Algorithm of Stability-Analysis-Based Feature Selection for NIR Calibration Transfer

Modeling of soluble solid content of PE-packaged blueberries based on near-infrared spectroscopy with back propagation neural network and partial least squares (BP-PLS) algorithm

Blueberries are a nutritious and popular berry worldwide. The physical and chemical properties of blueberries constantly change through the cycle of the supply chain (from harvest to sale). The purpose of this study was to develop a rapid method for detecting the properties of packaged blueberries based on near-infrared (NIR) spectroscopy. NIR was applied to quantitatively determine the soluble solid content (SSC) of polyethylene (PE)-packaged blueberries. An orthogonal partial least squares discriminant analysis model was established to show the correlation between spectral data and the measured SSC. Multiplicative scattering correction, standard normal variable, Savitzky-Golay convolution first derivative, and normalization (Normalize) were used for spectra preprocessing. Uninformative variables elimination, competitive adaptive reweighted sampling, and iteratively retaining informative variables were jointly used for wavelength optimization. NIR-based SSC prediction models for unpacked blueberries and PE-packaged blueberries were developed using partial least squares (PLS). The prediction model for PE-packaged samples (RP 2 = 0.876, root mean square error of prediction [RMSEP] = 0.632) had less precision than the model for unpacked samples (RP 2 = 0.953, RMSEP = 0.611). To reduce the effect of PE, the back propagation (BP) neural network and PLS were combined into the BP-PLS algorithm based on the residual learning algorithm. The model of BP-PLS (RP 2 = 0.947, RMSEP = 0.414) was successfully developed to improve the prediction accuracy of SSC for PE-packaged blueberries. The results suggested a promising way of using the BP-PLS method in tandem with NIR for the rapid detection of the SSC of PE-packaged blueberries. PRACTICAL APPLICATION: Most of the NIR-based research used unpacked blueberries as samples, while the use of packaged blueberries would provide researchers with a better understanding of the crucial factors at different phases of the blueberry supply chain (from harvest to sale). To meet market demands and minimize losses, NIR spectroscopy has been proven to be a rapid and nondestructive method for the determination of the SSC of PE-packaged blueberries. This study provides an effective method for monitoring the properties of blueberries in the entire supply chain.

A Model Transfer Method among Spectrometers Based on Improved Deep Autoencoder for Concentration Determination of Heavy Metal Ions by UV-Vis Spectra

Ultraviolet Visible (UV-Vis) spectroscopy detection technology has been widely used in quantitative analysis for its advantages of rapid and non-destructive determination. However, the difference of optical hardware severely restricts the development of spectral technology. Model transfer is one of the effective methods to establish models on different instruments. Due to the high dimension and nonlinearity of spectral data, the existing methods cannot effectively extract the hidden differences in spectra of different spectrometers. Thus, based on the necessity of spectral calibration model transfer between the traditional large spectrometer and the micro-spectrometer, a novel model transfer method based on improved deep autoencoder is proposed to realize spectral reconstruction between different spectrometers. Firstly, two autoencoders are used to train the spectral data of the master and slave instrument, respectively. Then, the hidden variable constraint is added to enhance the feature representation of the autoencoder, which makes the two hidden variables equal. Combined with a Bayesian optimization algorithm for the objective function, the transfer accuracy coefficient is proposed to characterize the model transfer performance. The experimental results show that after model transfer, the spectrum of the slave spectrometer is basically coincident with the master spectrometer and the wavelength shift is eliminated. Compared with the two commonly used direct standardization (DS) and piecewise direct standardization (PDS) algorithms, the average transfer accuracy coefficient of the proposed method is improved by 45.11% and 22.38%, respectively, when there are nonlinear differences between different spectrometers.

Rapid Determination of Cellulose and Hemicellulose Contents in Corn Stover Using Near-Infrared Spectroscopy Combined with Wavelength Selection

The contents of cellulose and hemicellulose (C and H) in corn stover (CS) have an important influence on its biochemical transformation and utilization. To rapidly detect the C and H contents in CS by near-infrared spectroscopy (NIRS), the characteristic wavelength selection algorithms of backward partial least squares (BIPLS), competitive adaptive reweighted sampling (CARS), BIPLS combined with CARS, BIPLS combined with a genetic simulated annealing algorithm (GSA), and CARS combined with a GSA were used to select the wavelength variables (WVs) for C and H, and the corresponding regression correction models were established. The results showed that five wavelength selection algorithms could effectively eliminate irrelevant redundant WVs, and their modeling performance was significantly superior to that of the full spectrum. Through comparison and analysis, it was found that CARS combined with GSA had the best comprehensive performance; the predictive root mean squared errors of the C and H regression model were 0.786% and 0.893%, and the residual predictive deviations were 3.815 and 12.435, respectively. The wavelength selection algorithm could effectively improve the accuracy of the quantitative analysis of C and H contents in CS by NIRS, providing theoretical support for the research and development of related online detection equipment.