GWLS-SVR模型的红枣树叶片叶绿素含量估算  

Chlorophyll Content Estimation of Jujube Leaves Based on GWLS-SVR Model

作　　者：尼格拉·吐尔逊 苏磊·乃比 高健 沈江龙 郑江华[1] 余丹林 Nigela Tuerxun;Sulei Naibi;GAO Jian;SHEN Jiang-long;ZHENG Jiang-hua;YU Dan-lin(College of Resources and Environmental Science,Xinjiang University,Urumqi 830046,China;College of Mathematics and System Science,Xinjiang University,Urumqi 830046,China;Institute of Modern Forestry,Xinjiang Academy of Forestry Sciences,Urumqi 830063,China;Department of Earth and Environmental Studies,Montclair State University,New Jersey 07043,USA)

机构地区：[1]新疆大学资源与环境科学学院,新疆乌鲁木齐830042 [2]新疆大学数学与系统科学学院,新疆乌鲁木齐830046 [3]新疆林业科学院现代林业研究所,新疆乌鲁木齐830063 [4]Department of Earth and Environmental Studies,Montclair State University,New Jersey 07043,USA

出　　处：《光谱学与光谱分析》2021年第6期1730-1736,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(41461035);新疆特色林果业典型树种分布遥感监测(2019);新疆青年科技创新人才工程项目(2017)资助。

摘　　要：叶绿素含量是红枣树光合作用能力、生长状况、营养状况的指示剂,不同地理位置种植的红枣树受到自然、人为等因素的影响,叶绿素含量分布有所不同,该研究实地测定了若羌县枣树叶片高光谱反射率及表征叶绿素含量的枣树叶片SPAD(soil plant analysis development)值。为了高效无损地估算红枣树叶片SPAD值,计算了红枣树叶片SPAD值全局莫兰指数,以SPAD值和高光谱波段之间的相关性为基础,通过C P统计量计算重要程度高的特征波段,运用地理加权最小二乘支持向量回归GWLS-SVR(geographically weighted least squares-support vector regression)模型对红枣树叶片SPAD值进行预测,与多元线性回归(MLR)、支持向量机回归(SVR)模型比较并探讨GWLS-SVR模型估算红枣树叶片SPAD值的能力。结果表明:(1)光谱一阶导数可以有效去除噪声并突出光谱信息尤其是492~510,542~543,642~652,657~670和682~692 nm区间内显著的提高了与SPAD值的相关性。(2)C P统计量方法能够有效的选择敏感区间的特征波段,进而提高模型估算精度,由统计量方法计算出原始光谱重要程度最高的两个变量为595与696 nm,光谱一阶导数的特征波段为688 nm。其中对于同一个敏感波段区间的波段组合总有单个波段的统计量低于多个波段组合的统计量,这可能是相近波段间的较强共线性导致的。(3)若羌县红枣树叶片SPAD值存在显著的空间聚集性,全局莫兰指数为0.1258(p<0.1),适合建立考虑空间位置的GWLS-SVR模型。(4)结合Bootstrap再抽样与t检验模型检验得到结合地理位置信息的GWLS-SVR模型总体上估算能力优于SVR和MLR模型,且结果高度显著(p<0.001),其中基于光谱一阶导数的GWLS-SVR模型为最优的红枣树叶片SPAD值估算模型(R 2为0.975,MSE为1.082),能够为高光谱定量反演红枣树SPAD值进而快速无损的监测红枣生长状况提供一定参考。Chlorophyll Contentis an indicator of the photosynthetic capacity,growth and nutritional status of jujube trees.The distribution of chlorophyll content is different in jujube trees planted in different geographical locations under the influence of natural and human-made factors.The Hyperspectral reflectance of jujube leaves and the SPAD value of jujube leaves that representing chlorophyll content in Ruoqiang county were measured on the spots.To estimate the SPAD value of jujube leaves efficiently and losslessly,the global Moran’s I of jujube SPAD value was calculated,The statistics was calculated based on the correlation between SPAD value and Hyperspectral bands to choose the most important characteristic bands.The GWLS-SVR(Geographically Weighted Least Squares-Support Vector Regression)model was used to predict the SPAD value and compared with multiple linear regression(MLR)and support vector regression(SVR)models,and explored the ability of the model to estimate the SPAD value of the jujube leaves.The results show that:(1)the First derivative of the spectrum can effectively remove the noise and highlight the spectral information,especially in the range of 492~510,542~543,642~652,657~670 and 682~692 nm,and significantly improve the correlation of the spectrum with SPAD value.(2)statistics method can effectively select the feature bands of the sensitive range,thus improves the model estimation accuracy.The two variables with the highest importance of the original spectrum were 595 and 696 nm,and the feature band of the first derivative of the spectrum was 688 nm.Among them,the statistics of a single band were always lower than those of multiple band combinations of the same sensitive band interval,which may be caused by the strong collinearity between the adjacent bands.(3)There was significant spatial aggregation on the SPAD value of jujube leaves in Ruoqiang county,the global Moran’s I was 0.1258(p<0.1),which is suitable for the establishment of GWLS-SVR model that considers the spatial location.(4)By comb

关 键 词：叶绿素含量 GWLS-SVR模型 高光谱 红枣树 Bootstrap再抽样 t检验 SPAD值 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]