首页|高光谱成像的黄瓜病虫害识别和特征波长提取方法

高光谱成像的黄瓜病虫害识别和特征波长提取方法

Identification of Cucumber Disease and Insect Pest Based on Hyperspectral Imaging

扫码查看
黄瓜霜霉病和斑潜蝇是制约黄瓜产业发展的严重病虫害.为实现黄瓜病虫害快速在线识别,采用高光谱成像和机器学习研究快速识别黄瓜霜霉病和斑潜蝇虫害的方法,为开发实用的基于多光谱成像的黄瓜病虫害快速识别设备奠定基础.使用高光谱成像系统采集黄瓜无症状叶片、霜霉病叶片、斑潜蝇虫害叶片的高光谱图像,在病斑区域选择若干个感兴趣区域(ROI),计算每个ROI的平均反射率数据作为叶片原始光谱数据.使用Kennard-Stone算法将光谱数据按照3∶1的比例划分为训练集和测试集.使用直接正交信号校正(DOSC)、多元散射校正(MSC)、移动窗口平均平滑(MA)3种方法对原始光谱数据进行预处理.采用空间迭代收缩法(VISSA)、竞争性自适应重加权算法(CARS)、迭代保留信息变量法(IRIV)、随机蛙跳算法(SFLA)对MA预处理后的光谱数据进行特征波长提取,分别提取出53、20、26、10个特征波长.然后使用连续投影算法(SPA)分别对特征波长光谱数据进行二次降维,最终VISSA-SPA提取的特征波长为455、536、615和726 nm;CARS-SPA提取的特征波长为452、501、548和578 nm;IRIV-SPA提取的特征波长为452、513、543和553 nm;SFLA-SPA提取的特征波长为462、484、500和550 nm.分别对全波段光谱数据、一次降维光谱数据、二次降维光谱数据进行支持向量机(SVM)、Elman神经网络、随机森林(RF)建模,结果表明,MA预处理后的全波段光谱数据所建模型识别效果最好,其中MA-RF模型测试集总分类精度(OA)达到97.89%,Kappa系数为0.97.采用一次降维光谱数据所建模型中,MA-VISSA-RF模型效果最好,测试集OA为98.19%,Kappa系数为0.97.采用二次降维光谱数据所建模型中,MA-IRIV-SPA-SVM模型效果最好,测试集OA为96.23%,Kappa系数为0.95.研究结果表明,使用高光谱成像技术识别黄瓜霜霉病和斑潜蝇虫害具有良好的效果,452、513、543和553 nm可以作为识别黄瓜霜霉病和斑潜蝇虫害的特征波长,为开发黄瓜病虫害快速识别设备提供了理论依据.
Cucumber downy mildew and libria sativa are serious diseases and insect pests that restrict the development of the cucumber industry.In order to realize the rapid identification of cucumber diseases and insect pests,hyperspectral imaging technology and machine learning were used to explore the characteristic wavelengths of cucumber diseases and insect pests,which laid a foundation for the development of practical cucumber diseases and insect pests identification equipment based on multispectral imaging.This study used a hyperspectral imaging system to collect hyperspectral images of asymptomatic leaves,downy mildew leaves and leaf miner-infected leaves.According to the size of the leaf spot area,several regions of interest(ROI)were selected in the spot area,and the average reflectance data of each ROI was calculated as the original spectral data of the leaf.The samples were divided into training sets and test sets in a 3∶1 ratio by the Kennard-Stone algorithm.Direct orthogonal signal correction(DOSC),multiplicative scatter correction(MSC)and moving average(MA)were used to preprocess the original spectral data.Variable iterative space shrinkage approach(VISSA),competitive adaptive reweight sampling method(CARS),iteratively retains informative variables(IRIV)and shuffled frog leaping algorithm(SFLA)were used to extract characteristic wavelengths,respectively obtains 53,20,26,10 characteristic wavelengths.Then,the successive projections algorithm(SPA)was used to perform secondary dimensionality reduction on the characteristic wavelength data,and finally,the characteristic wavelengths extracted by VISSA-SPA were 455,536,615,and 726 nm.The characteristic wavelengths extracted by CARS-SPA were 452,501,548 and 578 nm.The characteristic wavelengths extracted by IRIV-SPA were 452,513,543 and 553 nm.The characteristic wavelengths extracted by SFLA-SPA are 462,484,500 and 550 nm.Support vector machine(SVM),Elman neural network and random forest(RF)modeling were carried out for the full-band and characteristic wavelength data.The results showed that the full band spectral data preprocessed by MA had the best recognition effect,in which the OA of the MA-RF model reached 97.89%and the Kappa coefficient was 0.97.The MA-VISSA-RF model had the best effect among the models built by the data of the characteristic wavelength first extracted,with 98.19%OA and 0.97 Kappa coefficient.MA-IRIV-SPA-SVM model had the best effect among the models built by quadratic dimensionality reduction data,with OA 96.23%and Kappa coefficient 0.95.The results showed that hyperspectral imaging technology had a good effect on the identification of cucumber downy mildew and the insect pest,and 452,513,543,553 nm could be used as the characteristic wavelength for identification of cucumber downy mildew and the insect pest,providing a theoretical basis for developing cucumber disease and insect pest identification equipment,providing a theoretical basis for the development of cucumber disease and insect pest rapid identification equipment.

Hyperspectral imagingMachine learningCharacteristic wavelengthsDowny mildewLeafminer-infected

李杨、李翠玲、王秀、范鹏飞、李余康、翟长远

展开 >

江苏大学农业工程学院,江苏镇江 212013

北京市农林科学院智能装备技术研究中心,北京 100097

国家农业智能装备工程技术研究中心,北京 100097

高光谱成像 机器学习 特征波长 霜霉病 斑潜蝇虫害

河北省重点研发计划项目北京市农林科学院科技创新能力建设专项国家自然科学基金面上项目

21327408DKJCX2021040231971775

2024

光谱学与光谱分析
中国光学学会

光谱学与光谱分析

CSTPCD北大核心
影响因子:0.897
ISSN:1000-0593
年,卷(期):2024.44(2)
  • 16