河南农业科学 ›› 2023, Vol. 52 ›› Issue (11): 167-173.DOI: 10.15933/j.cnki.1004-3268.2023.11.018

• 农业信息与工程·农产品加工 • 上一篇    下一篇

基于深度学习的小麦倒伏自动分类方法研究

臧贺藏1,2,王从胜1,3,赵巧丽1,2,赵晴1,2,张杰1,2,李国强1,2,郑国清1,2   

  1. (1.河南省农业科学院农业经济与信息研究所,河南 郑州4 50002;2.农业农村部黄淮海智慧农业技术重点实验室,河南 郑州 450002;3.河南师范大学 计算机与信息工程学院,河南 新乡 453007)
  • 收稿日期:2023-07-17 出版日期:2023-11-15 发布日期:2023-12-05
  • 通讯作者: 郑国清(1964-),男,河南淅川人,研究员,博士,主要从事农业信息技术研究。E-mail:zgqzx@hnagri.org.cn
  • 作者简介:臧贺藏(1983-),女,河南驻马店人,副研究员,博士,主要从事作物表型鉴定研究。E-mail:zanghecang@163.com
  • 基金资助:
    河南省科技攻关计划项目(232102110272);河南省农业科学院自主创新项目(2023ZC063);河南省农业科学院科技创新团
    队项目(2022TD14)

Study on Automatic Classification Method of Wheat Lodging Based on Deep Learning

ZANG Hecang1,2,WANG Congsheng1,3,ZHAO Qiaoli1,2,ZHAO Qing1,2,ZHANG Jie1,2,LI Guoqiang1,2,ZHENG Guoqing1,2   

  1. (1.Institute of Agricultural Economy and Information,Henan Academy of Agricultural Sciences,Zhengzhou 450002,China;2.Huanghuaihai Key Laboratory of Intelligent Agricultural Technology,Ministry of Agriculture and Rural Areas,Zhengzhou 450002,China;3.College of Computer and Information Engineering,Henan Normal University,Xinxiang 453007,China)
  • Received:2023-07-17 Published:2023-11-15 Online:2023-12-05

摘要: 倒伏是制约小麦产量的关键因素。针对现有小麦倒伏区域统计费时费力、倒伏分类方法较为单一、模型预测精度低的问题,采用无人机遥感平台获取小麦倒伏可见光图像,自建小麦倒伏图像数据集,采用分割模型U-Net、PSPNet、DeepLabv3+和ACSNet自动提取小麦倒伏区域。结果表明,通过4种方法对小麦倒伏区域提取比较,ACSNet对小麦倒伏检测的分割效果较好,其精准率(Pre)、召回率(Rec)、Dice相关系数(DSC)和IoU指数分别为87.5%、91.7%、87.0%、88.6%,计算量(FLOPs)较低,具有较强的准确性和鲁棒性。ACSNet对小麦倒伏预测结果与真实结果接近,识别的平均相对误差为4.5%。表明通过ACSNet能够有效提取小麦倒伏信息,为无人机遥感评估小麦受灾面积及损失提供支撑。

关键词: 小麦, 倒伏, 无人机遥感, 深度学习, 分类方法

Abstract: Lodging is a key factor restricting wheat yield. Aiming at the problems of time⁃consuming and laborious statistics of wheat lodging area,single lodging classification method and low model prediction accuracy,this study used UAV remote sensing platform to obtain wheat lodging visible light image,self⁃built wheat lodging image data set,and used segmentation models U⁃Net,PSPNet,DeepLabv3+,ACSNet to automatically extract wheat lodging area.The results showed that ACSNet had a better segmentation effect on wheat lodging detection by comparing the four methods of wheat lodging area extraction,and its Pre,Rec,DSC and IoU were 87.5%,91.7%,87.0% and 88.6%,respectively,FLOPs were lower,with strong accuracy and robustness.After testing,the prediction results of wheat lodging by ACSNet were close to the real results,and the average relative error of recognition was 4.5%.Thus,ACSNet can effectively extract wheat lodging information,providing support for UAV remote sensing assessment of wheat disaster area and loss.

Key words: Wheat, Lodging, UAV remote sensing, Deep learning, Classification method

中图分类号: