近红外光谱法快速无损识别普通、高油、超高油玉米籽粒

Nondestructive Identification of Different Oil Content Maize Kernels by Near-Infrared Spectra

以75粒普通玉米、72粒高油玉米和73粒超高油玉米共计220粒玉米籽粒样品为实验材料，通过玉米籽粒的近红外光谱结合主成分信息提取技术建立了不同油分含量的玉米籽粒样品的BPANN识别模型。为考察模型的实际应用效果，连续10次随机划分样品集，每次在各类别玉米籽粒中随机选取4／5作为建模集，剩余1／5作为预测集，选择光谱信息的第2～15个主成分作为网络输入，样品以3个类别值-1，0，1作为目标输出，10次建模的学习识别率均达到100％。以所建BPANN模型对预测集样品进行分类识别，普通玉米、高油玉米和超高油玉米籽粒平均正确识别率分别为99．33％，97．88％和91．43％，总体正确识别率平均达到95％以上。研究结果表明BP人工神经网络近红外光谱法建立玉米籽粒识别模型可对不同油分含量的玉米籽粒进行快速、无损识别，对于玉米籽粒的选育工作具有一定的指导意义。另外还探讨了选择主成分建模对不同油分含量的玉米籽粒种类识别效果的影响，结果显示具有方差贡献率99％以上的光谱第一主成分参与建模，对模型预测效果有负影响，说明不同主成分包含的区分普通、高油与超高油玉米籽粒的分类信息不同，因此近红外光谱法建立样品分类识别模型时选择不同主成分建模是有必要的。

Using 220 maize single kernels, containing 75 common maize single kernels, 72 high-oil maize single kernels and 73 super high-oil maize single kernels as study materials, BPANN identification model was set up for maize single kernel with different oil content based on principal components of near infrared （NIR） spectra. Four fifths of the samples were randomly selected as training set and the other samples as prediction set. Fourteen principal components from the second to the fifteenth were selected as nets input and - 1, 0, 1 as nets output. Ten models were set up like this and the aecurate identification rate of all the training sets can reaeh 100%. For prediction sets, fifteen common corn grain samples had an average accurate identification rate of 99.33%, fourteen high-oil corn grain samples had an average accurate identification rate of 97. 88%, fourteen super high-oil corn grain samples had an average accurate identification rate of 91.43%, and total maize grains in prediction set had an average accurate identification rate of over 95 %. Results showed that NIR spectroscopy combined with BP-ANN teehnology could identify maize kernels fast and nondestruetively aeeording to oil content, which offered a very useful classification method for maize seed breeding. The effect of different principal component on BPANN models was also studied. Results told us that the first principal component with over 99% of variance eontribution had negative effect on the identification model. The predictive ability of identification models set up by different principal component was discriminatory, although the learning aecurate identification rates were all 100% So it is necessary to choose correlative principal component to set up identification model.