Using an Extreme Gradient Boosting Learner for Mapping Hydrogeochemical Parameters in Germany

Information on the spatial distribution of hydrogeochemical parameters is crucial for decision making. Machine learning based methods for the mapping of hydrogeochemical parameter concentrations have been already studied for many years to evolve from deterministic and geostatistical interpolation methods. However, the reflection of all relevant processes that the target variable depends on is often difficult to achieve, because of the mostly insufficient determination and/or availability of features. This is especially true if you limit yourself to freely accessible data.

In this study, we apply an extreme gradient boosting learner (XGB) to map major ion concentrations across Germany. The training data consist of water samples from approximately 50K observation wells across Germany and a wide range of environmental data as predictors. The water samples were collected between the 1950s and 2005 at anthropogenically undisturbed locations.

The environmental data includes hydrogeological units and parameters, soil type, lithology, digital elevation model (DEM) and DEM derived parameters etc. The values of these features at the respective water sample location were extracted on the basis of a polygon, approximately representing the area that has an impact on the target variable (ion concentration). For a comparison, different polygon shapes are used.

The model was set up as chained multioutput regression, meaning that the prediction of the previous model in a linear sequence of single-output models is used as input for the subsequent model.

The results are planned to serve for a comparison with state-of-the-art deep learning architectures.