高光谱结合主成分分析的苎麻品种识别

苎麻（Boehmeiria nivea L）是我国的特产，作为一种传统的纤维作物，一直有着较高的经济地位。开发一种基于高光谱的、新型高效的苎麻品种识别方法，有利于苎麻栽种、种质资源开发利用，为实现苎麻高产优质及麻田精准管理提供关键技术支撑，对提高苎麻产量和品质有重要意义。为了将高光谱技术应用于苎麻品种识别，采集了9个不同基因型苎麻品种，利用地物光谱仪测定苎麻叶片高光谱反射率，共1 458个叶片高光谱数据，利用主成分分析(PCA)对高光谱数据进行降维，探讨PCA最佳主因子个数的确定方法，比较不同主因子个数与不同判别分析(DA)方法--即线性判别分析(LDA)、二次判别分析(QDA)和马氏距离判别分析(MD-DA)组合，在建立基于叶片高光谱的苎麻品种识别模型中效果。对全波段的数据样本进行主成分分析之后，以2～20个主成分作为特征变量，分别建立LDA，QDA和MD-DA三种品种判别模型进行预测，以预测集正确率为评价标准，比较各种组合的效果。结果表明，若以累积贡献率≥85%为标准，选择2个主成分时，LDA，QDA和MD-DA三种判别模型预测集正确率分别为32.92%，38.48%和33.54%；以特征值≥1为标准，选择11个主成分时，三种判别模型预测集正确率分别为68.72%，87.04%和83.54%；若以预测集正确率为优先考虑标准，将主成分个数增加至20个时，三种判别模型正确率有较大提高，分别为84.98%，95.68%和95.27%。由此，得到如下结论：①利用PCA组合DA方法建立基于苎麻叶片高光谱的品种识别模型是可行的，但因子数不同、DA判别标准不同、组合方法不同效果差异非常大；②主因子个数对识别结果的影响较为明显，适当增加主成分个数可以显著提高模型判别正确率，因此不应局限于PCA特征值和方差累积贡献率的选择方法；③主因子个数相同时，三种判别标准中，QDA效果最好，LDA效果最差；④最佳组合是20个主成分+QDA方法，其数据维度大大降低（由全波段的2 031维降低20维），而预测集正确率为95.68%。

Identifying Ramie Variety by Combining the Hyperspectral Technology with the Principal Component Analysis

Ramie（Boehmeiria nivea L）is a special and traditional fiber crop in China, having higher economic status. Determining the hyperspectral reflectance of ramie leaves with the spectrometer and developing a hyperspectrum-based method of ramie variety identification of high efficiency will be beneficial for the cultivation of ramie, the development and utilization of germplasm resources as well as the provision of critical technological supports to realize the top quality and high production of ramie and the accurate management of ramie croplands, which are significant for improving ramie yield and quality. In order to apply the hyperspectral technology for identifying ramie varieties, total 1458 hyperspectral data on the ramie leaves coming from nine ramie varieties of different genotypes were collected. According to these data, we explored the using of the Principal Components Analysis(PCA) to reduce dimensions of the hyperspectral data and how to determine the best appropriate number of principal factors in the PCA. Further, we compared different combinations constituted by different principal factors and different Discriminant Analysis approaches, and the results of the ramie variety identifying models based on the hyperspectrum of ramie leaves were established. After the principal component analysis of the full-band data sample, with 2~20 principal components as the feature variables, we applied three discriminant models, namely the Linear Discriminant analysis(LDA), the Quadratic Discriminant Analysis(QDA), and the Mahalanobis Distance Discriminant Analysis, (MD-DA), to create variety discriminant models and used them to predict, and with the accuracy of the prediction set as the evaluation criteria, the effects of various combinations were compared. The results showed that when we used the cumulative contribution rate(≥85%) as the criteria and selected two principal components, the accuracies for the LDA, the QDA and the MD-DA prediction sets were respectively 32.92%, 38.48% and 33.54%; but, when we used the feature value(≥1) as the criteria, and selected eleven principal components, the accuracies for the prediction sets of above discriminant models were respectively 68.72%, 87.04% and 83.54%; and further, when we considered the accuracy of the prediction set as the preferential criteria and selected twenty principal components, the accuracies for above discriminant models were all significantly improved and were respectively 84.98%, 95.68% and 95.27%. Therefore, we can draw the following conclusions: (1) it is feasible to establish the ramie leaf-based hyperspectral variety identification model by combining the PCA and the DA, but there are big differences between results due to different numbers of factors, different DA criterias and different combination approaches; （2）The impact of the number of principal factors on the identification results are significant, and the appropriate adding of the principal components can notably improve the accuracies of corresponding models, thus it is not confined to how to select the feature values of the PCA and the accumulative variance contribution rate ; (3) When the numbers of principal factors are the same, among above three discriminant criteria, the effect of the QDA is the best while that of the LDA is the worst; (4) Twenty principal components and the QDA approach constitute the best combination, which makes data dimensions be hugely reduced, from 2031 dimensions of the full-band down to 20 dimensions, and the accuracy of the prediction set is 95.68%.