热带地理 ›› 2019, Vol. 39 ›› Issue (4): 482-491.doi: 10.13284/j.cnki.rddl.003153

• 专刊:无人机在生态学和地理学中的应用 • 上一篇    下一篇

基于无人机遥感的盛花期薇甘菊爆发点识别与监测

孙中宇1,荆文龙1,乔 曦2,杨 龙1   

  1. (1. 广东省地理空间信息技术与应用公共实验室//广州地理研究所,广州 510070; 2. 中国农业科学院 深圳农业基因组研究所,广东 深圳 518120)
  • 出版日期:2019-07-10 发布日期:2019-07-10
  • 通讯作者: 杨龙(1981—),男,陕西韩城人,博士,研究员,主要研究方向为恢复生态学,(E-mail)yanglong@gdas.ac.cn。
  • 作者简介:孙中宇(1986—),男,吉林榆树人,博士,副研究员,主要研究方向为恢复生态学和无人机遥感生态应用,(E-mail)sunzhyu@gdas.ac.cn;
  • 基金资助:

    广东省科技计划项目(2017A020216022;2018B030324002);国家重点研发计划(2017YFC1200105)

Identification and Monitoring of Blooming Mikania micrantha Outbreak Points Based on UAV Remote Sensing

Sun Zhongyu1,Jing Wenlong1,Qiao Xi2 and Yang Long1   

  1. (1. Guangdong Provincial Public Laboratory of Geospatial Information Technology and Application//Guangzhou Institute of Geography, Guangzhou 510070, China; 2. Shenzhen Institute of Agricultural Genomics, Chinese Academy of Agricultural Sciences, Shenzhen 518120, China)
  • Online:2019-07-10 Published:2019-07-10

摘要:

局域尺度上爆发点的识别与监测是薇甘菊(Mikania micrantha)入侵研究的一个难点,无人机遥感为此提供了新的研究手段。采用无人机搭载RGB相机获取研究地的正射影像,采用波段运算、影像分割和深度学习3种方法对盛花期薇甘菊的爆发点进行识别。结果表明:高分辨率的RGB拼接影像可直接用于目视识别薇甘菊的爆发点。过绿指数(EGI)、归一化过绿指数(NEGI)、蓝绿差异指数(BGDI)、绿红差异指数(GRDI)、归一化绿红差异指数(NGRDI)以及植被色素比值指数(PPR)均无法分离薇甘菊和其附主植物;但PPR指数可为面向对象的多尺度分割提供参数支持。面向对象的多尺度分割可自动识别薇甘菊的爆发点,但会低估其爆发面积。基于深度学习(Deeplab V3+)的自动识别方法,能准确识别薇甘菊的爆发点和爆发面积,测试集的平均交并比(mIoU)为78.46%,像素精度为88.62%。无人机遥感数据为局域尺度上的薇甘菊扩散机制研究提供了基础,也为薇甘菊入侵的监测、预警和精准防治提供了有力支撑。

关键词: 机器学习, 深度学习, 自动识别, 生态监测, 无人机遥感, 薇甘菊

Abstract:

The identification and monitoring of outbreak points on a local scale is a challenge in the study of Mikania micrantha invasion. In this study, orthophoto images of the research area were acquired by using a red, green, and blue (RGB) camera mounted on an unmanned aerial vehicle (UAV). Three methods, including band math, image segmentation, and deep learning, were tested in order to identify the blooming point of Mikania micrantha. Results showed that high-resolution RGB mosaic images could be directly used to visually identify Mikania micrantha outbreak points. Neither the over-green index (EGI), normalized over-green index (NEGI), blue-green differential index (BGDI), green-red differential index (GRDI), normalized green-red differential index (NGRDI), nor Plant Pigment Ratio index (PPR) could separate Mikania micrantha from its host plants. Though it should be noted that PPR could provide parameter support for multi-scale segmentation. Object-oriented multi-scale segmentation was able to automatically identify the eruption point of Mikania micrantha, though the area of eruption was underestimated. The automatic recognition method based on deep learning (Deeplab V3+) was also able to accurately identify the eruption point and area of Mikania micrantha. The average intersection ratio (mIoU) of the test set was 78.46% and the pixel accuracy was 88.62%. UAV remote sensing data provides a basis for the study of the Mikania micrantha diffusion mechanism on a local scale, as well as strong support for the monitoring, early warning, and precise control of Mikania micrantha invasions. Despite our advances, challenges still exist in system integration, sensor price, aviation control, and discipline interaction for the applications of a UAV remote sensing system in invasion ecology. Recognition algorithm, data quality, and phenological period are three key factors for the automatic identification accuracy of invasive plants. As such, the integration of information from multi-sensors could enhance identification precision. Building a database that includes spectrum, temperature, and height information of invasive plants is helpful for the automatic identification and quantification of these.

Key words: machine learning, deep learning, automatic identification, ecological monitoring, UAV remote sensing, Mikania micrantha