基于窄带光谱图像分析的小麦冠层植被指数测量方法研究  

Wheat canopy vegetation index measurement method based on narrow band spectral image analysis

作　　者：余洪锋[1] 徐焕良[2] 丁永前[2,3] 杨紫楠 窦祥林[1] 李庆[2] 关心桐 YU Hongfeng;XU Huanliang;DING Yongqian;YANG Zinan;DOU Xianglin;LI Qing;GUAN Xintong(College of Engineering,Nanjing Agricultural University,Nanjing 210031,China;College of Artifical Intelligence,Nanjing Agricultural University,Nanjing 210031,China;Collaborative Innovation Center for Modern Crop Production Co-sponsored by Province and Ministry,Nanjing Agricultural University,Nanjing 210095,China)

机构地区：[1]南京农业大学工学院,江苏南京210031 [2]南京农业大学人工智能学院,江苏南京210031 [3]南京农业大学现代作物生产省部共建协同创新中心,江苏南京210095

出　　处：《南京农业大学学报》2023年第1期189-199,共11页Journal of Nanjing Agricultural University

基　　金：国家重点研发计划项目(2016YFD070030403)。

摘　　要：[目的]针对小麦生长早期植被指数易受土壤背景干扰的问题,提出了一种基于窄带光谱图像分析的小麦植被指数测量方法。[方法]构建了多镜头结构的窄带光谱图像获取装置,实时获取656和770 nm的田间小麦窄带光谱图像。运用简单线性聚类(simple linear iterative clustering, SLIC)和VGG16(visual geometry group network 16)全卷积神经网络对小麦近红外窄带光谱图像进行超像素聚类和分类,把交并比(Qseg)、综合评价指标(F值)、精度(Precision)作为分割精度评价指标,分析传统阈值分割方法和本研究方法去土壤背景干扰的性能差异。去除土壤背景后的窄带光谱图像采用太阳光免白板标定方法计算植被指数,并与GreenSeeker RT200的实测数据进行对比分析,定性定量评价本研究方法去除土壤背景干扰的性能。[结果]试验共采集12个小麦品种、2个施氮水平、24块种植小区图像,Qseg、Precision和F值的平均值分别为90.41%、80.82%和72.73%,分割性能均优于传统的阈值分割方法。针对相同测试田块,GreenSeeker RT200测量的各小区归一化植被指数(normalized difference vegetation index, NDVI)变异系数的最大值、平均值和标准差分别为47.12%、33.61%、10.17%,而本测量方法的各小区NDVI的相应指标值分别为18.59%、9.61%、3.88%;当采样小区小麦封行后,本方法所提取的NDVI与GreenSeeker RT200的测量值具有较高的相关性,决定系数为0.895 9。[结论]该方法可以完成复杂土壤背景、大田光照变化条件下的小麦窄带光谱图像的冠层提取与植被指数测量,可为多镜头结构的作物冠层反射光谱仪的优化设计提供参考。[Objectives]Aiming at the problem that the vegetation index in the early stage of wheat growth is easily disturbed by soil background, a wheat canopy vegetation index measurement method based on narrow band spectral image analysis was proposed. [Methods]A narrow-band spectral image acquisition device with multi-lens structure was constructed to obtain 656 and 770 nm narrow-band spectral images of field wheat in real time. Simple linear iterative clustering(SLIC)and visual geometry group network 16(VGG16)full convolution neural network were used for super-pixel clustering and classification of wheat near-infrared narrow-band spectral images. Using intersection over union Qseg,comprehensive evaluation index(F value)and Precision as segmentation accuracy evaluation indexes, the performance difference between traditional threshold segmentation method and the proposed method to remove soil background interference was analyzed. The sunlight free whiteboard calibration method was used for the narrow-band spectral image after removing the soil background to calculate the vegetation index. The calculated vegetation index took the measured data of GreenSeeker RT200 as a reference to qualitatively and quantitatively evaluate the performance of the proposed method to remove the soil background interference. [Results]The images of 12 wheat varieties, two nitrogen application levels and 24 planting areas were collected. The average value of Qseg,Precision, F were 90.41%,80.82% and 72.73%,respectively. The segmentation performance was better than the traditional threshold segmentation method. For the same test field, the maximum variation coefficient, average value and standard deviation of normalized difference vegetation index(NDVI)of each area measured by GreenSeeker RT200 were 47.12%,33.61% and 10.17% respectively, while that of NDVI of each area measured by the proposed method were 18.59%,9.61% and 3.88%,respectively. Moreover, the vegetation index extracted by this method had a high correlation with the measurement result

关 键 词：小麦 窄带光谱图像 冠层分割 VGG16神经网络 超像素 植被指数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]