欢迎访问林业科学,今天是

林业科学 ›› 2023, Vol. 59 ›› Issue (7): 78-88.doi: 10.11707/j.1001-7488.LYKX20220577

• 研究论文 • 上一篇    下一篇

基于多光谱成像与机器学习的植物叶绿素含量反演

范学星1(),张慧春1,2,*(),邹义萍3,4,黄玉萍1,边黎明3   

  1. 1. 南京林业大学机械电子工程学院 南京 210037
    2. 南京林业大学林业资源高效加工利用协同创新中心 南京 210037
    3. 南京林业大学林学院 南京 210037
    4. 江苏青好景观园艺有限公司 南京 211225
  • 收稿日期:2022-08-28 出版日期:2023-07-25 发布日期:2023-09-08
  • 通讯作者: 张慧春 E-mail:1362454289@qq.com;njzhanghc@hotmail.com
  • 基金资助:
    国家自然科学基金项目(32171790,32101615,32171818);江苏省农业科技自主创新资金项目(CX[23]3126);江苏省重点研发计划现代农业项目(BE2021307);江苏省333工程项目(苏人20186)

Inversion of Plant Chlorophyll Content Based on Multispectral Imaging and Machine Learning

Xuexing Fan1(),Huichun Zhang1,2,*(),Yiping Zou3,4,Yuping Huang1,Liming Bian3   

  1. 1. College of Mechanical and Electronic Engineering, Nanjing Forestry University Nanjing 210037
    2. Co-Innovation Center of Efficient Processing and Utilization of Forest Resources, Nanjing Forestry University Nanjing 210037
    3. College of Forestry, Nanjing Forestry University Nanjing 210037
    4. Jiangsu Qinghao Ornamental Horticulture Co., Ltd. Nanjing 211225
  • Received:2022-08-28 Online:2023-07-25 Published:2023-09-08
  • Contact: Huichun Zhang E-mail:1362454289@qq.com;njzhanghc@hotmail.com

摘要:

目的: 利用多光谱图像光谱信息快速准确检测植物叶绿素含量,为植物生长监测、胁迫诊断与精确管理提供参考依据和技术指导。方法: 以冬青科中1年生苗期的大别山冬青和北美冬青为研究对象,构建一套基于多光谱相机RedEdge-MX近端提取植物表型信息的系统,采集成熟期和生长期2个品种冬青叶片的蓝色、绿色、红色、近红外、红边5个波段的多光谱图像,处理得到每个叶片各波段处的光谱反射率。将得到的光谱反射率与使用手持式叶绿素含量测定仪测得的叶绿素相对含量(SPAD)进行相关性分析,采用传统支持向量回归(SVR)算法与网格搜索算法(GS)、遗传算法(GA)和粒子群算法(PSO)进行反演建模,对比得到拟合度最高的反演模型,实现利用多光谱图像光谱信息快速准确检测植物叶绿素含量。结果: 对比传统SVR算法与优化后GS-SVR算法、GA-SVR算法和PSO-SVR算法得到反演模型的光谱反射率与SPAD相关性,其模型拟合度分别为 $ {R}_{1}^{2} $ =0.24,均方根误差RMSE1=0.160; $ {R}_{2}^{2} $ =0.72, $ {\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}_{2} $ =0.097; $ {R}_{3}^{2} $ =0.84, $ {\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}_{3} $ =0.073; $ {R}_{4}^{2} $ =0.91, $ {\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}_{4} $ =0.066。校正决定系数 ${R}_{\mathrm{adjusted1}}^{2}$ =0.23,平均绝对误差MAE1=0.119; ${R}_{\mathrm{adjusted2}}^{2}$ =0.71, $ {\mathrm{M}\mathrm{A}\mathrm{E}}_{2} $ =0.069; ${R}_{\mathrm{adjusted3}}^{2}$ =0.83, $ {\mathrm{M}\mathrm{A}\mathrm{E}}_{3} $ =0.050; ${R}_{\mathrm{adjusted4}}^{2}$ =0.87, $ {\mathrm{M}\mathrm{A}\mathrm{E}}_{4} $ =0.044。结论: 综合对比发现,优化后的PSO-SVR算法反演预测效果表现最佳。本研究通过采集植物多光谱图像,采用对比优化后的PSO-SVR算法得到5个波段光谱反射率与SPAD的最优反演模型,预测模型的准确性和鲁棒性增加,可以实现植物叶片叶绿素含量的快速检测。同时,本研究结果推广应用至遥感层面,可以实现对大面积区域内的植被叶绿素进行反演,能够为精确林业苗木生长监测、胁迫诊断与动态调控提供理论基础和技术支撑。

