Quantification of SOC and Clay Content Using Visible Near-Infrared Reflectance–Mid-Infrared Reflectance Spectroscopy With Jack-Knifing Partial Least Squares Regression

Abstract

A total of 125 soil samples were collected from a Danish field varying in soil texture from sandy to loamy. Visible near-infrared reflectance (Vis-NIR) and mid-infrared reflectance (MIR) spectroscopy combined with chemometric methods were used to predict soil organic carbon (SOC) and clay contents. The main objective of this study was to find the best model for predicting SOC and clay content in the sampled field using Vis-NIR, MIR, and the combination of Vis-NIR and MIR and using different model development techniques. The secondary objectives were (i) to use iterations of calculation to find the optimal number of replicates for MIR measurements based on the root mean square error of cross validation (RMSECV) and (ii) to apply partial least squares regression in combination with jack-knifing (JK) to identify the most important part of spectral variables and the best model for predicting SOC and clay content. The study showed that with repeated MIR measurements it was possible to improve RMSECV by 20%. The optimal number of repeated MIR measurements was between 3 and 4 for SOC and clay content. Comparing all the prediction results, the combination of MIR and Vis-NIR with the partial least squares regression–JK technique resulted in the lowest prediction errors (RMSECV_soc of 0.35% and RMSECV_clay of 1.05%). The average uncertainties of laboratory measurements were 0.39% and 1.86% for SOC and clay contents, respectively. All models had acceptable and—to a large extent—comparable margins of error. Partial least squares regression with JK simplified and enhanced the interpretation of the developed models because of a reduction in the number of variables used in the models.