This paper evaluates the performances of a neural network approach to estimate LAI from CYCLOPES and MODIS nadir normalized reflectance and LAI products. A data base was generated from these products over the BELMANIP sites during the 2001-2003 period. Data were aggregated at 3 km x 3 km, resampled at 1/16 days temporal frequency and filtered to reject outliers. VEGETATION and MODIS reflectances show very consistent values in the red, near infrared and short wave infrared bands. Neural networks were trained over part of this data base for each of the 6 MODIS biome classes to retrieve both MODIS and CYCLOPES LAI products. Results show very good performances of neural networks to estimate the original LAI products with an overall root mean square error (RMSE) around 0.5 for MODIS LAI from both MODIS and CYCLOPES normalized reflectances and a RMSE ranging between 0.12 (CYCLOPES reflectances) and 0.29 (MODIS reflectances) for CYCLOPES LAI. A drop of 15% of performance was found by training MODIS biome dependant algorithm by a single network over all the classes at the same time. More detailed analyses show that CYCLOPES and MODIS LAI values are very consistent for grasses and crops. Conversely, other biomes including shrubs, savanna, needleleaf and broadleaf forests show significant discrepancies, mainly due to differences between LAI definitions used between CYCLOPES (closer to effective LAI) and MODIS (closer to true LAI). However, products derived from the original CYCLOPES LAI products show a better agreement with both effective and true LAI ground measurements values. MODIS LAI products show more instability, partly because of the slightly shorter temporal resolution as compared to CYCLOPES. These results confirm the interest and versatility of neural networks for operational algorithms. This approach could be extended to other products or sensors, and may constitute a step forward for the fusion of data from several sensors, hence contributing to develop 'virtual constellations'.