Abstract: In order to improve the detection efficiency of egg freshness by visible near-infrared spectroscopy and develop an optimized predictive model, we optimized the modelling process by taking the advantages of a combination of wavelength feature selection and feature extraction. First derivative was used to preprocess the visible near-infrared transmittance spectral data in the range of 550–950 nm. Considering the influence of redundant spectral information on the model accuracy, a total of 45 sensitive characteristic wavelengths were selected from the preprocessed spectral data by competitive adaptive reweighted sampling (CARS) for support vector regression (SVR) modeling. The correlation coefficients of cross-validation (Rcv) and prediction (Rp) of the developed model were 0.880 5 and 0.888 9, and the root mean square errors of cross-validation (RMSECV) and prediction (RMSEP) were 8.59 and 8.42, respectively. In order to improve the calculation rate and the stability of the model, we used local tangent space alignment (LTSA) as a nonlinear feature extraction method to reprocess the selected characteristic wavelengths. In the new CARS-LTSA model, Rcv and Rp were 0.896 0 and 0.898 3, and RMSECV and RMSEP were 8.04 and 8.18. Compared with the CARS model, the CARS-LTSA model showed improved prediction accuracy and was simplified by eliminating 14 data dimensions. The results of this study illustrated that combined use of feature selection and feature extraction for visible near-infrared spectral data preprocessing not only improved the detection efficiency but also enhanced the accuracy of the predictive model and therefore could provide a reference method for the optimization of predictive modelling for detecting egg freshness based on infrared spectral data.

Key words: egg, visible near-infrared spectrum, feature selection, feature extraction, optimization, freshness