基于机器学习的春小麦叶片水分含量高光谱估算  

Hyperspectral Estimation of Spring Wheat Leaf Water Content Based on Machine Learning

作　　者：热依拉·艾合买提 吾木提·艾山江 阿不都艾尼·阿不里[4] 尼加提·卡斯木 RAHILA Ahmat;UMUT Hasan;ABDUGHENI Abliz;NIJAT Kasim(College of Biology and Geography Sciences,Yili Normal University,Yining,Xinjiang 835000,China;Institute of Applied Remote Sensing and Information Technology,College of Environmental and Resource Sciences,Zhejiang University,Hangzhou,Zhejiang 310058,China;Key Laboratory of Agricultural Remote Sensing and Information Systems,Zhejiang University,Hangzhou,Zhejiang,310058,China;College of Resources and Environmental Sciences,Xinjiang University,Urumqi,Xinjiang 830046,China)

机构地区：[1]伊犁师范大学生物与地理科学学院,新疆伊宁835000 [2]浙江大学农业遥感与信息技术应用研究所,杭州310058 [3]浙江省农业遥感与信息技术重点研究实验室,浙江杭州310058 [4]新疆大学资源与环境科学学院,新疆乌鲁木齐830046

出　　处：《麦类作物学报》2022年第5期640-648,共9页Journal of Triticeae Crops

基　　金：大学生创新训练项目(202010764004X)。

摘　　要：为了比较不同机器学习算法在干旱半干旱区春小麦叶片水分含量(leaf water content,LWC)遥感监测中的应用效果及筛选最佳波段组合,在田间尺度上,以春小麦冠层高光谱数据为基础,采用两波段组合形式,计算15种光谱参数(比值植被指数RVI、归一化植被指数NDVI、差值植被指数DVI和12种水分植被指数),通过对抽穗期叶片含水量与光谱参数拟合效果进行对比与分析,分别构建了基于机器学习[人工神经网络(artificial neural network,ANN)、K近邻(K-nearest neighbors,KNN)和支持向量回归(support vector regression,SVR)]和光谱参数的春小麦LWC反演模型,并对模型精度进行验证,以确定有效波段组合。结果表明,小麦抽穗期LWC与冠层高光谱反射率(R_(784~950))、12种水分植被指数均显著相关(P<0.01);波段组合形式有效地优化了两波段指数的波段组合,在800~1000 nm区间光谱参数(RVI_(1046,1057)、NDVI_(1272,1279)、DVI_(1272,1279))的波段组合计算明显提升了其对LWC的敏感性;在不同的机器学习算法中,基于两波段组合光谱参数的KNN算法所见模型对LWC的预测效果(r^(2)=0.64,RMSE=2.35,RPD=2.01)优于ANN、SVR两种算法。这说明两波段光谱指数和KNN算法在春小麦叶片水分含量的高光谱遥感估算中具有一定的优势。In arid and semi-arid regions,remote sensing monitoring of crop leaf water content(LWC)is essential for crop drought diagnosis and irrigation strategy formulation.At the field scale,based on the spring wheat canopy hyperspectral data,using the two-band combination form to calculate 15 kinds of spectral parameters(ratio vegetation index,RVI;normalized difference vegetation index,NDVI;difference vegetation index,DVI;12 kinds of water vegetation indeces),through the comparison and analysis of the fitting effect of leaf water content and spectral parameters at the heading stage,the machine learning artificial neural network,ANN,K-nearest neighbors,KNN and support vector regression(SVR)and spectral parameters of arid zone spring wheat leaf water content inversion model and verification of the estimated model to determine the effective band combination.The results showed that LWC at the heading stage,canopy hyperspectral reflectance(R_(784-950)),and 12 water vegetation indices were all significantly correlated(P<0.01);(Ⅱ)The combination of bands effectively optimized the two bands exponential band combination,the calculation effect of spectral parameter(RVI_(1046,1057),NDVI_(1272,1279),DVI_(1272,1279))band combination in the 800-1000 nm range had significantly improved its sensitivity to LWC;In different machine learning algorithms,based on the two-band combined spectrum,the predictive effect of parameter KNN algorithm on LWC(r^(2)=0.64,RMSE=2.35,RPD=2.01)was better than ANN and SVR.Therefore,the effective calculation of the two-band spectral index and the KNN algorithm have certain advantages in the field of hyperspectral remote sensing estimation of crop leaf moisture.

关 键 词：春小麦 机器学习 水分含量 植被指数 高光谱 

分 类 号：S512.1[农业科学—作物学] S314