Enhanced redox mapping at national scale of Denmark through integration of sediment color and groundwater chemistry in a machine learning framework

Redox conditions play a crucial role in determining the fate of geogenic and anthropogenic contaminants in groundwater, impacting ecosystem services vital for both the aquatic environment and human water supply. For example, investigating the reduction of nitrate underscores the importance of data on redox conditions since denitrification takes places in anoxic environments. Specifically, knowledge of the depth to the uppermost reduced layer, i.e., first redox interface, can inform water and land management by identifying agricultural areas vulnerable or robust to nitrate leaching. Assessing redox processes is complicated by geological heterogeneities, resulting in complexities of local to regional groundwater flow paths. Geospatial machine learning techniques have previously successfully mapped redox conditions based on sediment color or water chemistry observations. This study introduces a novel approach that combines both data sources to enhance understanding of subsurface redox conditions in Denmark. In the first step, depth to the first redox interface is mapped using sediment color information from 26,800 boreholes. This depth is derived from sediment color changes, transitioning from oxic to reduced colors of quaternary sediments. The mapping utilizes a regression-based gradient boosting with decision tree algorithm trained against sediment color data and 20 covariates, encompassing information on hydrogeology, lithology, topography, and hydrology. In the second step, the depth of the first redox interface is compared against groundwater chemistry to classify continuous and discontinuous redox conditions. Continuous conditions exhibit the absence of oxic groundwater below the first redox interface, while discontinuous conditions show oxic groundwater below the interface. This classification is performed using a gradient boosting with decision tree algorithm utilizing the same 20 covariate maps and 21,800 classified groundwater samples. Both models undergo comprehensive cross-validation and feature importance analysis. The depth to the first redox interface is modeled with a mean error of 0.001 m and a root-mean-squared error of 8.3 m. The continuous/discontinuous classification attains an accuracy of 69.5 %. Both variables are mapped at a 25 m spatial resolution at the national scale of Denmark. Results indicate a mean depth to the first redox interface of 9.4 m and a standard deviation of 5.7 m, with spatial patterns largely driven by the groundwater table. 66.0% of Denmark is classified as discontinuous, indicating complex redox conditions, predominantly collocated with moraine clay. These maps contribute significantly to understanding subsurface redox processes, supporting national-scale land and water management.