Excess nitrate leaching from agricultural areas, resulting in contamination of groundwater resources, is a global phenomenon. This is also the case in New Zealand, where recent increases in land use intensity associated with farming, have resulted in a substantial increase of nitrate levels in groundwater and fresh water systems. A key removal process for nitrate in the subsurface is denitrification, which is a microbial process, occurring under low oxygen conditions in the presence of a suitable electron donor.
Due to physical and chemical heterogeneity nitrate reduction may vary significantly, even within small catchments. In this work we integrate a machine learning approach within a physically based groundwater flow and nitrate transport model to incorporate geochemical uncertainty in the form of uncertain and spatially variable redox conditions. The aim of this work is to understand and quantify the uncertainty of model parameters used to numerically parameterize nitrate reduction in geochemically heterogeneous aquifers, and in particular to i) quantify the uncertainty related to the spatial delineation of uncertain and spatially variable groundwater redox status in an alluvial aquifer, ii) assess the uncertainty on the estimated nitrate reduction potential due to uncertainty in redox conditions, iii) assess the uncertainty of denitrification model parameters and how the uncertainty changes with increasing scale, and iv) analyze how parameter uncertainties affect model prediction uncertainties at various scales.