The generation of continuous land surface products from remote sensing observations is key to many environment applications. Nevertheless satellite products are often affected by poor quality or missing data due to instrumentation problems, cloud cover or snow cover. An innovative multi-sensor fusion approach for improving spatio-temporal continuity, consistency, temporal smoothness and accuracy of satellite products is here presented. The approach is based on the use of neural networks and temporal fitting techniques. It is applicable to any optical sensor and satellite product. In this study, the potential of this technique was demonstrated for restoring missing and low quality data in the MODIS leaf area index (LAI) product over the BELMANIP sites in the 2001-2003 period. MODIS and VEGETATION reflectance data were used. The developed FUSION LAI product showed an overall good agreement with the MODIS LAI standard product but resulted more continuous (reduction of 90% of the invalid LAI values), stable (smoother temporal evolution with a reduction of artefacts), consistent (improved monitoring of vegetation dynamics) and accurate (better agreement with ground measurements). The proposed approach thus provides products more appropriate for environmental prediction.