Numerous studies on the prediction of soil properties from visible and near-infrared spectroscopy, based on large libraries at county scale or small size soil libraries at local scales have been reported in literature. However, difficulties appear when large libraries are used to estimate the soil properties of a small area. The aim of this paper was to compare various strategies to predict soil properties of local samples using a French national database. Models were built: i) from the national database alone and ii) from the national database spiked with subsets of the local database. Two regression methods were tested: partial least square [PLS] and a local regression method (fast Fourier transform local weighted [FFT-LW]). No general rule was obtained in this study as the best strategy differed according to the property under study. It seems that when strong spectral features are related to the characteristic under study (as for CaCO3 content), the addition of local samples did not bring a decisive advantage over calibration based on a wide national database. There are three important and encouraging points of this work to emphasize: i) the evidence of the added value brought by the national library for the prediction of some soil properties over a local area, ii) the pertinence of spiking with local samples to the global database to reach accurate predictions, and iii) the interest of the FFT-LW non-linear method. As we examined only one local site with peculiar land-use and geologic characteristics, further researches are needed to elucidate the way in which these results depend on intrinsic properties of the local site samples and on the relationship between spectral features and considered soil properties.