灵武长枣VC含量的高光谱快速检测研究

采用可见-近红外高光谱成像技术结合化学计量学方法检测灵武长枣维生素C(VC)含量，探究一种全新的水果内部成分的快速无损检测方法。采用高效液相色谱法(HPLC)测得长枣的VC含量化学值，可见-近红外高光谱成像系统采集164个灵武长枣400～1 000 nm的高光谱图像，利用ENVI4.8软件提取图像的感兴趣区域(region of interest，ROI)，计算其平均光谱，获得光谱值，将化学值与光谱值通过The UnsecramblerX 10.4软件建立模型。利用蒙特卡洛交叉验证法剔除异常值，采用光谱理化值共生距离法(sample set partitioning based on joint x-y distance，SPXY)进行样本划分以提高模型的预测性能；对光谱采用移动平滑(moving average)、中值滤波(median filter)、归一化(normalize)、基线校准(baseline)、多元散射校正(multiple scattering correction，MSC)、去趋势(detrending)和标准正态变量变换(standard normal variate，SNV)等7种方法进行预处理；为进一步减少数据量，降低维度，提高运算速度，使用竞争性自适应加权算法(competitive adaptive reweighted sampling，CARS)、无信息变量消除算法(uninformative variable elimination ，UVE)和连续投影算法(successive projections algorithm，SPA)提取特征波长，以期实现以少数波段代替全波段；将全波段光谱(full spectrum，FS)以及CARS, UVE和SPA三种方法提取的特征波长分别建立偏最小二乘(partial least squares wavelength regression，PLSR)和支持向量机(support vector machine，SVM)模型，从而确定最优的建模模型。利用蒙特卡洛交叉验证法共剔除7个异常样本，采用SPXY法将剔除异常样本后的157个数据区分为校正集和预测集，校正集中样本个数为117，预测集中样本个数为40。将未经光谱预处理的建模结果与分别经过七种光谱预处理的建模结果相比，选择未经光谱预处理的数据进行后续分析；将未经光谱预处理的光谱值采用CARS，UVE，SPA方法进行提取特征波长，CARS共优选出406，415，487，631，636，655，660，665，670，684，689，694，723，732，747和881 nm下的光谱变量16个，利用CARS提取出的特征波长占总波长的12.8%；UVE共优选出406，415，627，631，636，651，655，660，665，670，675，679，684，689，694，699，703，708，742，747，751，756，761，766，771，775，780，785，790，795，919和924 nm下的32个特征波长，利用UVE提取出的特征波长占总波长的25.6%；SPA共优选出401，665，684 nm三个特征波长，利用SPA提取出的特征波长占总波长的2.4%。将全波段光谱与提取出的特征波长建立PLSR模型和SVM模型，对比模型结果显示UVE-SVM模型最优，其R2_c为0.847 1，R2_p为0.714 9，说明UVE有效地对光谱进行降维，简化了数据处理过程。本研究对高光谱成像技术在水果领域的应用进行了有益探索，探究了一种全新的灵武长枣VC含量的无损检测方法，相应建立的可见-近红外高光谱模型为其他水果成分的快速检测提供了理论基础。

A Rapid Evaluation of VC Content on Lingwu Long Jujube Using Hyperspectral Technique

In this paper, Lingwu Long Jujube VC content was regarded as the research object, and a combination of hyperspectral imaging technique with chemometrics method was used to explore a rapid and nondestructive detecting method for fruit internal components. Vitamin C content of Long jujube was measured by high performance liquid chromatography (HPLC). A total of 164 Lingwu long jujubes of hyperspectral images in region of 400~1 000 nm were acquired. Then spectral curves were obtained by ENVI 4.8 software from the region of interest (ROI). The models were built for chemical value and spectral data by UnsecramblerX 10.4 software. Outliers were to be eliminated by Monte Carlo cross validation method; Samples division was set partitioning based on joint X-Y distance(SPXY) method to improve the prediction performance of the model; The spectral's pretreatment was analyzed, such as Moving Average, Median Filter, Normalize, Baseline, multiple scatter correction (MSC), Detrending and standard normal variate (SNV) and so on; To reduce the amount and dimension of data, the feature wavelengths were extracted by competitive adaptive weighting algorithm (CARS), uninformative variable elimination ( UVE) and continuous feeding Shadow algorithm (SPA) ; Compared to the models of full spectrum (FS) and the feature wavelengths extracted by CARS and UVE of PLSR and SVM built, the optimal model was determined. A total of 7 abnormal samples were eliminated using Monte Carlo cross validation method. After eliminating abnormal sample data, the samples were divided into calibration set and prediction set by SPXY method, and calibration samples is 117, and prediction samples is 40. The spectral pretreatment were studied by the 7 methods. The results showed that the model effect without spectral pretreatment was the best, and its Rc was 0.8779, and RMSECV was 0.0481; Without a preprocessing method by CARS, UVE and SPA method to reduce the dimensions, a total of 16 feature wavelengths were selected by CARS, which were 415, 487, 406, 631 636, 655, 660, 665, 670, 684, 689, 694, 723, 732, 747 and 881 nm. A total of 32 feature wavelengths were selected by UVE, which were 415, 406, 627, 631, 636, 651, 655, 660, 665, 670, 675. 679, 684, 689, 694, 699, 703, 708, 742, 747, 751, 756, 761, 766, 771, 775, 780, 785, 790, 795, 919 and 924 nm. A total of 3 feature wavelengths were selected by SPA, which were 401, 665 and 684 nm. Comparing models of the full band spectrum with the models of extracted characteristic wavelengths of PLSR and SVM, the UVE-SVM model is the best, and its R2_c is 0.847 1 and R2_p is 0.714 9, which indicates that UVE effectively reduces the dimension of the spectrum and simplifies the data processing. This study explores the application of hyperspectral imaging technology in the field of fruit, explores a new method for nondestructive testing of Lingwu Long Jujube VC content, provides a theoretical basis for visible and near infrared hyperspectral model established for the rapid detection of other components of fruit.