The intensive use of pesticides is associated with significant risks to the quality of groundwater and surface water bodies. Using numerical models can be a relevant approach to identify such risks and to implement adapted mitigating strategies. The PESHMELBA model was recently developed to simulate water and pesticide transfers on small agricultural catchments. It focuses on representing different landscape structures (grass strips, ditches, hedges, etc.) and their impact on pesticide transfers. The long-term objective is, among other things, to make PESHMELBA an operational tool for watershed managers to identify an optimal landscape configuration with respect to pesticide transfer mitigation. This thesis aims to prepare such an operational use of PESHMELBA by quantifying and reducing the uncertainties associated with the variables simulated by the model.To this end, an uncertainty analysis and a sensitivity analysis are first carried out on several integrated variables and several time series produced by PESHMELBA.

Given the complexity of the model, several methods of sensitivity analysis are explored in order to identify the approach best suited to the large size of the input parameter space, the very high computational cost of a simulation and the spatialised nature of the variables. Sobol indices calculated by polynomial chaos decomposition, HSIC measures and importance measures from Random Forest are thus compared. The results provide knowledge on the functioning of the model by making the link between influential parameters and the physical processes involved. These results also make it possible to identify the input parameters for which it will be most relevant to reduce the uncertainty.In a second step, a methodology for uncertainty reduction targeting the input parameters identified as the most influential as well as several output variables (surface and subsurface moisture, concentration at the outlet) is developed. For this purpose, several ensemble data assimilation methods derived from the Kalman filter (EnKF, iEnKS and ES-MDA) are applied to the PESHMELBA model. First, satellite images of surface moisture are assimilated. The results show that in the considered scenario, such data allow an efficient correction of surface saturation moisture and water content, but that this correction does not propagate to subsurface variables and parameters. In this configuration, the influence of the ensemble size, frequency and error associated with observations is also explored to best assess the performance of the different approaches. Then, the joint assimilation of several types of observations with contrasting spatial and temporal resolutions (surface moisture images, point moisture profiles and weekly mean concentration at the outlet) is explored and the results obtained allow to establish an assimilation strategy adapted to the targeted variables and parameters.

Funding

• Salary: 100% INRAE (AQUA department + Riverly UR)
• Operational Costs: ECOPHYTO SPIRIT and LEFE-MANU MARQUISE projects

References

• A. A. Emerick and A. C. Reynolds. Computers & Geosciences, 2013. doi:10.1016/j.cageo.2012.03.011.
• K. Radišić, E. Rouzies, C. Lauvernet, and A. Vidard. Socio-Environmental Systems Modelling. doi:10.18174/sesmo.18570.
• E. Rouzies, C. Lauvernet, C. Barachet, T. Morel, F. Branger, I. Braud, and N. Carluer. Science of The Total Environment, 2019. doi:10.1016/j.scitotenv.2019.03.060.
• E. Rouzies, C. Lauvernet, B. Sudret, and A. Vidard. Geoscientific Model Development, 2023. doi:10.5194/gmd-16-3137-2023.
• E. Rouzies, C. Lauvernet, and A. Vidard. Hydrology and Earth System Sciences Discussions, 2024. doi:10.5194/hess-2024-219.

Cite the thesis

Emilie Rouzies. Quantification et réduction de l'incertitude dans un modèle de transfert de pesticides à l'échelle du bassin versant. Mathématiques [math]. Université Grenoble Alpes [2020-..], 2023. Français. ⟨NNT : 2023GRALM025⟩. ⟨tel-04659164v2⟩

Access manuscript on HAL thèses