提出了基于高光谱信息的大白菜种子品种分类识别方法。 利用近红外高光谱图像采集系统采集了八种共239个大白菜种子样本; 提取15 pixel 15 pixel感兴趣区域平均光谱反射率信息作为样本信息; 采用多元散射校正预处理方法对光谱进行消噪; 验证了Ada-Boost 算法、 极限学习机(extreme learning machine, ELM)、 随机森林(random forest, RF)和支持向量机(support vector machine, SVM)四种分类算法的分类判别效果。 为了简化输入变量, 通过载荷系数分析选取了10个大白菜种子品种分类判别的特征波长。 实验结果表明, 四种分类算法基于全波段的分类识别对81个预测样本的正确区分率均超过90%, 最优的分类判别模型为ELM和RF, 识别正确率达到了100%; 以10个特征波长的分类判别精度略有下降, 但输入变量大幅减少, 提高了信息处理效率, 其中最优分类判别模型为EW-ELM模型, 判别正确率为100%, 因此以载荷系数选取的特征波长是有效的。 利用高光谱结合机器学习对大白菜种子品种进行快速、 无损分类识别是可行的, 为大白菜种子批量化在线检测提供了一种新的方法。

The variety of Chinese cabbage seeds were recognized using hyperspectral imaging with 256 bands from 874 to 1 734 nm in the present paper. A total of 239 Chinese cabbage seed samples including 8 varieties were acquired by hyperspectral image system, 158 for calibration and the rest 81 for validation. A region of 15 pixel 15 pixel was selected as region of interest (ROI) and the average spectral information of ROI was obtained as sample spectral information. Multiplicative scatter correction was selected as pretreatment method to reduce the noise of spectrum. The performance of four classification algorithms including Ada-boost algorithm, extreme learning machine (ELM), random forest (RF) and support vector machine (SVM) were examined in this study. In order to simplify the input variables, 10 effective wavelengths (EMS) including 1 002, 1 005, 1 015, 1 019, 1 022, 1 103, 1 106, 1 167, 1 237 and 1 409 nm were selected by analysis of variable load distribution in PLS model. The reflectance of effective wavelengths was taken as the input variables to build effective wavelengths based models. The results indicated that the classification accuracy of the four models based on full-spectral were over 90%, the optimal models were extreme learning machine and random forest, and the classification accuracy achieved 100%. The classification accuracy of effective wavelengths based models declined slightly but the input variables compressed greatly, the efficiency of data processing was improved, and the classification accuracy of EW-ELM model achieved 100%. ELM performed well both in full-spectral model and in effective wavelength based model in this study, it was proven to be a useful tool for spectral analysis. So rapid and nondestructive recognition of Chinese cabbage seeds by hyperspectral imaging combined with machine learning is feasible, and it provides a new method for on line batch variety recognition of Chinese cabbage seeds.