling is not done for consecutive depth intervals. Integrating lithological and hydrochemical information therefore offers a more robust approach.

This study presents a two-stage machine learning (ML) framework to predict groundwater NRF positions across Flanders. Firstly, a logistic regression model (ML1) was developed to estimate oxidation probabilities for individual borehole layers using lithological characteristics (colour, texture, stratigraphy) and relative depth within the aquifer as predictors. The model reproduced the redox conditions (reduced or oxidized) in 72% of borehole layers, which were assessed from the hydrochemistry (dissolved oxygen, [Fe²⁺] and redox potential) of the associated filters for groundwater monitoring.[JV2] [ES3]   The derived probabilities were used to assess the likelihood that a boundary between two borehole layers is located at the NRF leading to a likelihood-depth profile. A separate likelihood profile was developed using the hydrochemistry data from the filters of a multilevel borehole. Multiplication of these two profiles yielded the most probable NRF position, resulting in NRF estimates at 1,902 locations.

Secondly, these positions were used to train a gradient-boosted regression tree model (XGBoost) to predict the NRF depth for any location across Flanders (ML2i). The training data were the NRF depths at the 1,902 multilevel observation wells.