In order to realize quality monitoring and spoilage early warning of tomatoes during storage based on threedimensional (3D) fluorescence information of the storage room gas, a strategy for selecting feature wavelengths based on wavelet packet decomposition was proposed. Firstly, Rayleigh scattering removal and Savitzky-Golay (SG) smoothing were carried out on the original data to eliminate the noise interference and the background emission spectrum without fluorescence information. Secondly, the emission spectrum corresponding to each excitation wavelength was decomposed by the wavelet packet method and the feature excitation wavelengths were selected by the energy of the lowest frequency band combined with its proportion in the total energy. Thirdly, the emission spectrum corresponding to each selected feature excitation wavelength was divided into different bands, each band was decomposed by the wavelet packet method, and the feature emission wavelengths were selected by the lowest wavelet packet frequency band energy. Fourthly, considering the different fluorescence information of tomatoes before and after spoilage, the benchmark of tomato spoilage could be determined by the total energy change of the feature excitation wavelengths; at the same time, the same benchmark was obtained by cluster analysis of spectral vectors at the feature emission wavelengths. Fifthly, the feature emission wavelengths were optimized according to the spoilage benchmark to reduce the number of the feature wavelengths and simplify the analysis. Finally, the feature emission wavelengths of the spoilage benchmark date were taken as the benchmark vector to calculate the Mahalanobis distance between it and fluorescence feature vectors before the benchmark date. The results showed that with the advance of storage, the Mahalanobis distance decreased, effectively describing the quality change of tomatoes during storage and ultimately realizing spoilage early warning.

In order to rapidly identify perilla species and avoid passing off, three-dimensional (3D) fluorescence spectral data of perilla from four regions in China were acquired. A feature selection strategy of fluorescence data based on wavelet packet decomposition fused with Fisher discriminant analysis (FDA) was proposed, and effective identification of the four species of perilla was implemented. First, the 3D fluorescence data were preprocessed by using Delaunay triangle interpolation to remove the adverse influence of Rayleigh scattering and Raman scattering; Savitzky-Golar (SG) convolutional smoothing was applied to smooth the data for the purpose of reducing the interference of noise. At the same time, the 3D fluorescence data were initially screened to remove emission wavelengths with fluorescence intensity less than 0.01. Second, the 3-layer sym4 wavelet packet decomposition of the emission spectrum corresponding to each excitation wavelength was performed, and the wavelet packet energy value of the lowest frequency band was calculated as the amount of spectral data characterization for each excitation wavelength. Third, FDA was used for discriminant analysis of these wavelet packet energy values, and the discrepancy information contained in them was fused to obtain the new variables generated by FDA; the first three FD variables with 99% cumulative discriminative power were selected as variables for the characterization of the discrepancy information of different species, and then a characterization strategy for the 3D fluorescence data was proposed. Finally, two pattern recognition algorithms, back propagation neural network (BPNN) and support vector machine (SVM), were used to analyze the characterization variables, and identification results were obtained with FDA + BPNN and FDA + SVM. A correct rate of 97.5% for the training set and 95% for the test set was observed with FDA + BPNN, and the correct rate obtained with FDA + SVM for both the training and test sets was 98.33%. These results showed that 3D fluorescence spectroscopy combined with wavelet packet decomposition, FDA and SVM algorithms could basically identify perilla from different regions, which will provide a basis for further research on perilla, such as quantitative detection of some active components.

In order to improve the accuracy and reliability of the prediction results of the vitamin C (VC) content in potatoes by hyperspectral imaging, a method for constructing input variables for predictive models based on Fisher discriminant analysis (FDA) separable data fusion was proposed. First, hyperspectral information of 200 potato samples was collected by hyperspectral imaging technology, and by comparing the modeling results obtained with the spectral data before and after preprocessing by 6 spectral preprocessing methods, multiplicative scatter correction (MSC) was determined as the optimal preprocessing method. Second, competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA) and CARS-SPA algorithm were used to extract the feature wavelengths, and 34 effective feature wavelengths were finally determined through comparative analysis. Third, the effective feature wavelengths were fused by FDA to achieve data separability, and the fused new variables were screened for their capacity to discriminate the differences among samples in order to determine the input variables for predictive models. Finally, predictive models using partial least squares (PLS) and back propagation neural network (BPNN) were established based on the variables selected before and after FDA fusion, and the results from these models were compared and analyzed. It was shown that the correlation coefficient of the BPNN model increased from 0.9726 to 0.9990, and the root mean square error (RMSE) reduced from 0.7723 to 0.1727 when the 34 effective feature wavelengths extracted by CARS were used for FDA fusion and the first three fused variables were used as the input variables, which not only greatly reduced data dimensionality, but also improved the accuracy of the detection results. Therefore, constructing input variables for the detection model based on FDA separable data fusion could improve the accuracy of the detection of potato VC content.

