Spatial prediction of pesticide occurrence in riparian buffer zones using machine learning

Riparian buffer zones can reduce pesticide transport from agricultural land to streams but may themselves become contaminated and expose non-target organisms to pesticides. Despite their ecological importance, pesticide occurrence in riparian vegetation and soil remains difficult to predict. Here, we developed a machine learning framework to predict the occurrence of individual pesticides in riparian buffer zones using physicochemical properties, German national pesticide sales data, spatially explicit land use, and topographic predictors. We trained the model on presence-absence data for 93 pesticides measured in vegetation and soil along transects between agricultural fields and streams. The model performance was high for both vegetation and soil. Across 20 cross-validation runs, overall accuracy averaged 0.94 ± 0.01 for vegetation and 0.93 ± 0.01 for soil, with macro average F1-scores of 0.85 ± 0.02 and 0.76 ± 0.02, respectively. Predictor importance analyses showed that national pesticide sales, followed by physicochemical properties, contributed most strongly to model predictions at the individual-predictor level. At the grouped level, physicochemical properties contributed most strongly to model performance in both vegetation and soil. Including adjacent field samples improved soil predictions, whereas stream samples did not improve soil or vegetation predictions. Our results show that pesticide occurrence in riparian buffer zones can be predicted from a combination of usage intensity, compound-specific environmental behavior, and landscape context. Therefore, this approach may support the targeted monitoring and management of non-target habitats in agricultural landscapes.