基于Vis-NIR光谱的柑橘叶片黄龙病检测及其光谱特性研究

Detection of Citrus Greening Based on Vis-NIR Spectroscopy and Spectral Feature Analysis

黄龙病作为柑橘类水果最具毁灭性的疾病之一,目前尚无有效的治愈手段,因此疾病预防成为已知的唯一有效方法。基于四种柑橘叶片(健康叶片、黄龙病叶片、铁缺乏叶片及氮缺乏叶片)VIS-NIR的反射光谱详细讨论了黄龙病的辨别方法以及在判别模型中光谱特征值的提取方法。在两类判别分析的特征值提取方法中,判别值(discriminability)运算的引入,为特征值提取提供了一个可靠依据,判别值越大表明光谱差异性越大。以被选特征值建立的Fisher线性判别分析模型,黄龙病与健康、铁缺乏、氮缺乏叶片的分类判别预测准确率分别都超过了90%,分类效果符合预期。最后,又讨论了分类树(classificationTree)在多类判别中的应用。通过对柑橘叶片原始反射谱,一阶导数谱及被选特征值分别建立分类模型,四种柑橘叶片平均预测准确度都超过88%,尤其是基于特征值的分类结果更是超过94%,验证了在多类判别中检测柑橘黄龙病的可行性及特征值提取的重要性。结合传统分类方法(k-NN,Bayesian)的结果分析,特征值作为输入变量的分类结果明显要优于原始光谱,证实了特征值选取的正确性,并为将来基于光谱特征值开发多光谱成像技术检测黄龙病打下坚实的基础。

In the present paper we discussed the methods of classification of citrus greening and extraction of spectral features based on the spectral reflectance of four different statuses of citrus leaves (healthy ,HLB ,iron deficiency and nitrogen deficien-cy) .Between two classes of classification ,the values of discriminability of different spectra were calculated to extract spectral features .The greater value of discriminability showed a bigger difference of the two spectra ,which means it would be easier to distinguish the two classes .By the Fisher linear discriminant analysis ,three classification models (HLB & healthy ,HLB & iron deficiency and HLB & nitrogen deficiency ) based on the spectral features yielded more than 90% accuracies ,which were better than expected .And at last ,we discussed the application of the classification tree in multi-class discriminant analysis and spectral features extraction .The models trained based on the original reflectance spectra ,first derivative and selected spectral features yielded more than 88% average accuracy ,and especially the model based on the spectral features yielded more than 94% average accuracies ,which verified the feasibility of detection of citrus greening in multi-class discriminant analysis and the importance of the spectral feature extraction .The results were compared based on classification tree ,k-NN and Bayesian classifiers .Adoption of spectral features as input variables was significantly superior to using the original spectrum ,which confirmed the validity of spectral feature selection .Spectral features could be used well for developing a multi-spectral imaging system to detect the citrus g reening .