基于高光谱和参数优化支持向量机的水稻施氮水平分类研究
作者:
作者单位:

(1.江西农业大学计算机与信息工程学院,江西南昌330045;2.江西农业大学软件学院 /江西省高等学校农业信息技术重点实验室,江西南昌330045)

作者简介:

罗建军(1995-),男,江西高安人,硕士,主要从事基于机器学习的水稻氮素营养诊断。E-mail:ljj1781891045@163.com。

通讯作者:

基金项目:

基金项目:国家自然科学基金项目(No.61562039,No.61762048,No.61862032);江西省教育厅科技项目(GJJ160374,GJJ170279)。


Classification of nitrogen application levels in rice based on hyperspectral and parameter optimized support vector machine
Author:
Affiliation:

(1.College of Computerand Information Engineering,Jiangxi Agricultural University,Nanchang 330045;2. College of Software,JiangxiAgricultural University/Key Laboratory of Agricultural Information Technology,Jiangxi Higher Education,Nanchang330045)

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
    摘要:

    为探索水稻氮素营养的快速、无损诊断方法以及构建基于高光谱技术的水稻氮素营养状况分类识别模型。本研究以 4种不同施氮水平的“中嘉早 17”水稻分蘖期顶部第三完全展开叶叶片(简称顶三叶)为研究对象,测定各叶片的可见光到近红外波段(350~ 2500 nm)内的光谱数据,对所获取的光谱数据进行平滑处理和归一化处理,以消除噪声及量纲的影响,并采用主成分分析(PCA)的方法进行数据降维至 22维,同时分别选用基于网格搜索算法、粒子群算法和遗传算法优化参数的支持向量机进行水稻氮素营养状况分类识别模型的建立。研究结果表明:1)不同施氮水平下的水稻叶片光谱反射率曲线走势大致相同,但不同施氮水平下 780~ 1 300、1 400~ 1 850及 1 900~ 2500 nm波段光谱反射率存在一定的差别;2)优化参数后的 SVM模型与默认参数下的 SVM模型相比,其训练集与测试集分类识别效果都要优于默认参数下的 SVM模型。其中,以遗传算法优化参数的 SVM模型识别分类效果最佳,训练集和测试集识别准确率分别为 99.375%、98.750%,测试集的4种施氮水平(施氮量从低到高)识别准确率分别为 100%、95%、100%和 100%。结果表明利用高光谱技术能够很好地进行水稻氮素营养状况的定性诊断研究。为快速水稻氮素营养诊断提供了一种新途径,为精确施氮提供了技术支撑和理论依据。

    Abstract:

    In order to explore a rapid and non-destructive diagnosis method of rice nitrogen nutrition and to construct aclassification and recognition model of rice nitrogen nutrition status based on hyperspectraltechnology,the top three fully expanded leaves of the“Zhongjiazao 17”rice at the tillering stage of four different nitrogen application levels are taken as theresearchobject,and the visible light to the near -infraredband(350~2 500nm)of each leaf ismeasured. The acquiredspectral data are smoothed and normalized to eliminate the effects of noise and dimension,and the principal componentanalysis(PCA)method is used to reduce the data to 22 dimensions. Support vector machines are selected based on gridsearch algorithm,particle swarm optimization algorithm and genetic algorithm to establish a classification and recognitionmodel of rice nitrogen nutrition status. The results showthat:1)the spectral reflectance curves of rice leaves underdifferent nitrogen application levels are roughly the same,but there are some differences in the spectral reflectance of the 780~1 300nm,1 400~ 1 850 nm and 1 900~2 500 nm bands under different nitrogen applicationlevels;2)Comparedwith the SVM model with default parameters,the SVM model with optimized parameters has better training and test setclassification and recognition performance than the SVM model with defaultparameters. Amongthem,the SVM model withgenetic algorithm optimized parameters has the best recognition and classificationeffect. The recognition accuracy of thetraining set and the test set is 99.375% and98.750%,respectively. The four nitrogen levels(low tohigh)of the test setareidentified. The accuracy rates are100%,95%,100%,and100%. The results show that the use of hyperspectraltechnology can well carry out qualitative diagnosis of nitrogen status inrice. It provides a new way for rapid nitrogen nutritiondiagnosis ofrice,and provides technical support and theoretical basis for accurate nitrogenapplication.

    参考文献
    相似文献
    引证文献
引用本文

罗建军,杨红云,路艳,万颖,孙爱珍,易文龙.基于高光谱和参数优化支持向量机的水稻施氮水平分类研究[J].中国土壤与肥料,2020,(5):250-257.

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
历史
  • 收稿日期:2019-11-17
  • 最后修改日期:
  • 录用日期:2020-02-02
  • 在线发布日期: 2020-11-06