An IGA-PLSP method for FT-NIR wavelength selection for measuring soluble solid content of citrus fruits

Abstract

Near-infrared spectroscopy is a useful technique for fast, noninvasive quality measurement of fruits. The research was aimed to select the optimal wavelengths from Fourier transform near-infrared reflectance (FT-NIR) spectroscopy for the soluble solid content (SSC) evaluation of citrus. A total of 220 Gongchuan citrus were analyzed in this experiment. Standard genetic algorithm, immune genetic algorithm (IGA), and IGA combined with partial least squares projection algorithm (IGA-PLSP) were used for the optimal wavelength selection. The partial least squares (PLS) and least squares support vector machine (LS-SVM) were used for the prediction of SSC. Prediction models using the optimal wavelengths (280 wavelengths) selected by IGA-PLSP significantly improved the prediction results compared with the models using the full spectra. The standard errors of prediction (RMSEP) were reduced by 14.8% and 34.0% for the PLS and LS-SVM model, respectively. In addition, the LS-SVM model based on the IGA-PLSP method achieved better prediction results (correlation coefficient of prediction or R_p = 0.923 and RMSEP = 0.66 °Brix) compared to the PLS model (R_p = 0.914 and RMSEP = 0.69 °Brix). This research demonstrated that the IGA-PLSP algorithm provides an effective means for the optimal wavelength selection, and improves the FT-NIR spectral prediction of citrus SSC.