The objective of this study is to provide a consistent archive of global FAPAR maps over the past 29 years. The product must be consistent with the ones developed in the recent years that take advantages from the better capabilities of recent sensors. Among these available products, we chose the CYCLOPES FAPAR. Indeed, it has been demonstrated through global LAI and FAPAR inter-comparison and validation exercises that this product provides (i) the best RMSE when compared to ground measurements and (ii) the smoothest temporal profiles, which is expected since vegetation structure variables vary continuously with time. We therefore applied the methodology suggested by Verger et al (2008) to train a neural network using NOAA16 AVHRR or VEGETATION reflectances as inputs and the CYCLOPES FAPAR as output over the 1999-2003 period. Then, to take into account the differences of spectral response of the different AVHRR sensors, a radiative transfer model was used to compute conversion coefficients between a given AVHRR spectral band to the corresponding NOAA16. The algorithm was applied to the 1981-2008 period over Europe. Results showed good performances, especially for non forest biomes when compared to ground measurements. Temporal profiles derived for the whole period over the European BELMANIP2 sites (ensemble of 3kmx3km homogeneous areas representative of the biome distribution over the Earth), showed a very good temporal consistency without observing accident (sudden increase or decrease) when transiting from one AVHRR sensor to another.