无人机遥感及深度学习在油菜冻害识别中的应用研究  

Application research of unmanned aerial vehicle remote sensing and deep learning in identification of freezing injury in rapeseed

作　　者：安谈洲 李俐俐 张瑞杰 李礼[2] 刘清云 乔江伟 王新发[1] 姚剑[2] AN Tan-zhou;LI Li-li;ZHANG Rui-jie;LI Li;LIU Qing-yun;QIAO Jiang-wei;WANG Xin-fa;YAO Jian(Oil Crops Research Institute,Chinese Academy of Agricultural Sciences,Wuhan 430062,China;School of Re-mote Sensing and Information Engineering,Wuhan University,Wuhan 430079,China;Xishui County Agricultural Technology Extension Center,Huanggang 421125,China)

机构地区：[1]中国农业科学院油料作物研究所,湖北武汉430062 [2]武汉大学遥感信息工程学院,湖北武汉430079 [3]浠水县农业技术推广中心,湖北黄冈421125

出　　处：《中国油料作物学报》2021年第3期479-486,共8页Chinese Journal of Oil Crop Sciences

基　　金：“十三五”国家重点研发计划项目(2018YFD0100601);湖北省技术创新专项:优质高产多抗油菜品种选育(2018ABA087)。

摘　　要：气象灾害是影响我国农作物产量及农业可持续发展的重要因素,低温冻害导致油菜生长发育迟缓和大幅度减产。便捷、精准地识别和评估油菜冻害不仅为精准农田管理提供依据而且对培育抗冻品种有重要意义。本研究借助低空无人机遥感技术,用大疆精灵无人机Phantom 4 Pro V2.0,搭载2000万像素RGB相机,对生长88天正处于越冬期的2052份油菜育种材料进行航拍。将拍摄后的图片经拼接、剪裁、评价、扩展后按8:2划分为训练集和验证集。然后将训练集图片分批输入自主建立的深度学习网络模型中提取冻害特征并优化网络模型,优化完成后用验证集图片对识别结果进行验证。结果表明深度学习网络模型对油菜是否发生冻害的整体识别精度达98.13%,Micro F1为98.13%、Macro F1为98.11%,Kappa系数为0.96,说明深度学习网络模型整体性能较好。本试验使用较低的成本实现了大规模油菜材料冻害情况快速、高效、准确的评估,弥补了传统调查方法的缺陷。同时本研究建立的成套油菜冻害鉴定方法可为后续抗冻性状机理研究及品种选育提供重要支撑,也可为其它类似冻害的逆境鉴定方法开发提供借鉴和参考。Meteorological disaster is a major factor that affects crop yield and agricultural sustainable development in China.Freezing injury could lead to growth retardation and yield reduction of rapeseed.Rapid and accurate identification and evaluation of rapeseed freezing injury not only can provide basic information for the accurate field management but also have important significance for rapeseed breeding with freezing tolerance.Here in this study,we employed the low-altitude UAV remote sensing technology,by using Dajiang Phantom 4 Pro V2.0 that carries a 20 million pixel RGB camera,photographed 2052 accessions of rapeseed breeding materials.These materials were in wintering growth stage of 88 days old.After splicing,clipping,evaluation and expansion,the images taken were divided into training set and verification set according to 8:2.The images of training set were input into the developed deep learning network model in batches to extract freezing injury features and optimize the network model.After optimization,the identification results were verified with the images of verification set.The results demonstrated that the overall identification accuracy of the deep learning network model was 98.13%,the Micro F1 was 98.13%,the Macro F1 was 98.11%,the Kappa coefficient was 0.96,indicating that this deep learning network model had reliable and fine overall performance.This experiment realized the rapid,efficient and accurate evaluation of freezing damage of large area rapeseed materials with relatively low cost,and made up for the defects of traditional investigation methods.At the same time,the complete set of freezing injury identification methods established in this study can provide important support for the subsequent research on the mechanism of freezing resistance and variety selection.What is more,it can also substantially support the development of phenotyping systems on other alike traits.

关 键 词：无人机遥感 深度学习 油菜 冻害 

分 类 号：S127[农业科学—农业基础科学] S565.4