| 摘要: |
| 【目的】研究猕猴桃叶片叶绿素含量的高光谱估算方法，为猕猴桃长势的遥感监测提供理论依据。【方法】以陕西杨凌蒋家寨村2018年不同生育期（初花期、幼果期、膨果期、壮果期、果实成熟期）的猕猴桃叶片为研究对象，分别测定其高光谱反射率和叶绿素含量（SPAD值），分析原始光谱和5个常见的植被指数（归一化植被指数、归一化叶绿素指数、改进的叶绿素吸收反射率指数、MERIS地面叶绿素指数、土壤调整指数）与叶绿素含量之间的相关关系，提取各生育期的特征波段，分别建立基于特征波段和植被指数的单波段叶绿素含量一元线性估算模型。利用主成分分析对原始光谱数据进行降维，将得到的主成分得分作为随机森林模型的输入变量，建立基于多波段信息的叶绿素含量多元估算模型，并对模型进行精度验证和分析。【结果】不同生育期猕猴桃叶片光谱反射率变化趋势基本一致，整体趋势为可见光波段反射率低，近红外波段反射率高；在可见光波段，光谱反射率随着叶绿素含量的升高而降低；在近红外波段，光谱反射率则随着叶绿素含量的增加而升高。通过相关性分析可知，初花期、幼果期、膨果期、壮果期、果实成熟期原始光谱的特征波段分别为729，548，707，707和712 nm，估算模型决定系数（R2）分别为0.18，0.85，0.54，0.85和0.82，其中初花期估算模型未通过显著性检验，其余生育期均通过极显著性检验。在5个常用植被指数中，初花期与叶绿素含量相关性最高的是归一化叶绿素指数（NPCI），但是估算模型决定系数R2只有0.1，未通过显著性检验；其他生育期与叶绿素含量相关性最高的是MERIS地面叶绿素指数（MTCI），所建立的估算模型拟合效果好，预测精度高。基于主成分分析和随机森林回归建立的不同生育期猕猴桃叶片叶绿素含量估算模型的R2在0.91~0.98，均通过极显著性检验，其拟合效果和预测精度远高于单波段一元线性回归和基于植被指数的一元线性回归模型，是估算猕猴桃叶片叶绿素含量的最优模型。【结论】基于主成分分析的随机森林模型包含了更完整的波段信息，对不同生育期猕猴桃叶片叶绿素含量具有较好的预测能力。 |
| 关键词: 猕猴桃 叶绿素含量 高光谱遥感技术 主成分分析 随机森林模型 |
| DOI： |
| 分类号: |
| 基金项目:国家“863”高新技术研究与发展计划项目（2013AA102401） |
|
| Hyperspectral estimation model of chlorophyll content in kiwifruit leaves |
|
CHEN Lan,CHANG Qingrui,GAO Yifan,et al
|
| Abstract: |
| 【Objective】A method for hyperspectral estimation of chlorophyll content in kiwifruit leaves was established to provide basis for remote sensing monitoring of kiwifruit growth.【Method】Kiwifruit leaves in different growth stages (initial flowering stage,young fruit stage,fruit filling stage,strong fruit stage and fruit ripening stage) in Jiangjiazhai village,Yangling,Shaanxi were selected,and their hyperspectral reflectance was measured.Chlorophyll content (SPAD value),original spectrum,five common vegetation indices (normalized vegetation index,normalized chlorophyll index,improved chlorophyll absorption reflectance index,MERIS ground chlorophyll index,soil adjustment index) and chlorophyll content were obtained.The correlation between growth bands of each growth period was extracted,and a one-band linear estimation model of single band chlorophyll content based on characteristic bands and vegetation indices was established. Principal component analysis was used to reduce the dimension of the original spectral data.The principal component score was used as the input variable of the random forest model.The multi-band information chlorophyll content multivariate estimation model was then established to verify the accuracy of the model.【Result】The spectral reflectance of kiwifruit leaves in different growth stages was basically the same.The overall trend showed low reflectivity in the visible light band and high reflectance in the near-infrared band.In the visible light range,the spectral reflectance decreased with the increase of chlorophyll content.In the near-infrared region,the spectral reflectance increased as the chlorophyll content increased.Through the correlation analysis,the characteristic bands of the original spectrum of the initial flowering stage,the young fruit stage,the fruiting stage,the strong fruit stage and the fruit ripening stage were 729,548,707,707 and 712 nm,respectively,and the estimated model determination coefficient (R2) were 0.18,0.85,0.54,0.85 and 0.82,respectively.The initial flowering estimation model failed the significance test,and the rest passed the extremely significant test.Among the five commonly used vegetation indices,the normalized chlorophyll index (NPCI) had the highest correlation with chlorophyll content at the initial flowering stage with determined coefficient R2 of only 0.1,which failed the significance test.In other growth stages,the MERIS ground chlorophyll index (MTCI) had the highest correlation with chlorophyll content.The proposed model had high prediction accuracy with good fitting.The R2 of chlorophyll content estimation model of kiwifruit leaves at different growth stages based on principal component analysis and random forest regression was 0.91-0.98,which passed the extremely significant test.The fitting results and prediction accuracy were much higher than those of single band linear regression model and single band linear regression model based on vegetation index.It was the optimum model for estimating chlorophyll content in leaves.【Conclusion】The stochastic forest model based on principal component analysis contained more complete band information and had better predictive ability for chlorophyll content in kiwifruit leaves at different growth stages. |
| Key words: kiwifruit leaves chlorophyll content hyperspectral remote sensing technology principal component analysis random forest model |
