Assimilation of remote sensing data into crop models is generally applied pixel by pixel, while the whole image data is available. Estimating the crop model parameters of a pixel independently from the neighboring ones generates at least four issues: (1) the method does not take into account the possible spatial structures although there are not necessarily existing or easy to quantify (2) the spatial properties of images, and even more of time series of images constitute an additional source of information to use, which is neglected in this case (3) the inverse problem is generally ill-posed when applied at the pixel level and (4) repeating the same action a great number of times is not computer efficient and provides sub-optimal solutions. When inverting models of significant size, computational cost becomes a clear limitation. Some constraints are sometimes added to ensure better spatial consistency through the regularization of the behavior of a pixel by that the neighboring ones. However, this ignores the spatial dependencies of the parameters to be estimated. We propose here to exploit some spatial structures of the parameters to reduce the size of the problem and make its inversion manageable. It is applied to process concurrently a set of pixels using variational assimilation using the adjoint model. This method was applied on a simple model of plant growth, the BONSAI model, which simulates LAI (Leaf Area Index) as a function of 6 parameters. The method proposed here assumes that the parameters are governed by spatial structures depending on several levels: the cultivar, the field, and the pixel, while some of them are assumed to be stable over the whole image. For example, cultivar parameters govern phenological stages. At a lower level, some parameters depend on the field, such as agricultural practices. Other parameters depend on the pixel level, such as soil parameters. Finally, to improve the robustness of the method and reduce the space of realization, the parameters to which the model is not sensitive were considered fixed to a default value on all plots and all cultivars. The constrained variational method has been tested on twin experiments (with virtual data) and on actual observations from the ADAM experiement and evaluated on (1) the quality of the estimation of the model input parameters, (2) the quality of LAI simulation and (3) its sensitivity on the frequency of observations, i.e. satellite revisit time. If it is not always relevant to assert that space constraints allow a better reproduction of the trajectory of LAI as compared to a conventional method when a lot of observations are available, the method allows obtaining an equally satisfactory result for a much lower computational cost. However, the spatial constraints allow more stable and robust results when the frequency of observations is relaxed as compared to the assimilation by pixel. In addition, this new method requires the minimization of a single cost function easier to control when there are divergence or local minimum. Finally, estimate of the model input parameters, which is generally the main objective of data assimilation, is much less dependent on the number of observations assimilated by using these spatial constraints.