Abstract
The frequent use of pesticides in agriculture leads to widespread contamination of various environmental compartments, especially surface waters. This contamination represents a major environmental and health risk to manage. In this context, spatially distributed mechanistic models represent valuable tools for predicting the impact of changes in agricultural practices on surface water quality at the territorial scale. However, their operational implementation to assess the risks of chronic or acute contamination over long periods and over extensive watersheds is limited, notably due to their high computational time. Thus, the general objective of this thesis is to propose an approach for spatially distributed modeling of pesticide hydrological transfers with a computation time compatible with operational use. To this end, a numerical simplification approach of an existing distributed mechanistic model was undertaken to reduce computational time while maintaining predictive capacity compatible with operational use in watersheds. The work is based on the spatially distributed model MHYDAS-Pesticide 1.0 (Crevoisier et al., 2021), and observed data from the Roujan vineyard watershed, in a Mediterranean context, as part of the OMERE Environmental Research Observatory (Molénat et al., 2018). The proposed numerical simplification approach consists of replacing the module for calculating water and solute fluxes in the soil at field scale of the spatially distributed mechanistic model with a statistical model. The hybrid mechanistic-statistical model obtained retains the spatial structure of the original model. First, the accuracy of the field-scale model of MHYDAS-Pesticide 1.0 in predicting the temporal variability of pesticide runoff contaminations at the outlet of an agricultural field was verified by comparison with 19 years of observations. Subsequently, a statistical meta-model, i.e., a simplified mathematical representation, of the field-scale model of MHYDAS Pesticide 1.0 was developed by statistical learning of a Long Short-Term Memory (Hochreiter and Schmidhuber, 1997) model. Then, a numerical simplification of MHYDAS-Pesticide 1.0 was assembled by coupling the field-scale statistical model with the initial distributed hydrological structure of MHYDAS Pesticide 1.0 to simulate the temporal dynamics of surface water fluxes and pesticide concentrations at the watershed scale. An application of the hybrid mechanistic-statistical model was conducted to assess the risk of surface water contamination by glyphosate at the outlet of a watershed based on the spatial distribution of pesticide application. The results obtained demonstrate the potential of the proposed modeling approach for the operational implementation of spatially distributed hydrological models. The hybrid mechanistic-statistical model obtained retains interesting functionalities of a distributed mechanistic hydrological model for operational use and presents reduced computation time (-76%). It can be used to evaluate both chronic and acute risks of surface water contamination by a variety of pesticide compounds depending on the spatial and temporal distribution of pesticide application in different watersheds. Furthermore, the predictive performance of the field-scale hybrid model is compatible with the analysis of acute and chronic contamination risks. Therefore, this thesis opens up perspectives for the development of operational tools for assessing the impact of agricultural practices on surface water quality at the watershed scale.