Improved NIR quantitative model of soluble solids titratable acid ratio and titratable acidity in strawberry based on LS-SVM

In order to improve performance of near infrared spectroscopy (NIR) models for quantitative analysis of soluble-solid-content-to-titratable-acidity ratio (SSC-to-TA) and titratable acidity (TA) in fresh strawberry, least squares-support vector machine (LS-SVM) with latent variables (LVs), extracted by partial least squares (PLS), as input were used to establish calibration models. And the performance were compared with PLS models. Three hundreds and eighteen fresh strawberry samples of three varieties including “Tianbao” (n=100), “Fengxiang” (n=100) and “Mingxing” (n=118) were analyzed. The spectral region used in this paper was 6000～12500 cm-1. The detector, scan times and resolution were Pbs, 64 and 8 cm-1 respectively. The internal gold background as the reference spectrum was scanned before samples spectra collection. The reference data of SSC values were measured by a digital refractometer with 0.02°Brix accuracy using temperature correction from 10 to 60 °C. And the TA data were obtained by an acid-base titration method according to the National Standard of the pepole’s republic of China. Before models construction Chauvenet rule was used to detect spectral outliers that should be removed from the sample set, and then concentration outliers were removed based on student residual and leverage values. Various mathematical signal treatments were used and compared when PLS models were constructed, including savitzky-golay smoothing (SG) (points of 5, 15 and 25), first and second derivative, multiplicative scatter correction (MSC), and the standard normal variate (SNV). But the PLS models with these pretreatments either for SSC-to-TA or for TA were deteriorated. The best PLS model was established using full bands raw spectra, with correlation coefficients of calibration, root mean square error of calibration and prediction (rc, RMSEC and RMSEP) of 0.430, 0.096%, and 0.096% for TA; of 0.688, 0.926, and 1.190 for SSC-to-TA, which showed a poor predictive accuracy. Ten LVs were extracted from raw spectra of full bands by PLS. The LS-SVM models with input of LVs from 1 to 10 were compared, and the LS-SVM model presenting the best performance was obtained when the first 10 LVs were inputted. The two step grid searching and leave-one-out cross validations were used to realize the global optimization of regularization parameter gamma (γ) and kernel parameter sig2 (σ2) of radial basis function (RBF). The best LS-SVM model was far superior to the best PLS. The optimal models were obtained by LS-SVM with the first 10 LVs as input, with rc, correlation coefficients of prediction (rp), RMSEC, RESEP and the residual predictive deviation (RPD) of 0.965, 0.967, 0.028%, 0.027% and 3.881 for TA; 0.980, 0.973, 0.258, 0.373 and 3.111 for SSC-to-TA. The results indicate that with LVs as input nonlinear methods of LS-SVM offers more effective quantitative capability for SSC-to-TA and TA in strawberry. Further studies with a larger size and more varieties of strawberry samples should be done to improve the specificity, prediction accuracy, and robustness of models.