利用可见近红外(Vis-NIR)高光谱成像技术对完好和损伤等级灵武长枣进行快速识别检测。 采用定量损伤装置得到损伤Ⅰ, Ⅱ, Ⅲ, Ⅳ和Ⅴ级的灵武长枣, 借助高光谱成像系统采集完好长枣和损伤长枣样本高光谱图像。 提取感兴趣区域(region of interest, ROI)并计算样本平均光谱值。 利用光谱-理化值共生距离算法(SPXY)将420个长枣样本按3∶1的比例划分校正集315个和预测集105个。 灵武长枣原始光谱建立偏最小二乘判别分析(PLS-DA)分类模型, 得到校正集和预测集准确率分别为72.70%和86.67%; 灵武长枣原始光谱数据采用移动平均(MA)、 卷积平滑(SG)、 多元散射校正(MSC)、 正交信号修正(OSC)、 基线校准(baseline)和去趋势(de-trending)等方法进行光谱预处理并建立PLS-DA分类判别模型。 通过分析比较, 得到MSC-PLS-DA为最优分类判别模型, 校正集准确率为76.19%, 预测集准确率为86.67%, 其中校正集比原始光谱建模准确率提高了3.49%, 预测集准确率较原始光谱建模结果未提高; 为了提高建模效果, 对灵武长枣原始光谱和预处理后的光谱分别采用连续投影算法(SPA)、 无信息变量消除(UVE)、 竞争性自适应加权抽样(CARS)和区间变量迭代空间收缩法(iVISSA)等算法提取特征波长, 建立PLS-DA分类判别模型, 结果表明, MSC-CARS-PLS-DA为最优模型组合, 校正集准确率为77.14%, 预测集准确率为89.52%, 建模准确率较原始光谱建模准确率分别提高了4.44%和2.85%。 结果表明, Vis-NIR高光谱成像技术结合MSC-CARS-PLS-DA模型可实现灵武长枣损伤等级的快速识别。

The visible near-infrared (Vis-NIR) hyperspectral imaging technology was used to identify the intact and damaged Lingwu long jujube rapidly. In this study, damage grades, including Ⅰ, Ⅱ, Ⅲ, Ⅳ and Ⅴ of Lingwu long jujubes were obtained by using quantitative damage devices. Hyperspectral images of intact and damaged samples were collected by using a hyperspectral imaging system. Region of interest (ROI) was extracted from the image and average spectral values of samples were calculated. Sample set partitioning based on joint x-y distance (SPXY) was used to divide all samples (420) into calibration sets (315) and prediction sets (105) in a ratio of 3∶1. The partial least squares discriminant analysis (PLS-DA) classification model was established for the original spectrum, and the accuracies of the calibration set and prediction set were 72.70% and 86.67%, respectively. The original spectrum of Lingwu long jujube was preprocessed by means of moving average (MA), Savitzky Golay (SG), multiplicative scatter correction (MSC), orthogonal signal corrections (OSC), baseline and de-trending. PLS-DA classification model was established after pretreatment. The results showed that in the PLS-DA classification model established by spectrum preprocessed by different pretreatment algorithms. Through analysis and comparison, it was found that MSC-PLS-DA was the optimal model combination. In the established classification discrimination model, the accuracies of the calibration set and prediction set were 76.19% and 86.67%, respectively. The accuracy of the calibration set was 3.49% higher than that of the original spectral modeling, and the accuracy of the prediction set was not higher than that of the original spectral modeling. Original spectral and spectral after pretreatment was used to extract feature wavelengths using the successive projections algorithm (SPA), uninformative variable elimination (UVE), competitive adaptive reweighted sampling (CARS) and interval variable iterative space shrinkage approach (iVISSA), and established the PLS-DA classification model based on the feature wavelengths. The results showed that MSC-CARS-PLS-DA was the optimal classification model, the accuracy of the calibration set was 77.14%, the accuracy of the prediction set was 89.52%. The modeling accuracy was improved by 4.44% and 2.85% respectively compared with the original spectral modeling accuracy. The above research showed that the Vis-NIR hyperspectral imaging technology combined with MSC-CARS-PLS-DA model could realize the rapid identification of lingwu jujube damage grade.