为探讨利用可见/近红外光谱进行小麦黑胚病快速无损检测的可行性, 以及基于主流机器学习算法, 寻找面向生产的小麦黑胚病优化识别模型, 利用自行研发的近红外光谱采集平台采集了579~1 099 nm波段23个品种共2760个小麦单籽粒的吸光度光谱数据, 采用标准正态变量变换(SNV)进行预处理之后分别经过SPA(successive projections algorithm), PCA(principal component analysis)等两种数据降维方法, 结合ELM(extreme learning machine), SVM(support vector machine), RF(random forest)和AdaBoost等四种分类方法, 分别构建SPA-SVM, SPA-ELM, SPA-RF, SPA-AdaBoost, PCA-SVM, PCA-ELM, PCA-RF, PCA-AdaBoost八种小麦黑胚病识别模型; 结果表明小麦黑胚籽粒的识别准确率达到93.3%~98.6%, 识别效果优于前人文献中利用近红外波段的识别效果; 其中SPA-SVM模型具有最高的识别率, PCA-AdaBoost模型具有更好的普适性。 将SPA-SVM模型和PCA-AdaBoost模型作为优选模型, 从生产实际出发, 分别对未感病+轻感病、 中感病+重感病籽粒进行了二分类识别, 对未感病, 轻感病+中感病、 重感病籽粒进行了三分类识别, 以及对未感病、 轻感病、 中感病、 重感病籽粒进行了四分类识别, 并深入分析了识别效果和产生原因。 总体来说, 小麦黑胚粒的识别准确率随分类程度的细化而下降, 二分类的识别模型可直接用于生产, 尽管三分类和四分类的感病粒识别效果较差, 但是对未感病粒的检出率则不受分类程度的影响, 识别率在87.2%以上, 符合生产需求。 综合来看, SPA-SVM模型分类效果优于PCA-AdaBoost模型, 可作为首选识别模型, 该研究为小麦籽粒黑胚病的在线批量快速检测提供了技术依据。

In order to explore the feasibility of detecting wheat kernel black tip (BT) disease and investigating an optimized classification model based on mainstream machine learning algorithms, a large amount of 2 760 wheat kernels spectral data of Vis/NIR bands (579~1 099 nm) were collected by self-made spectral acquisition platform. After pretreated with standard normal variate correction (SNV) of 600~1 045 nm bands, 7 kinds of data sets were established. Successive Projections Algorithm (SPA) and Principal Component Analysis (PCA) of spectral data dimensionality reduction methods, and four machine learning algorithms, Support Vector Machine (SVM), Extreme Learning Machine (ELM), Random Forest (RF) and AdaBoost, were adopted to develop eight classification models. Results showed that Vis/NIR spectrums combined with all the machine learning methods could be used to detect BT disease with accuracies ranging from 93.3% to 98.6%, which indicated that Vis/NIR would be the more effectively compared to NIR. As SPA-SVM possessed a high average classification accuracy and PCA-AdaBoost showed better generalization performance than other algorithms, considering practical purposes, these two algorithms were adopted as optimized models in 2-category classification, 3-category classification and 4-category classification for various degrees of BT detection. Results indicated that the classification accuracies declined gradually with the classification number increasing, but the detection accuracy of non-diseased wheat kernel tended to be stable with an accuracy of more than 87.2%. Taken together, SPA-SVM performed better than PAC-AdaBoost in wheat BT disease detection. The models and conclusions of this research are intended to lead to the streamlining of VIS/NIR spectroscopy in automated wheat black tip inspection as well as to provide criteria for high speed sorting.