Premium
Fine grinding is needed to maintain the high accuracy of mid‐infrared diffuse reflectance spectroscopy for soil property estimation
Author(s) -
Wijewardane Nuwan K.,
Ge Yufeng,
Sanderman Jonathan,
Ferguson Richard
Publication year - 2021
Publication title -
soil science society of america journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.836
H-Index - 168
eISSN - 1435-0661
pISSN - 0361-5995
DOI - 10.1002/saj2.20194
Subject(s) - grinding , diffuse reflectance infrared fourier transform , calibration , partial least squares regression , spectral line , goodness of fit , spectroscopy , soil test , environmental science , mathematics , biological system , computer science , materials science , remote sensing , soil science , statistics , soil water , chemistry , geology , physics , biochemistry , photocatalysis , astronomy , quantum mechanics , biology , composite material , catalysis
In mid‐infrared diffuse reflectance (MIR) soil spectroscopy, grinding is one major step that can have pronounced effects on spectra and model calibrations. The reported literature on the effects of fine grinding on spectroscopic model performance have been inconsistent, likely in part because of limitations in sample set and model calibrations in previous studies. This study was focused on answering the question whether fine grinding is necessary for MIR spectroscopy in order to minimize model uncertainty. The main goal of this study was to compare model performance with and without fine grinding for eight soil properties using two different modeling techniques: partial least squares regression (PLS) and artificial neural networks (ANN). Approximately 500 soil samples were extracted from a large MIR spectral library in the United States to obtain spectra at non‐fine ground (NG, <2 mm,) and fine‐ground (FG, <0.18 mm,) states. Performance of calibration models built using subsets of the 500 FG and 500 NG spectra were compared with models built using the entire FG spectral library ( n > 40,000). All the model calibrations and validations were repeated 100 times to evaluate the uncertainty of the model performances. The results showed that PLS performed similar to ANN for the smaller dataset, but the best model performance was obtained with the FG full spectral library with ANN models. Predictions on the FG spectra always outperformed predictions on the NG spectra in terms of goodness‐of‐fit and variance of statistics. Overall, this study confirmed the importance of fine grinding to ensure the best MIR spectroscopic model performance.