Pesticide transfers in agricultural catchments are responsible for diffuse but major risks to water quality. Landscape elements such as hedges, ditches or vegetative filter zones are of great influence as they act as discontinuities for water and pollutant flows. If they are integrated into a spatialized pesticide transfer model, we can assess how landscape organization structure impacts water quality. The PESHMELBA model (Pesticide and hydrology: modeling at the catchment scale) [Rouzies et al., 2019] simulates pesticide transfers and fate in agricultural catchments and explicitly couples different types of landscape elements (buffer zones, ditches, etc.). Its structure is modular which makes it suitable for exploring landscape management scenarios. Comparing these scenarios can be of interest for stakeholders in order to identify an optimal landscape organization that minimizes the impact of pesticides on water bodies. However, such use of PESHMELBA in an operational context cannot be considered without first quantifying and reducing the model uncertainties. In this study, we aim at performing ensemble data assimilation in PESHMELBA for reducing uncertainties on transfer simulation on a vineyard catchment located near Lyon (France). We also aim at estimating input parameters that would be set for the landscape management scenarios exploration. To meet both objectives, joint data assimilation abilities are investigated. Data to be assimilated are surface moisture images obtained from Sentinel-2. Assimilating satellite images brings additional challenges because observations for all pixels come from the same sensor. Observation errors cannot be considered as independent anymore and DA performances can be strongly affected if error correlations are not taken into account. Based on recent work by Chabot, Nodet, Papadakis and Vidard [2015] and Chabot, Nodet and Vidard [2020], Fourier and wavelet transformations are investigated to accurately describe such covariance matrix when performing the analysis step. This work summarizes our methodological findings on a simplified catchment composed of a reduced number of landscape elements. Synthetic images are produced and different filtering and smoothing approaches are compared on a twin experiment design. Preliminary results allowed for identifying the most appropriate DA for both variable correction and parameter estimation and can be extended to realistic case studies at a larger scale. Additionally, we show that image transformations are promising tools for taking into account observation error correlations. Considering the complexity of the PESHMELBA model, and of pesticide transfers in general with highly non-linear processes, data assimilation is a relevant way to quantify uncertainty which is crucial to develop decision-support tools.