First steps towards a data-driven groundwater vulnerability index for pesticides in Germany using probabilistic neural networks

The world largely relies on groundwater extraction for drinking water supply, which is also the case in Germany. In the EU, the Water Framework directive regulates the standards for a chemically good state of water bodies. Thresholds are often exceeded due to fertilizers and pesticides. Methods to assess groundwater vulnerability to contamination to chemical compounds are mainly index-based, GIS-overlay tools. Other approaches are process-based leaching models and statistical approaches. Commonly used index methods remain conceptual in nature and lack validation with monitoring data. Process-oriented approaches tend to focus on the soil layer. Statistical approaches remain underexplored. In the project FARM (Groundwater vulnerability assessment during authorisation procedure of pesticides), we aim to improve groundwater protection by developing a data-based vulnerability index that exploits the existing extensive data bases and covers all relevant environmental conditions and agricultural inputs.

The federal groundwater quality monitoring infrastructure and of water suppliers delivered data of about 26.000 sites which are sampled for about 500 different pesticides. For pesticide monitoring data in Germany, such an exhaustive national database is unprecedented. Given this vast data set, this project aims to apply a fully data-driven approach to identify previously unknown, relevant factors and their interactions that drive groundwater vulnerability to pesticides. We aim to investigate the complex interactions between subsurface (soil, hydrogeology) and surface (meteorology, land use, crop sequences, agricultural practices) site characteristics with the physical-chemical properties (mobility, persistence) of pesticides. The potential of this data set will be exploited by the development, testing and validation of a supervised a machine-learning (ML) model. After an initial feature selection procedure, a Bayesian convolutional neural net will be trained on groundwater quality data and the mentioned extensive catalogue of features. This set-up takes the uncertainty into account introduced by the large percentage of left-censored data (concentrations below limit of quantification of the analytical method). High interpretability of the ML-model is essential, identified factors need to be comprehensible and actionable for decision-makers. We are dealing with a highly heterogeneous, asymmetric monitoring data set and strong biases in many variables are expected. This project thus pioneers in assessing the potential and suitability, as well as limitations and pitfalls of training neural nets on the status-quo of groundwater quality monitoring in Germany. A second major outcome of the project are specific recommendations on adjustments of the national monitoring (spectra of sampled substances, sampling frequency and timings, addition or reduction of monitoring wells in specific areas).