Support vector regression coupled with wavelength selection as a robust analytical method

This paper assesses the support vector regression (SVR) as a robust alternative to partial least squares (PLS) in multivariate calibration using twelve public domain NIR spectroscopy datasets. It also proposes the use of the support vector regression - recursive feature elimination (SVR-RFE) algorithm to select the most informative wavelengths for SVR models. Models based on full spectra were built using SVR and PLS, while wavelength selection methods were carried out using SVR-RFE, interval PLS (iPLS), backward interval PLS (biPLS), synergy interval PLS (siPLS), and successive projection algorithm PLS (SPA-PLS). The prediction performance of tested methods was measured by means of the root mean squared error (RMSE), index of agreement (d-index) and R2 on the test set. SVR-based models yielded the best results in 8 out of 12 datasets, 4 of them using full spectra and 4 relying on SVR-RFE selected wavelengths. Statistical comparison was carried out for the wavelength selection algorithms using Friedman test, which pointed the SVR-RFE as a competitive technique when compared to the other algorithms. This study revealed SVR as a robust alternative to PLS, especially when SVR-RFE is employed for wavelength selection.