Comparison of multivariate models and variable selection algorithms for rapid analysis of the chemical composition of field crops

Abstract

This study evaluates the use of visible and near-infrared spectroscopy for rapid prediction of total carbon, total nitrogen, and total phosphorus concentrations in field crop samples. Two multivariate models (partial least squares regression and support vector machine regression) were compared. In addition, four spectral variable selection algorithms (competitive adaptive reweighted sampling, genetic algorithm, uninformative variable elimination, and variable importance for projection) were applied with support vector machine regression to determine the most accurate predictions. The results showed that support vector machine regression performed better than partial least squares regression for predicting the three chemical compositions. The combination of competitive adaptive reweighted sampling and support vector machine regression outperformed the other models for the predictions of total carbon and total nitrogen with high coefficients of determination of 0.91 and 0.90, respectively. For the determination of total phosphorus, the prediction accuracy of competitive adaptive reweighted sampling was comparable with the best result obtained from genetic algorithm with the coefficients of determination of 0.73 and 0.77, respectively. In conclusion, the support vector machine regression combined with competitive adaptive reweighted sampling has great potential to accurately determine the chemical composition of field crops using the visible and near-infrared spectroscopy.

Keywords

• model calibration
• support vector machine regression
• total carbon
• total nitrogen
• total phosphorus
• visible and near-infrared spectroscopy