Uncertainty assessment of nitrate reduction in heterogeneous aquifers under uncertain redox conditions

Abstract

Spatially distributed nitrate reactivity was estimated for the alluvial aquifer system within the Ruamāhanga catchment in New Zealand, according to the groundwater redox status, integrating machine learning and physically based modelling approaches. Redox classification was carried out for sampled groundwater, and linear discriminant analysis (LDA) was used to spatially discriminate between three redox classes across the catchment, using mappable physical parameters as predictive variables. The LDA model predictive uncertainty was used to quantify the spatially distributed geochemical uncertainty of the denitrification potential. Nitrate reduction was simulated as first order irreversible reaction using MT3DMS. Using a Monte Carlo approach, where random redox status realizations were aggregated to various model spatial discretization scales and each transport model realization was re-calibrated, we assessed the relationship between lower order nitrate reduction parameter statistics and aggregation scale. Our results indicate that both the average nitrate reduction parameter values and their standard deviations increase with increased spatial scale. This suggests that the parameters used to model denitrification as first-order reduction in geochemically heterogeneous environments, depend on the geochemical heterogeneity scale. This can have implications when knowledge gained at local scales needs to be applied for basin-scale assessment of effects. Similar were our findings with regards to the parameter-scale dependency on the model predictive uncertainty. Even though the average nitrate reduction across the model domain did not vary with redox scale, the standard deviation around the average almost doubled between the 250 and 5000 m scales.