柑橘黄龙病检测的近红外光谱集成建模方法  

Near-infrared Spectroscopy Based Ensembel Modeling Method for Huanglongbing Detection

作　　者：贺胜晖 李灵巧[1,2] 刘彤 刘振丙 杨辉华[1,2] HE Shenghui;LI Lingqiao;LIU Tong;LIU Zhenbing;YANG Huihua(Guilin University of Electronic Technology,Guilin 541004;Beijing University of Posts and Telecommunications,Beijing 100876;Guangzhou Xundong Network Technology CO.,Ltd,Guangzhou 510000)

机构地区：[1]桂林电子科技大学,广西桂林541004 [2]北京邮电大学,北京100876 [3]广州讯动网络科技有限公司,广东广州510000

出　　处：《分析科学学报》2020年第2期287-290,共4页Journal of Analytical Science

基　　金：国家自然科学基金(No.21365008,61562013)。

摘　　要：针对黄龙病检测问题,提出了一种集成了多特征提取模型和多分类器的柑橘黄龙病检测算法。将谱回归核判别分析和主成分分析并行融合进行特征提取,将偏最小二乘判别分析、决策树和支持向量机利用Stacking策略融合完成分类任务。基于3个主要柑橘品种共1620条近红外光谱数据,与单特征提取单分类器方法和多特征提取单分类器方法进行对比,集成分类模型的正确率可达98.52%,精度在98.57%以上,F2得分可达98.01%。实验结果表明,集成分类模型明显优于单特征提取单分类模型和多特征提取单分类模型,证明利用集成分类模型进行柑橘黄龙病的无损检测是可行的,为其他领域的光谱分类提供参考。To improve the robustness of algorithm,a citrus huanglongbing detection algorithm which ensemble multi-feature extraction model and multi-classification is proposed in this paper.The kernel discriminant analysis of spectral regression and principal component analysis are merged to extract features in parallel.The partial least squares discriminant analysis,decision tree and support vector machine are merged for classification task by Stacking method.Based on a total of 1620 near infrared spectroscopy from 3 main citrus varieties,the proposed method is compared with single-classifier method by single-feature extraction and single-classifier method by multi-feature extraction.Accuracy of ensemble classification model is 98.52%,precision of ensemble classification model can reach 98.57%,and F2-score of ensemble classification model is 98.01%.The experiment result show that ensemble classification model is significantly better than single-classifier method by single-feature extraction and single-classifier method by multi-feature extraction.It is proved that the non-destructive detection of citrus Huanglongbing is feasible by the ensemble classification model.

关 键 词：集成学习 黄龙病 近红外光谱 谱回归核判别分析 Stacking策略 

分 类 号：O657.33[理学—分析化学]