Qualitative Analysis of Soluble Solid Content and Firmness of Pear Based on Successive Projections Algorithm and Least Square Support Vector Machine

Abstract

Effective wavelengths in visible-near infrared spectra region were selected based on successive projections algorithm (SPA). The selected effective wavelengths were set as inputs of least square support vector machine (LS-SVM) for qualitative analysis of soluble solid content (SSC) and firmness in pear. In this study, one-hundred and sixty samples were selected as sample set. 120 pear samples were selected randomly for the calibration set, and the remaining 40 samples for the prediction set. 16 variables included 431, 434, 439, 443, 448, 535, 595, 635, 681, 728, 742, 998, 1129, 1403, 1506 and 1771 nm, and 6 effective variables included 409, 412, 415, 419, 478 and 773 nm for prediction of SSC and firmness were selected by SPA for building the models of SSC and firmness, respectively. And, SPA-LS-SVM models were also compared with full-spectral PLS models, full-spectral LS-SVM models and SPA-PLS models. The correlation coefficients (r) of SSC and firmness were 0.8560 and 0.8452, respectively. The root mean square error of prediction (RMSEP) of SSC and firmness were 0.4648 and 1.0041, respectively. The overall results showed that SPA can fast and effectively select the optimal wavelengths. The selecting process is simple and does not need abundant parameter debugging. The prediction performance of SPA-LS-SVM model is better than conventional linear PLS model because SPA-LS-SVM model can enough use the linear and non-linear information in pear. In addition, this study also indicated that only visible spectra region or only NIR region was should not be considered for the prediction of SSC of pear because the color variances had certain indirect and latent relationship with the chemical compositions of pear. In terms of firmness prediction, it also indicated that visible light spectra may be more important for firmness prediction of pear.