Multivariate methods with feature wavebands selection and stratified calibration for soil organic carbon content prediction by Vis-NIR spectroscopy

Visible-near-infrared (Vis-NIR) spectroscopy is increasingly used to predict soil organic carbon (SOC) content. However, the prediction accuracy of this technology is dependent on model selection and study scale. This study explored the roles of spectral variable selection and stratified calibration based on soil type in Vis-NIR spectroscopy for predicting SOC content at a provincial scale. A total of 490 samples, collected in Jiangxi Province (southeast China), were used for modeling with partial least squares regression, support vector machine, random forests, and back-propagation neural network (BPNN). The feature wavebands of soil samples were selected by competitive adaptive reweighted sampling (CARS), and a stratified calibration was conducted based on soil type. The results showed that CARS-based models outperformed models with full wavebands in predicting the SOC content. The CARS-BPNN model combined with stratified calibration showed the best prediction performance for total soils (validation set R-2 = .82, which was .21 higher than that of BPNN based on global calibration). This study established an accurate method to predict SOC content from provincial-scale spectral data using the CARS-BPNN model coupled with stratified calibration based on soil type.