Visible and near-infrared (VIS-NIR) spectroscopy is extensively used in the field of soil science to predict several soil properties, mostly in laboratory conditions. When measured in situ, contact probes are used, and, very often, time-consuming methods are applied to generate better spectra. Unfortunately, spectra obtained by these methods differ greatly from spectra remotely acquired. This study tried to address this issue by measuring reflectance spectra directly with a fibre optic or a 4° lens on bare untouched soils. C, N content and soil texture (sand, silt, and clay) prediction models were established using partial least-square (PLS) and support vector machine (SVM) regression. With spectral pre-processing, some satisfactory models were obtained, i.e., for C content (R2 = 0.57; RMSE = 0.09%) and for N content (R2 = 0.53; RMSE = 0.02%). Some models were improved when using moisture and temperature as auxiliary data for the modelling. Maps of C, N and clay content generated with laboratory and predicted values were presented. Based on this study, VIS-NIR spectra acquired with bare fibre optic and/or a 4° lens could be used to build prediction models in order to obtain basic preliminary information on soil composition at the field scale. The predicting maps seem suitable for a fast but rough field screening.
Keywords: PLS; SVM; field measurements; near-infrared spectroscopy; soil mapping; soil properties.