Regional predictions of soil organic carbon content from spectral reflectance measurements

Diffuse reflectance spectroscopy is used to overcome the limitations of conventional methods of soil analysis. The objective was to develop a regional prediction model of soil organic carbon content based on laboratory measurements of reflectance within the visible and near-infrared spectral ranges. To achieve this, principal component analysis was used in order to determine the chemical and physical variability of 64 soil samples collected from different sites in Brittany (France). This analysis allowed samples to be divided into both calibration and validation data sets with quite similar analytical properties. A partial least squares regression algorithm was then applied to model and predict the soil organic carbon content on the basis of its spectral reflectance within the visible and near-infrared domain (400–950 nm). Results revealed a high level of agreement between measured and predicted values with coefficients of determination, root mean-squared errors and relative prediction deviations of 0.91, 0.36% and 3.4 in cross-validation and of 0.83, 0.46% and 2.35 in prediction. The model proved to be valid over the range 0.90–5.20% of organic carbon content. Good predictions of the soil organic carbon content are therefore still possible by simply using a cheap spectrometer operating between 400 and 950 nm using a regional soil database which can be progressively enhanced.