Accuracy of Estimating Soil Properties with Mid‐Infrared Spectroscopy: Implications of Different Chemometric Approaches and Software Packages Related to Calibration Sample Size

Abstract

Core Ideas Selection of spectral regions was included in MIR predictions of soil properties. The R software outperformed a commercial chemometric software. SVMR and model averaging performed better than PLSR and artificial neural networks. With decreasing calibration sample size, the usefulness of SVMR over PLSR decreased. MIRS studies may focus more on representativeness, sample sizes, and variabilities. Different algorithms exist in various software programs for the estimation of soil properties using mid‐infrared (MIR) spectroscopy, with recommendations varying between different studies regarding which algorithm should be used. Objectives were to compare the performance of the commercial OPUS Quant 2 software, which uses partial least squares regression (PLSR) and a selection of spectral ranges, with the R software and to study the accuracy of different algorithms as a function of the information provided in the calibration. Contents of soil organic carbon (SOC), nitrogen, and texture for surface soils of an arable field were determined, and MIR were spectra recorded. Partial least squares regression used with either software was useful (ratio of performance to interquartile distance in the validation sample [RPIQV] >1.89) for an estimation of SOC, clay, and N contents but not for sand and silt. The wavenumber region selection concept used in OPUS was also implemented in R, and it proved useful for SOC (all algorithms) and total nitrogen (artificial neural networks, support vector machine regression [SVMR]) in the validation. Support vector machine regression generally slightly outperformed the other approaches and resulted in a successful estimation of sand content. The usefulness of SVMR over PLSR generally decreased with decreasing sample size used for the calibration (thus decreasing the information provided), and PLSR partly outperformed SVMR in the validation. Overall, this study indicates that there is no general superiority of a chemometric algorithm over PLSR independent of the information provided in the calibration sample.