柚类种质和品种资源繁多, 现有的柚类品种鉴别方法检测时间长, 费用高。 旨在利用高光谱成像技术探索主要柚类品种快速识别的可行性。 试验选用4个具有代表性的柚类品种, 利用高光谱成像技术, 采集240个叶片样本(60个/品种)上表面和下表面的高光谱图像。 高光谱图像标定后, 提取样本感兴趣区域平均光谱信息作为样本的光谱进行分析。 利用Kennard-Stone法将样本划分为校正集(192个)和验证集(48个)。 采用多元散射校正(MSC)和标准正态变量变换(SNV)对原始光谱曲线进行预处理后, 分别采用主成分分析 (PCA)和连续投影算法 (SPA )提取最佳主成分和有效波长, 并将其作为最小二乘支持向量机(LS-SVM)的输入变量, 建立基于叶片上表面和下表面光谱信息的PCA-LS-SVM和SPA-LS-SVM 模型。 结果显示, 基于叶片上表面光谱信息建立的PCA-LS-SVM和SPA-LS-SVM 模型对建模集样本的识别正确率分别为99.46%和98.44%, 对预测集样本的识别正确率均为95.83%。 基于叶片下表面光谱信息建立的PCA-LS-SVM和SPA-LS-SVM模型对建模集样本和预测集样本的识别正确率皆为100%。 表明, 利用高光谱成像技术结合PCA-LS-SVM和SPA-LS-SVM可实现柚类品种的快速鉴别, 叶片下表面光谱信息鉴别效果优于叶片上表面。 该研究为柚类的品种快速鉴别提供了一种新方法。

Existing methods for the identification of pummelo cultivars are usually time-consuming and costly, and are therefore inconvenient to be used in cases that a rapid identification is needed. This research was aimed at identifying different pummelo cultivars by hyperspectral imaging technology which can achieve a rapid and highly sensitive measurement. A total of 240 leaf samples, 60 for each of the four cultivars were investigated. Samples were divided into two groups such as calibration set (48 samples of each cultivar) and validation set (12 samples of each cultivar) by a Kennard-Stone-based algorithm. Hyperspectral images of both adaxial and abaxial surfaces of each leaf were obtained, and were segmented into a region of interest (ROI) using a simple threshold. Spectra of leaf samples were extracted from ROI. To remove the absolute noises of the spectra, only the date of spectral range 400～1 000 nm was used for analysis. Multiplicative scatter correction (MSC) and standard normal variable (SNV) were utilized for data preprocessing. Principal component analysis (PCA) was used to extract the best principal components, and successive projections algorithm (SPA) was used to extract the effective wavelengths. Least squares support vector machine (LS-SVM) was used to obtain the discrimination model of the four different pummelo cultivars. To find out the optimal values of σ2 and γ which were important parameters in LS-SVM modeling, Grid-search technique and Cross-Validation were applied. The first 10 and 11 principal components were extracted by PCA for the hyperspectral data of adaxial surface and abaxial surface, respectively. There were 31 and 21 effective wavelengths selected by SPA based on the hyperspectral data of adaxial surface and abaxial surface, respectively. The best principal components and the effective wavelengths were used as inputs of LS-SVM models, and then the PCA-LS-SVM model and the SPA-LS-SVM model were built. The results showed that 99.46% and 98.44% of identification accuracy was achieved in the calibration set for the PCA-LS-SVM model and the SPA-LS-SVM model, respectively, and a 95.83% of identification accuracy was achieved in the validation set for both the PCA-LS-SVM and the SPA-LS-SVM models, which were built based on the hyperspectral data of adaxial surface. Comparatively, the results of the PCA-LS-SVM and the SPA-LS-SVM models built based on the hyperspectral data of abaxial surface both achieved identification accuracies of 100% for both calibration set and validation set. The overall results demonstrated that use of hyperspectral data of adaxial and abaxial leaf surfaces coupled with the use of PCA-LS-SVM and the SPA-LS-SVM could achieve an accurate identification of pummelo cultivars. It was feasible to use hyperspectral imaging technology to identify different pummelo cultivars, and hyperspectral imaging technology provided an alternate way of rapid identification of pummelo cultivars. Moreover, the results in this paper demonstrated that the data from the abaxial surface of leaf was more sensitive in identifying pummelo cultivars. This study provided a new method for to the fast discrimination of pummelo cultivars.