%A LIAN Xiao-qin, CHEN Qun, TANG Shen-miao, WU Jing-zhu, WU Ye-lan, GAO Chao
%T Quantitative Analysis Method of Key Nutrients in Lanzhou Lily Based on NIR and SOM-RBF
%0 Journal Article
%D 2022
%J SPECTROSCOPY AND SPECTRAL ANALYSIS
%R 10.3964/j.issn.1000-0593(2022)07-2025-08
%P 2025-2032
%V 42
%N 07
%U {https://www.gpxygpfx.com/CN/abstract/article_12784.shtml}
%8 2022-07-01
%X In order to realize the rapid and nondestructive detection of key nutrients protein and polysaccharide of Lanzhou lily, near infrared spectroscopy (NIRS) of 59 Lanzhou lily powder samples were collected in the range of 12 000~4 000 cm^{-1}. Firstly, ten pretreatment methods of SG, Normalize, SNV, MSC, Detrend, OSC, SG+1D, SG+Normalize, SG+SNV and SG+Detrend were used to process the original spectral data, and the optimal pretreatment method was SG+Detrend, Detrend was the best pretreatment method for polysaccharide. Then, CARS, SPA and PCA were used to screen the characteristic wavelength of the preprocessed spectral data. Finally, the SPA algorithm was used to determine the best extraction method for protein and polysaccharideâ€™s characteristic wavelength. The results showed that the correlation coefficient *R*_{p} of the prediction set was 0.810 6, and the root mean square error of the prediction set RMSEP was 1.195 3 in the protein PLSR model established by SG+Detrend_SPA treatment. In the polysaccharide PLSR model established by the Detrend SPA treatment, the correlation coefficient *R*_{p} of the prediction set was 0.810 9, and the root means square error RMSEP of the prediction set was 2.094 6. Considering the limitation of precision of the classical PLSR nondestructive prediction model, SOM-RBF neural network nondestructive prediction model is proposed in this paper. Firstly, the SOM network is used to cluster the data samples, and then the number of clustering categories and clustering center obtained is used as the number of hidden layer nodes and the data center of hidden layer nodes of the RBF network to optimize the structural parameters of RBF. In the established protein SOM-RBF neural network model, the correlation coefficient *R*_{p} of the prediction set is 0.866 6, and the root means square error of the prediction set RMSEP is 1.038 5. In the SOM-RBF neural network model established for polysaccharides, the correlation coefficient *R*_{p} of the prediction set was 0.868 1, and the root means square error RMSEP of the prediction set was 1.799 4. Comparing-PLSR and SOM-RBF prediction results, the SOM-RBF neural network model was determined as the optimal modeling method. Finally, the optimal model was established based on SG+Detrend_SPA_SOM-RBF in protein detection. The correlation coefficient of the prediction set of the model was 5.6% higher than that of PLSR, and the root means square error of the prediction set was 0.156 8 lower than that of PLSR. In the detection of polysaccharides, the optimal model was established based on Detrend_SPA_SOM-RBF, and the correlation coefficient of the model was 5.72% higher than that of PLSR, and the root means square error of the model was 0.295 2 lower than that of PLSR. The results showed that NIR and SOM-RBF techniques could be used for the rapid and non-destructive detection of key nutrients, proteins and polysaccharides, and the results could provide a theoretical basis for the future rapid and non-destructive detection of nutrients in Lily of Lanzhou.