Detection of soil organic matter from laser-induced breakdown spectroscopy (LIBS) and mid-infrared spectroscopy (FTIR-ATR) coupled with multivariate techniques

• Published in: Geoderma Vol. 355; p. 113905
• Main Authors: Xu, Xuebin, Du, Changwen, Ma, Fei, Shen, Yazhen, Wu, Ke, Liang, Dong, Zhou, Jianmin
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2019
• Subjects: algorithms; atomic absorption spectrometry; calibration; Data fusion; Fourier transform infrared spectroscopy; FTIR-ATR spectroscopy; Laser-induced breakdown spectroscopy; least squares; monitoring; neural networks; precision agriculture; prediction; principal component analysis; reflectance spectroscopy; soil; soil fertility; soil nutrients; soil organic matter; spectral analysis; Spectral preprocessing; wavelet; Data fusion; FTIR-ATR spectroscopy; Laser-induced breakdown spectroscopy; Spectral preprocessing
• DOI: 10.1016/j.geoderma.2019.113905
• ISSN: 0016-7061

Spectroscopy is a useful method for soil monitoring because of its environmental friendliness, and its ability to produce rapid, nondestructive, simultaneous multi-element analysis. In this work, data fusion strategies for laser-induced breakdown spectroscopy (LIBS) and attenuated total reflectance Fourier-transform mid-infrared spectroscopy (FTIR-ATR), as well as a combination of multivariate calibration methods were investigated for prediction of soil organic matter (SOM) content. The root mean square error (RMSE) and residual prediction deviation (RPD) of the calibration and validation sets, systematic error, and residual assessment, were applied to evaluate the robustness and accuracy of these predictions. The results of a principal component analysis (PCA) indicated that baseline wander present in the spectral data could be effectively removed using morphological weighted penalized least squares (MPLS) and wavelet transform (WT) algorithms. The quantitative prediction ability of SOM content by a partial least squares regression (PLSR) model could be improved using principal component weighted mean (PCWM) and Euclidean distance weighted mean (EDWM) algorithms applied to parallel LIBS spectra. The prediction ability of SOM content was dramatically improved using mid-level data fusion based on the concatenation of latent variables of LIBS and FTIR-ATR spectra obtained by partial least squares algorithm. The considerable prediction accuracy and robustness were achieved using the PLSR model (RV2 = 0.792, RMSEV = 1.76 g kg−1, and RPDV = 2.16), the support vector regression (SVR) model (RV2 = 0.811, RMSEV = 1.68 g kg−1, and RPDV = 2.27), and the artificial neural network (ANN) model (RV2 = 0.830, RMSEV = 1.60 g kg−1, and RPDV = 2.39). The findings from this work suggest that the use of LIBS and FTIR-ATR spectra in combination with multivariate calibration can be a simple, fast, and nondestructive approach to monitor SOM. This strategy is potentially of great significance in the evaluation of soil fertility, the management of soil nutrients, and in guiding the agricultural production of precision agriculture. •Soil atomic spectra of laser-induced breakdown spectroscopy•Mid-infrared attenuated total reflectance spectra of soil•Fusions of soil molecular spectra and atomic spectra•Prediction of soil organic matter through optimized data fusion strategy and chemometric method