A Wavelet Component Selection Method for Multivariate Calibration of Near-Infrared Spectra Based on Information Entropy Theory

A new hybrid algorithm (EWPCS) was proposed for selecting appropriate wavelet packet components containing the variations of analyte as the input data of regression model based on wavelet packet transform (WPT) and information entropy theory. At first, WPT algorithm and its reconstruction algorithm are employed to split the raw spectra into different frequency components with the maximum levels. Then the information entropy of the differences between the raw spectra and each frequency component was calculated, showing the importance of each component. At last, based on an optimized threshold value determined by the performance of regression model, the wavelet packet components representing the features of analyte variation can be obtained according to the difference of information entropy. To validate EWPCS method, it was applied to measure the oil content of corn using near-infrared spectra. The results show that the prediction ability and robustness of models obtained with EWPCS and partial least squares regression can be significantly improved with the prediction errors decreasing by up to 43.2%, indicating that EWPCS algorithm is an effective way for preprocessing modeling of near-infrared spectra.