A Self-Adaptive Model for the Prediction of Soil Organic Matter Using Mid-Infrared Photoacoustic Spectroscopy

Fast quantification of soil organic matter (SOM) is important in crop production and soil fertility evaluation. Fourier-transform infrared (FTIR) spectroscopy has been widely utilized for rapid, cost-effective, and non-destructive SOM determination. However, the lack of accuracy has limited the application of FTIR spectroscopy to quantitative SOM prediction because the models are built from a typical database, resulting in large errors in new independent samples. In this study, using 933 paddy soil samples collected in Lishui, China, a self-adaptive model was designed for predicting SOM content, in conjunction with Fourier-transform mid-infrared photoacoustic spectroscopy (FTIR-PAS). The resulting FTIR-PAS spectra afforded abundant soil information, reflected in O-H, N-H, and C-H vibrations (4000-2800 cm-1), C=O and C-H vibrations (2500-1200 cm(-1)), and the fingerprint region (1200-500 cm(-1)). The self-adaptive model was established by: (i) identification of soil samples, selected by Euclidean distance, with soil spectra to similar the target (unknown) soil sample and ranking of the Euclidean distance values in ascending order; (ii) selection of the optimal parameters to build a partial least squares (PLS) model based on an optimal calibration sample subset; and (iii) prediction and validation of the unknown soil sample. The predictive capabilities of the self-adaptive model and conventional PLS model were compared; the self-adaptive and conventional PLS models had R-2 values of 0.9293 and 0.5796, root mean square errors of prediction of 1.65 and 3.26 g kg(-1), and ratios of percentage deviation (RPD) of 3.18 and 1.62, respectively. Therefore, the self-adaptive model showed greater potential for application, having significantly enhanced applicability while improving the accuracy of prediction.