The NOAA-AVHRR Normalised Difference Vegetation Index (NDVI) dataset is used to investigate the predictability of the vegetation cycle in an area centred on the Gourma region in Sahelian Mali at scales varying from 8 km(2) to 1024 km(2) over a period spanning from 1982 to 2004. The predictability of the vegetation cycle is analysed with a model based on a reconstruction approach that fully relies on the dataset. Two parameters deduced from the growth of the forecast error are considered: the horizon of effective predictability, H-E, which is the horizon at which a satisfying prediction can be effectively forecasted at a given level of error, and the level of noise. Predictability is therefore analysed with regard to the horizon of prediction and the spatial scale; the influence of the model's dimensions is also discussed. The analysis clearly indicates that the signal predictability increases, and the level of noise decreases with an expanding area. However, even though the signal is more regular, its complexity increases within the narrowing entangled trajectory, setting the level of error of any prediction at a minimum of 15%, which matches the level of noise characteristic of the AVHRR-NDVI data series. The forecasting error quickly increases with the horizon of prediction, setting the optimum horizon of predictability in the range of 2 to 4 decades, with high intra-annual variability. At the short horizon of 1 decade, a resolution of 16 km(2) is reasonable to achieve an accuracy of approximately 20%. At the longer horizon of 3 decades, only low resolutions (256 km(2) or lower) give an accuracy equal to or better than 35%.