关键词: 多光谱图像, 叶绿素含量, 表型信息, 粒子群优化算法, 支持向量回归

Abstract:

Objective: Rapid and accurate detecting chlorophyll content in plant leaves is an important step to explore the photosynthesis, nitrogen nutrition, stress status and yield prediction. In this study, multispectral image spectral information was used to quickly and accurately detect plant chlorophyll content, in order to provide reference and technical guidance for plant growth monitoring, stress diagnosis, and precise management. Method: In this study, the 1-year-old seedlings of Ilex dabieshanensis and Ilex verticillata were targeted. A multispectral camera, RedEdge-MX, was used to construct a system for extracting plant phenotypic information. The system collected multispectral images of the leaves of the two species at maturity and growth stages in five bands, including blue (B), green (G), red (R), near infrared (NIR), and red edge (RedEdge). The multispectral images were processed to obtain the spectral reflectance at each band of each leaf. Correlation analysis was conducted between the obtained spectral reflectance and the relative chlorophyll content (SPAD value) obtained by using a hand-held chlorophyll content analyzer. The traditional support vector regression (SVR) algorithm and the grid search (GS) algorithm, genetic algorithm (GA) and particle swarm optimization (PSO) were used for inversion modeling, respectively. The inversion models were compared, and a model with the highest fitting degree was obtained and it was able to quickly and accurately predict plant chlorophyll content using multispectral image spectral information. Result: The results showed that the correlation between the spectral reflectance and SPAD was obtained with the inversion model selected by comparing the traditional SVR algorithm with the optimized GS-SVR algorithm, GA-SVR algorithm and the PSO-SVR algorithm, the model fitting degrees were $ {R}_{1}^{2} $ =0.24 and (root mean-square error) $ {\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}_{1} $ =0.160; $ {R}_{2}^{2} $ =0.72, $ {\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}_{2} $ =0.097; $ {R}_{3}^{2} $ =0.84, $ {\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}_{3} $ =0.073; $ {R}_{4}^{2} $ =0.91, $ {\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}_{4} $ =0.066, respectively. Secondly, (adjusted R-squared) ${R}_{\mathrm{adjusted1}}^{2}$ =0.23, (mean absolute error) $ {\mathrm{M}\mathrm{A}\mathrm{E}}_{1} $ =0.119; ${R}_{\mathrm{adjusted2}}^{2}$ =0.71, $ {\mathrm{M}\mathrm{A}\mathrm{E}}_{2} $ =0.069; ${R}_{\mathrm{adjusted3}}^{2}$ =0.83, $ {\mathrm{M}\mathrm{A}\mathrm{E}}_{3} $ =0.050; ${R}_{\mathrm{adjusted4}}^{2}$ =0.87, $ {\mathrm{M}\mathrm{A}\mathrm{E}}_{4} $ =0.044. Conclusion: After comprehensive comparison, it is found that the optimized PSO-SVR algorithm has the best inversion prediction effect. In this study, by collecting multispectral images of plants, the optimized PSO-SVR algorithm is used to obtain the optimal inversion model of spectral reflectance and SPAD in five bands, which can realize the rapid detection of chlorophyll content in plant leaves. Combined with the multispectral imaging and machine learning algorithms, the accuracy and robustness of the models have been improved. Meanwhile, the results of this study can be extended to remote sensing level to realize the inversion of vegetation chlorophyll in a large area, providing theoretical basis and technical support for accurate forest seedling growth monitoring, stress diagnosis and dynamic regulation.

Key words: multispectral image, chlorophyll content, phenotypic information, particle swarm optimization algorithm, support vector regression

中图分类号: