摘要
作物病害类型的快速无损检测对提高作物品质和产量至关重要。传统的病害分类方法费时费力且不能实时检测。为此,利用高光谱进行大豆病害分类。以健康大豆为对照,灰斑病和细菌性斑点病两种病害为研究对象,获取三种类别叶片高光谱数据。基于高光谱曲线分析病害与健康叶片反射率的变化规律。采用主成分分析(PCA)和光谱指数(SI)两种单一方法进行病害有效信息提取,共使用30个SI。在此基础上,提出一种PCA与SI相结合的组合方法(PCA-SI),通过提取有效主成分(PC)及有效SI,将有效SI按得分情况分为两组(9SIs和18SIs),再分别对应每一个有效PC进行分组,形成病害光谱有效信息的变量集。采用三种方法分别进行病害有效信息的提取,基于提取后的光谱变量,采用最小二乘支持向量机(LSSVM)和支持向量机(SVM)两种分类器建立病害分类模型。以原始高光谱为基准,以病害分类正确率为指标,评价模型的病害分类性能及不同病害有效信息提取方法和分类器的有效性。结果表明:高光谱反射率具有可见光450~700 nm波段范围病害叶片高于健康叶片而近红外760~1 000 nm波段范围其特征完全相反的规律。采用单一PCA方法提取出了34个有效PC用于病害分类。基于PCA-SI组合方法提取出5个有效PC(PC1—PC5)和18个有效SI,将其进行分组得到10组变量,共计13组变量作为建模集。三种方法提取病害有效信息后的光谱变量均比原始高光谱具有更好的病害分类能力,提出的PCA-SI组合方法具有最优的病害有效信息提取能力,PC1-18SIs和PC4-18SIs为最优建模集,LSSVM分类器具有最优的分类性能。PC1-18SIs-LSSVM和PC4-18SIs-LSSVM模型为最优病害分类模型,训练集和预测集的总病害分类正确率分别为100%和98.85%,与原始高光谱分类模型相比,总分类能力分别提高了6.47%和21.74%,模型分类能力较好,可为病害实时无损分类识别提供参考。
Rapid and non-destructive detection of crop disease types are essential to improve crop quality and yield.Traditional disease classification methods are time-consuming and difficult to detect in real-time.Therefore,the classification of soybean diseases was carried out by the hyperspectral technique.In this paper,healthy soybean was used as the control,frogeye leaf spot and bacterial blight diseases were the research objects,and hyperspectral data of three types of leaves were obtained.Changes inthe reflectance of diseased and healthy leaves were analyzed based on hyperspectral curves.Two single methods,principal component analysis(PCA)and spectral index(SI),were used to extract effective disease information.A total of 30 SI were used.A combination method of PCA and SI(PCA-SI)was proposed on this basis.Extracting the effective principal component(PC)and the effective SI,which were divided into two groups(9SIs and 18SIs)according to the score,and then grouped corresponding to each effective PC respectively to form the variable set of effective information of the disease spectrum.Three methods were used to extract effective disease information respectively.Based on the extracted spectral variables,the least square support vector machine(LSSVM)and support vector machine(SVM)was used to establish the disease classification model.With the original hyperspectral as the benchmark and the accuracy of disease classification as the index,the disease classification performance of the model,the effective information extraction methods of different diseases and the effectiveness of the classifier were evaluated.The results showed that the hyperspectral reflectance of diseased leaves was higher than that of healthy leaves in the visible band of 450~700 nm,while the characteristics of diseased leaves were opposite in the near-infrared band of 760~1000 nm.A single PCA method was used to extract 34 effective PCS for disease classification.Based on the PCA-SI combination method,5 effective PCs(PC1—PC5)and 18 effective SIs were extracted and grouped to obtain 10 groups of variables,and 13 groups of variables were used as modeling sets.The spectral variables extracted by the three methods have better disease classification ability than the original hyperspectral,and the proposed PCA-SI combination method has the optimal disease-effective information extraction ability.PC1-18SIs and PC4-18SIs were the best modeling sets,and the LSSVM classifier performed the best classification.PC1-18SIs-LSSVM and PC4-18SIs-LSSVM models were the optimal disease classification models.The total disease classification accuracy of the training and prediction sets was 100%and 98.85%,respectively.Compared with the original hyperspectral classification model,the overall classification ability of these two models was improved by 6.47%and 21.74%,respectively,and the model classification ability was good.It can provide a reference for real-time and non-destructive classification and identification of diseases.
作者
刘爽
于海业
隋媛媛
孔丽娟
于占东
郭晶晶
乔建磊
LIU Shuang;YU Hai-ye;SUI Yuan-yuan;KONG Li-juan;YU Zhan-dong;GUO Jing-jing;QIAO Jian-lei(College of Horticulture,Jilin Agricultural University,Changchun 130118,China;School of Biological and Agricultural Engineering,Jilin University,Changchun 130022,China;College of Engineering and Technology,Jilin Agricultural University,Changchun 130118,China)
出处
《光谱学与光谱分析》
SCIE
EI
CAS
CSCD
北大核心
2023年第5期1550-1555,共6页
Spectroscopy and Spectral Analysis
基金
国家自然科学基金青年科学基金项目(32001418)
吉林省重点科技研发项目(20200403140SF)
吉林省现代农业产业技术体系建设项目(201922)资助。
关键词
大豆
病害分类
高光谱
Soybean
Disease classification
Hyperspectral data