Near infrared spectroscopic determination of soil organic matter based on dung beetle optimizer and support vector machine regression

Abstract

Learning performance, and generalization ability, of support vector machine regression (SVR) models used to predict soil organic matter (SOM), largely depend on the selection of correlation coefficients. Soil sample data from the state-owned Huangmian Forest Farm and Yachang Forest Farm in the Guangxi Zhuang Autonomous Region, were used to compare competitive adaptive reweighted sampling (CARS) with principal component analysis (PCA) and support vector regression (SVR) models based on full spectral data. In addition, the SVR model optimized via multi-strategy, improved dung beetle optimizer (MSDBO) and grey wolf optimizer (GWO) were compared. The result shows two spectral feature wavelength selection algorithms and two optimization algorithms could improve the determination coefficient, and reduced the root mean square error of prediction (RMSEP). The SVR model optimized by PCA and MSDBO illustrates the best generalization performance among all the evaluated models, the determination coefficient and RMSEP were 0.94 and 2.9 g·kg^-1 respectively. Mean relative error and mean absolute error were 10.0% and 2.2 respectively. The results show that the content of organic matter in soil can be accurately detected by SVR model optimized by input characteristics and parameters based on near infrared spectral data.