Comparison of Vis/SWNIR and NIR spectrometers combined with different multivariate techniques for estimating soil fertility parameters of calcareous topsoil in an arid climate

Development of a Soil Organic Matter Content Prediction Model Based on Supervised Learning Using Vis-NIR/SWIR Spectroscopy

“…A previous study reported that 1400 and 1900 nm are water-related wavelength regions, the absorption region at 1400 nm is the first overtone of O-H stretching (moisture adsorbed to the topsoil surface), and the ration absorption region at approximately 1900 nm is the combination of O-H stretching and H-O-H bending of trapped water molecules in the crystal lattice [ 23 ]. However, it was recently reported that in the case of an air-dried soil sample, a peak at around 1400 nm might occur because of the first overtone of C-H functional groups related to organic matter [ 17 ]. A spectral absorption peak of approximately 2200 nm depended on phenolic O-H, amide N-H, amine N-H, and aliphatic C-H SOM groups [ 20 , 23 , 38 ].…”

“…To maximize the covariance between x and y, the PLSR algorithm incorporates successive regression and compression steps to obtain a set of orthogonal factors called latent variables (LVs). The validation method of leave-one-out cross-validation (LOOCV) was used to evaluate the PLSR model for the calibration set [ 17 , 23 ]. Optimal factors for PLSR models were obtained by minimizing root mean square error of cross-validation (RMSECV), using the LOOCV technique [ 29 , 30 ].…”

“…Regression coefficients (b-coefficients) were applied to determine the effective wavelength band from the PLSR calibration. When the b-coefficient value exceeded the thresholds, which were set to the standard deviation of their values, the corresponding wavelength was considered significant [ 17 , 23 , 31 ].…”

“…Furthermore, it can detect specific spectra from various substances in the corresponding wavelength band (350–2500 nm) using intrinsic components and can be applied for the qualitative and quantitative analysis of soil components in combination with algorithms such as partial least square regression (PLSR), partial least square discriminant analysis (PLS-DA), support vector regression (SVR), and artificial neural networks (ANN) [ 9 , 14 , 15 , 16 ]. However, when using Vis-NIR reflectance spectroscopy in topsoil, a generally complex environmental component, it is difficult to distinguish the absorption patterns of specific components within it because of overlapping absorption and low concentrations of topsoil components [ 17 ]. For this reason, analyzing soil spectral data and selecting an appropriate pre-processing technique and algorithm is a key step in improving the performance of soil composition prediction models [ 16 ].…”

“…These techniques are mathematical procedures that can transform reflectance measurements, eliminate variability in light scattering effects, and improve spectral features [ 18 ]. In many previous studies, predictive models of SOM, carbon, and nitrogen were developed using the PLSR model, which extracted complex absorption patterns and demonstrated high levels of efficiency when investigating a correlation between patterns and soil [ 17 , 19 , 20 , 21 , 22 ]. Owing to the relationship between spectral data and soil characteristics rarely being linear, studies applying neural networks and machine learning-based ANN, SVR, and random forest have been conducted [ 5 , 23 , 24 ].…”