Study of wavelet denoising in apple's charge-coupled device near-infrared spectroscopy

Discrete wavelet transform was used to eliminate the noise in the charge-coupled device near-infrared (CCD-NIR) spectra of apple. The influence of three parameters (wavelet function, decomposition level, and threshold) on the predictive ability of the calibration model was investigated. The result showed that the db, sym, and bior wavelet families performed well, while the coif, dmey, and haar wavelets were not able to denoise effectively. The best decomposition level was 2. The threshold selection rules of the default, Birge-Massart, and Penalty had good denoising results, while SURE, Sqtwolog, Heuristic SURE, and Minimax set all detailed coefficients to zero due to their high threshold values. The best denoising result was obtained with the combination of the bior3.3 wavelet function, two levels of decomposition, default threshold selection rule, and the soft thresholding method. The optimal model of soluble solids content was constructed. The relative standard deviation of prediction decreased from 7.79 to 5.82% after wavelet denoising.