摘要
对感染黑斑病的刺五加叶片进行光谱特性研究,能为药用植物病害的早期筛选与精准治疗提供重要研究资料。实验目的,运用高光谱成像技术实现植物病害的自动监督分类与识别。实验过程,首先使用高光谱成像系统在可见光波段(380~960 nm)内采集刺五加黑斑病的叶片样本,光谱数据经过去除亮暗噪声和平滑预处理后,再经过主成分分析实现数据降维,继而运用基于不同核函数的支持向量机法建立分类模型,最后利用总体分类精度、Kappa系数等因子评价不同核函数对分类器性能的影响。根据叶片表面的特征将其分为四类样本:健康亮部、健康暗部、轻度病害和重度病害等。对比各类样本的光谱可知,刺五加的健康样本在540 nm波长存在一个明显峰值,在620~680 nm光谱曲线急剧上升;而病害样本的光谱反射率呈现缓慢且平稳的上升趋势,上述特征能够将图像空间上反射强度接近的健康亮部和严重病害完全区分开。经对比发现前四个主成分(PC1,PC2,PC3,PC4)在分类表达上存在差异,主要表现为PC1含有的信息多,能够较好地区分各类样本;PC2则出现健康亮部和严重病害的交叉混淆;PC3是对于PC2的补充,能基本完整地表达轻微病害;PC4的贡献率仅有0.19%,依然能够准确地识别严重病害。不同主成分分量在表达各类样本特征中存在的差异能够作为复杂样本分类的参考依据。对比四种核函数对支持向量机分类器性能的影响,结果显示线性核函数的识别过程受光强反射的影响较大,Sigmoid核函数的训练精度易受数据集大小的影响,在识别健康亮或暗,以及轻微病害上均存在一定的误差,多项式核函数与径向基核函数的效果较好,其中,多项式核函数的精度更高,为92.77%。研究表明,利用高光谱成像技术能够准确地识别刺五加的健康叶片和患病叶片,为实现自动诊断药用植物叶片病害提供新方法。
Taking the leaves of Acanthopanax acanthopanax infected with black spot disease as an example,the study of plant disease detection by spectral technology provides a research basis for early screening and precise treatment of medicinal plant diseases.The experiment aimed to realize the supervised classification and identification of plant diseases by hyperspectral imaging technology.The experimental procedure is as follows:First,the leaf samples of Acanthopanax japonicus were collected using the hyperspectral imaging technique.After the spectral data were preprocessed by removing light and dark noise and smoothing,the data dimension was reduced using principal component analysis.Then,a support vector machine(SVM)based on different kernel functions was used to establish a classification model separately.Finally,the overall classification accuracy,Kappa coefficient and other factors were used to evaluate different kernel functions’influence on the classifier performance.According to the leaf’s surface characteristics,the leaf was divided into four kinds of samples:healthy bright part,healthy dark part,mild disease and severe disease.It can be seen that the healthy sample of Acanthopanax senticosus had a significant peak at 540 nm,and the spectral curve rose sharply at 620~680 nm;while the spectral reflectance of disease samples showed a slow and steady rising trend.The above features could completely distinguish healthy samples with close reflection intensity from serious disease samples on the image.After comparison,it was found that the first four principal components(PC1,PC2,PC3,PC4)have certain differences in the classification results.The main differences were that PC1 contains much information and can better distinguish various samples;PC2 showed a cross-confusion between bright healthy samples and seriously diseased samples;PC3 was a supplement to PC2,which can find mild diseased areas;PC4 contribution rate was only 0.19%,and it could still accurately identify serious diseased areas.The differences of pri
作者
赵森
付芸
崔江南
鲁烨
杜旭东
李永亮
ZHAO Sen;FU Yun;CUI Jiang-nan;LU Ye;DU Xu-dong;LI Yong-liang(School of Electro-Optical Engineering,Changchun University of Science and Technology,Changchun 130022,China)
出处
《光谱学与光谱分析》
SCIE
EI
CAS
CSCD
北大核心
2021年第6期1898-1904,共7页
Spectroscopy and Spectral Analysis
基金
国家自然科学基金项目(61675035)
吉林省重点科技计划项目(20170204015GX,20180201049YY)资助。
关键词
可见光谱
植物病害
支持向量机
核函数
Visible spectrum
Plant diseases
Support Vector Machine
Kernel function
