Technological advances have opened the possibility to have Earth observation imagery combining both high spatial resolution with high revisit frequency. The monitoring of land surface processes will greatly benefit from this type of data if biophysical variables such as Leaf Area Index (LAI) can be retrieved from it. However, the large amounts of data that will become available call for methods satisfying constraint of (i) portability across different landscapes, (ii) independence from field measurements and (iii) computational efficiency. Currently, the most suitable technique to approach this objective is neural network inversion of radiative transfer. This paper quantifies the performance of such approach for a specific crop (winter wheat) all along the growing season. By re-defining the retrieved biophysical variable as Green Area Index (GAI) and correcting the typical saturation effect that affects the estimation of higher values, this study shows that physically-based retrieval can yield similar results to those obtained by empirical methods such as stepwise linear regression (with an relative RMSE of 24.7% and 22.8% respectively). The research further shows how a trained neural network for a specific sensor can produce a GAI map from imagery acquired from different viewing conditions without a significant loss in accuracy.