Least Squares-Support Vector Machine-Based Analysis of Near-Infrared Spectra with Techniques of Dimension Reduction and Parameter Optimization

To improve the training efficiency of least squares-support vector machine (LS-SVM) method, a new algorithm was proposed for developing the multivariate regression model using near-infrared (NIR) spectra and named as PCA-PSO-LS-SVM. Coupled with principal component analysis (PCA) and particle swarm optimization (PSO), this algorithm can take advantage of spectral dimension reduction and parameter optimization of LS-SVM model. In PCA-PSO-LS-SVM algorithm, PCA algorithm is firstly employed to reduce the dimension of raw spectra for simplicity based on the cumulative contribution rate of variation of each principle component. Then, to get the calibration model with better prediction precision, the strategies of grid searching and PSO searching are used to optimize the parameters of LS-SVM model in calibration set. At last, the prediction model can be constructed using the optimum parameters of LS-SVM obtained in calibration phase. To validate the PCA-PSO-LS-SVM algorithm, it was applied to measure the oil content of corn samples. The experimental results show that the PCA-based dimension reduction of NIR spectra has negligible effect on the performance of regression model. In addition, compared with the PLS algorithm, the PCA-PSO-LS-SVM algorithm can greatly improve the prediction precision of model by up to 69.3%, indicating that it is an efficient tool for NIR spectra regression.