摘要： 高光谱成像可将图像和光谱相结合，同时获得目标对象的图像和光谱信息，已在农产品定性和定量分析检测方面得到广泛利用。利用可见-近红外高光谱成像结合化学计量学方法对贮藏期内灵武长枣果糖含量进行无损检测。采用高效液相色谱测量长枣果糖含量的化学值，可见-近红外高光谱系统采集长枣的高光谱图像，提取每个样本感兴趣区域的平均光谱；建立长枣贮藏期的径向基核函数支持向量机（radial basis kernel function support vector machine，RBF-SVM）模型；分别选用正交信号校正法（orthogonal signal correction，OSC）、多元散射校正（multiplicative scatter correction，MSC）、中值滤波（median-filter，MF）、卷积平滑（savitzky-golay，SG）、归一化（normalization，Nor）、高斯滤波（gaussian-filter，GF）和标准正态变换（standard normalized variate，SNV）等方法对原始光谱进行预处理；为减少数据量，降低维度，提高运算速度，采用反向区间偏最小二乘法（backward interval partial least squares，BiPLS）、间隔随机蛙跳算法（interval random frog，IRF）和竞争性自适应加权算法（competitive adaptive reweighted sampling，CARS）对光谱数据提取特征变量；建立全波段和特征波段的偏最小二乘回归（partial least squares regression，PLSR）和主成分回归（principle component regression，PCR）长枣果糖含量预测模型。结果表明：RBF-SVM判别模型校正集准确率为98.04%，预测集准确率为97.14%，能很好地预测长枣的贮藏期；利用BiPLS, IRF及CARS进行降维处理，提取特征波长个数为100, 63和23，占原光谱数据的80%，50.4%和18.4%；为简化模型运算过程并提高模型精度，采用CARS算法对BiPLS及IRF算法所选取的特征波长进行二次筛选，分别优选出18和15个特征波长，占原光谱数据的14.4%和12%，显著减少特征波长数；将全波段光谱与提取出的特征波长分别建立长枣果糖含量的PLSR及PCR预测模型，优选出CARS提取特征波长建立的PLSR模型效果最优，其中校正集的相关系数Rc=0.854 4，均方根误差RMSEC=0.005 3，预测集的相关系数Rp=0.830 3，均方根误差RMSEP=0.005 7，说明CARS有效地对光谱进行降维，简化了数据处理过程。研究表明，利用可见-近红外高光谱成像结合化学计量学方法及计算机编程，可以有效的实现灵武长枣果糖含量的快速无损分析，为灵武长枣内部品质的检测提供理论依据。
关键词：高光谱成像技术；果糖；贮藏期；高效液相色谱法；偏最小二乘回归
Abstract：Hyperspectral imaging technique which is a non-destructive method combines image and spectral techniques to obtain image and spectral information of target objects’ and qualitative and quantitative analysis using spectral data has been widely used in the field of agricultural product testing. This paper uses visible/near-infrared spectroscopic imaging technique combined with chemometrics methods to achieve the non-destructive detection of fructose content of Lingwu long jujube during storage. The chemical value of jujube fructose was determined by High performance liquid chromatography (HPLC), and the hyperspectral images of long jujubes were collected using near-infrared hyperspectral system, and the average spectral data for each sample area of interest were extracted. Support Vector Machine With RBF Nucleus (RBF-SVM) Model for establishing storage time of long jujube. Orthogonal Signal Correction (OSC), Multiple Scatter Correction (MSC), Median Filter (MF), Savitzky-Golay (SG), Normalize (Nor), Gaussian filter (GF) and Standard Normalized Variate (SNV) were used to preprocess the original spectral data. To reduce the amount and dimension of data, the characteristic wavelengths were extracted by Backward interval Partial Least Squares (BiPLS), Interval Random Frog(IRF) and Competitive Adaptive Reweighted Sampling (CARS); the partial least squares regression( PLSR) model and principle component regression (PCR) were established based on full spectra and characteristic wavelengths for predicting fructose of Lingwu long jujube. The results indicated that the accuracy of the RBF-SVM model calibration set was 98.04%, and the accuracy of the prediction set was 97.14%, which could well predict the storage time of the jujube; The BiPLS, IRF and CARS methods were used to select characteristic wavelengths with 100, 63 and 23 from 125 wavelengths, respectively. In order to simplify the model and improve the accuracy of prediction of the model, the CARS algorithm was used to perform secondary extracted characteristic wavelengths of BiPLS and IRF and select characteristic wavelengths with 18 and 15, respectively, which significantly reduced the number of characteristic wavelengths. Comparing models of the full band spectrum with the models of extracted characteristic wavelengths of PLSR and PCR, PLSR model based on the characteristic variables selected by CARS was the best, and correlation coefficient of Calibration set (Rc) and root-mean-square error of Calibration set (RMSEC) of the model were 0.854 4 and 0.005 3, and correlation coefficient of prediction (Rp) and root-mean-square error of prediction set (RMSEP) of the model were 0.830 3 and 0.005 7, respectively, which indicated that CARS effectively reduced the dimension of the spectrum and simplified the data processing. The results showed that visible/near-infrared hyperspectral imaging technique combined with chemometrics methods and computer programming can effectively detect fructose content of Lingwu long jujube rapidly and non-destructively, providing a theoretical basis for the detection of internal quality of Lingwu long jujube.
Key words：Hyperspectral imaging technique; Fructose; Storage; High performance liquid chromatography; Partial least squares regression
万国玲，刘贵珊，何建国，杨晓玉，程丽娟，张 翀. 贮藏期内灵武长枣果糖含量的高光谱预测[J]. 光谱学与光谱分析, 2019, 39(10): 3261-3266.
WAN Guo-ling, LIU Gui-shan, HE Jian-guo, YANG Xiao-yu, CHENG Li-juan, ZHANG Chong. Prediction of Fructose Content of Lingwu Long Jujube During Storage Using Hyperspectral Imaging Technique. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2019, 39(10): 3261-3266.
[1] Simeone M L F, Parrella R A C, Schaffert R E, et al. Microchemical Journal, 2017, 134: 125.
[2] Das B, Sahoo R N, Pargal S, et al. Spectrochim Acta Part A: Molecular and Biomolecular Spectroscopy, 2017, 192: 41.
[3] Mo C, Kim M S, Kim G, et al. Biosystems Engineering, 2017, 159: 10.
[4] SUN Jing-tao, MA Ben-xue, DONG Juan, et al(孙静涛, 马本学, 董 娟, 等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2017, 37(7): 2184.
[5] Oliveira G A D, Castilhos F D, Bureau S. Food Research International, 2014, 60(6): 154.
[6] Hu W, Sun D W, Blasco J. Postharvest Biology & Technology, 2017, 125: 168.
[7] Guo Y, Ni Y, Kokot S. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2016, 153: 79.
[8] FENG Di, JI Jian-wei, ZHANG Li, et al(冯 迪, 纪建伟, 张 莉, 等). Chinese Journal of Luminescence(发光学报), 2017, 38(6): 799.
[9] FENG Yu-qin, WU Long-guo, HE Jian-guo, et al(冯愈钦, 吴龙国, 何建国，等). Chinese Journal of Luminescence(发光学报), 2016, 37(8): 1014.
[10] Zou Xiaobo, Xu Yiwei, Chen Wu, et al. Mod. Food Sci. Technol.，2015, 32(9): 263.
[11] Siedliska A, Baranowski P, Zubik M, et al. Journal of Food Engineering, 2017, 215: 61.
[12] Siedliska A, Baranowski P, Zubik M, et al. Postharvest Biology & Technology, 2018, 139: 115.
