近红外高光谱成像结合特征波长筛选识别小麦赤霉病瘪粒

沈广辉 曹瑶瑶 刘馨 徐剑宏 史建荣 LEE Yin-won

摘要： 为实现小麦赤霉病瘪粒快速识别，本研究使用主成分分析（Principal component analysis， PCA）结合最大类间方差法（Otsu）对小麦高光谱图像进行背景分割，以赤霉病瘪粒识别正确率为评价指标，探究判别分析方法与竞争性自适应权重取样法（Competitive adaptive reweighted sampling， CARS）的最佳组合方式。结果显示，基于全谱段构建的偏最小二乘判别分析（Partial least squares discrimination analysis， PLS-DA）和支持向量机判别分析（Support vector machine discriminant analysis， SVM-DA）模型预测精度相同，外部验证集健康籽粒和赤霉病瘪粒识别正确率分别为95.2%和100.0%;基于CARS筛选的8个特征波长构建的CARS-PLS-DA模型外部验证集健康籽粒和赤霉病瘪粒识别正确率均为100.0%，预测精度高于CARS-SVM-DA模型，可有效实现赤霉病瘪粒的快速识别。研究结果将为谷物仓储和加工过程中赤霉病瘪粒高通量快速识别提供理论依据和技术支撑。

关键词： 高光谱成像;赤霉病瘪粒;近红外光谱;无损检测

中图分类号： S123;TP391.41  文献标识码： A  文章编号： 1000-4440（2021）02-0509-08

Abstract： In order to realize rapid identification of unfilled grain from wheat infected by Fusarium， principal component analysis （PCA） combined with Otsu algorithm was used for background segmentation of wheat hyperspectral imaging. The compound mode of discriminant analysis method and competitive adaptive reweighted sampling （CARS） method were optimized based on the identification accuracy of Fusarium damaged kernels. The results indicated that， the predication accuracy of partial least squares discrimination analysis （PLS-DA） model and support vector machine discriminant analysis （SVM-DA） model constructed based on full spectrum were the same， and the recognition accuracy of healthy and Fusarium damaged kernels in the external validation set were 95.2% and 100.0%， respectively. The recognition accuracy of healthy and Fusarium damaged kernels were both 100.0% in the external validation set of CARS-PLS-DA model which was built based on eight characteristic wavelengths selected by CARS algorithm， and the prediction accuracy was higher than CARS-SVM-DA model， and could rapidly identify Fusarium damaged kernels effectively. The results can provide theoretical basis and technical support for the high throughput and rapid detection of Fusarium damaged kernels during grain storage and processing.

Key words： hyperspectral imaging;Fusarium damaged kernels;near infrared spectroscopy;non-destructive detection

小麥赤霉病是由亚洲镰刀菌和禾谷镰刀菌侵染引发的真菌病害，赤霉病发生过程中会产生有毒次级代谢产物脱氧雪腐镰刀菌烯醇（Deoxynivalenol， DON），又称呕吐毒素，其不仅会破坏小麦的细胞组织结构，降低出粉率，误食被DON污染饲料的家畜会导致呕吐、拒食、腹泻、出血甚至死亡，对人体也有较大毒性[1-3]。近年来，受极端气候等环境因素影响，小麦赤霉病爆发频率增加，DON污染风险不断加剧，已成为制约中国及世界小麦产品质量安全的主要风险因子[4-5]。DON结构性质稳定，在谷物加工过程中难以消除，中国规定小麦及全麦粉中DON的最大残留限量标准为1 000 μg/kg。……