Abstract:

In order to rapidly and accurately identificate rice germplasm resources， an efficient identification method was developed based on hyperspectral analysis. Taking 9 rice varieties in the South China rice region as experimental samples， the hyperspectral reflectances of 2 700 seeds were obtained by spectrometer， and principal component analysis （PCA） was used to reduce the dimensionality of the hyperspectral data. To explore the optimal number of principal components in PCA， this paper compared the effect of combining different principal component numbers and discriminant analysis methods （linear discriminant， quadratic discriminant， and Markov distance discriminant） in establishing a rice variety recognition model based on seed hyperspectral data. Principal component analysis on full band data samples was performed， 3 variety discrimination models for prediction were established using 2~20 principal components as feature variables and the accuracy of the prediction set as an evaluation indicator， and their effects were compared. The results showed that if taking the cumulative contribution rate≥85% as evaluation criterion， 2 principal components were selected， and the accuracy rates of the 3 models prediction sets were 32.14%， 38.69% and 33.73%， respectively； when using eigenvalues≥1 as standard， 11 principal components were selected， and the accuracy rates of the prediction sets were 68.21%， 87.33%， and 83.18%， respectively； considering the accuracy of the model， 20 principal components were selected and the accuracy of the prediction set was 84.99%， 95.71%， and 95.32%， respectively. The rice hyperspectral variety recognition model established using principal component analysis and discriminant analysis methods was feasible， but the different number of principal components and DA method evaluation standards resulted in significantly different recognition effectiveness. When the number of principal components was same， the quadratic discriminant analysis method had the best recognition effect among 3 discriminant standards. The best combination was 20 principal components+quadratic discriminant analysis method， and the accuracy of the prediction was 95.71%. The research on rice variety recognition based on hyperspectral technology and principal component analysis could quickly identify different rice varieties and had high application value.

Key words: rice seeds, hyperspectral technology, principal component analysis, variety identification

摘要：

为快速精准识别水稻种质资源，开发一种基于高光谱的高效识别方法。以9种稻种为试验样品，利用光谱仪测定2 700颗种子的高光谱反射率，利用主成分分析法（principal component analysis，PCA）对高光谱数据进行降维处理。为探讨PCA最佳的主成分个数，比较了不同主成分个数与判别分析法（线性判别、二次判别和马氏距离判别）组合在基于种子高光谱数据建立水稻品种识别模型中的效果。对全波段数据样本进行主成分分析，以2~20个主成分个数作为特征变量，以预测集正确率为评价指标，建立3个品种判别模型，并比较其预测效果。结果表明，以累积贡献率≥85%为评价标准，选择2个主成分，3种模型预测集正确率分别为32.14%、38.69%和33.73%；以特征值≥1为标准，选择11个主成分，预测集正确率为68.21%、87.33%和83.18%；若考虑模型的正确率，选择20个主成分，预测集正确率分别为84.99%、95.71%和95.32%。利用主成分分析+判别分析方法的稻种高光谱品种识别模型可行，但主成分个数不同，判别分析法的评价标准不同，识别效果差异大。当主成分个数相同时，3种判别标准中二次判别分析方法的识别效果最佳，最佳组合为20个主成分个数+二次判别分析法，预测集正确率为95.71%。研究结果表明基于高光谱技术与主成分分析的识别方法可快速识别不同稻种，具有较高的应用价值。

关键词: 稻种, 高光谱技术, 主成分分析, 品种识别