Agricultural monitoring and yield forecasting can greatly benefit from coupling crop growth models with remote sensing information through data assimilation. Remote sensing observations can be used to estimate biophysical variables, such as the Green Area Index (GAI), which is a key variable in the photosynthetic processus of the canopy. For crop growth monitoring, high observation frequency is mandatory, especially when anomalies due to climatic variability must be detected. Wide geographic coverage is a further requisite to monitor specific crops at regional/continental scales. Nowadays, instruments that can satisfy these requirements have coarse spatial resolutions like MODIS (>250m). Although technological improvements are bound to provide finer spatial resolution data with more frequent revisit times, coarse instruments will retain a valuable interest since they provide a long time record. In many areas of the world, a 250m spatial resolution is of the same order of magnitude as the size of the crop fields. To make use of MODIS data for crop growth monitoring, it is required to select pixels whose observational footprint falls within the target crop specific fields. The task is compounded by the complexity of the MODIS observation support: (i) it is twice as large across track than along track; (ii) it varies along the swath within a single image; and (iii) it varies from image to image. The approach proposed here involves modelling the point spread function (PSF) and taking into account the observation geometry to assure the adequacy between the pixel footprint and the targeted crop. A neural network retrieval method is then used to derive GAI from MODIS reflectance time series. By comparing the results with GAI retrieved from concurrent fine spatial resolution time series, the quality of the coarse spatial resolution estimations can be quantified with respect to MODIS observational parameters such as the view zenith angle (VZA) and the observation coverage (OBSCOV). Undetected clouds or residual atmospheric effects can further reduce the quality of the GAI estimations. These outliers are removed by exploiting the expected temporal behaviour of the GAI profile. A semi-empirical mechanistic model, relating GAI to thermal time by way of a simple mathematical relationship, is adjusted on the punctual GAI estimations using a robust fitting method. This entire methodology is demonstrated on a study site in Fundulea (Romania) which, thanks to the ADAM ( Assimilation de Données par Agro-Modélisation) project, provides an extensive dataset of 20m SPOT images well distributed along the 2001 winter wheat growing season.