本文用MPA傅立叶变换近红外光谱仪研究了单粒水稻种子的漫反射光谱特征, 发现颖壳对光谱扫描有影响, 为了尽量消除这种影响, 保证光谱的代表性, 应对其上下表面分别进行多次光谱采集, 取它们的平均光谱。本文利用种子品种特有的光谱特性, 结合不同光谱预处理方法建立了多个聚类分析模型, 比较它们对杂交F1代种子“03S/0412”和其父本种子“0412”的鉴别效果。结果显示: 选择4000～8900 cm-1光谱范围, 通过“无预处理”、“矢量归一化”、“二阶导数（25点平滑）”和“二阶导数（25点平滑）+矢量归一化”建立的模型校正集正确率分别为52.4%、65.2%、75.2%和100%, 可见, 相比无预处理, 经过各种预处理后正确率都有提高, 其中“二阶导数（25点平滑）+矢量归一化”建立的模型最好, 用该模型对预测集预测, 分类正确率也为100%, 具有很好的预测性能。这说明近红外光谱技术可用于单粒水稻种子品种真伪性鉴别。

In this paper, MPA FT-NIR spectroscopy was used to study the diffuse reflection spectrum characteristics of the single grain of rice seeds. The results showed the chaff had influence on spectral scanning, so the spectra of the top and bottom surfaces should be collected respectively with many times to obtain their average spectrum in order to eliminate the effect as far as possible and ensure the representation of the spectrum. Combined with different spectral pretreatment methods, the unique spectral characteristics of a seed variety were used to establish several clustering analysis models and compared their identification effects for F1 seeds “03S/0412” and paternal seeds “0412”. The results showed the models’ classification accuracy of calibration set were 52.4%, 65.2%, 75.2% and 100% respectively using “no pretreatment”, “vector normalization”, “second derivative (25 points smoothing)”, “second derivative (25 points smoothing) plus vector normalization” in spectral region between 4000 cm-1～8900 cm-1. So compared with no pretreatment, the accuracy had increased after various pretreatments, among them the model was best using “second derivative (25 points smoothing) plus vector normalization”, its classification accuracy of prediction set was also 100%, indicating it had a very good prediction performance. It suggested that the near infrared spectroscopy could be used in variety authenticity identification for the single grain of rice seeds.