水是植物正常生长发育必不可缺的元素之一, 能够快速检测并获取植物叶片水分, 对田间作物灌溉生产管理和作物的生理需水特性研究等具有重要的意义。 利用RedEdge-M型号多光谱相机, 以不同生育期的55组玉米叶片作为试验对象, 在光线充足且无阴影遮挡的环境下对试验玉米叶片样本进行拍摄, 拍摄过程中通过直连下行光传感器来消除太阳高度角对光谱反射的影响, 每组玉米叶片样本经过拍摄可得到蓝、 绿、 红、 近红外和红边等5个波段的TIFF图像。 借助图像处理软件ENVI5.3构建玉米叶片样本兴趣区域(ROI), 以ROI范围内玉米叶片样本的平均反射光谱作为该样本的反射光谱来减小镜头边缘减光现象带来的误差。 参照标准白板出厂时提供的专属标定反射率、 白板ROI范围内的平均反射光谱和玉米叶片样本白板ROI范围内的平均反射光谱, 比值换算得到各组玉米叶片5个波段处的光谱反射率。 同时利用YLS-D型号植物营养测定仪, 采用五点取样法选择玉米叶片的5个区域测取玉米叶片样本的水厚度平均值作为叶片含水量的测量指标。 随机选取43组玉米叶片样本得出的光谱反射率作为训练样本, 采用BP神经网络建立基于多光谱图像的玉米叶片含水量反演模型, 并融合莱文贝格-马夸特理论(Levenberg-Marquardt, L-M)进行经典神经网络现有缺点的改进。 输入神经元数目为5个, 即蓝、 绿、 红、 近红外和红边等5个波段图像对应的反射率, 输出神经元为1个, 即玉米叶片含水量。 剩余12组玉米叶片作为验证样本用于模型反演数据的相关性分析, 结果表明, 利用多光谱图像光谱信息并结合基于Levenberg-Marquardt方法改进后BP神经网络玉米叶片含水量反演模型, 模型反演的拟合相关系数能达到0.896 37, 12组验证集中玉米叶片含水量参考值和反演值的相关系数r达到0.894 8, 反演结果比较理想。 可以实现对玉米叶片含水量的快速准确检测, 对精准农业的推广和应用提供了方法和参考依据。

Water is part of the essential elements for the normal growth and development of plants. The ability to detect and obtain plant leaf moisture quickly is of great importance to the study of crop irrigation production management and the physiological water demand characteristics of crops in the field. Using RedEdge-M multispectral camera, 55 groups of maize leaves at different growth stages were selected as the test objects, and the test maize leaf samples were photographed in a mellow light environment without shading. During the photographing process, the influence of solar elevation angle on spectral reflection was eliminated by directly connecting down light sensors, and TIFF images in 5 bands of blue, green, red, near-infrared and red edges were obtained by photographing each group of maize leaf samples. With the help of image processing software ENVI 5.3, the region of interest (ROI) of maize leaf samples was constructed, and the average reflection spectrum of maize leaf samples within the ROI range was used as the reflection spectrum of the samples to reduce the error caused by lens edge dimming phenomenon. According to the calibration reflectivity of the standard white board, the average reflection spectrum in the ROI range of the white board and the average reflection spectrum in the ROI range of the maize leaf sample white board, the ratio was converted to obtain the spectral reflectivity of each group of maize leaves at five bands. At the same time, using YLS-D chlorophyll meter, using five-point sampling method. The average water thickness of maize leaf samples was measured in five areas of maize leaf as the measurement index of leaf water content. Randomly selected spectral reflectance of 43 sets of maize leaf samples as training samples, using BP neural network to build an inversion model of maize leaf water content, which was based on multi-spectral image, and the Levenberg-Marquardt method was introduced to improve the existing shortcomings of classical neural network. The number of input neurons was 5, that is, the reflectance corresponding to the five and images of blue, green, red, near-infrared and red-edged, and the output neurons were one, that is, the moisture content of the maize leaves. The remaining 12 sets of samples were invoked as verification samples for correlation verification analysis of model inversion data. The results showed that the multispectral image spectral information combined with the improved BP neural network based on the Levenberg-Marquardt method can be utilized to retrieve the water content of the maize leaf. The fitting correlation coefficient of the model inversion can reach 0.896 37. As a verification of the 12 groups of maize leaves moisture reference value and the inversion value of the correlation coefficient R2 reaches 0.894 8, the inversion result is ideal. It can realize the rapid and accurate detection of the moisture content of maize leaves, and provides a method and reference for the promotion and application of precision agriculture.