Abstract:
Objective The survey and assessment of soil organic carbon (SOC) on large scale require a large number of soil samples while conventional schemes of analysis are expensive and time-consuming. The spectra of visible near-infrared (VNIR) and mid-infrared (MIR) or data fusion of both can be used to predict SOC concentrations rapidly. However, there is still a lack of investigation on soil core samples with a data fusion approach.
Method In this study, a total of 3755 soil samples from 677 soil cores with VNIR and MIR spectra and SOC concentrations were taken from the global soil spectral library. The spectra were pre-processed using Savitzky-Golay smoothing and first-order derivation, and divided into calibration and validation datasets by Kennard-Stone algorithm. The partial least-squares regression and random forest were used to develop an estimation of SOC models, respectively.
Result We found that MIR showed better performance than VNIR, however, the accuracy with VNMIR was lower than that of MIR but better than VNIR.
Conclusion We therefore conclude that data fusion using VNIR and MIR might not be more accurate than single spectra on the prediction of SOC concentrations from soil cores. The MIR spectrum is a better option for estimation of SOC concentrations on a large scale rapidly and accurately.
下载: