使用高光谱成像技术对田间杂草进行表征和识别  

Characterization and identification of weeds in the field using hyperspectral imaging

作　　者：柳莹莹 尹雁玲 刘江华 尹勇 储涛涛 江阳 LIU Yingying;YIN Yanling;LIU Jianghua;YIN Yong;CHU Taotao;JIANG Yang(Colege of Mechanical and Electrical Engineering,Tarim University,Alar,Xinjiang 843300;College of Life Science and Technology,Tarim University,Alar,Xinjiang 843300;Service Center of Comprehensive Affairs,Shangjiahe Town,Xinbin Manzu Autonomous County,Fushun,Liaoning 113000)

机构地区：[1]塔里木大学机械电气化工程学院,新疆阿拉尔843300 [2]塔里木大学生命科学与技术学院,新疆阿拉尔843300 [3]新宾满族自治县上夹河镇综合事务服务中心,辽宁抚顺113000

出　　处：《塔里木大学学报》2024年第3期89-97,共9页Journal of Tarim University

基　　金：黑龙江省北方寒地现代农业装备技术重点实验室2022年度开放课题项目(55200511)。

摘　　要：进行杂草表征和分类对于智慧农业杂草清除与田间管理而言至关重要,然而杂草识别精度较低的问题尚未得到深入探究。本研究以田间杂草为研究对象,建立了农业地区7种杂草物种的高光谱数据库。首先,使用地面高光谱相机采集了高光谱图像,提取并分析了每个物种的代表性光谱曲线、光谱剖面特征,并对各物种进行主成分分析,揭示了不同杂草物种的差异性。采用多元散射校正、归一化、一阶和二阶差分求导对原始光谱数据进行预处理。最后通过建立支持向量机与一维卷积神经网络分类模型,用于高光谱图像中识别杂草。结果表明,MSC-1DCNN模型的分类效果最佳,不同物种的用户准确性为95.71%~100%。该研究不仅为了解杂草物种的高光谱特征以及杂草管理提供了有力的保障,而且也有助于自动化除草机器人的研制与无人农场的实施。Carrying out weed characterisation and classification is crucial for weed removal and field management in smart agriculture,however,the problem of low accuracy of weed identification has not been explored in depth.In this study,a hyperspectral database of seven types of weed species in agricultural areas was established using field weeds as research objects.Firstly,hyperspectral images were collected using a ground-based hyperspectral camera and representative spectral curves,spectral profile features for each species were selected and analysed,and principal componet analysis was carried out for each species to reveal differences between weed species.The raw spectral data were then preprocessed using multiple scattering correction,normalisation,and first-order and second-order difference derivation.Finally,the support vector machine with a one dimensional convolution neural networks classification model was established and applied in identifying weeds in the hyperspectral images.The results suggested that the MSC-1DCNN model shows the best classification effect,with user accuracies ranging from 95.71%to 100%.This research not only provides strong hyperspectral characterisation of weed species and weed management,but also contributes to the development of automated weeding robots and the implementation of unmanned farms.

关 键 词：杂草 高光谱 卷积神经网络 田间管理 

分 类 号：S451.0[农业科学—植物保护]