Assessing the impact of the spatial arrangement of agricultural practices on pesticide runoff in small catchments: Combining hydrological modeling and supervised learning

Pesticide runoff is an important contributor to surface-water contamination. Transfer conditions at the field scale are well known, but our understanding is still quite limited in assessing water contamination by pesticides at the catchment scale. In this study, we focused on the spatial arrangement of agricultural practices in small catchments (<1 km2) and its impact on surface-water quality with respect to pesticide losses. An innovative method, the conjugation of mechanistic modeling and supervised-learning, was used to simplify this complex situation of pesticide transfer at the catchment scale.At the catchment level, the connectivity between each field and the hydraulic network is a major factor in determining the transfer of pesticides from fields to the outlet. We propose a connectivity indicator that assesses the contribution of each field to the hydraulic network runoff flow, which can be interpreted as the impact of each field on the catchment water quality. The connectivity indicator was inferred from the results of a mechanistic model simulation. First, we constructed a database from simulations using a proven hydrological model (MHYDAS) with defined scenarios. Then, two-thirds of the data were processed by the REPtree method to yield a regression tree. The rules obtained by this tree formed the basis for the connectivity indicator. The tree was then validated using the last third of the simulated data and the connectivity indicator was applied to a small wine-growing catchment in Roujan, France. This indicator required several types of data for its calculation: a spatial mapping of the studied watershed (topography, field characteristics, and network distribution) and an estimation of the spatial distribution of saturated hydraulic conductivity. The indicator developed here can be used to carry out risk assessments for diffuse pollution in diagnostics as well as test-mitigation strategies. The assessment results can be represented as a risk map to facilitate their interpretation.