Determination of Soil Organic Matter and Total Nitrogen from Visible Near-Infrared Spectroscopy by Multivariate Models and Variable Selection Techniques

Abstract

The status of soil nutrient content is a fundamental factor affecting changes in soil quality, influencing the growth conditions and yield levels of crops. The practicality of combining visible and near-infrared spectroscopy to evaluate soil organic matter (SOM) and total nitrogen (TN) may give an alternative to soil physicochemical examination in the laboratory, which is laborious and contaminative. A total of 394 Ferralsols soil samples were gathered from navel orange orchards located in the province of Jiangxi, China. To enhance the spectrum information, the spectra were preprocessed using five different techniques, including lg(1/R), multiplicative scatter correction, standard normal variate), detrending and Savitzky-Golay smoothing. Four variable selection algorithms—competitive adaptive reweighted sampling, successive projections algorithm, random frog, and genetic algorithm – were combined with three multivariate methods—partial least squares regression, multiple linear regression, and least squares support vector machine. The most efficient strategy combines LSSVM calibration methods with GA and lg(1/R) preprocessing. It generates values for the determination coefficient of prediction, root mean square error of prediction, and residual predictive deviation that are as follows: 0.8948, 0.1597, and 3.0949, respectively, for SOM; and 0.9129, 0.0021, and 3.4014, respectively, for TN. The results indicate that this method can accurately determine the SOM and TN in agricultural land soil, facilitating the timely adjustment of soil management measures.