engleski

COMPARISION OF MULTIVARIATE CALIBRATION METHODS FOR PREDICTING SOIL ORGANIC CARBON WITH VIS-NIR SPECTROSCOPY

This study aimed to compare four different regression methods, including partial least squares regression (PLSR), principal component regression (PCR), boosted regression trees (BRT) and support vector machines regression (SVMR) for predicting organic carbon content in soil samples with visible-infrared (Vis-NIR) diffuse reflectance spectroscopy (DRS). The reflectance spectra of a total of 238 topsoils (0-25 cm) heterogeneous samples were measured in a laboratory using a portable Terra Spec 4 Hi-Res Mineral Spectrometer with a wavelength range of 350-2500 nm. The analysed soils belong to various soil types used in agricultural production with SOC content ranging from 0.29 to 37.50 g C kg-1 . The estimation accuracy was determined with the coefficient of determination (R 2 ), the ratio of performance to deviation (RPD) and the root mean square error of prediction (RMSEP). The results of this study showed that the SVMR and the PLSR models obtained good predictions for soil organic carbon with R2 values of 0.87 and 0.86, RMSEP 2.48 and 2.58 g C kg-1 and RPD values of 2.5 and 2.4, respectively. The PCR and BRT showed a noticeable decrease in prediction accuracy with R2 values of 0.77 and 0.72, RMSEP 3.49 and 3.71 g C kg-1 and RPD values of 1.6 and 1.5, respectively indicating fair predictions. Our study confirmed the potential of VisNIR spectroscopy for estimation of SOC content and showed that the estimation accuracy of the analysed calibration methods decreased in the following order: SVMR>PLSR>PCR>BRT.

BRT, Chemometrics, PCR, PLSR, SVMR

nije evidentirano