种子活性受到存储条件的影响很大。 收集了真实情况下受到不同存储条件影响的种子, 通过发芽实验验证了其成活率存在差异。 再从中选择适量的种子样本, 采集其单颗种子的可见-近红外反射光谱, 运用不同的光谱预处理技术, 结合不同的机器学习建模手段, 以区分不同成活率的种子。 比较了不同的光谱预处理方法, 比如标准反射光谱校正、 多元散射校正等。 从识别准确度的角度, 认为标准反射光谱校正的方法, 能够很大程度上提升不同存活率种子的光谱差异性, 从而经过机器学习判断达到更高的识别准确度。 同时比较了支持向量机、 K邻近和距离判别分析等机器监督学习建模方法, 发现利用标准反射光谱校正的方法结合距离判别分析, 能够对种子样本实现完全准确的判定。 更进一步, 为了满足实际运用中快速识别的要求, 将高分辨率的光谱数据压缩成为低分辨率多通道带通光谱数据, 这样可以大大降低的光谱数据长度, 节约各种机器学习器在训练和判断中所用的时间。 使用简化过后的多通道带通光谱数据判定种子存活率, 其识别准确度仍然接近90%。 充分说明了, 利用多通道宽带光谱数据, 并选择合适的机器学习建模方式, 足以满足实际选种产业的一般性需求, 有潜力作为未来粮种成活率快速鉴别的技术手段。 还采用了多种带通宽度以简化光谱, 分析比较不同带通宽度对识别精度的影响。 总体来说由于带宽增大, 数据量减少, 识别速度更快, 但是识别精度降低。 从10～50 nm改变光谱带宽, 标准反射校正后的简化光谱的识别精度从87.50%下降到58.75%。 在实际运用中, 需要权衡识别速率和预期识别精度, 合理的选择带宽。 验证了根据简化后的可见近红外反射光谱, 能够较快速且准确的识别水稻种子存活率, 为以后的基于带通滤波片的快速种子存活率识别奠定了基础。

The seed vigor was greatly affected by the storage condition. This work collected seeds that were affected by different storage conditions under real circumstance, and verified their germination rate difference by germination experiment. After that, we took some samples from them and measured the visible-near-infrared reflection spectra of each single grain. By adapting different spectral preprocessing techniques, combined with several machine learning modeling methods, we tried to distinguish the seeds according to their germination rates. In our experiments, some spectral pretreatment methods were compared, such as standard reflection spectrum correction. It also compared supervised machine learning modeling method such as support vector machine, K-near neighbor and discriminant analysis. From the perspective of recognition accuracy, we believed that standard reflection spectrum correction method can greatly improve the spectral difference of seeds with different survival rates, and thus achieve higher recognition accuracy through machine learning. At the same time, we compared the supervised machine learning modeling methods such as support vector, k near neighbor and distance discriminant analysis, and find that the standard reflection spectrum correction method combined with distance discriminant analysis can achieve hundred-percent accuracy of predicting different seeds category. Furthermore, in order to meet the requirements of rapid identification in practical applications, in the experiment we compressed high-resolution spectral data into low-resolution multi-channel band-pass spectral data, which can greatly reduce the spectral data length and save the time spent in training and classifying of various machine learner. The prediction accuracy of models trained by those simplified spectra data is still close to 90%. It fully demonstrates that the use of multi-channel broadband spectral data combining with the selection of appropriate machine learning modeling methods is sufficient to meet the general needs of the actual seed selection industry, and it is a potential technique for rapid identification of rice grain survival rates in the future. The experiment also used a variety of bandpass widths to simplify the spectra, and analyze and compare the effects of different bandpass widths on the recognition accuracy. In general, due to the increase in bandwidth, the length of data is reduced, and the recognition speed is faster, but the recognition accuracy is decreased. In the experiment, we changed the spectral bandwidth from 10 to 50 nm, and the recognition accuracy of the simplified spectrum after standard reflection correction decreased from 87.50% to 58.75%. In practical use, it is necessary to balance the recognition rate and the expected recognition accuracy, and select a reasonable bandwidth. This study verified that the simplified near-infrared reflectance spectroscopy can quickly and accurately identify the survival rate of rice seeds, which lays a foundation for rapid seed survival rate recognition technique based on bandpass filters in the future.