您好，欢迎访问浙江省农业科学院 机构知识库！
浙江省农业科学院机构知识库

全部

  • 全部
  • 标题
  • 学者
  • 机构
  • 关键词

Weakly supervised dual-mask marginal segmentation and variable path planning method for bean weed based on UAV remote sensing

文献类型： 外文期刊

作者： Zhang, Jianlin 1 ; Lu, Xiangyu 1 ; Yang, Rui 1 ; Xu, Hongxing 2 ; Huai, Yan 3 ; Liu, Fei 1 ;

作者机构： 1.Zhejiang Univ, Coll Biosyst Engn & Food Sci, Hangzhou 310058, Peoples R China

2.Zhejiang Acad Agr Sci, Inst Plant Protect & Microbiol, State Key Lab Managing Biot & Chem Threats Qual &, Hangzhou 310021, Peoples R China

3.Zhejiang Agr Technol Extens Ctr, Hangzhou 310020, Peoples R China

关键词： Variable path planning; Marginal segmentation; Teacher-to-student; Deep learning; Image processing

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：8.9； 五年影响因子：9.3 ）

ISSN： 0168-1699

年卷期： 2025 年 230 卷

页码：

收录情况： SCI

摘要： Field path planning based on UAV remote sensing is the foundation for the autonomous navigation of agricultural weeding robots and other vehicles. Traditional path planning methods are limited by the accuracy of predefined navigation lines and do not consider the variation in workload at different field locations during vehicle operation. This study presents a variable navigation path planning process that integrates the bean weed marginal segmentation technology. The proposed process consists of five main steps. Firstly, constructing a dataset for detecting of bean and weed using manual semi-automatic annotation method; secondly, developing a bean and weed detection model using a teacher-to-student training approach; thirdly, integrating model detection outcomes with the dual-mask pixel-level registration method to achieve bean and weed marginal detection; fourth, semantic segmentation of crop ridges and differential methods are used to extract basic navigation lines. Finally, a variable path planning method is designed by combining the marginal weed feature distribution prescription with the basic navigation lines. To test the algorithm, experiments were conducted using ridge-cultivated beans. The results show that the mIoU, mPA, and Acc values of the best-performing student model, Segformer-b0-S, reached 79.28 %, 88.95 %, and 95.36 % respectively, significantly reducing the annotation workload and successfully achieving the marginal segmentation of bean weed. The mean, median, and Std deviation indicators of the crop ridge navigation line based on Deeplabv3+_mobilenet are 6.07, 4.37, and 8.10 cm, respectively. The study indicates that the developed variable navigation path planning process has achieved satisfactory results, laying an important foundation for the advancement of agricultural weeding robots.

  • 相关文献
作者其他论文 更多>>
置顶
购物车
反馈

© 京ICP备09089781号-29 主办单位：浙江省农业科学院

学术资源： 中国农业科学院 农业专业知识服务系统 农科联盟资源共建共享平台 中国农业科技文献与信息服务平台 中国农业期刊集成服务平台