The objectives of this study were to (i) evaluate the ability of Vis-NIR spectroscopy to predict soil organic carbon (SOC) and CaCO3 content in the heterogeneous agricultural soils from Dalmatia, Croatia and (ii) to compare the performance of two multivariate calibration techniques: partial least square regression (PLSR) and support vector machine regression (SVMR). The reflectance spectra of a total of 250 top-soils (0-25 cm) samples were collected in the laboratory using a portable Terra Spec 4 Hi-Res Mineral Spectrometer with a wavelength range 350-2500 nm. The coefficient of determination (R2), the residual prediction deviation (RPD) and the root mean square error (RMSE) were used for the model evaluation. The CaCO3 prediction derived by PLSR and SVMR with R2 (0.86 and 0.88) and RPD (2.67 and 2.82), respectively are considered good prediction models. The SOC prediction with SVMR (R2 0.84 and RPD 2.43) indicates good prediction and approximate quantitative prediction with PLSR with R2 of 0.78 and RPD of 1.94. Our results showed that (i) CaCO3 and SOC estimations were obtained with acceptable accuracy using Vis-NIR spectroscopy, (ii) the SVMR method produced more accurate estimations of selected soil properties compared to LSR, and (iii) Vis-NIR spectroscopy, in combination with SWMR can be recommended as a rapid and inexpensive method for screening of the CaCO3 and SOC content.
Aleksandra Bensa, University of Zagreb, Faculty of Agriculture, Soil Science Department, Svetošimunska 25, 10000 Zagreb, Croatia, firstname.lastname@example.org