Deciphering complex signals: What can science learn from environmental pesticide monitoring?

Environmental and water resources agencies run comprehensive monitoring programs of pesticides and their transformation products in surface and groundwater systems to determine water pollution and to comply with the reporting obligations of the EU. Besides, institutions carry out additional monitoring programs serving different purposes, including a review of the registration process, finding evidence for risks that have been underrated so far, determining application errors or improper handling, and deriving recommendations for agricultural and water resources management. However, these programs are often considered to be of marginal value for research. Among others, usually pesticide application and management data are scarce, the input of other sources like, e.g., deposition of trifluoroacetate (TFA), is unknown, soil properties exhibit enormous but purely known spatial heterogeneity, and knowledge about transformation pathways of the active ingredients is limited.

Thus, there is urgent need for developing a blueprint for the analysis of such monitoring data that makes maximum use of the information but avoiding pitfalls of unjustified basis assumptions. We present an approach based on the analysis of a 4.5 years monitoring program with monthly sampling in twenty shallow groundwater wells in the Federal State of Schleswig-Holstein in North Germany. Agricultural management data were available for part of the capture zones of some wells but were not complete. Thus, a forward modelling was not possible. Instead, in a first step we aimed at assessing the effects of vadose zone and aquifer properties, filter screen depth, and weather conditions on the observed spatial and temporal patterns of the concentration of pesticide and transformation products (TP). Canonical correlation analysis of time courses of solute concentration and of groundwater head at the twenty groundwater wells revealed very close resemblance between both. In fact, groundwater head dynamics proved to be a very powerful predictor of pesticide and TP dynamics. This provides clear evidence that most of the observed dynamics reflects transient immobilisation and later remobilisation in the vadose zone rather than direct effects of pesticide application.

In a next step, support vector machine models were set up separately for various substances. They explained more than 90% of the total variance for most substances. There were some cases of characteristic deviation between the observed and simulated concentration that could be ascribed to recent applications of the respective active ingredients. In most cases, however, there was clear evidence for a long-term stock of substances being occasionally flushed to the groundwater during short episodes. In regard to TFA our analysis revealed strong indications for a major and increasing contribution of deposition from non-agricultural sources in peri-urban regions. We conclude that analysis of the residuals of the support vector machine models is a powerful tool for making efficient use of monitoring data, even in face of incomplete data about the boundary conditions.