Predictions of soil organic carbon using laboratory-based hyperspectral data in the northern Tianshan mountains, China.

This paper presents a study dealing with soil organic carbon (SOC) estimation of soil through the combination of soil spectroscopy and multivariate stepwise linear regression. Soil samples were collected in the three sub-regions, dominated by brown calcic soil, in the northern Tianshan Mountains, China. Spectral measurements for all soil samples were performed in a controlled laboratory environment by a portable ASD FieldSpec FR spectrometer (350-2,500 nm). Twelve types of transformations were applied to the soil reflectance to remove the noise and to linearize the correlation between reflectance and SOC content. Based on the spectral reflectance and its derivatives, hyperspectral models can be built using correlation analysis and multivariable statistical methods. The results show that the main response range of soil organic carbon is between 400 and 750 nm. Correlation analysis indicated that SOC has stronger correlation with the second derivative than with the original reflectance and other transformations data. The two models developed with laboratory spectra gave good predictions of SOC, with root mean square error (RMSE) <5.0. The use of the full visible near-infrared spectral range gave better SOC predictions than using visible separately. The multivariate stepwise linear regression of second derivate model (model A) is optimal for estimating SOC content, with a determination coefficient of 0.894 and RMSE of 0.322. The results of this research study indicated that, for the grassland regions, combining soil spectroscopy and mathematical statistical methods does favor accurate prediction of SOC.