叶片多理化参数的高光谱遥感与深度学习估算  

Estimation of Leaf Physical and Chemical Parameters Based on Hyperspectral Remote Sensing and Deep Learning Technologies

作　　者：岳继博 冷梦蝶 田庆久[2] 郭伟[1] 刘杨[3] 冯海宽[4] 乔红波 YUE Ji-bo;LENG Meng-die;TIAN Qing-jiu;GUO Wei;LIU Yang;FENG Hai-kuan;QIAO Hong-bo(College of Information and Management Science,Henan Agricultural University,Zhengzhou 450002,China;International Institute for Earth System Science,Nanjing University,Nanjing 210023,China;Key Lab of Smart Agriculture System,Ministry of Education,China Agricultural University,Beijing 100083,China;Key Laboratory of Quantitative Remote Sensing in Agriculture of Ministry of Agriculture China,Beijing Research Center for Information Technology in Agriculture,Beijing 100097,China)

机构地区：[1]河南农业大学信息与管理科学学院,河南郑州450002 [2]南京大学国际地球系统科学研究所,江苏南京210023 [3]中国农业大学智慧农业系统集成研究教育部重点实验室,北京100083 [4]北京市农林科学院信息技术研究中心农业农村部农业遥感机理与定量遥感重点实验室,北京100097

出　　处：《光谱学与光谱分析》2024年第10期2873-2883,共11页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(42101362);河南省科技攻关计划项目(232102111123)资助。

摘　　要：准确的植物叶片理化参数对于监测植物生长状况至关重要。随着深度学习技术的迅速应用,结合深度学习和高光谱遥感技术的植物叶片理化参数分析应用潜力巨大;然而,现阶段结合深度学习和高光谱遥感技术在植物多叶片理化参数的联合估算研究尚少。该研究旨在挖掘结合高光谱遥感技术和深度学习技术开展高精度的多植被叶片理化参数(叶绿素含量、类胡萝卜素含量、水分含量、蛋白质含量和碳基成分含量)联合估算的潜力。首先,通过利用新型PROSPECT-PRO辐射传输模型模拟分析,确定了多个植被叶片理化参数的敏感光谱区域,并设计了LeafTraitNet模型;然后,基于Lopex93数据开展LeafTr aitNet模型训练和验证,取得了高精度的叶片参数估算结果。得到以下结论:(1)基于PROSPECT-PRO辐射传输模型辅助实测开展植被光谱特征选择十分必要;叶绿素(约434和约676 nm)和类胡萝卜素(约445 nm)两类色素的光谱吸收峰均主要位于可见光区域;然而,其与叶片光谱的相关系数绝对值最大的点却不是各自的吸收峰位置,这可能是因为叶绿素和类胡萝卜素对光谱吸收的相互影响。(2)水分的吸收峰主要位于950~2500 nm范围内,这与叶片蛋白质和碳基成分的吸收区域重叠,因此削弱了后者的高光谱遥感估算精度。基于PROSPECT-PRO辐射传输模型和Lopex93数据集的叶片参数相关性分析结果表明,叶片水分含量与950~2500 nm范围内叶片光谱反射率相关系数绝对值接近1,而叶片蛋白质和碳基成分含量与950~2500 nm范围内光谱反射率相关系数较低。(3)三种传统统计回归方法和LeafTraitNet模型的叶片理化参数估算精度可以基于其估算总nRMSE而排序为:Le afTraitNet(总nRMSE=0.84)<RF(总nRMSE=1.59)<MLP(总nRMSE=1.73)<MLR(总nRMS E=1.74)。研究显示基于深度学习的LeafTraitNet模型有潜力提供远高于传统统计回归模型的植物叶片理化参数估算结Plant leaf physical and chemical parameters,such as leaf chlorophyll content,carotenoid content,water content,protein content,and Carbone-based constituents content,are crucial for accurately monitoring plant growth status.In recent years,with the rapid development of deep learning technology in vegetation remote sensing,the combined use of deep learning and hyperspectral remote sensing for plant leaf parameters estimation has been widely applied;however,currently,few leaf parameters estimation works based on the combination of deep learning and hyperspectral remote sensing technology have been conducted.This study explores the possibility of estimating leaf chlorophyll,carotenoid,water,protein,and Carbone-based constituent content by combining hyperspectral remote sensing and deep learning techniques.The main work of this paper is to propose a leaf physical and chemical parameter estimation method based on hyperspectral remote sensing and deep learning.Firstly,this study determines the sensitive spectral regions of multiple vegetation leaf physical and chemical parameters based on the PROSPECT-PRO radiative transfer model.Then,we designed a LeafTraitNet deep learning model;the LeafTraitNet model is trained and tested based on the lobex93 dataset,and a high-precision leaf parameter estimation result is obtained.The conclusions of this study are as follows:(1)It is vital to select leaf spectral absorption features based on the PROSPECT-PRO radiative transfer model.The leaf chlorophyll(434 and 676 nm)and carotenoids(445 nm)spectral absorption regions are located in the visible bands.However,the absorption regions with the most significant correlation coefficients(absolute values)are not their maximum spectral absorption bands,which the mutual influence of leaf chlorophyll and carotenoid absorptions may cause.(2)The leaf water spectral absorption regions are mainly located in the bands 950~2500 nm,which overlaps with the spectral absorption regions of leaf protein and carbon-based component content,thus weakening th

关 键 词：深度学习 高光谱遥感 叶片蛋白质含量 叶片叶绿素含量 叶片类胡萝卜素含量 

分 类 号：S127[农业科学—农业基础科学]