Forecasting sunflower grain yield a few weeks before crop harvesting is of strategic interest for cooperatives that collect and store grains. With such information, they can optimize their logistics and thus reduce the financial and environmental costs of grain storage. To provide these predictions, data assimilation approaches involving the crop model SUNFLO are used. The methods are based on the re-estimation of soil conditions and on the sequential update of crop model states using an ensemble Kalman filter. They combine the simulation of the crop model and time series of leaf area index (LAI) derived from remote sensors and extracted over 281 fields near Toulouse, France. A sensitivity analysis is used to identify the most relevant model inputs to consider into the data assimilation process. Results show that data assimilation leads to statistically significant better predictions than the simulation alone (from an RMSE of 9.88 q.ha(-1) to an RMSE 7.49 q.ha(-1)). Significant improvement is achieved by relying on smoothed LAI rather than raw LAI. Nevertheless, there is still an over estimation of the grain yield that can be partially explained by the limiting factors observed on the fields and the forecast yield still need improvements to meet the required applications' accuracy.