全球变化知识资源中心

Information about soil organic carbon fractions is important in understanding the vulnerability of soil carbon to climate change and land management. Soil organic carbon can be divided into fractions that are labile and others that are more stable. All existing methods to fractionate soil organic carbon are time consuming and complex. Near-infrared reflectance spectroscopy (NIRS) is a rapid analytical technique. In this study we evaluated and optimized the use of NIRS to predict soil organic carbon fractions with the constraint that the carbon fractions (labile and stabilized) should add up to 100% (total organic carbon content). We used samples from two datasets from agricultural and forest sites in Germany (dataset A) and Europe (dataset B). Samples were fractionated by two different methods (density and physical-chemical fractionation) as reference methods. Soil samples were scanned in the NIR range and calibration models were developed using partial least squares regression. The key to improving model performance was the log-ratio transformation proposed by Aitchison for compositional data that enabled us to model the fractions as dependent variables. Traditional methods for the optimization of NIRS models, such as selection of wavelength range and pretreatment of spectra, showed no effective reduction in error. With the constraint that both fractions add to 100% (log-ratio transformation), the ratio of performance to deviation (RPD) increased from 1.6 to 2.8 for the labile C fraction and from 1.5 to 3.2 for the stabilized C fraction. Root mean square error of cross-validation of the labile fraction was 4.3 g C kg(-1) for dataset A and 2.7 gC kg(-1) for dataset B, which corresponds to an R-2 of 0.88-0.80 and RPD of 2.9-2.2, respectively. The models performed equally well for different soil textures and land-use types. With log-ratio transformation, more precise calibrations of NIRS for C fractions could be obtained.