Use of VIS-NIRS for land management classification with a support vector machine and prediction of soil organic carbon and other soil properties
The objective of this research was to investigate the effects of a long-term experiment on soil spectral properties and to develop prediction models of these properties (soil organic carbon (SOC), N, pH, Hh, P2O5, K2O, Ca, Mg, K, and Na content) from texturally homogeneous samples (loamy sand). To this aim, chemometric techniques, such as partial least square (PLS) regression and support vector machine (SVM) classification, were used. Our results show that visible and near infrared spectroscopy (VIS-NIRS) is suitable for the prediction of properties of texturally homogeneous samples. The effects of fertilizer applications were sufficient to modify the soil chemical composition to a range suitable for using VIS-NIRS for calibration and modeling purposes. The best results were obtained for SOC and N content prediction using the full dataset with cross-validation (r² = 0.76, RMSECV = 0.04, RPD = 2.02 and r² = 0.81, RMSECV = 0.01, RPD = 2.20, respectively) and with an independent validation dataset (r² = 0.70, RMSEP = 0.04, RPD = 1.80 and r² = 0.73, RMSEP = 0.03, RPD = 1.22, for SOC and N content, respectively). The use of fertilizers and the type of crop rotation appear to have a significant impact on soil spectral properties; the SVM methodology with a linear kernel function was able to classify soil samples as functions of the applied doses of organic and inorganic fertilizers with 75% accuracy with cross-validation and the type of crop rotation with more than 90% accuracy with full validation of separate datasets.