Comparing different multivariate calibration methods for the determination of soil organic carbon pools with visible to near infrared spectroscopy

A successful determination of spectrally active soil components with visible and near infrared reflectance spectroscopy (VIS-NIRS, 400-2500 nm) depends on the selection of an adequate multivariate calibration technique. In this study, the contents of thermolabile organic carbon (C_375 °C), the inert organic C fraction (C_inert) and the sum of both (total soil organic carbon, OC_tot) were estimated with three different methods: partial least squares regression (PLSR) as common standard tool, a combination of PLSR with a genetic algorithm (GA-PLSR) for spectral feature selection, and support vector machine regression (SVMR) with non-linear fitting capacities. The objective was to explore whether these methods show differences concerning their ability to predict soil organic carbon pools from VIS-NIR data. For this analysis, we used both measured spectra and also spectra successively blurred with uniformly distributed white noise. Soil sampling was performed in a floodplain (grassland plots) near Osnabrück (Germany) and comprised a total of 149 samples (109 calibration samples, 40 validation samples); spectral readings were taken in the laboratory with a fibre-optics ASD FieldSpec II Pro FR spectroradiometer.In the external validation, differences between the calibration methods were rather small, none of the applied techniques emerged to be the fittest with superior prediction accuracies. For C_375 °C and OC_tot, all approaches provided reliable estimates with r² (coefficient of determination) greater than 0.85 and RPD values (defined as ratio of standard deviation of measurements to standard error of prediction) greater than 2.5. For C_inert, accuracies dropped to r² < 0.50 and RPD < 1.5; after the removal of two extreme values (n = 38) results improved at best (GA-PLSR) to r² = 0.80 and RPD = 1.98. The noise experiment revealed different responses of the studied approaches. For PLSR and GA-PLSR, increasing spectral noise resulted in successively reduced r² and RPD values. By contrast, SVMR kept high coefficients of determination even at high levels of noise, but increasing noise caused severely biased estimates, so that regression models were less accurate than those of PLSR and GA-PLSR.