In this study, hyperspectral technology was used to solve the problem of identifying the geographical origin of spearmint. First, 375 spearmint leaf samples from 5 geographical origins were selected for hyperspectral data collection, and the hyperspectral data were preprocessed by multiplicative scatter correction (MSC) and analyzed by principal component analysis (PCA); the principal component variables were used to construct Wilks Λ statistics, which were then ranked in an increasing order. Then, the weight coefficient curves of the principal components corresponding to the first three smallest Wilks Λ values were drawn at each wavelength; a total of 37 feature wavelengths were obtained, namely the peak and valley wavelengths in the coefficient curves. Next, Fisher discriminant analysis (FDA) was used to construct the input variables for a model for graphical origin identification of spearmint. At last, a support vector machine (SVM) model and a back propagation neural network (BPNN) model for identifying the geographical origin of spearmint were constructed. The results indicated that the SVM model outperformed the BPNN model, with discrimination accuracy of 99.67% and 98.67% in the training and test sets, respectively. Therefore, the SVM model, constructed using PCA combined with Wilks Λ statistics, can effectively identify the geographical origin of spearmint. In this model, the extracted feature wavelengths are not affected by the number of physicochemical indexes, making the model robustness.

This study aims to rapidly and accurately identify the origin of spearmint. A feature wavelength extraction was proposed to combine the sequential fusion of Kernel Principal Component Analysis (KPCA) with Kernel Locality Preserving Projections (KLPP) and Wilks' Λ statistics. Effective discrimination was realized among the five origins of spearmint. Firstly, the three-dimensional fluorescence data was collected from the 300 spearmint samples in five locations. The interferences of Rayleigh and Raman scattering were removed from the original spectra by triangular interpolation. Savitzky-Golay (SG) was used to smooth and then eliminate the noise. Secondly, the preprocessed three-dimensional fluorescence spectrum matrix was merged into the data matrix, according to the emission wavelength. KPCA, KPCA + KLPP, KPCA + Wilks Λ statistics, and KPCA + KLPP + Wilks Λ statistics were used to extract the feature wavelengths of the three-dimensional fluorescence spectrum matrix after preprocessing. Among them, KPCA and KPCA + Wilks Λ statistics were used to extract 40 wavelengths of feature emission and eight wavelengths of feature excitation, respectively. KPCA+KLPP and KPCA + KLPP + Wilks Λ statistics were to extract 50 wavelengths of feature emission and 10 wavelengths of feature excitation, respectively. Thirdly, the spectral values of feature emission wavelength were converted into row vectors, according to the wavelength order of feature excitation. Four 300-row spectral matrices of feature wavelength were then obtained from the 300 samples. Furthermore, Fisher discriminant analysis (FDA) was applied to the spectral matrix of feature wavelengths for data separable fusion for the separable fisher discriminant (FD) variables. The first four FD variables were selected with a cumulative discriminant ability of 99% as the input vectors of the discriminant model for the spearmint origins. Finally, the support vector machine (SVM) was used to analyze the four FD variables. The FDA + SVM identification model showed that there was a four-feature wavelength extraction. The correct rates for the training set were 92.00%, 97.78%, 95.11%, and 100%, respectively, and the correct rates for the test set were 92.00%, 96.00%, 94.67%, and 100%, respectively. The results show that both KLPP and Wilks Λ statistics improved the performance of KPCA. The sequential fusion significantly improved the overall feature extraction, in order to better remove the redundant information of spectral data for the high correct of identification. Therefore, the sequential fusion of KPCA + KLPP + Wilks Λ statistics + FDA + SVM can be expected to effectively identify the origin of spearmint. The extraction of feature wavelength with sequential fusion of KPCA + KLPP + Wilks Λ statistics can effectively reduce the redundancy of three-dimensional fluorescence spectral data for the informative features of the original fluorescence data. Better identification was achieved in the five kinds of spearmint origins. In addition, these research findings can lay a foundation for the subsequent quantitative analysis of the important components in spearmint by three-dimensional fluorescence spectroscopy.